Patent Grant
8898068
Embodiments according to the invention are related to an audio decoder for providing a decoded audio information on the basis of an encoded audio information, an audio encoder for providing an encoded audio information on the basis of an input audio information, a method for providing a decoded audio information on the basis of an encoded audio information, a method for providing an encoded audio information on the basis of an input audio information and a computer program.
Embodiments according to the invention are related to an improved spectral noiseless coding, which can be used in an audio encoder or decoder, like, for example, a so-called unified-speech-and-audio coder (USAC).
In the following, the background of the invention will be briefly explained in order to facilitate the understanding of the invention and the advantages thereof. During the past decade, big efforts have been put on creating the possibility to digitally store and distribute audio contents with good bitrate efficiency. One important achievement on this way is the definition of the International Standard ISO/IEC 14496-3. Part 3 of this Standard is related to an encoding and decoding of audio contents, and subpart 4 of part 3 is related to general audio coding. ISO/IEC 14496 part 3, subpart 4 defines a concept for encoding and decoding of general audio content. In addition, further improvements have been proposed in order to improve the quality and/or to reduce the required bit rate.
According to the concept described in said Standard, a time-domain audio signal is converted into a time-frequency representation. The transform from the time-domain to the time-frequency-domain is typically performed using transform blocks, which are also designated as “frames”, of time-domain samples. It has been found that it is advantageous to use overlapping frames, which are shifted, for example, by half a frame, because the overlap allows to efficiently avoid (or at least reduce) artifacts. In addition, it has been found that a windowing should be performed in order to avoid the artifacts originating from this processing of temporally limited frames.
By transforming a windowed portion of the input audio signal from the time-domain to the time-frequency domain, an energy compaction is obtained in many cases, such that some of the spectral values comprise a significantly larger magnitude than a plurality of other spectral values. Accordingly, there are, in many cases, a comparatively small number of spectral values having a magnitude, which is significantly above an average magnitude of the spectral values. A typical example of a time-domain to time-frequency domain transform resulting in an energy compaction is the so-called modified-discrete-cosine-transform (MDCT).
The spectral values are often scaled and quantized in accordance with a psychoacoustic model, such that quantization errors are comparatively smaller for psychoacoustically more important spectral values, and are comparatively larger for psychoacoustically less-important spectral values. The scaled and quantized spectral values are encoded in order to provide a bitrate-efficient representation thereof.
For example, the usage of a so-called Huffman coding of quantized spectral coefficients is described in the International Standard ISO/IEC 14496-3:2005(E), part 3, subpart 4.
However, it has been found that the quality of the coding of the spectral values has a significant impact on the required bitrate. Also, it has been found that the complexity of an audio decoder, which is often implemented in a portable consumer device, and which should therefore be cheap and of low power consumption, is dependent on the coding used for encoding the spectral values.
In view of this situation, there is a need for a concept for an encoding and decoding of an audio content, which provides for an improved trade-off between bitrate-efficiency and resource efficiency.
According to an embodiment, an audio decoder for providing a decoded audio information on the basis of an encoded audio information may have: an arithmetic decoder for providing a plurality of decoded spectral values on the basis of an arithmetically-encoded representation of the spectral values included in the encoded audio information; and a frequency-domain-to-time-domain converter for providing a time-domain audio representation using the decoded spectral values, in order to acquire the decoded audio information; wherein the arithmetic decoder is configured to select a mapping rule describing a mapping of a code value of the arithmetically-encoded representation of spectral values onto a symbol code representing one or more of the decoded spectral values or at least a portion of one or more of the decoded spectral values in dependence on a context state described by a numeric current context value; and wherein the arithmetic decoder is configured to determine the numeric current context value in dependence on a numeric previous context value and in dependence on a plurality of previously-decoded spectral values, wherein the arithmetic decoder is configured to modify a number representation of the numeric previous context value, describing a context state for the decoding of one or more previously decoded spectral values, in dependence on a context subregion value describing a subregion of a context, to acquire a number representation of a numeric current context value describing a context state for the decoding of one or more spectral values to be decoded.
According to another embodiment, an audio encoder for providing an encoded audio information on the basis of an input audio information may have: an energy-compacting time-domain-to-frequency-domain converter for providing a frequency-domain audio representation on the basis of a time-domain representation of the input audio information, such that the frequency-domain audio representation includes a set of spectral values; and an arithmetic encoder configured to encode a spectral value or a preprocessed version thereof, using a variable length codeword, wherein the arithmetic encoder is configured to map one or more spectral values, or a value of a most significant bitplane of one or more spectral values, onto a code value, wherein the encoded audio information includes a plurality of variable length codewords, wherein the arithmetic encoder is configured to select a mapping rule describing a mapping of one or more spectral values, or of a value of a most significant bitplane of one or more spectral values, onto a code value in dependence on a context state described by a numeric current context value; and wherein the arithmetic encoder is configured to determine the numeric current context value in dependence on a numeric previous context value and in dependence on a plurality of previously-encoded spectral values, wherein the arithmetic encoder is configured to modify a number representation of the numeric previous context value, describing a context state for the encoding of one or more previously-encoded spectral values, in dependence on a context subregion value describing a subregion of a context, to acquire a number representation of a numeric current context value describing a context state for the encoding of one or more spectral values to be encoded.
According to another embodiment, a method for providing a decoded audio information on the basis of an encoded audio information may have the steps of: providing a plurality of decoded spectral values on the basis of an arithmetically-encoded representation of the spectral values included in the encoded audio information; and providing a time-domain audio representation using the decoded spectral values, in order to acquire the decoded audio information; wherein providing the plurality of decoded spectral values includes selecting a mapping rule describing a mapping of a code value of the arithmetically encoded representation of spectral values onto a symbol code representing one or more of the decoded spectral values, or at least a portion of one or more of the decoded spectral values in dependence on a context state described by a numeric current context value; and wherein the numeric current context value is determined in dependence on a numeric previous context value and in dependence on a plurality of previously decoded spectral values, wherein a number representation of the numeric previous context value, describing a context state for the decoding of one or more previously decoded spectral values, is modified in dependence on a context subregion value describing a subregion of a context, to acquire a number representation of a numeric current context value, describing a context state for the decoding of one or more spectral values to be decoded.
According to another embodiment, a method for providing an encoded audio information on the basis of an input audio information may have the steps of: providing a frequency-domain audio representation on the basis of a time-domain representation of the input audio information using an energy-compacting time-domain-to-frequency-domain conversion, such that the frequency-domain audio representation includes a set of spectral values; and arithmetically encoding a spectral value, or a preprocessed version thereof, using a variable-length codeword, wherein a spectral value or a value of a most significant bitplane of a spectral value is mapped onto a code value; wherein a mapping rule describing a mapping of one or more spectral values, or of a most significant bitplane of one or more spectral values, onto a code value is selected in dependence on a context state described by a numeric current context value; and wherein the numeric current context value is determined in dependence on a numeric previous context value and in dependence on a plurality of previously-encoded spectral values; wherein a number representation of the numeric previous context value, describing a context state for the encoding of one or more previously encoded spectral values, is modified in dependence on a context subregion value describing a subregion of a context, to acquire a number representation of a numeric current context value describing a context state for the encoding of one or more spectral values to be encoded; wherein the encoded audio information includes a plurality of variable-length codewords.
Another embodiment may have a computer program for performing the method according to claim 17 when the computer program runs on a computer.
Another embodiment may have a computer program for performing the method according to claim 18 when the computer program runs on a computer.
An embodiment according to the invention creates an audio decoder for providing a decoded audio information on the basis of an encoded audio information. The audio decoder comprises an arithmetic decoder for providing a plurality of decoded spectral values on the basis of an arithmetically encoded representation of the spectral values. The audio decoder also comprises a frequency-domain-to-time-domain converter for providing a time-domain audio representation using the decoded spectral values, in order to obtain the decoded audio information. The arithmetic decoder is configured to select a mapping rule describing a mapping of a symbol value onto a symbol code (which symbol code typically describes a spectral value or a plurality of spectral values or a most-significant bit plane of a spectral value or of a plurality of spectral values) in dependence on a context state described by a numeric current context value. The arithmetic decoder is configured to determine the numeric current context value in dependence on a plurality of previously-decoded spectral values. The arithmetic decoder is configured to modify a number representation of a numeric previous context value, describing a context state associated with one or more previously-decoded spectral values (or, more precisely, describing the context state for the decoding of said one or more previously-decoded spectral values), in dependence on a context subregion value, to obtain a number representation of a numeric current context value describing a context state associated with one or more spectral values to be decoded (or, more precisely, describing the context state for the decoding of said one or more spectral values to be decoded).
This embodiment according to the invention is based on the finding that it is computationally very efficient to modify a number representation of a numeric previous context value in dependence on a context subregion value, to obtain a number representation of a numeric current context value, because a complete re-computation of the numeric current context value can be avoided. Rather, correlations between the numeric previous context value and the numeric current context value can be exploited in order to keep the computational effort comparatively small. It has been found that a large variety of possibilities exist for the modification of the number representation of the numeric previous context value, including a combination of a re-scaling of the number representation of a numeric previous context value, an addition of a context subregion value or a value derived therefrom (like, for example, a bit-shifted version of a context sub-region value) to the number representation of the numeric previous context value or to a processed number representation of the numeric previous context value, a replacement of a portion of the number representation (rather than the entire number representation) of the numeric previous context value in dependence on the context subregion value, etc. Thus, maintaining at least a portion of a number representation of the numeric previous context value (possibly, in a shifted version) allows to significantly reduce the computational effort for the update of the numeric context value.
In a preferred embodiment, the arithmetic decoder is configured to provide the number representation of the numeric current context value such that portions of the number representation having different numeric weights are determined by different context subregion values. Accordingly, an iterative update of the numeric context value, to derive the numeric current context value from the numeric previous context value, can be done with small computational effort, while omitting a loss of information.
In a preferred embodiment, the number representation is a binary number representation of a single numeric current context value. Preferably, a first subset of bits of the binary number representation is determined by a first context subregion value associated with one or more previously-decoded spectral values, and a second subset of bits of the binary number representation is determined by a second context subregion value associated with one or more previously-decoded spectral values, wherein the bits of the first subset of bits comprise a different numeric weight than the bits of the second subset of bits. It has been found that such a representation is well-suited for the iterative derivation of the numeric current context value from the numeric previous context value.
In a preferred embodiment, the arithmetic decoder is configured to modify a bit-wise masked subset of information bits of the number representation of the numeric previous context value, or of a bit-shifted version of the number representation of the numeric previous context value, in dependence on a context subregion value which has not been considered for the derivation of the numeric previous context value, in order to obtain the number representation of the numeric current context value. By performing a bit-wise masking of the number representation of the numeric previous context value, or by bit-shifting the number representation of the numeric previous context value, it can be achieved that portions of a context which are no longer as relevant as before, are removed from the numeric context value and, preferably, are replaced by other portions of the context which are more relevant in the current context. A bit-wise masking of a subset of information bits of the number representation of the numeric previous context value allows to replace a portion of the numeric previous context value in dependence on a context subregion value, which, in turn, allows to consider a portion of the context which has not yet been considered previously. Moreover, a shift operation reflects the fact that there is some overlap between previously-decoded spectral values used to determine the previous context (i.e. the context used for decoding the previous tuple of spectral values) and previously-decoded spectral values used to determine the current context (i.e. the context for the decoding of the spectral values to be currently decoded). Moreover, the shift operations also reflect the fact that the frequency relation (for example, equal in frequency, larger in frequency by one frequency bin, etc.) of the previously-decoded spectral values with respect to spectral values decoded using the numeric previous context value is different from the frequency relationship of the previously-decoded spectral values with respect to the spectral values to be decoded using the numeric current context value.
In a preferred embodiment, the arithmetic decoder is configured to bit-shift the number representation of the numeric previous context value, such that numeric weights of subsets of bits associated with different context subregion values are modified, in order to obtain the number representation of the numeric current context value. Accordingly, the shift of the frequency position between the one or more spectral values decoded using the numeric previous context value and the one or more spectral values to be decoded using the numeric current context value can be reflected in the numeric context value in an efficient manner. Moreover, a shift operation typically can be performed with low computational effort using a standard microprocessor.
In a preferred embodiment, the arithmetic decoder is configured to bit-shift the number representation of the numeric previous context value, such that a subset of bits, which are associated with a context subregion value, is deleted from the number representation, in order to obtain the number representation of the numeric current context value. Accordingly, a double functionality can be provided by a single shift operation, namely the consideration of a change of the frequency position and the consideration of the fact that some spectral values (represented by a context subregion value) which has been used to obtain the numeric previous context value, are no longer needed to obtain the numeric current context value.
In a preferred embodiment, the arithmetic decoder is configured to modify a first subset of bits of a binary number representation of a numeric previous context value, or of a bit-shifted version of a binary number representation of a numeric previous context value, in dependence on a context subregion value, and to leave second subsets of bits of the binary number representation of the numeric previous context value, or of the bit-shifted version of the binary number representation of the numeric previous context value, unchanged, to derive the binary number representation of the numeric current context value from the binary number representation of the numeric previous context value by selectively modifying one or more subsets of bits associated with context sub-regions considered for the decoding of the previously-decoded spectral values (decoded using the numeric previous context value) and not considered for the decoding of spectral values to be decoded using the numeric current context value. This concept has proven to be particularly efficient.
In a preferred embodiment, the arithmetic decoder is configured to provide the number representation of the numeric current context value such that a subset of least-significant bits of the number representation of the numeric current context value describes a context subregion value, which context subregion value is used for a decoding of spectral values for which a context state is defined by the numeric current context value, and which context subregion value is not used for a decoding of spectral values for which a context state is defined by a numeric subsequent context value (e.g. a numeric context value derived from the numeric current context value). This approach allows to derive the numeric current context value from the numeric previous context value (and to derive the numeric subsequent context value from the numeric current context value) using a shift operation, as the least-significant bits of the number representation can easily be shifted out. Moreover, it has also been found that it is appropriate to allocate a small numeric weight to such context subregion values which are relevant for the numeric previous context value, but no longer relevant for the numeric current context value (or, equivalently, which are relevant for the numeric current context value, but no longer relevant for the numeric subsequent context value), because this allows for an efficient mapping of the numeric (current) context value onto a mapping rule index value.
In a preferred embodiment, the arithmetic decoder is configured to evaluate at least one table to determine whether the numeric current context value is identical to a table context value (for example, a significant state value) described by an entry of the table or lies within an interval described by entries of the table, and to derive a mapping rule index value describing a selected mapping rule in dependence on a result of an evaluation of the at least one table. It has been found that a numeric (current) context value, which is constructed and updated as described above, is well-suited for such a mapping onto a mapping rule index value.
In a preferred embodiment, the arithmetic decoder is configured to check whether a sum of a plurality of context subregion values is smaller than or equal to a predetermined sum threshold value, and to selectively modify the numeric current context value in dependence on a result of the check. It has been found that such an additional selective modification of the numeric current context value is well-suited to efficiently introduce meaningful context information into the numeric current context value without any conflict with respect to the concept for an updating of the numeric context value.
In a preferred embodiment, the arithmetic decoder is configured to check whether a sum of a plurality of context subregion values, which context subregion values are associated with a same temporal portion of the audio content as the one or more spectral values to be decoded using a context state defined by the numeric current context value, and which context subregion values are associated with lower frequencies than the one or more spectral values to be decoded using the context state defined by the numeric current context value, is smaller than or equal to a predetermined sum threshold value, and to selectively modify the numeric current context value in dependence on a result of the check. It has been found that such a check for identifying the presence of a region of comparatively small spectral values provides a valuable additional information.
In a preferred embodiment, the arithmetic decoder is configured to sum absolute values of a first plurality of previously decoded spectral values in order to obtain a first context subregion value associated with the first plurality of previously decoded spectral values, and to sum absolute values of a second plurality of previously decoded spectral values in order to obtain a second context subregion value associated with the second plurality of previously decoded spectral values. Accordingly, different context subregion values can be obtained.
In a preferred embodiment, the automatic decoder is configured to limit the context subregion values, such that the context subregion values are representable using a true subset of information bits of the number representation of the numeric previous context value. It has been found that a limitation of the context subregion values does not have a significant detrimental effect on the information content of the context subregion values. However, such a limitation brings along the advantage that a number of bits required to represent the context subregion value can be kept reasonably small, which has a positive impact on the memory demand. Also, the limitation of the context sub-region values facilitates the iterative update of the numeric context value.
Another embodiment according to the invention creates an audio encoder for providing an encoded audio information on the basis of input audio information. The audio encoder comprises an energy-compacting time-domain-to-frequency-domain converter for providing a frequency-domain audio representation on the basis of a time-domain representation of the input audio information, such that the frequency-domain audio representation comprises a set of spectral values. The audio encoder also comprises a arithmetic encoder configured to encode a spectral value, or a preprocessed version thereof, or—equivalently—a plurality of spectral values or a preprocessed version thereof, using a variable length codeword. The arithmetic encoder is configured to map a spectral value, or a value of a most significant bit plane of a spectral value, onto a code value. The arithmetic encoder is configured to select a mapping rule describing a mapping of a spectral value, or of a value of a most significant bit plane of a spectral value, onto a code value in dependence on a context state described by a numeric current context value. The arithmetic encoder is configured to determine the numeric current context value in dependence on a plurality of previously encoded spectral values. The arithmetic encoder is configured to modify a number representation of a numeric previous context value, describing a context state associated with one or more previously-encoded spectral values (or, more precisely, describing the context state for the encoding of said one or more previously-encoded spectral values), in dependence on a context subregion value, to obtain a number representation of a numeric current context value describing a context state associated with one or more spectral values to be encoded (or, more precisely, describing the context state for the encoding of said one or more spectral values to be encoded).
The audio encoder is based on the same findings as the audio decoder. Also, the audio encoder may be supplemented by the functionalities discussed with respect to the audio decoder.
Another embodiment according to the invention creates a method for providing a decoded audio information on the basis of an encoded audio information.
Another embodiment according to the invention creates a method for providing an encoded audio information on the basis of an input audio information.
Another embodiment according to the invention creates a computer program for performing one of said methods.
Embodiments of the present invention will be detailed subsequently referring to the appended drawings, in which:
a shows a pseudo-program-code representation of an algorithm “arith_map_context( )” for mapping a context;
b shows a pseudo-program-code representation of another algorithm “arith_map_context( )” for mapping a context;
c shows a pseudo-program-code representation of an algorithm “arith_get_context( )” for obtaining a context state value;
d shows a pseudo-program-code representation of another algorithm “arith_get_context( )” for obtaining a context state value;
e shows a pseudo-program-code representation of an algorithm “arith_get_pk( )” for deriving a cumulative-frequencies-table index value “pki” from a state value (or a state variable);
f shows a pseudo-program-code representation of another algorithm “arith_get_pk( )” for deriving a cumulative-frequencies-table index value “pki” from a state value (or a state variable);
g shows a pseudo-program-code representation of an algorithm “arith_decode( )” for arithmetically decoding a symbol from a variable length codeword;
h shows a first part of a pseudo-program-code representation of another algorithm “arith_decode ( )” for arithmetically decoding a symbol from a variable length codeword;
i shows a second part of a pseudo-program-code representation of the other algorithm “arith_decode( )” for arithmetically decoding a symbol from a variable length codeword;
j shows a pseudo-program-code representation of an algorithm for deriving absolute values a,b of spectral values from a common value m;
k shows a pseudo-program-code representation of an algorithm for entering the decoded values a,b into an array of decoded spectral values;
l shows a pseudo-program-code representation of an algorithm “arith_update_context( )” for obtaining a context subregion value on the basis of absolute values a,b of decoded spectral values;
m shows a pseudo-program-code representation of an algorithm “arith_finish( )” for filling entries of an array of decoded spectral values and an array of context subregion values;
n shows a pseudo-program-code representation of another algorithm for deriving absolute values a,b of decoded spectral values from a common value m;
o shows a pseudo-program-code representation of an algorithm “arith_update_context( )” for updating an array of decoded spectral values and an array of context subregion values;
p shows a pseudo-program-code representation of an algorithm “arith_save_context( )” for filling entries of an array of decoded spectral values and entries of an array of context subregion values;
q shows a legend of definitions;
r shows another legend of definitions;
a shows a syntax representation of a unified-speech-and-audio-coding (USAC) raw data block;
b shows a syntax representation of a single channel element;
c shows a syntax representation of a channel pair element;
d shows a syntax representation of an “ICS” control information;
e shows a syntax representation of a frequency-domain channel stream;
f shows a syntax representation of arithmetically coded spectral data;
g shows a syntax representation for decoding a set of spectral values;
h shows another syntax representation for decoding a set of spectral values;
i shows a legend of data elements and variables;
j shows another legend of data elements and variables;
a shows a schematic representation of a context for a state calculation, as it is used in accordance with working draft 4 of the USAC Draft Standard;
b shows an overview of the tables as used in the arithmetic coding scheme according to working draft 4 of the USAC Draft Standard;
a shows a schematic representation of a context for a state calculation, as it is used in embodiments according to the invention;
b shows an overview of the tables as used in the arithmetic coding scheme according to the present invention;
a shows a graphical representation of a read-only memory demand for the noiseless coding scheme according to the present invention, and according to working draft 5 of the USAC Draft Standard, and according to the AAC (advanced audio coding) Huffman Coding;
b shows a graphical representation of a total USAC decoder data read-only memory demand in accordance with the present invention and in accordance with the concept according to working draft 5 of the USAC Draft Standard;
FIGS. 21(1) and 21(2) show a table representation of a content of a table “ari_lookup_m[600]”;
FIGS. 22(1) to 22(4) show a table representation of a content of a table “ari_hash_m[600]”;
FIGS. 23(1) to 23(8) show a table representation of a content of a table “ari_cf_m[96][17]”; and
1. Audio Encoder According to
The arithmetic encoder 730 is configured to map a spectral value, or a value of a most-significant bit-plane of a spectral value, onto a code value (i.e. onto a variable-length codeword) in dependence on a context state. The arithmetic encoder is configured to select a mapping rule describing a mapping of a spectral value, or of a most-significant bit-plane of a spectral value, onto a code value, in dependence on a (current) context state. The arithmetic encoder is configured to determine the current context state, or a numeric current context value describing the current context state, in dependence on a plurality of previously-encoded (preferably, but not necessarily, adjacent) spectral values. For this purpose, the arithmetic encoder is configured to evaluate a hash-table, entries of which define both significant state values amongst the numeric context values and boundaries of intervals of numeric context values, wherein a mapping rule index value is individually associated to a numeric (current) context value being a significant state value, and wherein a common mapping rule index value is associated to different numeric (current) context values lying within an interval bounded by interval boundaries (wherein the interval boundaries are preferably defined by the entries of the hash table).
As can be seen, the mapping of a spectral value (of the frequency-domain audio representation 722), or of a most-significant bit-plane of a spectral value, onto a code value (of the encoded audio information 712), may be performed by a spectral value encoding 740 using a mapping rule 742. A state tracker 750 may be configured to track the context state. The state tracker 750 provides an information 754 describing the current context state. The information 754 describing the current context state may preferably take the form of a numeric current context value. A mapping rule selector 760 is configured to select a mapping rule, for example, a cumulative-frequencies-table, describing a mapping of a spectral value, or of a most-significant bit-plane of a spectral value, onto a code value. Accordingly, the mapping rule selector 760 provides the mapping rule information 742 to the spectral value encoding 740. The mapping rule information 742 may take the form of a mapping rule index value or of a cumulative-frequencies-table selected in dependence on a mapping rule index value. The mapping rule selector 760 comprises (or at least evaluates) a hash-table 752, entries of which define both significant state values amongst the numeric context values and boundaries and intervals of numeric context values, wherein a mapping rule index value is individually associated to a numeric context value being a significant state value, and wherein a common mapping rule index value is associated to different numeric context values lying within an interval bounded by interval boundaries. The hash-table 762 is evaluated in order to select the mapping rule, i.e. in order to provide the mapping rule information 742.
To summarize the above, the audio encoder 700 performs an arithmetic encoding of a frequency-domain audio representation provided by the time-domain-to-frequency-domain converter. The arithmetic encoding is context-dependent, such that a mapping rule (e.g. a cumulative-frequencies-table) is selected in dependence on previously encoded spectral values. Accordingly, spectral values adjacent in time and/or frequency (or, at least, within a predetermined environment) to each other and/or to the currently-encoded spectral value (i.e. spectral values within a predetermined environment of the currently encoded spectral value) are considered in the arithmetic encoding to adjust the probability distribution evaluated by the arithmetic encoding. When selecting an appropriate mapping rule, numeric context current values 754 provided by a state tracker 750 are evaluated. As typically the number of different mapping rules is significantly smaller than the number of possible values of the numeric current context values 754, the mapping rule selector 760 allocates the same mapping rules (described, for example, by a mapping rule index value) to a comparatively large number of different numeric context values. Nevertheless, there are typically specific spectral configurations (represented by specific numeric context values) to which a particular mapping rule should be associated in order to obtain a good coding efficiency.
It has been found that the selection of a mapping rule in dependence on a numeric current context value can be performed with particularly high computational efficiency if entries of a single hash-table define both significant state values and boundaries of intervals of numeric (current) context values. It has been found that this mechanism is well-adapted to the requirements of the mapping rule selection, because there are many cases in which a single significant state value (or significant numeric context value) is embedded between a left-sided interval of a plurality of non-significant state values (to which a common mapping rule is associated) and a right-sided interval of a plurality of non-significant state values (to which a common mapping rule is associated). Also, the mechanism of using a single hash-table, entries of which define both significant state values and boundaries of intervals of numeric (current) context values can efficiently handle different cases, in which, for example, there are two adjacent intervals of non-significant state values (also designated as non-significant numeric context values) without a significant state value in between. A particularly high computational efficiency is achieved due to a number of table accesses being kept small. For example, a single iterative table search is sufficient in most embodiments in order to find out whether the numeric current context value is equal to any of the significant state values, or in which of the intervals of non-significant state values the numeric current context value lays. Consequently, the number of table accesses which are both, time-consuming and energy-consuming, can be kept small. Thus, the mapping rule selector 760, which uses the hash-table 762, may be considered as a particularly efficient mapping rule selector in terms of computational complexity, while still allowing to obtain a good encoding efficiency (in terms of bitrate).
Further details regarding the derivation of the mapping rule information 742 from the numeric current context value 754 will be described below.
2. Audio Decoder According to
The arithmetic decoder 820 comprises a spectral value determinator 824, which is configured to map a code value of the arithmetically-encoded representation 821 of spectral values onto a symbol code representing one or more of the decoded spectral values, or at least a portion (for example, a most-significant bit-plane) of one or more of the decoded spectral values. The spectral value determinator 824 may be configured to perform a mapping in dependence on a mapping rule, which may be described by a mapping rule information 828a. The mapping rule information 828a may, for example, take the form of a mapping rule index value, or of a selected cumulative-frequencies-table (selected, for example, in dependence on a mapping rule index value).
The arithmetic decoder 820 is configured to select a mapping rule (e.g. a cumulative-frequencies-table) describing a mapping of code values (described by the arithmetically-encoded representation 821 of spectral values) onto a symbol code (describing one or more spectral values, or a most-significant bit-plane thereof) in dependence on a context state (which may be described by the context state information 826a). The arithmetic decoder 820 is configured to determine the current context state (described by the numeric current context value) in dependence on a plurality of previously-decoded spectral values. For this purpose, a state tracker 826 may be used, which receives an information describing the previously-decoded spectral values and which provides, on the basis thereof, a numeric current context value 826a describing the current context state.
The arithmetic decoder is also configured to evaluate a hash-table 829, entries of which define both significant state values amongst the numeric context values and boundaries of intervals of numeric context values, in order to select the mapping rule, wherein a mapping rule index value is individually associated to a numeric context value being a significant state value, and wherein a common mapping rule index value is associated to different numeric context values lying within an interval bounded by interval boundaries. The evaluation of the hash-table 829 may, for example, be performed using a hash-table evaluator which may be part of the mapping rule selector 828. Accordingly, a mapping rule information 828a, for example, in the form of a mapping rule index value, is obtained on the basis of the numeric current context value 826a describing the current context state. The mapping rule selector 828 may, for example, determine the mapping rule index value 828a in dependence on a result of the evaluation of the hash-table 829. Alternatively, the evaluation of the hash-table 829 may directly provide the mapping rule index value.
Regarding the functionality of the audio signal decoder 800, it should be noted that the arithmetic decoder 820 is configured to select a mapping rule (e.g. a cumulative-frequencies-table) which is, on average, well adapted to the spectral values to be decoded, as the mapping rule is selected in dependence on the current context state (described, for example, by the numeric current context value), which in turn is determined in dependence on a plurality of previously-decoded spectral values. Accordingly, statistical dependencies between adjacent spectral values to be decoded can be exploited. Moreover, the arithmetic decoder 820 can be implemented efficiently, with a good trade-off between computational complexity, table size, and coding efficiency, using the mapping rule selector 828. By evaluating a (single) hash-table 829, entries of which describe both significant state values and interval boundaries of intervals of non-significant state values, a single iterative table search may be sufficient in order to derive the mapping rule information 828a from the numeric current context value 826a. Accordingly, it is possible to map a comparatively large number of different possible numeric (current) context values onto a comparatively smaller number of different mapping rule index values. By using the hash-table 829, as described above, it is possible to exploit the finding that, in many cases, a single isolated significant state value (significant context value) is embedded between a left-sided interval of non-significant state values (non-significant context values) and a right-sided interval of non-significant state values (non-significant context values), wherein a different mapping rule index value is associated with the significant state value (significant context value), when compared to the state values (context values) of the left-sided interval and the state values (context values) of the right-sided interval. However, usage of the hash-table 829 is also well-suited for situations in which two intervals of numeric state values are immediately adjacent, without a significant state value in between.
To conclude, the mapping rule selector 828, which evaluates the hash-table 829, brings along a particularly good efficiency when selecting a mapping rule (or when providing a mapping rule index value) in dependence on the current context state (or in dependence on the numeric current context value describing the current context state), because the hashing mechanism is well-adapted to the typical context scenarios in an audio decoder.
Further details will be described below.
3. Context Value Hashing Mechanism According to
In the following, a context hashing mechanism will be disclosed, which may be implemented in the mapping rule selector 760 and/or the mapping rule selector 828. The hash-table 762 and/or the hash-table 829 may be used in order to implement said context value hashing mechanism.
Taking reference now to
As can be seen, a hash-table entry “ari_hash_m[i1]” describes an individual (true) significant state having a numeric context value of c1. As can be seen, the mapping rule index value mriv1 is associated to the individual (true) significant state having the numeric context value c1. Accordingly, both the numeric context value c1 and the mapping rule index value mriv1 may be described by the hash-table entry “ari_hash_m[i1]”. An interval 932 of numeric context values is bounded by the numeric context value c1, wherein the numeric context value c1 does not belong to the interval 932, such that the largest numeric context value of interval 932 is equal to c1−1. A mapping rule index value of mriv4 (which is different from mriv1) is associated with the numeric context values of the interval 932. The mapping rule index value mriv4 may, for example, be described by the table entry “ari_lookup_m[i1−1]” of an additional table “ari_lookup_m”.
Moreover, a mapping rule index value mriv2 may be associated with numeric context values lying within an interval 934. A lower bound of interval 934 is determined by the numeric context value c1, which is a significant numeric context value, wherein the numeric context value c1 does not belong to the interval 932. Accordingly, the smallest value of the interval 934 is equal to c1+1 (assuming integer numeric context values). Another boundary of the interval 934 is determined by the numeric context value c2, wherein the numeric context value c2 does not belong to the interval 934, such that the largest value of the interval 934 is equal to c2−1. The numeric context value c2 is a so-called “improper” numeric context value, which is described by a hash-table entry “ari_hash_m[i2]”. For example, the mapping rule index value mriv2 may be associated with the numeric context value c2, such that the numeric context value associated with the “improper” significant numeric context value c2 is equal to the mapping rule index value associated with the interval 934 bounded by the numeric context value c2. Moreover, an interval 936 of numeric context value is also bounded by the numeric context value c2, wherein the numeric context value c2 does not belong to the interval 936, such that the smallest numeric context value of the interval 936 is equal to c2+1. A mapping rule index value mriv3, which is typically different from the mapping rule index value mriv2, is associated with the numeric context values of the interval 936.
As can be seen, the mapping rule index value mriv4, which is associated to the interval 932 of numeric context values, may be described by an entry “ari_lookup_m[i1−1]” of a table “ari_lookup_m”, the mapping rule index mriv2, which is associated with the numeric context values of the interval 934, may be described by a table entry “ari_lookup_m[i1]” of the table “ari_lookup_m”, and the mapping rule index value mriv3 may be described by a table entry “ari_lookup_m[i2]” of the table “ari_lookup_m”. In the example given here, the hash-table index value i2, may be larger, by 1, than the hash-table index value i1.
As can be seen from
Moreover, the evaluation of the hash-table “ari_hash_m” may be used to obtain a hash-table index value (for example, i1−1, i1 or i2). Thus, the mapping rule selector 760, 828 may be configured to obtain, by evaluating a single hash-table 762, 829 (for example, the hash-table “ari_hash_m”), a hash-table index value (for example, i1−1, i1 or i2) designating a significant state value (e.g., c1 or c2) and/or an interval (e.g., 932,934,936) and an information as to whether the numeric current context value is a significant context value (also designated as significant state value) or not.
Moreover, if it is found in the evaluation of the hash-table 762, 829, “ari_hash_m”, that the numeric current context value is not a “significant” context value (or “significant” state value), the hash-table index value (for example, i1−1, i1 or i2) obtained from the evaluation of the hash-table (“ari_hash_m”) may be used to obtain a mapping rule index value associated with an interval 932, 934, 936 of numeric context values. For example, the hash-table index value (e.g., i1−1, i1 or i2) may be used to designate an entry of an additional mapping table (for example, “ari_lookup_m”), which describes the mapping rule index values associated with the interval 932, 934, 936 within which the numeric current context value lies.
For further details, reference is made to the detailed discussion below of the algorithm “arith_get_pk” (wherein there are different options for this algorithm “arith_get_pk( )”, examples of which are shown in
Moreover, it should be noted that the size of the intervals may differ from one case to another. In some cases, an interval of numeric context values comprises a single numeric context value. However, in many cases, an interval may comprise a plurality of numeric context values.
4. Audio Encoder According to
The audio encoder 1000 is configured to receive an input audio information 710 and to provide, on the basis thereof, an encoded audio information 712. The audio encoder 1000 comprises an energy-compacting time-domain-to-frequency-domain converter 720, which is configured to provide a frequency-domain representation 722 on the basis of a time-domain representation of the input audio information 710, such that the frequency-domain audio representation 722 comprises a set of spectral values. The audio encoder 1000 also comprises an arithmetic encoder 1030 configured to encode a spectral value (out of the set of spectral values forming the frequency-domain audio representation 722), or a pre-processed version thereof, using a variable-length codeword to obtain the encoded audio information 712 (which may comprise, for example, a plurality of variable-length codewords).
The arithmetic encoder 1030 is configured to map a spectral value, or a plurality of spectral values, or a value of a most-significant bit-plane of a spectral value or of a plurality of spectral values, onto a code value (i.e. onto a variable-length codeword) in dependence on a context state. The arithmetic encoder 1030 is configured to select a mapping rule describing a mapping of a spectral value, or of a plurality of spectral values, or of a most-significant bit-plane of a spectral value or of a plurality of spectral values, onto a code value in dependence on a context state. The arithmetic encoder is configured to determine the current context state in dependence on a plurality of previously-encoded (preferably, but no necessarily adjacent) spectral values. For this purpose, the arithmetic encoder is configured to modify a number representation of a numeric previous context value, describing a context state associated with one or more previously-encoded spectral values (for example, to select a corresponding mapping rule), in dependence on a context sub-region value, to obtain a number representation of a numeric current context value describing a context state associated with one or more spectral values to be encoded (for example, to select a corresponding mapping rule).
As can be seen, the mapping of a spectral value, or of a plurality of spectral values, or of a most-significant bit-plane of a spectral value or of a plurality of spectral values, onto a code value may be performed by a spectral value encoding 740 using a mapping rule described by a mapping rule information 742. A state tracker 750 may be configured to track the context state. The state tracker 750 may be configured to modify a number representation of a numeric previous context value, describing a context state associated with an encoding of one or more previously-encoded spectral values, in dependence on a context sub-region value, to obtain a number representation of a numeric current context value describing a context state associated with an encoding of one or more spectral values to be encoded. The modification of the number representation of the numeric previous context value may, for example, be performed by a number representation modifier 1052, which receives the numeric previous context value and one or more context sub-region values and provides the numeric current context value. Accordingly, the state tracker 1050 provides an information 754 describing the current context state, for example, in the form of a numeric current context value. A mapping rule selector 1060 may select a mapping rule, for example, a cumulative-frequencies-table, describing a mapping of a spectral value, or of a plurality of spectral values, or of a most-significant bit-plane of a spectral value or of a plurality of spectral values, onto a code value. Accordingly, the mapping rule selector 1060 provides the mapping rule information 742 to the spectral encoding 740.
It should be noted that, in some embodiments, the state tracker 1050 may be identical to the state tracker 750 or the state tracker 826. It should also be noted that the mapping rule selector 1060 may, in some embodiments, be identical to the mapping rule selector 760, or the mapping rule selector 828.
To summarize the above, the audio encoder 1000 performs an arithmetic encoding of a frequency-domain audio representation provided by the time-domain-to-frequency-domain converter. The arithmetic encoding is context dependent, such that a mapping rule (e.g. a cumulative-frequencies-table) is selected in dependence on previously-encoded spectral values. Accordingly, spectral values adjacent in time and/or frequency (or at least within a predetermined environment) to each other and/or to the currently-encoded spectral value (i.e. spectral values within a predetermined environment of the currently-encoded spectral value) are considered in the arithmetic encoding to adjust the probability distribution evaluated by the arithmetic encoding.
When determining the numeric current context value, a number representation of a numeric previous context value, describing a context state associated with one or more previously-encoded spectral values, is modified in dependence on a context sub-region value, to obtain a number representation of a numeric current context value describing a context state associated with one or more spectral values to be encoded. This approach allows avoiding a complete re-computation of the numeric current context value, which complete re-computation consumes a significant amount of resources in conventional approaches. A large variety of possibilities exist for the modification of the number representation of the numeric previous context value, including a combination of a re-scaling of a number representation of the numeric previous context value, an addition of a context sub-region value or a value derived therefrom to the number representation of the numeric previous context value or to a processed number representation of the numeric previous context value, a replacement of a portion of the number representation (rather than the entire number representation) of the numeric previous context value in dependence on the context sub-region value, and so on. Thus, typically the numeric representation of the numeric current context value is obtained on the basis of the number representation of the numeric previous context value and also on the basis of at least one context sub-region value, wherein typically a combination of operations are performed to combine the numeric previous context value with a context sub-region value, such as for example, two or more operations out of an addition operation, a subtraction operation, a multiplication operation, a division operation, a Boolean-AND operation, a Boolean-OR operation, a Boolean-NAND operation, a Boolean NOR operation, a Boolean-negation operation, a complement operation or a shift operation. Accordingly, at least a portion of the number representation of the numeric previous context value is typically maintained unchanged (except for an optional shift to a different position) when deriving the numeric current context value from the numeric previous context value. In contrast, other portions of the number representation of the numeric previous context value are changed in dependence on one or more context sub-region values. Thus, the numeric current context value can be obtained with a comparatively small computational effort, while avoiding a complete re-computation of the numeric current context value.
Thus, a meaningful numeric current context value can be obtained, which is well-suited for the use by the mapping rule selector 1060.
Consequently, an efficient encoding can be achieved by keeping the context calculation sufficiently simple.
5. Audio Decoder According to
The audio decoder 1100 is configured to receive an encoded audio information 810 and to provide, on the basis thereof, a decoded audio information 812. The audio decoder 1100 comprises an arithmetic decoder 1120 that is configured to provide a plurality of decoded spectral values 822 on the basis of an arithmetically-encoded representation 821 of the spectral values. The audio decoder 1100 also comprises a frequency-domain-to-time-domain converter 830 which is configured to receive the decoded spectral values 822 and to provide the time-domain audio representation 812, which may constitute the decoded audio information, using the decoded spectral values 822, in order to obtain a decoded audio information 812.
The arithmetic decoder 1120 comprises a spectral value determinator 824, which is configured to map a code value of the arithmetically-encoded representation 821 of spectral values onto a symbol code representing one or more of the decoded spectral values, or at least a portion (for example, a most-significant bit-plane) of one or more of the decoded spectral values. The spectral value determinator 824 may be configured to perform the mapping in dependence on a mapping rule, which may be described by a mapping rule information 828a. The mapping rule information 828a may, for example, comprise a mapping rule index value, or may comprise a selected set of entries of a cumulative-frequencies-table.
The arithmetic decoder 1120 is configured to select a mapping rule (e.g., a cumulative-frequencies-table) describing a mapping of a code value (described by the arithmetically-encoded representation 821 of spectral values) onto a symbol code (describing one or more spectral values) in dependence on a context state, which context state may be described by the context state information 1126a. The context state information 1126a may take the form of a numeric current context value. The arithmetic decoder 1120 is configured to determine the current context state in dependence on a plurality of previously-decoded spectral values 822. For this purpose, a state tracker 1126 may be used, which receives an information describing the previously-decoded spectral values. The arithmetic decoder is configured to modify a number representation of numeric previous context value, describing a context state associated with one or more previously decoded spectral values, in dependence on a context sub-region value, to obtain a number representation of a numeric current context value describing a context state associated with one or more spectral values to be decoded. A modification of the number representation of the numeric previous context value may, for example, be performed by a number representation modifier 1127, which is part of the state tracker 1126. Accordingly, the current context state information 1126a is obtained, for example, in the form of a numeric current context value. The selection of the mapping rule may be performed by a mapping rule selector 1128, which derives a mapping rule information 828a from the current context state information 1126a, and which provides the mapping rule information 828a to the spectral value determinator 824.
Regarding the functionality of the audio signal decoder 1100, it should be noted that the arithmetic decoder 1120 is configured to select a mapping rule (e.g., a cumulative-frequencies-table) which is, on average, well-adapted to the spectral value to be decoded, as the mapping rule is selected in dependence on the current context state, which, in turn, is determined in dependence on a plurality of previously-decoded spectral values. Accordingly, statistical dependencies between adjacent spectral values to be decoded can be exploited.
Moreover, by modifying a number representation of a numeric previous context value describing a context state associated with a decoding of one or more previously decoded spectral values, in dependence on a context sub-region value, to obtain a number representation of a numeric current context value describing a context state associated with a decoding of one or more spectral values to be decoded, it is possible to obtain a meaningful information about the current context state, which is well-suited for a mapping to a mapping rule index value, with comparatively small computational effort. By maintaining at least a portion of a number representation of the numeric previous context value (possibly in a bit-shifted or a scaled version) while updating another portion of the number representation of the numeric previous context value in dependence on the context sub-region values which have not been considered in the numeric previous context value but which should be considered in the numeric current context value, a number of operations to derive the numeric current context value can be kept reasonably small. Also, it is possible to exploit the fact that contexts used for decoding adjacent spectral values are typically similar or correlated. For example, a context for a decoding of a first spectral value (or of a first plurality of spectral values) is dependent on a first set of previously-decoded spectral values. A context for decoding of a second spectral value (or a second set of spectral values), which is adjacent to the first spectral value (or the first set of spectral values) may comprise a second set of previously-decoded spectral values. As the first spectral value and the second spectral value are assumed to be adjacent (e.g., with respect to the associated frequencies), the first set of spectral values, which determine the context for the coding of the first spectral value, may comprise some overlap with the second set of spectral values, which determine the context for the decoding of the second spectral value. Accordingly, it can easily be understood that the context state for the decoding of the second spectral value comprises some correlation with the context state for the decoding of the first spectral value. A computational efficiency of the context derivation, i.e. of the derivation of the numeric current context value, can be achieved by exploiting such correlations. It has been found that the correlation between context states for a decoding of adjacent spectral values (e.g., between the context state described by the numeric previous context value and the context state described by the numeric current context value) can be exploited efficiently by modifying only those parts of the numeric previous context value which are dependent on context sub-region values not considered for the derivation of the numeric previous context state, and by deriving the numeric current context value from the numeric previous context value.
To conclude, the concepts described herein allow for a particularly good computational efficiency when deriving the numeric current context value.
Further details will be described below.
6. Audio Encoder According to
The audio encoder 1200 is configured to receive an input audio information 710 and to provide, on the basis thereof, an encoded audio information 712. The audio encoder 1200 comprises an energy-compacting time-domain-to-frequency-domain converter 720 which is configured to provide a frequency-domain audio representation 722 on the basis of a time-domain audio representation of the input audio information 710, such that the frequency-domain audio representation 722 comprises a set of spectral values. The audio encoder 1200 also comprises an arithmetic encoder 1230 configured to encode a spectral value (out of the set of spectral values forming the frequency-domain audio representation 722), or a plurality of spectral values, or a pre-processed version thereof, using a variable-length codeword to obtain the encoded audio information 712 (which may comprise, for example, a plurality of variable-length codewords.
The arithmetic encoder 1230 is configured to map a spectral value, or a plurality of spectral values, or a value of a most-significant bit-plane of a spectral value or of a plurality of spectral values, onto a code value (i.e. onto a variable-length codeword), in dependence on a context state. The arithmetic encoder 1230 is configured to select a mapping rule describing a mapping of a spectral value, or of a plurality of spectral values, or of a most-significant bit-plane of a spectral value or of a plurality of spectral values, onto a code value, in dependence on the context state. The arithmetic encoder is configured to determine the current context state in dependence on a plurality of previously-encoded (preferably, but not necessarily, adjacent) spectral values. For this purpose, the arithmetic encoder is configured to obtain a plurality of context sub-region values on the basis of previously-encoded spectral values, to store said context sub-region values, and to derive a numeric current context value associated with one or more spectral values to be encoded in dependence on the stored context sub-region vales. Moreover, the arithmetic encoder is configured to compute the norm of a vector formed by a plurality of previously encoded spectral values, in order to obtain a common context sub-region value associated with the plurality of previously-encoded spectral values.
As can be seen, the mapping of a spectral value, or of a plurality of spectral values, or of a most-significant bit-plane of a spectral value or of a plurality of spectral values, onto a code value may be performed by a spectral value encoding 740 using a mapping rule described by a mapping rule information 742. A state tracker 1250 may be configured to track the context state and may comprise a context sub-region value computer 1252, to compute the norm of a vector formed by a plurality of previously encoded spectral values, in order to obtain a common context sub-region values associated with the plurality of previously-encoded spectral values. The state tracker 1250 is also preferably configured to determine the current context state in dependence on a result of said computation of a context sub-region value performed by the context sub-region value computer 1252. Accordingly, the state tracker 1250 provides an information 1254, describing the current context state. A mapping rule selector 1260 may select a mapping rule, for example, a cumulative-frequencies-table, describing a mapping of a spectral value, or of a most-significant bit-plane of a spectral value, onto a code value. Accordingly, the mapping rule selector 1260 provides the mapping rule information 742 to the spectral encoding 740.
To summarize the above, the audio encoder 1200 performs an arithmetic encoding of a frequency-domain audio representation provided by the time-domain-to-frequency-domain converter 720. The arithmetic encoding is context-dependent, such that a mapping rule (e.g., a cumulative-frequencies-table) is selected in dependence on previously-encoded spectral values. Accordingly, spectral values adjacent in time and/or frequency (or, at least, within a predetermined environment) to each other and/or to the currently-encoded spectral value (i.e. spectral values within a predetermined environment of the currently encoded spectral value) are considered in the arithmetic encoding to adjust the probability distribution evaluated by the arithmetic encoding.
In order to provide a numeric current context value, a context sub-region value associated with a plurality of previously-encoded spectral values is obtained on the basis of a computation of a norm of a vector formed by a plurality of previously-encoded spectral values. The result of the determination of the numeric current context value is applied in the selection of the current context state, i.e. in the selection of a mapping rule.
By computing the norm of a vector formed by a plurality of previously-encoded spectral values, a meaningful information describing a portion of the context of the one or more spectral values to be encoded can be obtained, wherein the norm of a vector of previously encoded spectral values can typically be represented with a comparatively small number of bits. Thus, the amount of context information, which needs to be stored for later use in the derivation of a numeric current context value, can be kept sufficiently small by applying the above discussed approach for the computation of the context sub-region values. It has been found that the norm of a vector of previously encoded spectral values typically comprises the most significant information regarding the state of the context. In contrast, it has been found that the sign of said previously encoded spectral values typically comprises a subordinate impact on the state of the context, such that it makes sense to neglect the sign of the previously decoded spectral values in order to reduce the quantity of information to be stored for later use. Also, it has been found that the computation of a norm of a vector of previously-encoded spectral values is a reasonable approach for the derivation of a context sub-region value, as the averaging effect, which is typically obtained by the computation of the norm, leaves the most important information about the context state substantially unaffected. To summarize, the context sub-region value computation performed by the context sub-region value computer 1252 allows for providing a compact context sub-region information for storage and later re-use, wherein the most relevant information about the context state is preserved in spite of the reduction of the quantity of information.
Accordingly, an efficient encoding of the input audio information 710 can be achieved, while keeping the computational effort and the amount of data to be stored by the arithmetic encoder 1230 sufficiently small.
7. Audio Decoder According to
The audio decoder 1300 is configured to receive an encoded audio information 810 and to provide, on the basis thereof, a decoded audio information 812. The audio decoder 1300 comprises an arithmetic decoder 1320 that is configured to provide a plurality of decoded spectral values 822 on the basis of an arithmetically-encoded representation 821 of the spectral values. The audio decoder 1300 also comprises a frequency-domain-to-time-domain converter 830 which is configured to receive the decoded spectral values 822 and to provide the time-domain audio representation 812, which may constitute the decoded audio information, using the decoded spectral values 822, in order to obtain a decoded audio information 812.
The arithmetic decoder 1320 comprises a spectral value determinator 824 which is configured to map a code value of the arithmetically-encoded representation 821 of spectral values onto a symbol code representing one or more of the decoded spectral values, or at least a portion (e.g. a most-significant bit-plane) of one or more of the decoded spectral values. The spectral value determinator 824 may be configured to perform a mapping in dependence on a mapping rule, which is described by a mapping rule information 828a. The mapping rule information 828a may, for example, comprise a mapping rule index value, or a selected set of entries of a cumulative-frequencies-table.
The arithmetic decoder 1320 is configured to select a mapping rule (e.g., a cumulative-frequencies-table) describing a mapping of a code value (described by the arithmetically-encoded representation 821 of spectral values) onto a symbol code (describing one or more spectral values) in dependence on a context state (which may be described by the context state information 1326a). The arithmetic decoder 1320 is configured to determine the current context state in dependence on a plurality of previously-decoded spectral values 822. For this purpose, a state tracker 1326 may be used, which receives an information describing the previously-decoded spectral values. The arithmetic decoder is also configured to obtain a plurality of context sub-region values on the basis of previously-decoded spectral values and to store said context sub-region values. The arithmetic decoder is configured to derive a numeric current context value associated with one or more spectral values to be decoded in dependence on the stored context sub-region values. The arithmetic decoder 1320 is configured to compute the norm of a vector formed by a plurality of previously decoded spectral values, in order to obtain a common context sub-region value associated with the plurality of previously-decoded spectral values.
The computation of the norm of a vector formed by a plurality of previously-encoded spectral values, in order to obtain a common context sub-region value associated with the plurality of previously decoded spectral values, may, for example, be performed by the context sub-region value computer 1327, which is part of the state tracker 1326. Accordingly, a current context state information 1326a is obtained on the basis of the context sub-region values, wherein the state tracker 1326 preferably provides a numeric current context value associated with one or more spectral values to be decoded in dependence on the stored context sub-region values. The selection of the mapping rules may be performed by a mapping rule selector 1328, which derives a mapping rule information 828a from the current context state information 1326a, and which provides the mapping rule information 828a to the spectral value determinator 824.
Regarding the functionality of the audio signal decoder 1300, it should be noted that the arithmetic decoder 1320 is configured to select a mapping rule (e.g., a cumulative-frequencies-table) which is, on average, well-adapted to the spectral value to be decoded, as the mapping rule is selected in dependence on the current context state, which, in turn, is determined in dependence on a plurality of previously-decoded spectral values. Accordingly, statistical dependencies between adjacent spectral values to be decoded can be exploited.
However, it has been found that it is efficient, in terms of memory usage, to store context sub-region values, which are based on the computation of a norm of a vector formed on a plurality of previously decoded spectral values, for later use in the determination of the numeric context value. It has also been found that such context sub-region values still comprise the most relevant context information. Accordingly, the concept used by the state tracker 1326 constitutes a good compromise between coding efficiency, computational efficiency and storage efficiency.
Further details will be described below.
8. Audio Encoder According to
In the following, an audio encoder according to an embodiment of the present invention will be described.
The audio encoder 100 is configured to receive an input audio information 110 and to provide, on the basis thereof, a bitstream 112, which constitutes an encoded audio information. The audio encoder 100 optionally comprises a preprocessor 120, which is configured to receive the input audio information 110 and to provide, on the basis thereof, a pre-processed input audio information 110a. The audio encoder 100 also comprises an energy-compacting time-domain to frequency-domain signal transformer 130, which is also designated as signal converter. The signal converter 130 is configured to receive the input audio information 110, 110a and to provide, on the basis thereof, a frequency-domain audio information 132, which preferably takes the form of a set of spectral values. For example, the signal transformer 130 may be configured to receive a frame of the input audio information 110, 110a (e.g. a block of time-domain samples) and to provide a set of spectral values representing the audio content of the respective audio frame. In addition, the signal transformer 130 may be configured to receive a plurality of subsequent, overlapping or non-overlapping, audio frames of the input audio information 110, 110a and to provide, on the basis thereof, a time-frequency-domain audio representation, which comprises a sequence of subsequent sets of spectral values, one set of spectral values associated with each frame.
The energy-compacting time-domain to frequency-domain signal transformer 130 may comprise an energy-compacting filterbank, which provides spectral values associated with different, overlapping or non-overlapping, frequency ranges. For example, the signal transformer 130 may comprise a windowing MDCT transformer 130a, which is configured to window the input audio information 110, 110a (or a frame thereof) using a transform window and to perform a modified-discrete-cosine-transform of the windowed input audio information 110, 110a (or of the windowed frame thereof). Accordingly, the frequency-domain audio representation 132 may comprise a set of, for example, 1024 spectral values in the form of MDCT coefficients associated with a frame of the input audio information.
The audio encoder 100 may further, optionally, comprise a spectral post-processor 140, which is configured to receive the frequency-domain audio representation 132 and to provide, on the basis thereof, a post-processed frequency-domain audio representation 142. The spectral post-processor 140 may, for example, be configured to perform a temporal noise shaping and/or a long term prediction and/or any other spectral post-processing known in the art. The audio encoder further comprises, optionally, a scaler/quantizer 150, which is configured to receive the frequency-domain audio representation 132 or the post-processed version 142 thereof and to provide a scaled and quantized frequency-domain audio representation 152.
The audio encoder 100 further comprises, optionally, a psycho-acoustic model processor 160, which is configured to receive the input audio information 110 (or the post-processed version 110a thereof) and to provide, on the basis thereof, an optional control information, which may be used for the control of the energy-compacting time-domain to frequency-domain signal transformer 130, for the control of the optional spectral post-processor 140 and/or for the control of the optional scaler/quantizer 150. For example, the psycho-acoustic model processor 160 may be configured to analyze the input audio information, to determine which components of the input audio information 110, 110a are particularly important for the human perception of the audio content and which components of the input audio information 110, 110a are less important for the perception of the audio content. Accordingly, the psycho-acoustic model processor 160 may provide control information, which is used by the audio encoder 100 in order to adjust the scaling of the frequency-domain audio representation 132, 142 by the scaler/quantizer 150 and/or the quantization resolution applied by the scaler/quantizer 150. Consequently, perceptually important scale factor bands (i.e. groups of adjacent spectral values which are particularly important for the human perception of the audio content) are scaled with a large scaling factor and quantized with comparatively high resolution, while perceptually less-important scale factor bands (i.e. groups of adjacent spectral values) are scaled with a comparatively smaller scaling factor and quantized with a comparatively lower quantization resolution. Accordingly, scaled spectral values of perceptually more important frequencies are typically significantly larger than spectral values of perceptually less important frequencies.
The audio encoder also comprises an arithmetic encoder 170, which is configured to receive the scaled and quantized version 152 of the frequency-domain audio representation 132 (or, alternatively, the post-processed version 142 of the frequency-domain audio representation 132, or even the frequency-domain audio representation 132 itself) and to provide arithmetic codeword information 172a on the basis thereof, such that the arithmetic codeword information represents the frequency-domain audio representation 152.
The audio encoder 100 also comprises a bitstream payload formatter 190, which is configured to receive the arithmetic codeword information 172a. The bitstream payload formatter 190 is also typically configured to receive additional information, like, for example, scale factor information describing which scale factors have been applied by the scaler/quantizer 150. In addition, the bitstream payload formatter 190 may be configured to receive other control information. The bitstream payload formatter 190 is configured to provide the bitstream 112 on the basis of the received information by assembling the bitstream in accordance with a desired bitstream syntax, which will be discussed below.
In the following, details regarding the arithmetic encoder 170 will be described. The arithmetic encoder 170 is configured to receive a plurality of post-processed and scaled and quantized spectral values of the frequency-domain audio representation 132. The arithmetic encoder comprises a most-significant-bit-plane-extractor 174, or even from two spectral values, which is configured to extract a most-significant bit-plane m from a spectral value. It should be noted here that the most-significant bit-plane may comprise one or even more bits (e.g. two or three bits), which are the most-significant bits of the spectral value. Thus, the most-significant bit-plane extractor 174 provides a most-significant bit-plane value 176 of a spectral value.
Alternatively, however, the most significant bit-plane extractor 174 may provide a combined most-significant bit-plane value m combining the most-significant bit-planes of a plurality of spectral values (e.g., of spectral values a and b). The most-significant bit-plane of the spectral value a is designated with m. Alternatively, the combined most-significant bit-plane value of a plurality of spectral values a,b is designated with m.
The arithmetic encoder 170 also comprises a first codeword determinator 180, which is configured to determine an arithmetic codeword acod_m [pki][m] representing the most-significant bit-plane value m. Optionally, the codeword determinator 180 may also provide one or more escape codewords (also designated herein with “ARITH_ESCAPE”) indicating, for example, how many less-significant bit-planes are available (and, consequently, indicating the numeric weight of the most-significant bit-plane). The first codeword determinator 180 may be configured to provide the codeword associated with a most-significant bit-plane value m using a selected cumulative-frequencies-table having (or being referenced by) a cumulative-frequencies-table index pki.
In order to determine as to which cumulative-frequencies-table should be selected, the arithmetic encoder preferably comprises a state tracker 182, which is configured to track the state of the arithmetic encoder, for example, by observing which spectral values have been encoded previously. The state tracker 182 consequently provides a state information 184, for example, a state value designated with “s” or “t” or “c”. The arithmetic encoder 170 also comprises a cumulative-frequencies-table selector 186, which is configured to receive the state information 184 and to provide an information 188 describing the selected cumulative-frequencies-table to the codeword determinator 180. For example, the cumulative-frequencies-table selector 186 may provide a cumulative-frequencies-table index “pki” describing which cumulative-frequencies-table, out of a set of 96 cumulative-frequencies-tables, is selected for usage by the codeword determinator. Alternatively, the cumulative-frequencies-table selector 186 may provide the entire selected cumulative-frequencies-table or a sub-table to the codeword determinator. Thus, the codeword determinator 180 may use the selected cumulative-frequencies-table or sub-table for the provision of the codeword acod_m[pki][m] of the most-significant bit-plane value m, such that the actual codeword acod_m[pki][m] encoding the most-significant bit-plane value m is dependent on the value of m and the cumulative-frequencies-table index pki, and consequently on the current state information 184. Further details regarding the coding process and the obtained codeword format will be described below.
It should be noted, however, that in some embodiments, the state tracker 182 may be identical to, or take the functionality of, the state tracker 750, the state tracker 1050 or the state tracker 1250. It should also be noted that the cumulative-frequencies-table selector 186 may, in some embodiments, be identical to, or take the functionality of, the mapping rule selector 760, the mapping rule selector 1060, or the mapping rule selector 1260. Moreover, the first codeword determinator 180 may, in some embodiments, be identical to, or take the functionality of, the spectral value encoding 740.
The arithmetic encoder 170 further comprises a less-significant bit-plane extractor 189a, which is configured to extract one or more less-significant bit-planes from the scaled and quantized frequency-domain audio representation 152, if one or more of the spectral values to be encoded exceed the range of values encodeable using the most-significant bit-plane only. The less-significant bit-planes may comprise one or more bits, as desired. Accordingly, the less-significant bit-plane extractor 189a provides a less-significant bit-plane information 189b. The arithmetic encoder 170 also comprises a second codeword determinator 189c, which is configured to receive the less-significant bit-plane information 189d and to provide, on the basis thereof, 0, 1 or more codewords “acod_r” representing the content of 0, 1 or more less-significant bit-planes. The second codeword determinator 189c may be configured to apply an arithmetic encoding algorithm or any other encoding algorithm in order to derive the less-significant bit-plane codewords “acod_r” from the less-significant bit-plane information 189b.
It should be noted here that the number of less-significant bit-planes may vary in dependence on the value of the scaled and quantized spectral values 152, such that there may be no less-significant bit-plane at all, if the scaled and quantized spectral value to be encoded is comparatively small, such that there may be one less-significant bit-plane if the current scaled and quantized spectral value to be encoded is of a medium range and such that there may be more than one less-significant bit-plane if the scaled and quantized spectral value to be encoded takes a comparatively large value.
To summarize the above, the arithmetic encoder 170 is configured to encode scaled and quantized spectral values, which are described by the information 152, using a hierarchical encoding process. The most-significant bit-plane (comprising, for example, one, two or three bits per spectral value) of one or more spectral values, is encoded to obtain an arithmetic codeword “acod_m[pki][m]” of a most-significant bit-plane value m. One or more less-significant bit-planes (each of the less-significant bit-planes comprising, for example, one, two or three bits) of the one or more spectral values are encoded to obtain one or more codewords “acod_r”. When encoding the most-significant bit-plane, the value m of the most-significant bit-plane is mapped to a codeword acod_m[pki][m]. For this purpose, 96 different cumulative-frequencies-tables are available for the encoding of the value m in dependence on a state of the arithmetic encoder 170, i.e. in dependence on previously-encoded spectral values. Accordingly, the codeword “acod_m[pki][m]” is obtained. In addition, one or more codewords “acod_r” are provided and included into the bitstream if one or more less-significant bit-planes are present.
Reset Description
The audio encoder 100 may optionally be configured to decide whether an improvement in bitrate can be obtained by resetting the context, for example by setting the state index to a default value. Accordingly, the audio encoder 100 may be configured to provide a reset information (e.g. named “arith_reset_flag”) indicating whether the context for the arithmetic encoding is reset, and also indicating whether the context for the arithmetic decoding in a corresponding decoder should be reset.
Details regarding the bitstream format and the applied cumulative-frequency tables will be discussed below.
9. Audio Decoder According to
In the following, an audio decoder according to an embodiment of the invention will be described.
The audio decoder 200 is configured to receive a bitstream 210, which represents an encoded audio information and which may be identical to the bitstream 112 provided by the audio encoder 100. The audio decoder 200 provides a decoded audio information 212 on the basis of the bitstream 210.
The audio decoder 200 comprises an optional bitstream payload de-formatter 220, which is configured to receive the bitstream 210 and to extract from the bitstream 210 an encoded frequency-domain audio representation 222. For example, the bitstream payload de-formatter 220 may be configured to extract from the bitstream 210 arithmetically-coded spectral data like, for example, an arithmetic codeword “acod_m[pki][m]” representing the most-significant bit-plane value m of a spectral value a, or of a plurality of spectral values a, b, and a codeword “acod_r” representing a content of a less-significant bit-plane of the spectral value a, or of a plurality of spectral values a, b, of the frequency-domain audio representation. Thus, the encoded frequency-domain audio representation 222 constitutes (or comprises) an arithmetically-encoded representation of spectral values. The bitstream payload deformatter 220 is further configured to extract from the bitstream additional control information, which is not shown in
The audio decoder 200 comprises an arithmetic decoder 230, which is also designated as “spectral noiseless decoder”. The arithmetic decoder 230 is configured to receive the encoded frequency-domain audio representation 220 and, optionally, the state reset information 224. The arithmetic decoder 230 is also configured to provide a decoded frequency-domain audio representation 232, which may comprise a decoded representation of spectral values. For example, the decoded frequency-domain audio representation 232 may comprise a decoded representation of spectral values, which are described by the encoded frequency-domain audio representation 220.
The audio decoder 200 also comprises an optional inverse quantizer/rescaler 240, which is configured to receive the decoded frequency-domain audio representation 232 and to provide, on the basis thereof, an inversely-quantized and resealed frequency-domain audio representation 242.
The audio decoder 200 further comprises an optional spectral pre-processor 250, which is configured to receive the inversely-quantized and resealed frequency-domain audio representation 242 and to provide, on the basis thereof, a pre-processed version 252 of the inversely-quantized and resealed frequency-domain audio representation 242. The audio decoder 200 also comprises a frequency-domain to time-domain signal transformer 260, which is also designated as a “signal converter”. The signal transformer 260 is configured to receive the pre-processed version 252 of the inversely-quantized and resealed frequency-domain audio representation 242 (or, alternatively, the inversely-quantized and resealed frequency-domain audio representation 242 or the decoded frequency-domain audio representation 232) and to provide, on the basis thereof, a time-domain representation 262 of the audio information. The frequency-domain to time-domain signal transformer 260 may, for example, comprise a transformer for performing an inverse-modified-discrete-cosine transform (IMDCT) and an appropriate windowing (as well as other auxiliary functionalities, like, for example, an overlap-and-add).
The audio decoder 200 may further comprise an optional time-domain post-processor 270, which is configured to receive the time-domain representation 262 of the audio information and to obtain the decoded audio information 212 using a time-domain post-processing. However, if the post-processing is omitted, the time-domain representation 262 may be identical to the decoded audio information 212.
It should be noted here that the inverse quantizer/rescaler 240, the spectral pre-processor 250, the frequency-domain to time-domain signal transformer 260 and the time-domain post-processor 270 may be controlled in dependence on control information, which is extracted from the bitstream 210 by the bitstream payload deformatter 220.
To summarize the overall functionality of the audio decoder 200, a decoded frequency-domain audio representation 232, for example, a set of spectral values associated with an audio frame of the encoded audio information, may be obtained on the basis of the encoded frequency-domain representation 222 using the arithmetic decoder 230. Subsequently, the set of, for example, 1024 spectral values, which may be MDCT coefficients, are inversely quantized, resealed and pre-processed. Accordingly, an inversely-quantized, resealed and spectrally pre-processed set of spectral values (e.g., 1024 MDCT coefficients) is obtained. Afterwards, a time-domain representation of an audio frame is derived from the inversely-quantized, resealed and spectrally pre-processed set of frequency-domain values (e.g. MDCT coefficients). Accordingly, a time-domain representation of an audio frame is obtained. The time-domain representation of a given audio frame may be combined with time-domain representations of previous and/or subsequent audio frames. For example, an overlap-and-add between time-domain representations of subsequent audio frames may be performed in order to smoothen the transitions between the time-domain representations of the adjacent audio frames and in order to obtain an aliasing cancellation. For details regarding the reconstruction of the decoded audio information 212 on the basis of the decoded time-frequency domain audio representation 232, reference is made, for example, to the International Standard ISO/IEC 14496-3, part 3, sub-part 4 where a detailed discussion is given. However, other more elaborate overlapping and aliasing-cancellation schemes may be used.
In the following, some details regarding the arithmetic decoder 230 will be described. The arithmetic decoder 230 comprises a most-significant bit-plane determinator 284, which is configured to receive the arithmetic codeword acod_m [pki][m] describing the most-significant bit-plane value m. The most-significant bit-plane determinator 284 may be configured to use a cumulative-frequencies table out of a set comprising a plurality of 96 cumulative-frequencies-tables for deriving the most-significant bit-plane value m from the arithmetic codeword “acod_m [pki][m]”.
The most-significant bit-plane determinator 284 is configured to derive values 286 of a most-significant bit-plane of one of more spectral values on the basis of the codeword acod_m. The arithmetic decoder 230 further comprises a less-significant bit-plane determinator 288, which is configured to receive one or more codewords “acod_r” representing one or more less-significant bit-planes of a spectral value. Accordingly, the less-significant bit-plane determinator 288 is configured to provide decoded values 290 of one or more less-significant bit-planes. The audio decoder 200 also comprises a bit-plane combiner 292, which is configured to receive the decoded values 286 of the most-significant bit-plane of one or more spectral values and the decoded values 290 of one or more less-significant bit-planes of the spectral values if such less-significant bit-planes are available for the current spectral values. Accordingly, the bit-plane combiner 292 provides decoded spectral values, which are part of the decoded frequency-domain audio representation 232. Naturally, the arithmetic decoder 230 is typically configured to provide a plurality of spectral values in order to obtain a full set of decoded spectral values associated with a current frame of the audio content.
The arithmetic decoder 230 further comprises a cumulative-frequencies-table selector 296, which is configured to select one of the 96 cumulative-frequencies tables in dependence on a state index 298 describing a state of the arithmetic decoder. The arithmetic decoder 230 further comprises a state tracker 299, which is configured to track a state of the arithmetic decoder in dependence on the previously-decoded spectral values. The state information may optionally be reset to a default state information in response to the state reset information 224. Accordingly, the cumulative-frequencies-table selector 296 is configured to provide an index (e.g. pki) of a selected cumulative-frequencies-table, or a selected cumulative-frequencies-table or sub-table itself, for application in the decoding of the most-significant bit-plane value m in dependence on the codeword “acod_m”.
To summarize the functionality of the audio decoder 200, the audio decoder 200 is configured to receive a bitrate-efficiently-encoded frequency-domain audio representation 222 and to obtain a decoded frequency-domain audio representation on the basis thereof. In the arithmetic decoder 230, which is used for obtaining the decoded frequency-domain audio representation 232 on the basis of the encoded frequency-domain audio representation 222, a probability of different combinations of values of the most-significant bit-plane of adjacent spectral values is exploited by using an arithmetic decoder 280, which is configured to apply a cumulative-frequencies-table. In other words, statistic dependencies between spectral values are exploited by selecting different cumulative-frequencies-tables out of a set comprising 96 different cumulative-frequencies-tables in dependence on a state index 298, which is obtained by observing the previously-computed decoded spectral values.
It should be noted that the state tracker 299 may be identical to, or may take the functionality of, the state tracker 826, the state tracker 1126, or the state tracker 1326. The cumulative-frequencies-table selector 296 may be identical to, or may take the functionality of, the mapping rule selector 828, the mapping rule selector 1128, or the mapping rule selector 1328. The most significant bit-plane determinator 284 may be identical to, or may take the functionality of, the spectral value determinator 824.
10. Overview of the Tool of Spectral Noiseless Coding
In the following, details regarding the encoding and decoding algorithm, which is performed, for example, by the arithmetic encoder 170 and the arithmetic decoder 230, will be explained.
Focus is placed on the description of the decoding algorithm. It should be noted, however, that a corresponding encoding algorithm can be performed in accordance with the teachings of the decoding algorithm, wherein mappings between encoded and decoded spectral values are inversed, and wherein the computation of the mapping rule index value is substantially identical. In an encoder, the encoded spectral values take over the place of the decoded spectral values. Also, the spectral values to be encoded take over the place of the spectral values to be decoded.
It should be noted that the decoding, which will be discussed in the following, is used in order to allow for a so-called “spectral noiseless coding” of typically post-processed, scaled and quantized spectral values. The spectral noiseless coding is used in an audio encoding/decoding concept (or in any other encoding/decoding concept) to further reduce the redundancy of the quantized spectrum, which is obtained, for example, by an energy compacting time-domain-to-frequency-domain transformer. The spectral noiseless coding scheme, which is used in embodiments of the invention, is based on an arithmetic coding in conjunction with a dynamically adapted context.
In some embodiments according to the invention, the spectral noiseless coding scheme is based on 2-tuples, that is, two neighbored spectral coefficients are combined. Each 2-tuple is split into the sign, the most-significant 2-bits-wise-plane, and the remaining less-significant bit-planes. The noiseless coding for the most-significant 2-bits-wise-plane m uses context dependent cumulative-frequencies-tables derived from four previously decoded 2-tuples. The noiseless coding is fed by the quantized spectral values and uses context dependent cumulative-frequencies-tables derived from four previously decoded neighboring 2-tuples. Here, neighborhood in both time and frequency is taken into account, as illustrated in
For example, the arithmetic coder 170 produces a binary code for a given set of symbols and their respective probabilities (i.e. in dependence on the respective probabilities). The binary code is generated by mapping a probability interval, where the set of symbols lie, to a codeword.
The noiseless coding of the remaining less-significant bit-plane r uses a single cumulative-frequencies-table. The cumulative frequencies correspond for example to a uniform distribution of the symbols occurring in the less-significant bit-planes, i.e. it is expected there is the same probability that a 0 or a 1 occurs in the less-significant bit-planes.
In the following, another short overview of the tool of spectral noiseless coding will be given. Spectral noiseless coding is used to further reduce the redundancy of the quantized spectrum. The spectral noiseless coding scheme is based on an arithmetic coding, in conjunction with a dynamically adapted context. The noiseless coding is fed by the quantized spectral values and uses context dependent cumulative-frequencies-tables derived from, for example, four previously decoded neighboring 2-tuples of spectral values. Here, neighborhood, in both time and frequency, is taken into account as illustrated in
The arithmetic coder produces a binary code for a given set of symbols and their respective probabilities. The binary code is generated by mapping a probability interval, where the set of symbols lies, to a codeword.
11. Decoding Process
11.1 Decoding Process Overview
In the following, an overview of the process of the coding of a spectral value will be given taking reference to
The process of decoding a plurality of spectral values comprises an initialization 310 of a context. Initialization 310 of the context comprises a derivation of the current context from a previous context, using the function “arith_map_context(N, arith_reset_flag)”. The derivation of the current context from a previous context may selectively comprise a reset of the context. Both the reset of the context and the derivation of the current context from a previous context will be discussed below.
The decoding of a plurality of spectral values also comprises an iteration of a spectral value decoding 312 and a context update 313, which context update 313 is performed by a function “arith_update_context(i, a,b)” which is described below. The spectral value decoding 312 and the context update 312 are repeated lg/2 times, wherein lg/2 indicates the number of 2-tuples of spectral values to be decoded (e.g., for an audio frame), unless a so-called “ARITH_STOP” symbol is detected. Moreover, the decoding of a set of lg spectral values also comprises a signs decoding 314 and a finishing step 315.
The decoding 312 of a tuple of spectral values comprises a context-value calculation 312a, a most-significant bit-plane decoding 312b, an arithmetic stop symbol detection 312c, a less-significant bit-plane addition 312d, and an array update 312e.
The state value computation 312a comprises a call of the function “arith_get_context(c,i,N)” as shown, for example, in
The most-significant bit-plane decoding 312b comprises an iterative execution of a decoding algorithm 312ba, and a derivation 312bb of values a,b from the result value m of the algorithm 312ba. In preparation of the algorithm 312ba, the variable lev is initialized to zero.
The algorithm 312ba is repeated, until a “break” instruction (or condition) is reached. The algorithm 312ba comprises a computation of a state index “pki” (which also serves as a cumulative-frequencies-table index) in dependence on the numeric current context value c, and also in dependence on the level value “esc_nb” using a function “arith_get_pk( )”, which is discussed below (and embodiments of which are shown, for example, in
Subsequently, a most-significant bit-plane value m may be obtained by executing a function “arith_decode( )”, taking into consideration the selected cumulative-frequencies-table (described by the variable “cum_freq” and the variable “cfl”). When deriving the most-significant bit-plane value m, bits named “acod_m” of the bitstream 210 may be evaluated (see, for example,
The algorithm 312ba also comprises checking whether the most-significant bit-plane value m is equal to an escape symbol “ARITH_ESCAPE”, or not. If the most-significant bit-plane value m is not equal to the arithmetic escape symbol, the algorithm 312ba is aborted (“break” condition) and the remaining instructions of the algorithm 312ba are then skipped. Accordingly, execution of the process is continued with the setting of the value b and of the value a at step 312bb. In contrast, if the decoded most-significant bit-plane value m is identical to the arithmetic escape symbol, or “ARITH_ESCAPE”, the level value “lev” is increased by one. The level value “esc_nb” is set to be equal to the level value “lev”, unless the variable “lev” is larger than seven, in which case, the variable “esc_nb” is set to be equal to seven. As mentioned, the algorithm 312ba is then repeated until the decoded most-significant bit-plane value m is different from the arithmetic escape symbol, wherein a modified context is used (because the input parameter of the function “arith_get_pk( )” is adapted in dependence on the value of the variable “esc_nb”).
As soon as the most-significant bit-plane is decoded using the one time execution or iterative execution of the algorithm 312ba, i.e. a most-significant bit-plane value m different from the arithmetic escape symbol has been decoded, the spectral value variable “b” is set to be equal to a plurality of (e.g. 2) more significant bits of the most-significant bit-plane value m, and the spectral value variable “a” is set to the (e.g. 2) lowermost bits of the most-significant bit-plane value m. Details regarding this functionality can be seen, for example, at reference numeral 312bb.
Subsequently, it is checked in step 312c, whether an arithmetic stop symbol is present. This is the case if the most-significant bit-plane value m is equal to zero and the variable “lev” is larger than zero. Accordingly, an arithmetic stop condition is signaled by an “unusual” condition, in which the most-significant bit-plane value m is equal to zero, while the variable “lev” indicates that an increased numeric weight is associated to the most-significant bit-plane value m. In other words, an arithmetic stop condition is detected if the bitstream indicates that an increased numeric weight, higher than a minimum numeric weight, should be given to a most-significant bit-plane value which is equal to zero, which is a condition that does not occur in a normal encoding situation. In other words, an arithmetic stop condition is signaled if an encoded arithmetic escape symbol is followed by an encoded most significant bit-plane value of 0.
After the evaluation whether there is an arithmetic stop condition, which is performed in the step 212c, the less-significant bit planes are obtained, for example, as shown at reference numeral 212d in
In the decoding of the one or more least-significant bit planes (if any) an algorithm 212da is iteratively performed, wherein a number of executions of the algorithm 212da is determined by the variable “lev”. It should be noted here that the first iteration of the algorithm 212da is performed on the basis of the values of the variables a, b as set in the step 212bb. Further iterations of the algorithm 212da are be performed on the basis of updated variable values of the variable a, b.
At the beginning of an iteration, a cumulative-frequencies table is selected. Subsequently, an arithmetic decoding is performed to obtain a value of a variable r, wherein the value of the variable r describes a plurality of less-significant bits, for example one less-significant bit associated with the variable a and one less-significant bit associated with the variable b. The function “ARITH_DECODE” is used to obtain the value r, wherein the cumulative frequencies table “arith_cf_r” is used for the arithmetic decoding.
Subsequently, the values of the variables a and b are updated. For this purpose, the variable a is shifted to the left by one bit, and the least-significant bit of the shifted variable a is set the value defined by the least-significant bit of the value r. The variable b is shifted to the left by one bit, and the least-significant bit of the shifted variable b is set the value defined by bit 1 of the variable r, wherein bit 1 of the variable r has a numeric weight of 2 in the binary representation of the variable r. The algorithm 412ba is then repeated until all least-significant bits are decoded.
After the decoding of the less-significant bit-planes, an array “x_ac_dec” is updated in that the values of the variables a,b are stored in entries of said array having array indices 2*i and 2*i+1.
Subsequently, the context state is updated by calling the function “arith_update_context(i,a,b)”, details of which will be explained below taking reference to
Subsequent to the update of the context state, which is performed in step 313, algorithms 312 and 313 are repeated, until running variable i reaches the value of lg/2 or an arithmetic stop condition is detected.
Subsequently, a finish algorithm “arith_finish( )” is performed, as can be seen at reference number 315. Details of the finishing algorithm “arith_finish( )” will be described below taking reference to
Subsequent to the finish algorithm 315, the signs of the spectral values are decoded using the algorithm 314. As can be seen, the signs of the spectral values which are different from zero are individually coded. In the algorithm 314, signs are read for all of the spectral values having indices i between i=0 and i=lg−1 which are non-zero. For each non-zero spectral value having a spectral value index i between i=0 and i=lg−1, a value (typically a single bit) s is read from the bitstream. If the value of s, which is read from the bit stream is equal to 1, the sign of said spectral value is inverted. For this purpose, access is made to the array “x_ac_dec”, both to determine whether the spectral value having the index i is equal to zero and for updating the sign of the decoded spectral values. However, it should be noted that the signs of the variables a, b are left unchanged in the sign decoding 314.
By performing the finish algorithm 315 before the signs decoding 314, it is possible to reset all necessary bins after an ARITH_STOP symbol.
It should be noted here that the concept for obtaining the values of the less-significant bit-planes is not of particular relevance in some embodiments according to the present invention. In some embodiments, the decoding of any less-significant bit-planes may even be omitted. Alternatively, different decoding algorithms may be used for this purpose.
11.2 Decoding Order According to
In the following, the decoding order of the spectral values will be described.
The quantized spectral coefficients “x_ac_dec[ ]” are noiselessly encoded and transmitted (e.g. in the bitstream) starting from the lowest-frequency coefficient and progressing to the highest-frequency coefficient.
Consequently, the quantized spectral coefficients “x_ac_dec[ ]” are noiselessly decoded starting from the lowest-frequency coefficient and progressing to the highest-frequency coefficient. The quantized spectral coefficients are decoded by groups of two successive (e.g. adjacent in frequency) coefficients a and b gathering in a so-called 2-tuple (a,b) (also designated with {a,b}). It should be noted here that the quantized spectral coefficients are sometimes also designated with “qdec”.
The decoded coefficients “x_ac_dec[ ]” for a frequency-domain mode (e.g., decoded coefficients for an advanced audio coding, for example, obtained using a modified-discrete-cosine transform, as discussed in ISO/IEC 14496, part 3, sub-part 4) are then stored in an array “x_ac_quant[g][win][sfb][bin]”. The order of transmission of the noiseless coding codewords is such that when they are decoded in the order received and stored in the array, “bin” is the most rapidly incrementing index, and “g” is the most slowly incrementing index. Within a codeword, the order of decoding is a,b.
The decoded coefficients “x_ac_dec[ ]” for the transform coded-excitation (TCX) are stored, for example, directly in an array “x_tcx_invquant[win][bin]”, and the order of the transmission of the noiseless coding codeword is such that when they are decoded in the order received and stored in the array “bin” is the most rapidly incrementing index, and “win” is the most slowly incrementing index. Within a codeword, the order of the decoding is a, b. In other words, if the spectral values describe a transform-coded-excitation of the linear-prediction filter of a speech coder, the spectral values a, b are associated to adjacent and increasing frequencies of the transform-coded-excitation. Spectral coefficients associated to a lower frequency are typically encoded and decoded before a spectral coefficient associated with a higher frequency.
Notably, the audio decoder 200 may be configured to apply the decoded frequency-domain representation 232, which is provided by the arithmetic decoder 230, both for a “direct” generation of a time-domain audio signal representation using a frequency-domain-to-time-domain signal transform and for an “indirect” provision of a time-domain audio signal representation using both a frequency-domain-to-time-domain decoder and a linear-prediction-filter excited by the output of the frequency-domain-to-time-domain signal transformer.
In other words, the arithmetic decoder, the functionality of which is discussed here in detail, is well-suited for decoding spectral values of a time-frequency-domain representation of an audio content encoded in the frequency-domain, and for the provision of a time-frequency-domain representation of a stimulus signal for a linear-prediction-filter adapted to decode (or synthesize) a speech signal encoded in the linear-prediction-domain. Thus, the arithmetic decoder is well-suited for use in an audio decoder which is capable of handling both frequency-domain encoded audio content and linear-predictive-frequency-domain encoded audio content (transform-coded-excitation-linear-prediction-domain mode).
11.3 Context Initialization According to
In the following, the context initialization (also designated as a “context mapping”), which is performed in a step 310, will be described.
The context initialization comprises a mapping between a past context and a current context in accordance with the algorithm “arith_map_context( )”, a first example of which is shown in
As can be seen, the current context is stored in a global variable “q[2][n_context]” which takes the form of an array having a first dimension of 2 and a second dimension of “n_context”. A past context may optionally (but not necessarily) be stored in a variable “qs[n_context]” which takes the form of a table having a dimension of “n_context” (if it is used).
Taking reference to the example algorithm “arith_map_context” in
Taking reference to the example of
A more complicated mapping is performed if the number of spectral values associated to the current audio frame is different from the number of spectral values associated to the previous audio frame. However, details regarding the mapping in this case are not particularly relevant for the key idea of the present invention, such that reference is made to the pseudo program code of
Moreover, an initialization value for the numeric current context value c is returned by the function “arith_map_context( )”. This initialization value is, for example, equal to the value of the entry “q[0][0]” shifted to the left by 12-bits. Accordingly, the numeric (current) context value c is properly initialized for an iterative update.
Moreover,
To summarize the above, the flag “arith_reset_flag” determines if the context must be reset. If the flag is true, a reset sub-algorithm 500a of the algorithm “arith_map_context( )” is called. Alternatively, however, if the flag “arith_reset_flag” is inactive (which indicates that no reset of the context should be performed), the decoding process starts with an initialization phase where the context element vector (or array) q is updated by copying and mapping the context elements of the previous frame stored in q[1][ ] into q[0][ ]. The context elements within q are stored on 4-bits per 2-tuple. The copying and/or mapping of the context element are performed in a sub-algorithm 500b.
In the example of
11.4 State Value Computation According to
In the following, the state value computation 312a will be described in more detail.
A first example algorithm will be described taking reference to
It should be noted that the numeric current context value c (as shown in
Regarding the computation of the state value, reference is also made to
However, it should be noted that some of these spectral values, which are not used for the “regular” or “normal” computation of the context for decoding the spectral values of the tuple 420 may, nevertheless, be evaluated for the detection of a plurality of previously-decoded adjacent spectral values which fulfill, individually or taken together, a predetermined condition regarding their magnitudes. Details regarding this issue will be discussed below.
Taking reference now to
It should be noted that the function “arith_get_context(c,i,N)” receives, as input variables, an “old state context”, which may be described by a numeric previous context value c. The function “arith_get_context(c,i,N)” also receives, as an input variable, an index i of a 2-tuple of spectral values to decode. The index i is typically a frequency index. An input variable N describes a window length of a window, for which the spectral values are decoded.
The function “arith_get_context(c,i,N)” provides, as an output value, an updated version of the input variable c, which describes an updated state context, and which may be considered as a numeric current context value. To summarize, the function “arith_get_context(c,i,N)” receives a numeric previous context value c as an input variable and provides an updated version thereof, which is considered as a numeric current context value. In addition, the function “arith_get_context” considers the variables i, N, and also accesses the “global” array q[ ][ ].
Regarding the details of the function “arith_get_context(c,i,N)”, it should be noted that the variable c, which initially represents the numeric previous context value in a binary form, is shifted to the right by 4-bits in a step 504a. Accordingly, the four least significant bits of the numeric previous context value (represented by the input variable c) are discarded. Also, the numeric weights of the other bits of the numeric previous context values are reduced, for example, a factor of 16.
Moreover, if the index i of the 2-tuple is smaller than N/4−1, i.e. does not take a maximum value, the numeric current context value is modified in that the value of the entry q[0][i+1] is added to bits 12 to 15 (i.e. to bits having a numeric weight of 212, 213, 214, and 215) of the shifted context value which is obtained in step 504a. For this purpose, the entry q[0][i+1] of the array q[ ][ ] (or, more precisely, a binary representation of the value represented by said entry) is shifted to the left by 12-bits. The shifted version of the value represented by the entry q[0][i+1] is then added to the context value c, which is derived in the step 504a, i.e. to a bit-shifted (shifted to the right by 4-bits) number representation of the numeric previous context value. It should be noted here that the entry q[0][i+1] of the array q[ ][ ] represents a sub-region value associated with a previous portion of the audio content (e.g., a portion of the audio content having time index t0−1, as defined with reference to
A selective addition of the entry q[0][i+1] of the array q[ ][ ] (shifted to the left by 12-bits) is shown at reference numeral 504b. As can be seen, the addition of the value represented by the entry q[0][i+1] is naturally only performed if the frequency index i does not designate a tuple of spectral values having the highest frequency index i=N/4−1.
Subsequently, in a step 504c, a Boolean AND-operation is performed, in which the value of the variable c is AND-combined with a hexadecimal value of 0xFFF0 to obtain an updated value of the variable c. By performing such an AND-operation, the four least-significant bits of the variable c are effectively set to zero.
In a step 504d, the value of the entry q[1][i−1] is added to the value of the variable c, which is obtained by step 504c, to thereby update the value of the variable c. However, said update of the variable c in step 504d is only performed if the frequency index i of the 2-tuple to decode is larger than zero. It should be noted that the entry q[1][i−1] is a context sub-region value based on a tuple of previously-decoded spectral values of the current portion of the audio content for frequencies smaller than the frequencies of the spectral values to be decoded using the numeric current context value. For example, the entry q[1][i−1] of the array q[ ][ ] may be associated with the tuple 430 having time index t0 and frequency index i−1, if it is assumed that the tuple 420 of spectral values is to be decoded using the numeric current context value returned by the present execution of the function “arith_get_context(c,i,N)”.
To summarize, bits 0, 1, 2, and 3 (i.e. a portion of four least-significant bits) of the numeric previous context value are discarded in step 504a by shifting them out of the binary number representation of the numeric previous context value. Moreover, bits 12, 13, 14, and 15 of the shifted variable c (i.e. of the shifted numeric previous context value) are set to take values defined by the context sub-region value q[0][i+1] in the step 504b. Bits 0, 1, 2, and 3 of the shifted numeric previous context value (i.e. bits 4, 5, 6, and 7 of the original numeric previous context value) are overwritten by the context sub-region value q[1][i−1] in steps 504c and 504d.
Consequently, it can be said that bits 0 to 3 of the numeric previous context value represent the context sub-region value associated with the tuple 432 of spectral values, bits 4 to 7 of the numeric previous context value represent the context sub-region value associated with a tuple 434 of previously decoded spectral values, bits 8 to 11 of the numeric previous context value represent the context sub-region value associated with the tuple 440 of previously-decoded spectral values and bits 12 to 15 of the numeric previous context value represent a context sub-region value associated with the tuple 450 of previously-decoded spectral values. The numeric previous context value, which is input into the function “arith_get_context(c,i,N)”, is associated with a decoding of the tuple 430 of spectral values.
The numeric current context value, which is obtained as an output variable of the function “arith_get_context(c,i,N)”, is associated with a decoding of the tuple 420 of spectral values. Accordingly, bits 0 to 3 of the numeric current context values describe the context sub-region value associated with the tuple 430 of the spectral values, bits 4 to 7 of the numeric current context value describe the context sub-region value associated with the tuple 440 of spectral values, bits 8 to 11 of the numeric current context value describe the numeric sub-region value associated with the tuple 450 of spectral value and bits 12 to 15 of the numeric current context value described the context sub-region value associated with the tuple 460 of spectral values. Thus, it can be seen that a portion of the numeric previous context value, namely bits 8 to 15 of the numeric previous context value, are also included in the numeric current context value, as bits 4 to 11 of the numeric current context value. In contrast, bits 0 to 7 of the current numeric previous context value are discarded when deriving the number representation of the numeric current context value from the number representation of the numeric previous context value.
In a step 504e, the variable c which represents the numeric current context value is selectively updated if the frequency index i of the 2-tuple to decode is larger than a predetermined number of, for example, 3. In this case, i.e. if i is larger than 3, it is determined whether the sum of the context sub-region values q[1][i−3], q[1][i−2], and q[1][i−1] is smaller than (or equal to) a predetermined value of, for example, 5. If it is found that the sum of said context sub-region values is smaller than said predetermined value, a hexadecimal value of, for example, 0x10000, is added to the variable c. Accordingly, the variable c is set such that the variable c indicates if there is a condition in which the context sub-region values q[1][i−3], q[1][i−3], and q[1][i−1] comprise a particularly small sum value. For example, bit 16 of the numeric current context value may act as a flag to indicate such a condition.
To conclude, the return value of the function “arith_get_context(c,i,N)” is determined by the steps 504a, 504b, 504c, 504d, and 504e, where the numeric current context value is derived from the numeric previous context value in steps 504a, 504b, 504c, and 504d, and wherein a flag indicating an environment of previously decoded spectral values having, on average, particularly small absolute values, is derived in step 504e and added to the variable c. Accordingly, the value of the variable c obtained steps 504a, 504b, 504c, 504d is returned, in a step 504f, as a return value of the function “arith_get_context(c,i,N)”, if the condition evaluated in step 504e is not fulfilled. In contrast, the value of the variable c, which is derived in steps 504a, 504b, 504c, and 504d, is incremented by the hexadecimal value of 0x10000 and the result of this increment operation is returned, in the step 504e, if the condition evaluated in step 540e is fulfilled.
To summarize the above, it should be noted that the noiseless decoder outputs 2-tuples of unsigned quantized spectral coefficients (as will be described in more detail below). At first the state c of the context is calculated based on the previously decoded spectral coefficients “surrounding” the 2-tuple to decode. In a preferred embodiment, the state (which is, for example, represented by a numeric context value) is incrementally updated using the context state of the last decoded 2-tuple (which is designated as a numeric previous context value), considering only two new 2-tuples (for example, 2-tuples 430 and 460). The state is coded on 17-bits (e.g., using a number representation of a numeric current context value) and is returned by the function “arith_get_context( )”. For details, reference is made to the program code representation of
Moreover, it should be noted that a pseudo program code of an alternative embodiment of a function “arith_get_context( )” is shown in
11.5 Mapping Rule Selection
In the following, the selection of a mapping rule, for example, a cumulative-frequencies-table which describes a mapping of a codeword value onto a symbol code, will be described. The selection of the mapping rule is made in dependence on a context state, which is described by the numeric current context value c.
11.5.1 Mapping Rule Selection Using the Algorithm According to
In the following, the selection of a mapping rule using the function “arith_get_pk(c)” will be described. It should be noted that the function “arith_get_pk( )” is called at the beginning of the sub-algorithm 312ba when decoding a code value “acod_m” for providing a tuple of spectral values. It should be noted that the function “arith_get_pk(c)” is called with different arguments in different iterations of the algorithm 312b. For example, in a first iteration of the algorithm 312b, the function “arith_get_pk(c)” is called with an argument which is equal to the numeric current context value c, provided by the previous execution of the function “arith_get_context(c,i,N)” at step 312a. In contrast, in further iterations of the sub-algorithm 312ba, the function “arith_get_pk(c)” is called with an argument which is the sum of the numeric current context value c provided by the function “arith_get_context(c,i,N)” in step 312a, and a bit-shifted version of the value of the variable “esc_nb”, wherein the value of the variable “esc_nb” is shifted to the left by 17-bits. Thus, the numeric current context value c provided by the function “arith_get_context(c,i,N)” is used as an input value of the function “arith_get_pk( )” in the first iteration of the algorithm 312ba, i.e. in the decoding of comparatively small spectral values. In contrast, when decoding comparatively larger spectral values, the input variable of the function “arith_get_pk( )” is modified in that the value of the variable “esc_nb”, is taken into consideration, as is shown in
Taking reference now to
Taking reference to
Subsequently, a search 506b is performed to identify an index value which designates an entry of the table “ari_hash_m”, such that the value of the input variable c of the function “arith_get_pk( )” lies within an interval defined by said entry and an adjacent entry.
In the search 506b, a sub-algorithm 506ba is repeated, while a difference between the variables “i_max” and “i_min” is larger than 1. In the sub-algorithm 506ba, the variable i is set to be equal to an arithmetic mean of the values of the variables “i_min” and “i_max”. Consequently, the variable i designates an entry of the table “ari_hash_m[ ]” in a middle of a table interval defined by the values of the variables “i_min” and “i_max”. Subsequently, the variable j is set to be equal to the value of the entry “ari_hash_m[i]” of the table “ari_hash_m[ ]”. Thus, the variable j takes a value defined by an entry of the table “ari_hash_m[ ]”, which entry lies in the middle of a table interval defined by the variables “i_min” and “i_max”. Subsequently, the interval defined by the variables “i_min” and “i_max” is updated if the value of the input variable c of the function “arith_get_pk( )” is different from a state value defined by the uppermost bits of the table entry “j=ari_hash_m[i]” of the table “ari_hash_m[ ]”. For example, the “upper bits” (bits 8 and upward) of the entries of the table “ari_hash_m[ ]” describe significant state values. Accordingly, the value “j>>8” describes a significant state value represented by the entry “j=ari_hash_m[i]” of the table “ari_hash_m[ ]” designated by the hash-table-index value i. Accordingly, if the value of the variable c is smaller than the value “j>>8”, this means that the state value described by the variable c is smaller than a significant state value described by the entry “ari_hash_m[i]” of the table “ari_hash_m[ ]”. In this case, the value of the variable “i_max” is set to be equal to the value of the variable i, which in turn has the effect that a size of the interval defined by “i_min” and “i_max” is reduced, wherein the new interval is approximately equal to the lower half of the previous interval. If it found that the input variable c of the function “arith_get_pk( )” is larger than the value “j>>8”, which means that the context value described by the variable c is larger than a significant state value described by the entry “ari_hash_m[i]” of the array “ari_hash_m[ ]”, the value of the variable “i_min” is set to be equal to the value of the variable i. Accordingly, the size of the interval defined by the values of the variables “i_min” and “i_max” is reduced to approximately a half of the size of the previous interval, defined by the previous values of the variables “i_min” and “i_max”. To be more precise, the interval defined by the updated value of the variable “i_min” and by the previous (unchanged) value of the variable “i_max” is approximately equal to the upper half of the previous interval in the case that the value of the variable c is larger than the significant state value defined by the entry “ari_hash_m[i]”.
If, however, it is found that the context value described by the input variable c of the algorithm “arith_get_pk( )” is equal to the significant state value defined by the entry “ari_hash_m[i]” (i.e. c==(j>>8)), a mapping rule index value defined by the lower most 8-bits of the entry “ari_hash_m[i]” is returned as the return value of the function “arith_get_pk( )” (instruction “return (j&0xFF)”).
To summarize the above, an entry “ari_hash_m[i]”, the uppermost bits (bits 8 and upward) of which describe a significant state value, is evaluated in each iteration 506ba, and the context value (or numeric current context value) described by the input variable c of the function “arith_get_pk( )” is compared with the significant state value described by said table entry “ari_hash_m[i]”. If the context value represented by the input variable c is smaller than the significant state value represented by the table entry “ari_hash_m[i]”, the upper boundary (described by the value “i_max”) of the table interval is reduced, and if the context value described by the input variable c is larger than the significant state value described by the table entry “ari_hash_m[i]”, the lower boundary (which is described by the value of the variable “i_min”) of the table interval is increased. In both of said cases, the sub-algorithm 506ba is repeated, unless the size of the interval (defined by the difference between “i_max” and “i_min”) is smaller than, or equal to, 1. If, in contrast, the context value described by the variable c is equal to the significant state value described by the table entry “ari_hash_m[i]”, the function “arith_get_pk( )” is aborted, wherein the return value is defined by the lower most 8-bits of the table entry “ari_hash_m[i]”.
If, however, the search 506b is terminated because the interval size reaches its minimum value (“i_max−“i_min” is smaller than, or equal to, 1), the return value of the function “arith_get_pk( )” is determined by an entry “ari_lookup_m[i_max]” of a table “ari_lookup_m[ ]”, which can be seen at reference numeral 506c. Accordingly, the entries of the table “ari_hash_m[ ]” define both significant state values and boundaries of intervals. In the sub-algorithm 506ba, the search interval boundaries “i_min” and “i_max” are iteratively adapted such that the entry “ari_hash_m[i]” of the table “ari_hash_m[ ]”, a hash table index i of which lies, at least approximately, in the center of the search interval defined by the interval boundary values “i_min” and “i_max”, at least approximates a context value described by the input variable c. It is thus achieved that the context value described by the input variable c lies within an interval defined by “ari_hash_m[i_min]” and “ari_hash_m[i_max]” after the completion of the iterations of the sub-algorithm 506ba, unless the context value described by the input variable c is equal to a significant state value described by an entry of the table “ari_hash_m[ ]”.
If, however, the iterative repetition of the sub-algorithm 506ba is terminated because the size of the interval (defined by “i_max−i_min”) reaches or exceeds its minimum value, it is assumed that the context value described by the input variable c is not a significant state value. In this case, the index “i_max”, which designates an upper boundary of the interval, is nevertheless used. The upper value “i_max” of the interval, which is reached in the last iteration of the sub-algorithm 506ba, is re-used as a table index value for an access to the table “ari_lookup_m”. The table “ari_lookup_m[ ]” describes mapping rule index values associated with intervals of a plurality of adjacent numeric context values. The intervals, to which the mapping rule index values described by the entries of the table “ari_lookup_m[ ]” are associated, are defined by the significant state values described by the entries of the table “ari_hash_m[ ]”. The entries of the table “ari_hash_m” define both significant state values and interval boundaries of intervals of adjacent numeric context values. In the execution of the algorithm 506b, it is determined whether the numeric context value described by the input variable c is equal to a significant state value, and if this is not the case, in which interval of numeric context values (out of a plurality of intervals, boundaries of which are defined by the significant state values) the context value described by the input variable c is lying. Thus, the algorithm 506b fulfills a double functionality to determine whether the input variable c describes a significant state value and, if it is not the case, to identify an interval, bounded by significant state values, in which the context value represented by the input variable c lies. Accordingly, the algorithm 506e is particularly efficient and requires only a comparatively small number of table accesses.
To summarize the above, the context state c determines the cumulative-frequencies-table used for decoding the most-significant 2-bits-wise plane m. The mapping from c to the corresponding cumulative-frequencies-table index “pki” as performed by the function “arith_get_pk( )”. A pseudo program code representation of said function “arith_get_pk( )” has been explained taking reference to
To further summarize the above, the value m is decoded using the function “arith_decode( )” (which is described in more detail below) called with the cumulative-frequencies-table “arith_cf_m[pki][ ]”, where “pki” corresponds to the index (also designated as mapping rule index value) returned by the function “arith_get_pk( )”, which is described with reference to
11.5.2 Mapping Rule Selection Using the Algorithm According to
In the following, another embodiment of a mapping rule selection algorithm “arith_get_pk( )” will be described with reference to
The algorithm “arith_get_pk( )” according to
The algorithm “arith_get_pk( )” provides, as an output variable, a variable “pki”, which describes and index of a probability distribution (or probability model) associated to a state of the context described by the input variable c. The variable “pki” may, for example, be a mapping rule index value.
The algorithm according to
However, different step sizes, e.g. different contents of the array “i_diff[ ]” may actually be chosen, wherein the contents of the array “i_diff[ ]” may naturally be adapted to a size of the hash-table “ari_hash_m[i]”.
It should be noted that the variable “i_min” is initialized to take a value of 0 right at the beginning of the algorithm “arith_get_pk( )”.
In an initialization step 508a, a variable s is initialized in dependence on the input variable c, wherein a number representation of the variable c is shifted to the left by 8 bits in order to obtain the number representation of the variable s.
Subsequently, a table search 508b is performed, in order to identify a hash-table-index-value “i_min” of an entry of the hash-table “ari_hash_m[ ]”, such that the context value described by the context value c lies within an interval which is bounded by the context value described by the hash-table entry “ari_hash_m[i_min]” and a context value described by another hash-table entry “ari_hash_m” which other entry “ari_hash_m” is adjacent (in terms of its hash-table index value) to the hash-table entry “ari_hash_m[i_min]” Thus, the algorithm 508b allows for the determining of a hash-table-index-value “i_min” designating an entry “j=ari_hash_m[i_min]” of the hash-table “ari_hash_m[ ]”, such that the hash-table entry “ari_hash_m[i_min]” at least approximates the context value described by the input variable c.
The table search 508b comprises an iterative execution of a sub-algorithm 508ba, wherein the sub-algorithm 508ba is executed for a predetermined number of, for example, nine iterations. In the first step of the sub-algorithm 508ba, the variable i is set to a value which is equal to a sum of a value of a variable “i_min” and a value of a table entry “i_diff[k]”. It should be noted here that k is a running variable, which is incremented, starting from an initial value of k=0, with each iteration of the sub-algorithm 508ba. The array “i_diff[ ]” defines predetermine increment values, wherein the increment values decrease with increasing table index k, i.e. with increasing numbers of iterations.
In a second step of the sub-algorithm 508ba, a value of a table entry “ari_hash_m[ ]” is copied into a variable j. Preferably, the uppermost bits of the table-entries of the table “ari_hash_m[ ]” describe a significant state values of a numeric context value, and the lowermost bits (bits 0 to 7) of the entries of the table “ari_hash_m[ ]” describe mapping rule index values associated with the respective significant state values.
In a third step of the sub-algorithm 508ba, the value of the variable S is compared with the value of the variable j, and the variable “i_min” is selectively set to the value “i+1” if the value of the variable s is larger than the value of the variable j. Subsequently, the first step, the second step, and the third step of the sub-algorithm 508ba are repeated for a predetermined number of times, for example, nine times. Thus, in each execution of the sub-algorithm 508ba, the value of the variable “i_min” is incremented by i_diff[ ]+1, if, and only if, the context value described by the currently valid hash-table-index i_min+i_diff[ ] is smaller than the context value described by the input variable c. Accordingly, the hash-table-index-value “i_min” is (iteratively) increased in each execution of the sub-algorithm 508ba if (and only if) the context value described by the input variable c and, consequently, by the variable s, is larger than the context value described by the entry “ari_hash_m[i=i_min+diff[k]]”.
Moreover, it should be noted that only a single comparison, namely the comparison as to whether the value of the variable s is larger than the value of the variable j, is performed in each execution of the sub-algorithm 508ba. Accordingly, the algorithm 508ba is computationally particularly efficient. Moreover, it should be noted that there are different possible outcomes with respect to the final value of the variable “i_min” For example, it is possible that the value of the variable “i_min” after the last execution of the sub-algorithm 512ba is such that the context value described by the table entry “ari_hash_m[i_min]” is smaller than the context value described by the input variable c, and that the context value described by the table entry “ari_hash_m[i_min+1]” is larger than the context value described by the input variable c. Alternatively, it may happen that after the last execution of the sub-algorithm 508ba, the context value described by the hash-table-entry “ari_hash_m[i_min−1]” is smaller than the context value described by the input variable c, and that the context value described by the entry “ari_hash_m[i_min]” is larger than the context value described by the input variable c. Alternatively, however, it may happen that the context value described by the hash-table-entry “ari_hash_m[i_min]” is identical to the context value described by the input variable c.
For this reason, a decision-based return value provision 508c is performed. The variable j is set to take the value of the hash-table-entry “ari_hash_m[i_min]” Subsequently, it is determined whether the context value described by the input variable c (and also by the variable s) is larger than the context value described by the entry “ari_hash_m[i_min]” (first case defined by the condition “s>j”), or whether the context value described by the input variable c is smaller than the context value described by the hash-table-entry “ari_hash_m[i_min]” (second case defined by the condition “c<j>>8”), or whether the context value described by the input variable c is equal to the context value described by the entry “ari_hash_m[i_min]” (third case).
In the first case, (s>j), an entry “ari_lookup_m[i_min+1]” of the table “ari_lookup_m[ ]” designated by the table index value “i_min+1” is returned as the output value of the function “arith_get_pk( )”. In the second case (c<(j>>8)), an entry “ari_lookup_m[i_min]” of the table “ari_lookup_m[ ]” designated by the table index value “i_min” is returned as the return value of the function “arith_get_pk( )”. In the third case (i.e. if the context value described by the input variable c is equal to the significant state value described by the table entry “ari_hash_m[i_min]”), a mapping rule index value described by the lowermost 8-bits of the hash-table entry “ari_hash_m[i_min]” is returned as the return value of the function “arith_get_pk( )”.
To summarize the above, a particularly simple table search is performed in step 508b, wherein the table search provides a variable value of a variable “i_min” without distinguishing whether the context value described by the input variable c is equal to a significant state value defined by one of the state entries of the table “ari_hash_m[ ]” or not. In the step 508c, which is performed subsequent to the table search 508b, a magnitude relationship between the context value described by the input variable c and a significant state value described by the hash-table-entry “ari_hash_m[i_min]” is evaluated, and the return value of the function “arith_get_pk( )” is selected in dependence on a result of said evaluation, wherein the value of the variable “i_min”, which is determined in the table evaluation 508b, is considered to select a mapping rule index value even if the context value described by the input variable c is different from the significant state value described by the hash-table-entry “ari_hash_m[i_min]”.
It should further be noted that the comparison in the algorithm should preferably (or alternatively) be done between the context index (numeric context value) c and j=ari_hash_m[i]>>8. Indeed, each entry of the table “ari_hash_m[ ]” represents a context index, coded beyond the 8th bits, and its corresponding probability model coded on the 8 first bits (least significant bits). In the current implementation, we are mainly interested in knowing whether the present context c is greater than ari_hash_m[i]>>8, which is equivalent to detecting if s=c<<8 is also greater than ari_hash_m[ i].
To summarize the above, once the context state is calculated (which may, for example, be achieved using the algorithm “arith_get_context(c,i,N)” according to
11.6 Arithmetic Decoding
11.6.1 Arithmetic Decoding Using the Algorithm According to
In the following, the functionality of the function “arith_decode( )” will be discussed in detail with reference to
It should be noted that the function “arith_decode( )” uses the helper function “arith_first_symbol (void)”, which returns TRUE, if it is the first symbol of the sequence and FALSE otherwise. The function “arith_decode( )” also uses the helper function “arith_get_next_bit(void)”, which gets and provides the next bit of the bitstream.
In addition, the function “arith_decode( )” uses the global variables “low”, “high” and “value”. Further, the function “arith_decode( )” receives, as an input variable, the variable “cum_freq[ ]”, which points towards a first entry or element (having element index or entry index 0) of the selected cumulative-frequencies-table or cumulative-frequencies sub-table. Also, the function “arith_decode( )” uses the input variable “cfl”, which indicates the length of the selected cumulative-frequencies-table or cumulative-frequencies sub-table designated by the variable “cum_freq[ ]”.
The function “arith_decode( )” comprises, as a first step, a variable initialization 570a, which is performed if the helper function “arith_first_symbol( )” indicates that the first symbol of a sequence of symbols is being decoded. The value initialization 550a initializes the variable “value” in dependence on a plurality of, for example, 16 bits, which are obtained from the bitstream using the helper function “arith_get_next_bit”, such that the variable “value” takes the value represented by said bits. Also, the variable “low” is initialized to take the value of 0, and the variable “high” is initialized to take the value of 65535.
In a second step 570b, the variable “range” is set to a value, which is larger, by 1, than the difference between the values of the variables “high” and “low”. The variable “cum” is set to a value which represents a relative position of the value of the variable “value” between the value of the variable “low” and the value of the variable “high”. Accordingly, the variable “cum” takes, for example, a value between 0 and 216 in dependence on the value of the variable “value”.
The pointer p is initialized to a value which is smaller, by 1, than the starting address of the selected cumulative-frequencies-table.
The algorithm “arith_decode( )” also comprises an iterative cumulative-frequencies-table-search 570c. The iterative cumulative-frequencies-table-search is repeated until the variable cfl is smaller than or equal to 1. In the iterative cumulative-frequencies-table-search 570c, the pointer variable q is set to a value, which is equal to the sum of the current value of the pointer variable p and half the value of the variable “cfl”. If the value of the entry *q of the selected cumulative-frequencies-table, which entry is addressed by the pointer variable q, is larger than the value of the variable “cum”, the pointer variable p is set to the value of the pointer variable q, and the variable “cfl” is incremented. Finally, the variable “cfl” is shifted to the right by one bit, thereby effectively dividing the value of the variable “cfl” by 2 and neglecting the modulo portion.
Accordingly, the iterative cumulative-frequencies-table-search 570c effectively compares the value of the variable “cum” with a plurality of entries of the selected cumulative-frequencies-table, in order to identify an interval within the selected cumulative-frequencies-table, which is bounded by entries of the cumulative-frequencies-table, such that the value cum lies within the identified interval. Accordingly, the entries of the selected cumulative-frequencies-table define intervals, wherein a respective symbol value is associated to each of the intervals of the selected cumulative-frequencies-table. Also, the widths of the intervals between two adjacent values of the cumulative-frequencies-table define probabilities of the symbols associated with said intervals, such that the selected cumulative-frequencies-table in its entirety defines a probability distribution of the different symbols (or symbol values). Details regarding the available cumulative-frequencies-tables will be discussed below taking reference to
Taking reference again to
The algorithm “arith_decode” also comprises an adaptation 570e of the variables “high” and “low”. If the symbol value represented by the variable “symbol” is different from 0, the variable “high” is updated, as shown at reference numeral 570e. Also, the value of the variable “low” is updated, as shown at reference numeral 570e. The variable “high” is set to a value which is determined by the value of the variable “low”, the variable “range” and the entry having the index “symbol−1” of the selected cumulative-frequencies-table. The variable “low” is increased, wherein the magnitude of the increase is determined by the variable “range” and the entry of the selected cumulative-frequencies-table having the index “symbol”. Accordingly, the difference between the values of the variables “low” and “high” is adjusted in dependence on the numeric difference between two adjacent entries of the selected cumulative-frequencies-table.
Accordingly, if a symbol value having a low probability is detected, the interval between the values of the variables “low” and “high” is reduced to a narrow width. In contrast, if the detected symbol value comprises a relatively large probability, the width of the interval between the values of the variables “low” and “high” is set to a comparatively large value. Again, the width of the interval between the values of the variable “low” and “high” is dependent on the detected symbol and the corresponding entries of the cumulative-frequencies-table.
The algorithm “arith_decode( )” also comprises an interval renormalization 570f, in which the interval determined in the step 570e is iteratively shifted and scaled until the “break”-condition is reached. In the interval renormalization 570f, a selective shift-downward operation 570fa is performed. If the variable “high” is smaller than 32768, nothing is done, and the interval renormalization continues with an interval-size-increase operation 570fb. If, however, the variable “high” is not smaller than 32768 and the variable “low” is greater than or equal to 32768, the variables “values”, “low” and “high” are all reduced by 32768, such that an interval defined by the variables “low” and “high” is shifted downwards, and such that the value of the variable “value” is also shifted downwards. If, however, it is found that the value of the variable “high” is not smaller than 32768, and that the variable “low” is not greater than or equal to 32768, and that the variable “low” is greater than or equal to 16384 and that the variable “high” is smaller than 49152, the variables “value”, “low” and “high” are all reduced by 16384, thereby shifting down the interval between the values of the variables “high” and “low” and also the value of the variable “value”. If, however, neither of the above conditions is fulfilled, the interval renormalization is aborted.
If, however, any of the above-mentioned conditions, which are evaluated in the step 570fa, is fulfilled, the interval-increase-operation 570fb is executed. In the interval-increase-operation 570fb, the value of the variable “low” is doubled. Also, the value of the variable “high” is doubled, and the result of the doubling is increased by 1. Also, the value of the variable “value” is doubled (shifted to the left by one bit), and a bit of the bitstream, which is obtained by the helper function “arith_get_next_bit” is used as the least-significant bit. Accordingly, the size of the interval between the values of the variables “low” and “high” is approximately doubled, and the precision of the variable “value” is increased by using a new bit of the bitstream. As mentioned above, the steps 570fa and 570fb are repeated until the “break” condition is reached, i.e. until the interval between the values of the variables “low” and “high” is large enough.
Regarding the functionality of the algorithm “arith_decode( )”, it should be noted that the interval between the values of the variables “low” and “high” is reduced in the step 570e in dependence on two adjacent entries of the cumulative-frequencies-table referenced by the variable “cum_freq”. If an interval between two adjacent values of the selected cumulative-frequencies-table is small, i.e. if the adjacent values are comparatively close together, the interval between the values of the variables “low” and “high”, which is obtained in the step 570e, will be comparatively small. In contrast, if two adjacent entries of the cumulative-frequencies-table are spaced further, the interval between the values of the variables “low” and “high”, which is obtained in the step 570e, will be comparatively large.
Consequently, if the interval between the values of the variables “low” and “high”, which is obtained in the step 570e, is comparatively small, a large number of interval renormalization steps will be executed to re-scale the interval to a “sufficient” size (such that neither of the conditions of the condition evaluation 570fa is fulfilled). Accordingly, a comparatively large number of bits from the bitstream will be used in order to increase the precision of the variable “value”. If, in contrast, the interval size obtained in the step 570e is comparatively large, only a smaller number of repetitions of the interval normalization steps 570fa and 570fb will be required in order to renormalize the interval between the values of the variables “low” and “high” to a “sufficient” size. Accordingly, only a comparatively small number of bits from the bitstream will be used to increase the precision of the variable “value” and to prepare a decoding of a next symbol.
To summarize the above, if a symbol is decoded, which comprises a comparatively high probability, and to which a large interval is associated by the entries of the selected cumulative-frequencies-table, only a comparatively small number of bits will be read from the bitstream in order to allow for the decoding of a subsequent symbol. In contrast, if a symbol is decoded, which comprises a comparatively small probability and to which a small interval is associated by the entries of the selected cumulative-frequencies-table, a comparatively large number of bits will be taken from the bitstream in order to prepare a decoding of the next symbol.
Accordingly, the entries of the cumulative-frequencies-tables reflect the probabilities of the different symbols and also reflect a number of bits required for decoding a sequence of symbols. By varying the cumulative-frequencies-table in dependence on a context, i.e. in dependence on previously-decoded symbols (or spectral values), for example, by selecting different cumulative-frequencies-tables in dependence on the context, stochastic dependencies between the different symbols can be exploited, which allows for a particular bitrate-efficient encoding of the subsequent (or adjacent) symbols.
To summarize the above, the function “arith_decode( )”, which has been described with reference to
To summarize the above, the arithmetic decoder is an integer implementation using the method of tag generation with scaling. For details, reference is made to the book “Introduction to Data Compression” of K. Sayood, Third Edition, 2006, Elsevier Inc.
The computer program code according to
11.6.2 Arithmetic Decoding Using the Algorithm According to
h and 5i show a pseudo program code representation of another embodiment of the algorithm “arith_decode( )”, which can be used as an alternative to the algorithm “arith_decode” described with reference to
It should be noted that both the algorithms according to
To summarize, the value m is decoded using the function “arith_decode( )” called with the cumulative-frequencies-table “arith_cf_m[pki][ ]” wherein “pki” corresponds to the index returned by the function “arith_get_pk( )”. The arithmetic coder (or decoder) is an integer implementation using the method of tag generation with scaling. For details, reference is made to the Book “Introduction to Data Compression” of K. Sayood, Third Edition, 2006, Elsevier Inc. The computer program code according to
11.7 Escape Mechanism
In the following, the escape mechanism, which is used in the decoding algorithm “values_decode( )” according to
When the decoded value m (which is provided as a return value of the function “arith_decode( )”) is the escape symbol “ARITH_ESCAPE”, the variables “lev” and “esc_nb” are incremented by 1, and another value m is decoded. In this case, the function “arith_get_pk( )” is called once again with the value “c+esc_nb<<17 as input argument, where the variable “esc_nb” describes the number of escape symbols previously decoded for the same 2-tuple and bounded to 7.
To summarize, if an escape symbol is identified, it is assumed that the most-significant bit-plane value m comprises an increased numeric weight. Moreover, current numeric decoding is repeated, wherein a modified numeric current context value “c+esc_nb<<17” is used as an input variable to the function “arith_get_pk( )”. Accordingly, a different mapping rule index value “pki” is typically obtained in different iterations of the sub-algorithm 312ba.
11.8 Arithmetic Stop Mechanism
In the following, the arithmetic stop mechanism will be described. The arithmetic stop mechanism allows for the reduction of the number of required bits in the case that the upper frequency portion is entirely quantized to 0 in an audio encoder.
In an embodiment, an arithmetic stop mechanism may be implemented as follows: Once the value m is not the escape symbol, “ARITH_ESCAPE”, the decoder checks if the successive m forms an “ARITH_ESCAPE” symbol. If the condition “esc_nb>0&&m==0” is true, the “ARITH_STOP” symbol is detected and the decoding process is ended. In this case, the decoder jumps directly to the “arith_finish( )” function which will be described below. The condition means that the rest of the frame is composed of 0 values.
11.9 Less-Significant Bit-Plane Decoding
In the following, the decoding of the one or more less-significant bit-planes will be described. The decoding of the less-significant bit-plane, is performed, for example, in the step 312d shown in
11.9.1 Less-Significant Bit-Plane Decoding According to
Taking reference now to
Subsequently, an arithmetic decoding of the least-significant bit-plane values r is repeated, wherein the number of repetitions is determined by the value of the variable “lev”. A least-significant bit-plane value r is obtained using the function “arith_decode”, wherein a cumulative-frequencies-table adapted to the least-significant bit-plane decoding is used (cumulative-frequencies-table “arith_cf_r”). A least-significant bit (having a numeric weight of 1) of the variable r describes a less-significant bit-plane of the spectral value represented by the variable a, and a bit having a numeric weight of 2 of the variable r describes a less-significant bit of the spectral value represented by the variable b. Accordingly, the variable a is updated by shifting the variable a to the left by 1 bit and adding the bit having the numeric weight of 1 of the variable r as the least significant bit. Similarly, the variable b is updated by shifting the variable b to the left by one bit and adding the bit having the numeric weight of 2 of the variable r.
Accordingly, the two most-significant information carrying bits of the variables a,b are determined by the most-significant bit-plane value m, and the one or more least-significant bits (if any) of the values a and b are determined by one or more less-significant bit-plane values r.
To summarize the above, it the “ARITH_STOP” symbol is not met, the remaining bit planes are then decoded, if any exist, for the present 2-tuple. The remaining bit-planes are decoded from the most-significant to the least-significant level by calling the function “arith_decode( )” lev number of times with the cumulative frequencies table “arith_cf_r[ ]”. The decoded bit-planes r permit the refining of the previously-decoded value m in accordance with the algorithm, a pseudo program code of which is shown in
11.9.2 Less-Significant Bit Band Decoding According to
Alternatively, however, the algorithm a pseudo program code representation of which is shown in
11.10 Context Update
11.10.1 Context Update According to
In the following, operations used to complete the decoding of the tuple of spectral values will be described, taking reference to
Taking reference now to
Subsequently, the context “q” is also updated for the next 2-tuple. It should be noted that this context update also has to be performed for the last 2-tuple. This context update is performed by the function “arith_update_context( )”, a pseudo program code representation of which is shown in
Taking reference now to
It should be noted here that the entry “q[1][i]” of the array “q[ ][ ]” may be considered as a context sub-region value, because it describes a sub-region of the context which is used for a subsequent decoding of additional spectral values (or tuples of spectral values).
It should be noted here that the summation of the absolute values a and b of the two currently decoded spectral values (signed versions of which are stored in the entries “x_ac_dec[2*i]” and “x_ac_dec[2*i+1]” of the array “x_ac_dec[ ]”), may be considered as the computation of a norm (e.g. a L1 norm) of the decoded spectral values.
It has been found that context sub-region values (i.e. entries of the array “q[ ][ ]”), which describe a norm of a vector formed by a plurality of previously decoded spectral values are particularly meaningful and memory efficient. It has been found that such a norm, which is computed on the basis of a plurality of previously decoded spectral values, comprises meaningful context information in a compact form. It has been found that the sign of the spectral values is typically not particularly relevant for the choice of the context. It has also been found that the formation of a norm across a plurality of previously decoded spectral values typically maintains the most important information, even though some details are discarded. Moreover, it has been found that a limitation of the numeric current context value to a maximum value typically does not result in a severe loss of information. Rather, it has been found that it is more efficient to use the same context state for significant spectral values which are larger than a predetermined threshold value. Thus, the limitation of the context sub-region values brings along a further improvement of the memory efficiency. Furthermore, it has been found that the limitation of the context sub-region values to a certain maximum value allows for a particularly simple and computationally efficient update of the numeric current context value, which has been described, for example, with reference to
Moreover, it has been found that a limitation of the context sub-region values to values between 1 and 15, brings along a particularly good compromise between accuracy and memory efficiency, because 4 bits are sufficient in order to store such a context sub-region value.
However, it should be noted that in some other embodiments, a context sub-region value may be based on a single decoded spectral value only. In this case, the formation of a norm may optionally be omitted.
The next 2-tuple of the frame is decoded after the completion of the function “arith_update_context” by incrementing i by 1 and by redoing the same process as described above, starting from the function “arith_get_context( )”.
When lg/2 2-tuples are decoded within the frame, or with the stop symbol according to “ARITH_ESCAPE” occurs, the decoding process of the spectral amplitude terminates and the decoding of the signs begins.
Details regarding the decoding of the signs have been discussed with reference to
Once all unsigned quantized spectral coefficients are decoded, the according sign is added. For each non-null quantized value of “x_ac_dec” a bit is read. If the read bit value is equal to 0, the quantized value is positive, nothing is done and the signed value is equal to the previously-decoded unsigned value. Otherwise (i.e. if the read bit value is equal to 1), the decoded coefficient (or spectral value) is negative and the two's complement is taken from the unsigned value. The sign bits are read from the low to the higher frequencies. For details, reference is made to
The decoding is finished by calling the function “arith_finish( )”. The remaining spectral coefficients are set to 0. The respective context states are updated correspondingly.
For details, reference is made to
The function “arith_finish” also receives, as an input value, a vector “x_ac_dec” of decoded spectral values, or at least a reference to such a vector of decoded spectral coefficients.
The function “arith_finish” is configured to set the entries of the array (or vector) “x_ac_dec”, for which no spectral values have been decoded due to the presence of an arithmetic stop condition, to 0. Moreover, the function “arith_finish” sets context sub-region values “q[1][i]”, which are associated with spectral values for which no value has been decoded due to the presence of an arithmetic stop condition, to a predetermined value of 1. The predetermined value of 1 corresponds to a tuple of the spectral values wherein both spectral values are equal to 0.
Accordingly, the function “arith_finish( )” allows to update the entire array (or vector) “x_ac_dec[ ]” of spectral values and also the entire array of context sub-region values q[1][i], even in the presence of an arithmetic stop condition.
11.10.2 Context Update According to
In the following, another embodiment of the context update will be described taking reference to
The next 2-tuple of the frame is then decoded by incrementing i by 1 and calling the function arith_decode( ). If the lg/2 2-tuples were already decoded with the frame, or if the stop symbol “ARITH_STOP” occurred, the function “arith_finish( )” is called. The context is saved and stored in the array (or vector) “qs” for the next frame. A pseudo program code of the function “arith_save_context( )” is shown in
Once all unsigned quantized spectral coefficients are decoded, the sign is then added. For each non-quantized value of “qdec”, a bit is read. If the read bit value is equal to 0, the quantized value is positive, nothing is done and the signed value is equal to the previously-decoded unsigned value. Otherwise, the decoded coefficient is negative and the two's complement is taken from the unsigned vale. The signed bits are read from the low to the high frequencies.
11.11 Summary of Decoding Process
In the following, the decoding process will briefly be summarized. For details, reference is made to the above discussion and also to
The decoded coefficients “x_ac_dec[ ]” for the frequency-domain (i.e. for a frequency-domain mode) are then stored in the array “x_ac_quant[g][win][sfb][bin]”. The order of transmission of the noiseless coding codewords is such that when they are decoded in the order received and stored in the array, “bin” is the most rapidly incrementing index and “g” is the most slowly incrementing index. Within a codeword, the order of decoding is a, then b. The decoded coefficients “x_ac_dec[ ]” for the “TCX” (i.e. for an audio decoding using a transform-coded excitation) are stored (for example, directly) in the array “x_tcx_invquant[win][bin]” and the order of the transmission of the noiseless coding codewords is such that when they are decoded in the order received and stored in the array, “bin” is the most rapidly incrementing index and “win” is the most slowly incrementing index. Within a codeword, the order of decoding is a, then b.
First, the flag “arith_reset_flag” determines if the context must be reset. If the flag is true, this is considered in the function “arith_map_context”.
The decoding process starts with an initialization phase where the context element vector “q” is updated by copying and mapping the context elements of the previous frame stored in “q[1][ ]” into “q[0][ ]”. The context elements within “q” are stored on a 4-bits per 2-tuple. For details, reference is made to the pseudo program code of
The noiseless decoder outputs 2-tuples of unsigned quantized spectral coefficients. At first, the state c of the context is calculated based on the previously-decoded spectral coefficients surrounding the 2-tuple to decode. Therefore, the state is incrementally updated using the context state of the last decoded 2-tuple considering only two new 2-tuples. The state is decoded on 17-bits and is returned by the function “arith_get_context”. A pseudo program code representation of the set function “arith_get_context” is shown in
The context state c determines the cumulative-frequencies-table used for decoding the most significant 2-bit-wise-plane m. The mapping from c to the corresponding cumulative-frequencies-table index “pki” is performed by the function “arith_get_pk( )”. A pseudo program code representation of the function “arith_get_pk( )” is shown in
The value m is decoded using the function “arith_decode( )” called with the cumulative-frequencies-table, “arith_cf_m[pki][ ]”, where “pki” corresponds to the index returned by “arith_get_pk( )”. The arithmetic coder (and decoder) is an integer implementation using a method of tag generation with scaling. The pseudo program code according to
When the decoded value m is the escape symbol “ARITH_ESCAPE”, the variables “lev” and “esc_nb” are incremented by 1 and another value m is decoded. In this case, the function “get_pk( )” is called once again with the value “c+esc_nb<<17” as input argument, where “esc_nb” is the number of escape symbols previously decoded for the same 2-tuple and bounded to 7.
Once the value m is not the escape symbol “ARITH_ESCAPE”, the decoder checks if the successive m forms an “ARITH_STOP” symbol. If the condition “(esc_nb>0&&m==0)” is true, the “ARITH_STOP” symbol is detected and the decoding process is ended. The decoder jumps directly to the sign decoding described afterwards. The condition means that the rest of the frame is composed of 0 values.
If the “ARITH_STOP” symbol is not met, the remaining bit-planes are then decoded, if any exist, for the present 2-tuple. The remaining bit-planes are decoded from the most-significant to the least-significant level, by calling “arith_decode( )” lev number of times with the cumulative-frequencies-table “arith_cf_r[ ]”. The decoded bit-planes r permit the refining of the previously-decoded value m, in accordance with the algorithm a pseudo program code of which is shown in
The context “q” is also updated for the next 2-tuple. It should be noted that this context update has to also be performed for the last 2-tuple. This context update is performed by the function “arith_update_context( )”, a pseudo program code representation of which is shown in
The next 2-tuple of the frame is then decoded by incrementing i by 1 and by redoing the same process as described as above, starting from the function “arith_get_context( )”. When lg/2 2-tuples are decoded within the frame, or when the stop symbol “ARITH_STOP” occurs, the decoding process of the spectral amplitude terminates and the decoding of the signs begins.
The decoding is finished by calling the function “arith_finish( )”. The remaining spectral coefficients are set to 0. The respective context states are updated correspondingly. A pseudo program code representation of the function “arith_finish” is shown in
Once all unsigned quantized spectral coefficients are decoded, the according sign is added. For each non-null quantized value of “x_ac_dec”, a bit is read. If the read bit value is equal to 0, the quantized value is positive, and nothing is done, and the signed value is equal to the previously decoded unsigned value. Otherwise, the decoded coefficient is negative and the two's complement is taken from the unsigned value. The signed bits are read from the low to the high frequencies.
11.12 Legends
q shows a legend of the definitions which is related to the algorithms according to
r shows a legend of the definitions which is related to the algorithms according to
12. Mapping Tables
In an embodiment according to the invention, particularly advantageous tables “ari_lookup_m”, “ari_hash_m”, and “ari_cf_m” are used for the execution of the function “arith_get_pk( )” according to
12.1 Table “ari_hash_m[600]” According to
A content of a particularly advantageous implementation of the table “ari_hash_m”, which is used by the function “arith_get_pk”, a first embodiment of which was described with reference to
Furthermore, it should be noted that the table entries of the table “ari_hash_m[ ]” according to
It should further be noted that the most-significant 24-bits of the table entries of the table “ari_hash_m” represent certain significant state values, while the least-significant 8-bits represent mapping rule index values “pki”. Thus, the entries of the table “ari_hash_m[ ]” describe a “direct hit” mapping of a context value onto a mapping rule index value “pki”.
However, the uppermost 24-bits of the entries of the table “ari_hash_m[ ]” represent, at the same time, interval boundaries of intervals of numeric context values, to which the same mapping rule index value is associated. Details regarding this concept have already been discussed above.
12.2 Table “ari_lookup_m” According to
A content of a particularly advantageous embodiment of the table “ari_lookup_m” is shown in the table of
It should be noted that the entries of the table “ari_lookup_m[600]” are listed in an ascending order of the table index “i” (e.g. “i_min” or “i_max”) between 0 and 599. The term “0x” indicates that the table entries are described in a hexadecimal format. Accordingly, the first table entry “0x02” corresponds to the table entry “ari_lookup_m[0]” having table index 0 and the last table entry “0x5E” corresponds to the table entry “ari_lookup_m[599]” having table index 599.
It should also be noted that the entries of the table “ari_lookup_m[ ]” are associated with intervals defined by adjacent entries of the table “arith_hash_m[ ]”. Thus, the entries of the table “ari_lookup_m” describe mapping rule index values associated with intervals of numeric context values, wherein the intervals are defined by the entries of the table “arith_hash_m”.
12.3. Table “ari_cf_m[96][17]” According to
As can be seen from
Within a sub-block (e.g. a sub-block 2310 or 2312, or a sub-block 2396), a first value describes a first entry of a cumulative-frequencies-table (having an array index or table index of 0), and a last value describes a last entry of a cumulative-frequencies-table (having an array index or table index of 16).
Accordingly, each sub-block 2310, 2312, 2396 of the table representation of
12.4 Table “ari_cf_r[ ]” According to
The four entries of said table are shown in
13. Performance Evaluation and Advantages
The embodiments according to the invention use updated functions (or algorithms) and an updated set of tables, as discussed above, in order to obtain an improved tradeoff between computational complexity, memory requirement, and coding efficiency.
Generally speaking, the embodiments according to the invention create an improved spectral noiseless coding. Embodiments according to the present invention describe an enhancement of the spectral noiseless coding in USAC (unified speech and audio encoding).
Embodiments according to the invention create an updated proposal for the CE on improved spectral noiseless coding of spectral coefficients, based on the schemes as presented in the MPEG input papers m16912 and m17002. Both proposals were evaluated, potential short-comings eliminated and the strengths combined.
As in m16912 and m17002, the resulting proposal is based on the original context based arithmetic coding scheme as the working draft 5 USAC (the draft standard on unified speech and audio coding), but can significantly reduce memory requirements (random access memory (RAM) and read-only memory (ROM)) without increasing the computational complexity, while maintaining coding efficiency. In addition, a lossless transcoding of bitstreams according to the working draft 3 of the USAC Draft Standard and according to the working draft 5 of the USAC Draft Standard was proven to be possible. Embodiments according to the invention aim at replacing the spectral noiseless coding scheme as used in working draft 5 of the USAC Draft Standard.
The arithmetic coding scheme described herein is based on the scheme as in the reference model 0 (RM0) or the working draft 5 (WD) of the USAC Draft Standard. Spectral coefficients in frequency or in time model a context. This context is used for the selection of cumulative-frequencies-tables for the arithmetic encoder. Compared to the working draft 5 (WD), the context modeling is further improved and the tables holding the symbol probabilities were re-trained. The number of different probability models was increased from 32 to 96.
Embodiments according to the invention reduce the table sizes (data ROM demand) to 1518 words of length 32-bits or 6072-bytes (WD 5: 16, 894.5 words or 67,578-bytes). The static RAM demand is reduced from 666 words (2,664 bytes) to 72 words (288 bytes) per core coder channel. At the same time, it fully preserves the coding performance and can even reach a gain of approximately 1.29 to 1.95% compared to the overall data rate over all 9 operating points. All working draft 3 and working draft 5 bitstreams can be transcoded in a lossless manner, without affecting the bit reservoir constraints.
In the following, a brief discussion of the coding concepts according to working draft 5 of the USAC Draft Standard will be provided to facilitate the understanding of the advantages of the concept described herein. Subsequently, some preferred embodiments according to the invention will be described.
In USAC working draft 5, a context based arithmetic coding scheme is used for noiseless coding of quantized spectral coefficients. As context, the decoded spectral coefficients are used, which are previous in frequency and time. In working draft 5, a maximum number of 16 spectral coefficients are used as context, 12 of them being previous in time. Also, spectral coefficients used for the context and to be decoded, are grouped as 4-tuples (i.e. 4 spectral coefficients neighbored in frequency, see
For the complete working draft 5 noiseless coding scheme, a memory demand (read-only memory (ROM)) of 16894.5 words (67578 byte) is required. Additionally, 666 words (2664 byte) of static RAM per core-coder channel are required to store the states for the next frame. The table representation of
It should be noted here that in regards to the noiseless coding, working drafts 4 and 5 of the USAC draft standard are the same. Both use the same noiseless coder.
A total memory demand of a complete USAC WD5 decoder is estimated to be 37000 words (148000-byte) for data ROM without program code and 10000 to 17000 words for the static RAM. It can clearly be seen that the noiseless coder tables consume approximately 45% of the total data ROM demand. The largest individual table already consumes 4096 words (16384-byte).
It has been found that both, the size of the combination of all of the tables and the large individual tables exceed typical cache sizes as provided by a fixed point processors used in consumer portable devices, which is in a typical range of 8 to 32 Kbyte (e.g. ARM9e, TI C64XX, etc). This means that the set of tables can probably not be stored in the fast data RAM, which enables a quick random access to the data. This causes the whole decoding process to slow down.
Moreover, it has been found that current successful audio coding technology such as HE-AAC has been proven to be implementable on most mobile devices. HE-AAC uses a Huffman entropy coding scheme with a table size of 995 words. For details, reference is made to ISO/IEC JTC1/SC29/WG11 N2005, MPEG98, February 1998, San José, “Revised Report on Complexity of MPEG-2 AAC2”.
At the 90th MPEG Meeting, in MPEG input papers m16912 and m17002, two proposals were presented which aimed at reducing the memory requirements and improving the encoding efficiency of the noiseless coding scheme. By analyzing both proposals, the following conclusions could be drawn.
In the course of these evaluations, it was identified that moving from a 4-tuple to a 1-tuple coding scheme had a significant impact on the computational complexity: a reduction of the coding dimension increases by the same factor the number of symbols to code. This means for the reduction from 4-tuples to 1-tuples that the operations needed to determine the context, access the hash-tables and decode the symbol have to be performed four times more often than before. Together with a more sophisticated algorithm for the context determination, this led to an increment in computational complexity by a factor of 2.5 or x.xxPCU.
In the following, the proposed new scheme according to the embodiments of the present invention will briefly be described.
To overcome the issue of memory footprint and the computational complexity, an improved noiseless coding scheme is proposed to replace the scheme as in working draft 5 (WD5). The main focus in the development was put on reducing memory demand, while maintaining the compression efficiency and not increasing the computational complexity. More specifically, the target was to reach a good (or even the best) trade-off in the multi-dimension complexity space of compression performance, complexity and memory requirements.
The new coding scheme proposal borrows the main feature of the WD5 noiseless encoder, namely the context adaptation. The context is derived using previously-decoded spectral coefficients, which come as in WD5 from both, the past and the present frame (wherein a frame may be considered as a portion of the audio content). However, the spectral coefficients are now coded by combining two coefficients together to form a 2-tuple. Another difference lays in the fact that the spectral coefficients are now split into three parts, the sign, the more-significant bits or most-significant bits (MSBs) and the less-significant bits or least-significant bits (LSBs). The sign is coded independently from the magnitude which is further divided into two parts, the most-significant bits (or more significant bits) and the rest of the bits (or less-significant bits), if they exist. The 2-tuples for which the magnitude of the two elements is lower or equal to 3 are coded directly by the MSBs coding. Otherwise, an escape codeword is transmitted first for signaling any additional bit-plane. In the base version, the missing information, the LSBs and the sign, are both coded using uniform probability distribution. Alternatively, a different probability distribution may be used.
The table size reduction is still possible, since:
In the following, some details regarding the MSBs coding will be described. As already mentioned, one of the main differences between WD5 of the USAC Draft Standard, a proposal submitted at the 90th MPEG Meeting and the current proposal is the dimension of the symbols. In WD5 of the USAC Draft Standard, 4-tuples were considered for the context generation and the noiseless coding. In a proposal submitted at the 90th MPEG Meeting, 1-tuples were used instead for reducing the ROM requirements. In the course of development, the 2-tuples were found to be the best compromise for reducing the ROM requirements, without increasing the computational complexity. Instead of considering four 4-tuples for the context innovation, now four 2-tuples are considered. As shown in
The table size reduction is due to three main factors. First, only probabilities for 17 symbols need to be stored (i.e. {[0;+3], [0;+3]}+ESC symbol). Grouping tables (i.e. egroups, dgroups, and dgvectors) are no longer required. Finally, the size of the hash-table was reduced by performing an appropriate training.
Although the dimension was reduced from four to two, the complexity was maintained to the range as in WD5 of the USAC Draft Standard. It was achieved by simplifying both the context generation and the hash-table access.
The different simplifications and optimizations were done in a manner that the coding performance was not affected, and even slightly improved. It was achieved mainly by increasing the number of probability models from 32 to 96.
In the following, some details regarding the LSBs coding will be described. The LSBs are coded with a uniform probability distribution in some embodiments. Compared to WD5 of the USAC Draft Standard, the LSBs are now considered within 2-tuples instead of 4-tuples.
In the following some details regarding the sign coding will be explained. The sign is coded without using the arithmetic core-coder for the sake of complexity reduction. The sign is transmitted on 1-bit only when the corresponding magnitude is non-null. 0 means a positive value and 1 means a negative value.
In the following, some details regarding the memory demand will be explained. The proposed new scheme exhibits a total ROM demand of at most 1522.5 new words (6090-bytes). For details, reference is made to the table of
Further on, the amount of information required for the context derivation in the next frame (static ROM) is also reduced. In WD5 of the USAC Draft Standard, the complete set of coefficients (a maximum of 1152 coefficients) with a resolution of typically 16-bits additional to a group index per 4-tuple of a resolution 10-bits needed to be stored, which sums up to 666 words (2664-bytes) per core-coder channel (complete USAC WD4 decoder: approximately 10000 to 17000 words). The new scheme reduces the persistent information to only 2-bits per spectral coefficient, which sums up to 72 words (288-byte) in total per core-coder channel. The demand on the static memory can be reduced by 594 words (2376-byte).
In the following, some details regarding the possible increase of coding efficiency will be described. Decoding efficiency of embodiments according to the new proposal was compared against the reference quality bitstreams according to working draft 3 (WD3) and WD5 of the USAC Draft Standard. The comparison was performed by means of a transcoder, based on a reference software decoder. For details regarding said comparison of the noiseless coding according to WD3 or WD5 of the USAC Draft Standard and the proposed coding scheme, reference is made to
Also, the memory demand in embodiments according to the invention was compared to embodiments according to the WD3 (or WD5) of the USAC Draft Standard.
The coding efficiency is not only maintained, but slightly increased. For details, reference is made to the table of
Details on average bit rates per operating mode can be found in the table of
Moreover,
In the following, some details regarding the computational complexity will be described. The reduction of the dimensionality of the arithmetic coding usually leads to an increase of the computational complexity. Indeed, reducing the dimension by a factor of two will make the arithmetic coder routines call twice.
However, it has been found that this increase of complexity can be limited by several optimizations introduced in the proposed new coding scheme according to the embodiments of the present invention. The context generation was greatly simplified in some embodiments according to the invention. For each 2-tuple, the context can be incrementally updated from the last generated context. The probabilities are stored now on 14 bits instead of 16 bits which avoids 64-bits operations during the decoding process. Moreover, the probability model mapping was greatly optimized in some embodiments according to the invention. The worst case was drastically reduced and is limited to 10 iterations instead of 95.
As a result, the computational complexity of the proposed noiseless coding scheme was kept in the same range as in WD 5. A “pen and paper” estimate was performed by different versions of the noiseless coding and is recorded in the table of
To summarize the above, it can be seen that embodiments according to the present invention provide a particularly good trade-off between computational complexity, memory requirements and coding efficiency.
14. Bitstream Syntax
14.1 Payloads of the Spectral Noiseless Coder
In the following, some details regarding the payloads of the spectral noiseless coder will be described. In some embodiments, there is a plurality of different coding modes, such as, for example, a so-called “linear-prediction-domain” coding mode and a “frequency-domain” coding mode. In the linear-prediction-domain coding mode, a noise shaping is performed on the basis of a linear-prediction analysis of the audio signal, and a noise-shaped signal is encoded in the frequency-domain. In the frequency-domain coding mode a noise shaping is performed on the basis of a psychoacoustic analysis and a noise shaped version of the audio content is encoded in the frequency-domain.
Spectral coefficients from both the “linear-prediction-domain” coded signal and the “frequency-domain” coded signal are scalar quantized and then noiselessly coded by an adaptively context dependent arithmetic coding. The quantized coefficients are gathered together into 2-tuples before being transmitted from the lowest frequency to the highest frequency. Each 2-tuple is split into a sign s, the most significant 2-bits-wise-plane m, and the remaining one or more less-significant bit-planes r (if any). The value m is coded according to a context defined by the neighboring spectral coefficients. In other words, m is coded according to the coefficients neighborhood. The remaining less-significant bit-planes r are entropy coded without considering the context. By means of m and r, the amplitude of these spectral coefficients can be reconstructed on the decoder side. For all non-null symbols, the signs s is coded outside the arithmetic coder using 1-bit. In other words, the values m and r form the symbols of the arithmetic coder. Finally, the signs s, are coded outside of the arithmetic coder using 1-bit per non-null quantized coefficient.
A detailed arithmetic coding procedure is described herein.
14.2 Syntax Elements
In the following, the bitstream syntax of a bitstream carrying the arithmetically-encoded spectral information will be described taking reference to
a shows a syntax representation of so-called USAC raw data block (“usac_raw_data_block( )”).
The USAC raw data block comprises one or more single channel elements (“single_channel_element( )”) and/or one or more channel pair elements (“channel_pair_element( )”).
Taking reference now to
c shows a syntax representation of a channel pair element. A channel pair element comprises core mode information (“core_mode( )”, “core_mode1”). In addition, the channel pair element may comprise a configuration information “ics_info( )”. Additionally, depending on the core mode information, the channel pair element comprises a linear-prediction-domain channel stream or a frequency-domain channel stream associated with a first of the channels, and the channel pair element also comprises a linear-prediction-domain channel stream or a frequency-domain channel stream associated with a second of the channels.
The configuration information “ics_info( )”, a syntax representation of which is shown in
A frequency-domain channel stream (“fd_channel_stream( )”), a syntax representation of which is shown in
The arithmetically-coded spectral data (“ac_spectral_data( )”), a syntax representation of which is shown in
In the following, the structure of the arithmetically encoded data-block will be described taking reference to
The context for the encoding of the current set (e.g., 2-tuple) of spectral values is determined in accordance with the context determination algorithm shown at reference numeral 660. Details with respect to the context determination algorithm have been explained above, taking reference to
If, however, one or more less-significant bit-planes are required (in addition to the most-significant bit-plane) for a proper representation of the spectral values, this is signaled by using one or more arithmetic escape codewords (“ARITH_ESCAPE”). Thus, it can be generally said that for a spectral value, it is determined how many bit-planes (the most-significant bit-plane and, possibly, one or more additional less-significant bit-planes) are required. If one or more less-significant bit-planes are required, this is signaled by one or more arithmetic escape codewords “acod_m[pki][ARITH_ESCAPE]”, which are encoded in accordance with a currently selected cumulative-frequencies-table, a cumulative-frequencies-table-index of which is given by the variable “pki”. In addition, the context is adapted, as can be seen at reference numerals 664, 662, if one or more arithmetic escape codewords are included in the bitstream. Following the one or more arithmetic escape codewords, an arithmetic codeword “acod_m[pki][m]” is included in the bitstream, as shown at reference numeral 663, wherein “pki” designates the currently valid probability model index (taking the context adaptation caused by the inclusion of the arithmetic escape codewords into consideration) and wherein m designates the most-significant bit-plane value of the spectral value to be encoded or decoded (wherein m is different from the “ARITH_ESCAPE” codeword).
As discussed above, the presence of any less-significant bit-plane results in the presence of one or more codewords “acod_r[r]”, each of which represents 1 bit of a least-significant bit-plane of a first spectral value and each of which also represents 1 bit of a least-significant bit-plane of a second spectral value. The one or more codewords “acod_r[r]” are encoded in accordance with a corresponding cumulative-frequencies-table, which may, for example, be constant and context-independent. However, different mechanisms for the selection of the cumulative-frequencies-table for the decoding of the one or more codewords “acod_r[r]” are possible.
In addition, it should be noted that the context is updated after the encoding of each tuple of spectral values, as shown at reference numeral 668, such that the context is typically different for encoding and decoding two subsequent tuples of spectral values.
i shows a legend of definitions and help elements defining the syntax of the arithmetically encoded data-block.
Moreover, an alternative syntax of the arithmetic data “arith_data( )” is shown in
To summarize the above, a bitstream format has been described, which may be provided by the audio encoder 100 and which may be evaluated by the audio decoder 200. The bitstream of the arithmetically encoded spectral values is encoded such that it fits the decoding algorithm discussed above.
In addition, it should be generally noted that the encoding is the inverse operation of the decoding, such that it can generally be assumed that the encoder performs a table lookup using the above-discussed tables, which is approximately inverse to the table lookup performed by the decoder. Generally, it can be said that a man skilled in the art who knows the decoding algorithm and/or the desired bitstream syntax will easily be able to design an arithmetic encoder, which provides the data defined in the bitstream syntax and required by an arithmetic decoder.
Moreover, it should be noted that the mechanisms for determining the numeric current context value and for deriving a mapping rule index value may be identical in an audio encoder and an audio decoder, because it is typically desired that the audio decoder uses the same context as the audio encoder, such that the decoding is adapted to the encoding.
15. Implementation Alternatives
Although some aspects have been described in the context of an apparatus, it is clear that these aspects also represent a description of the corresponding method, where a block or device corresponds to a method step or a feature of a method step. Analogously, aspects described in the context of a method step also represent a description of a corresponding block or item or feature of a corresponding apparatus. Some or all of the method steps may be executed by (or using) a hardware apparatus, like for example, a microprocessor, a programmable computer or an electronic circuit. In some embodiments, some one or more of the most important method steps may be executed by such an apparatus.
The inventive encoded audio signal can be stored on a digital storage medium or can be transmitted on a transmission medium such as a wireless transmission medium or a wired transmission medium such as the Internet.
Depending on certain implementation requirements, embodiments of the invention can be implemented in hardware or in software. The implementation can be performed using a digital storage medium, for example a floppy disk, a DVD, a Blue-Ray, a CD, a ROM, a PROM, an EPROM, an EEPROM or a FLASH memory, having electronically readable control signals stored thereon, which cooperate (or are capable of cooperating) with a programmable computer system such that the respective method is performed. Therefore, the digital storage medium may be computer readable.
Some embodiments according to the invention comprise a data carrier having electronically readable control signals, which are capable of cooperating with a programmable computer system, such that one of the methods described herein is performed.
Generally, embodiments of the present invention can be implemented as a computer program product with a program code, the program code being operative for performing one of the methods when the computer program product runs on a computer. The program code may for example be stored on a machine readable carrier.
Other embodiments comprise the computer program for performing one of the methods described herein, stored on a machine readable carrier.
In other words, an embodiment of the inventive method is, therefore, a computer program having a program code for performing one of the methods described herein, when the computer program runs on a computer.
A further embodiment of the inventive methods is, therefore, a data carrier (or a digital storage medium, or a computer-readable medium) comprising, recorded thereon, the computer program for performing one of the methods described herein. The data carrier, the digital storage medium or the recorded medium are typically tangible and/or non-transitionary.
A further embodiment of the inventive method is, therefore, a data stream or a sequence of signals representing the computer program for performing one of the methods described herein. The data stream or the sequence of signals may for example be configured to be transferred via a data communication connection, for example via the Internet.
A further embodiment comprises a processing means, for example a computer, or a programmable logic device, configured to or adapted to perform one of the methods described herein.
A further embodiment comprises a computer having installed thereon the computer program for performing one of the methods described herein.
A further embodiment according to the invention comprises an apparatus or a system configured to transfer (for example, electronically or optically) a computer program for performing one of the methods described herein to a receiver. The receiver may, for example, be a computer, a mobile device, a memory device or the like. The apparatus or system may, for example, comprise a file server for transferring the computer program to the receiver.
In some embodiments, a programmable logic device (for example a field programmable gate array) may be used to perform some or all of the functionalities of the methods described herein. In some embodiments, a field programmable gate array may cooperate with a microprocessor in order to perform one of the methods described herein. Generally, the methods are preferably performed by any hardware apparatus.
The above described embodiments are merely illustrative for the principles of the present invention. It is understood that modifications and variations of the arrangements and the details described herein will be apparent to others skilled in the art. It is the intent, therefore, to be limited only by the scope of the impending patent claims and not by the specific details presented by way of description and explanation of the embodiments herein.
16. Conclusions
To conclude, embodiments according to the invention comprise one or more of the following aspects, wherein the aspects may be used individually or in combination.
a) Context state hashing mechanism
b). Incremental Context Update
c). Context Derivation
Some other technologies, which are applied in embodiments according to the invention, are described in non-pre-published patent applications PCT EP2101/065725, PCT EP2010/065726, and PCT EP 2010/065727. Moreover, in some embodiments according to the invention, a stop symbol is used. Moreover, in some embodiments, only the unsigned values are considered for the context.
However, the above-mentioned non-pre-published International patent applications disclose aspects which are still in use in some embodiments according to the invention.
For example, an identification of a zero-region is used in some embodiments of the invention. Accordingly, a so-called “small-value-flag” is set (e.g., bit 16 of the numeric current context value c).
In some embodiments, the region-dependent context computation may be used. However, in other embodiments, a region-dependent context computation may be omitted in order to keep the complexity and the size of the tables reasonably small.
Moreover, the context hashing using a hash function is an important aspect of the invention. The context hashing may be based on the two-table concept which is described in the above-referenced non-pre-published International patent applications. However, specific adaptations of the context hashing may be used in some embodiments in order to increase the computational efficiency. Nevertheless, in some other embodiments according to the invention, the context hashing which is described in the above-referenced non-pre-published International patent applications may be used.
Moreover, it should be noted that the incremental context hashing is rather simple and computationally efficient. Also, the context-independence from the sign of the values, which is used in some embodiments of the invention, helps to simplify the context, thereby keeping the memory requirements reasonably low.
In some embodiments of the invention, a context derivation using the sum of two spectral values and a context limitation is used. These two aspects can be combined. Both aim to limit the context order by conveying the most meaningful information from the neighborhood.
In some embodiments, a small-value-flag is used which may be similar to an identification of a group of a plurality of zero values.
In some embodiments according to the invention, an arithmetic stop mechanism is used. The concept is similar to the usage of a symbol “end-of-block” in JPEG, which has a comparable function. However, in some embodiments of the invention, the symbol (“ARITH_STOP”) is not included explicitly in the entropy coder. Instead, a combination of already existing symbols, which could not occur previously, is used, i.e. “ESC+0”. In other words, the audio decoder is configured to detect a combination of existing symbols, which are not normally used for representing a numeric value, and to interpret the occurrence of such a combination of already existing symbols as an arithmetic stop condition.
An embodiment according to the invention uses a two-table context hashing mechanism.
To further summarize, some embodiments according to the invention may comprise one or more of the following four main aspects.
To further conclude, one aspect of embodiments according to the present invention lies in an incremental context update. Embodiments according to the invention comprise an efficient concept for the update of the context, which avoids the extensive calculations of the working draft (for example, of the working draft 5). Rather, simple shift operations and logic operations are used in some embodiments. The simple context update facilitates the computation of the context significantly.
In some embodiments, the context is independent from the sign of the values (e.g., the decoded spectral values). This independence of the context from the sign of the values brings along a reduced complexity of the context variable. This concept is based on the finding that a neglect of the sign in the context does not bring along a severe degradation of the coding efficiency.
According to an aspect of the invention, the context is derived using the sum of two spectral values. Accordingly, the memory requirements for storage of the context are significantly reduced. Accordingly, the usage of a context value, which represents the sum of two spectral values, may be considered as advantageous in some cases.
Also, the context limitation brings along a significant improvement in some cases. In addition to the derivation of the context using the sum of two spectral values, the entries of the context array “q” are limited to a maximum value of “0xF” in some embodiments, which in turn results in a limitation of the memory requirements. This limitation of the values of the context array “q” brings along some advantages.
In some embodiments, a so-called “small value flag” is used. In obtaining the context variable c (which is also designated as a numeric current context value), a flag is set if the values of some entries “q[1][i−3]” to “q[1][i−1]” are very small. Accordingly, the computation of the context can be performed with high efficiency. A particularly meaningful context value (e.g. numeric current context value) can be obtained.
In some embodiments, an arithmetic stop mechanism is used. The “ARITH_STOP” mechanism allows for an efficient stop of the arithmetic encoding or decoding if there are only zero values left. Accordingly, the coding efficiency can be improved at moderate costs in terms of complexity.
According to an aspect of the invention, a two-table context hashing mechanism is used. The mapping of the context is performed using an interval-division algorithm evaluating the table “ari_hash_m” in combination with a subsequent lookup table evaluation of the table “ari_lookup_m”. This algorithm is more efficient than the WD3 algorithm.
In the following, some additional details will be discussed.
It should be noted here that the tables “arith_hash_m[600]” and “arith_lookup_m[600]” are two distinct tables. The first is used to map a single context index (e.g. numeric context value) to a probability model index (e.g., mapping rule index value) and the second is used for mapping a group of consecutive contexts, delimited by the context indices in “arith_hash_m[ ]”, into a single probability model.
It should further be noted that table “arith_cf_msb[96][16]” may be used as an alternative to the table “ari_cf_m[96][17]”, even though the dimensions are slightly different. “ari_cf_m[ ][ ]” and “ari_cf_msb[ ][ ]” may refer to the same table, as the 17th coefficients of the probability models are always zero. It is sometimes not taken into account when counting the required space for storing the tables.
To summarize the above, some embodiments according to the invention provide a proposed new noiseless coding (encoding or decoding), which engenders modifications in the MPEG USAC working draft (for example, in the MPEG USAC working draft 5). Said modifications can be seen in the enclosed figures and also in the related description.
As a concluding remark, it should be noted that the prefix “ari” and the prefix “arith” in names of variables, arrays, functions, and so on, are used interchangeably.
While this invention has been described in terms of several embodiments, there are alterations, permutations, and equivalents which fall within the scope of this invention. It should also be noted that there are many alternative ways of implementing the methods and compositions of the present invention. It is therefore intended that the following appended claims be interpreted as including all such alterations, permutations and equivalents as fall within the true spirit and scope of the present invention.
This application is a continuation of copending International Application No. PCT/EP2011/050273, filed Jan. 11, 2011, which claims priority from U.S. application Ser. No. 61/294,357, filed Jan. 12, 2010, each of which is incorporated herein in its entirety by this reference thereto.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5222189 | Fielder | Jun 1993 | A |
| 5388181 | Anderson et al. | Feb 1995 | A |
| 5659659 | Kolesnik et al. | Aug 1997 | A |
| 6029126 | Malvar | Feb 2000 | A |
| 6061398 | Satoh et al. | May 2000 | A |
| 6075471 | Kimura et al. | Jun 2000 | A |
| 6217234 | Dewar et al. | Apr 2001 | B1 |
| 6424939 | Herre et al. | Jul 2002 | B1 |
| 6538583 | Hallmark et al. | Mar 2003 | B1 |
| 6646578 | Au | Nov 2003 | B1 |
| 6864813 | Horie | Mar 2005 | B2 |
| 7079057 | Kim et al. | Jul 2006 | B2 |
| 7088271 | Marpe et al. | Aug 2006 | B2 |
| 7132964 | Tsuru | Nov 2006 | B2 |
| 7262721 | Jeon et al. | Aug 2007 | B2 |
| 7283073 | Chen | Oct 2007 | B2 |
| 7304590 | Park | Dec 2007 | B2 |
| 7330139 | Kim et al. | Feb 2008 | B2 |
| 7365659 | Hoffmann et al. | Apr 2008 | B1 |
| 7516064 | Vinton et al. | Apr 2009 | B2 |
| 7528749 | Otsuka et al. | May 2009 | B2 |
| 7528750 | Kim et al. | May 2009 | B2 |
| 7554468 | Xu | Jun 2009 | B2 |
| 7617110 | Kim et al. | Nov 2009 | B2 |
| 7656319 | Yu et al. | Feb 2010 | B2 |
| 7660720 | Oh et al. | Feb 2010 | B2 |
| 7714753 | Lu | May 2010 | B2 |
| 7777654 | Chang et al. | Aug 2010 | B2 |
| 7808406 | He et al. | Oct 2010 | B2 |
| 7821430 | Sakaguchi et al. | Oct 2010 | B2 |
| 7839311 | Bao et al. | Nov 2010 | B2 |
| 7840403 | Mehrotra et al. | Nov 2010 | B2 |
| 7864083 | Mahoney et al. | Jan 2011 | B2 |
| 7903824 | Faller et al. | Mar 2011 | B2 |
| 7932843 | Demircin et al. | Apr 2011 | B2 |
| 7948409 | Wu et al. | May 2011 | B2 |
| 7979271 | Bessette | Jul 2011 | B2 |
| 7982641 | Su et al. | Jul 2011 | B1 |
| 7991621 | Oh et al. | Aug 2011 | B2 |
| 8018996 | Chiba et al. | Sep 2011 | B2 |
| 8149144 | Mittal et al. | Apr 2012 | B2 |
| 8224658 | Lei et al. | Jul 2012 | B2 |
| 8301441 | Vos | Oct 2012 | B2 |
| 8321210 | Grill et al. | Nov 2012 | B2 |
| 8682645 | Taleb et al. | Mar 2014 | B2 |
| 20020016161 | Dellien et al. | Feb 2002 | A1 |
| 20030093451 | Chuang et al. | May 2003 | A1 |
| 20030206582 | Srinivasan et al. | Nov 2003 | A1 |
| 20040044527 | Thumpudi et al. | Mar 2004 | A1 |
| 20040044534 | Chen et al. | Mar 2004 | A1 |
| 20040114683 | Schwarz et al. | Jun 2004 | A1 |
| 20040184544 | Kondo et al. | Sep 2004 | A1 |
| 20050088324 | Fuchigami et al. | Apr 2005 | A1 |
| 20050117652 | Schwarz et al. | Jun 2005 | A1 |
| 20050192799 | Kim et al. | Sep 2005 | A1 |
| 20050203731 | Oh et al. | Sep 2005 | A1 |
| 20050231396 | Dunn | Oct 2005 | A1 |
| 20050289063 | Lecomte et al. | Dec 2005 | A1 |
| 20060047704 | Gopalakrishnan | Mar 2006 | A1 |
| 20060173675 | Ojanpera et al. | Aug 2006 | A1 |
| 20060232452 | Cha et al. | Oct 2006 | A1 |
| 20060238386 | Huang et al. | Oct 2006 | A1 |
| 20060284748 | Kim et al. | Dec 2006 | A1 |
| 20070016427 | Thumpudi et al. | Jan 2007 | A1 |
| 20070036228 | Tseng | Feb 2007 | A1 |
| 20070094027 | Vasilache | Apr 2007 | A1 |
| 20070126853 | Ridge et al. | Jun 2007 | A1 |
| 20070282603 | Bessette et al. | Dec 2007 | A1 |
| 20080094259 | Yu et al. | Apr 2008 | A1 |
| 20080133223 | Son et al. | Jun 2008 | A1 |
| 20080243518 | Oraevsky et al. | Oct 2008 | A1 |
| 20080267513 | Sankaran | Oct 2008 | A1 |
| 20090157785 | Reznik et al. | Jun 2009 | A1 |
| 20090190780 | Nagaraja et al. | Jul 2009 | A1 |
| 20090192790 | El-Maleh et al. | Jul 2009 | A1 |
| 20090192791 | El-Maleh et al. | Jul 2009 | A1 |
| 20090234644 | Reznik et al. | Sep 2009 | A1 |
| 20090299756 | Davis et al. | Dec 2009 | A1 |
| 20090299757 | Guo et al. | Dec 2009 | A1 |
| 20100007534 | Girardeau, Jr. | Jan 2010 | A1 |
| 20100070284 | Oh et al. | Mar 2010 | A1 |
| 20100088090 | Ramabadran | Apr 2010 | A1 |
| 20100256980 | Oshikiri et al. | Oct 2010 | A1 |
| 20100262420 | Herre et al. | Oct 2010 | A1 |
| 20100324912 | Choo et al. | Dec 2010 | A1 |
| 20110137661 | Morii et al. | Jun 2011 | A1 |
| 20110153333 | Bessette et al. | Jun 2011 | A1 |
| 20110238426 | Fuchs et al. | Sep 2011 | A1 |
| 20110320196 | Choo et al. | Dec 2011 | A1 |
| 20120033886 | Balster et al. | Feb 2012 | A1 |
| 20120069899 | Mehrotra et al. | Mar 2012 | A1 |
| 20120195375 | Wuebbolt | Aug 2012 | A1 |
| 20120207400 | Sasai et al. | Aug 2012 | A1 |
| 20120215525 | Jiang et al. | Aug 2012 | A1 |
| 20120245947 | Neuendorf et al. | Sep 2012 | A1 |
| 20120265540 | Fuchs et al. | Oct 2012 | A1 |
| 20120278086 | Fuchs et al. | Nov 2012 | A1 |
| 20120330670 | Fuchs et al. | Dec 2012 | A1 |
| 20130010983 | Disch et al. | Jan 2013 | A1 |
| 20130013301 | Subbaraman et al. | Jan 2013 | A1 |
| 20130013322 | Fuchs et al. | Jan 2013 | A1 |
| 20130013323 | Subbaraman et al. | Jan 2013 | A1 |
| Number | Date | Country |
|---|---|---|
| 101015216 | Aug 2007 | CN |
| 101160618 | Apr 2008 | CN |
| 2005223533 | Aug 2005 | JP |
| 2008506987 | Mar 2008 | JP |
| 2009518934 | May 2009 | JP |
| 2013507808 | Mar 2013 | JP |
| 200537436 | Nov 2005 | TW |
| 200746871 | Dec 2007 | TW |
| 200818123 | Apr 2008 | TW |
| I302664 | Nov 2008 | TW |
| 200935403 | Aug 2009 | TW |
| 200947419 | Nov 2009 | TW |
| WO-2006006936 | Jan 2006 | WO |
| WO-2007066970 | Jun 2007 | WO |
| WO-2008150141 | Dec 2008 | WO |
| WO 2011048098 | Apr 2011 | WO |
| WO 2011048099 | Apr 2011 | WO |
| WO 2011048100 | Apr 2011 | WO |
| WO-2011042366 | Apr 2011 | WO |
| Entry |
|---|
| Oger, M.; Ragot, S.; Antonini, M.; , “Transform Audio Coding with Arithmetic-Coded Scalar Quantization and Model-Based Bit Allocation,” Acoustics, Speech and Signal Processing, 2007. ICASSP 2007. IEEE International Conference on , vol. 4, No., pp. IV-545-IV-548, Apr. 15-20, 2007. |
| Neuendorf, M.; Gournay, P.; Multrus, M.; Lecomte, J.; Bessette, B.; Geiger, R.; Bayer, S.; Fuchs, G.; Hilpert, J.; Rettelbach, N.; Salami, R.; Schuller, G.; Lefebvre, R.; Grill, B., “Unified speech and audio coding scheme for high quality at low bitrates,” Acoustics, Speech and Signal Processing, 2009. ICASSP 2009. |
| Min Lu; Shuhua Zhang; Weibei Dou, “Dual-mode switching used for unified speech and audio codec,” Audio Language and Image Processing (ICALIP), 2010 International Conference on , vol., No., pp. 700,704, Nov. 23-25, 2010. |
| Sang-Wook Shin; Chang-Heon Lee; Hyen-O Oh; Hong-Goo Kang, “Designing a unified speech/audio codec by adopting a single channel harmonic source separation module,” Acoustics, Speech and Signal Processing, 2008. ICASSP 2008. IEEE International Conference on , vol., No., pp. 185,188, Mar. 31, 2008-Apr. 4, 2008. |
| MPEG2008/M15867, Fraunhofer IIS, VoiceAge Corp., “Detailed Technical Description of Reference Model 0 of the CfP on Unified Speech and Audio Coding (USAC),” International Organisation for Standardisation ISO/IEC JTC1/SC29/WG11 Coding of Moving Pictures and Audio, Oct. 2008, Busan, South Korea. |
| Subpart 4: General Audio Coding (GA)—AAC, TwinVQ, BSAC, ISO/IEC 14496-3:2005, Dec. 2005, pp. 1-344. |
| Imm, et al., “Lossless Coding of Audio Spectral Coeeficients using Selective Bitplane Coding”, Proc. 9th Int'l Symposium on Communications and Information Technology, IEEE, Sep. 2009, pp. 525-530. |
| Meine, et al., “Improved Quantization and Lossless Coding for Subband Audio Coding”, 118th AES Convention, vol. 1-4, XP040507276, May 2005, pp. 1-9. |
| Neuendorf, Max et al., “A Novel Scheme for Low Bitrate Unified Speech and Audio Coding—MPEG RMO”, “A Novel Scheme for Low Bitrate Unified Speech and Audio Coding—MPEG RMO”, XP040508995, AES 126th Convention, Paper 7713, Munich, Germany, May 2009, 13 pages. |
| Quackenbush, et al., Revised Report on Complexity of MPEG-2 AAC Tool, ISO/IEC JTC1/SC29/WG11 N2957 MPEG99, Melbourne, Oct. 1999, (Based Upon ISO/IEC JTC1/SC29/WG11 N2005, MPEG98, Feb. 1998, San José), pp. 1-17. |
| Sayood, K , “Introduction to Data Compression”, Third Edition, Chapter 4 “Arithmetic Coding,” 2006, Elsevier Inc., pp. 81-97. |
| Meine, Nikolaus et al.: “Improved Quantization and lossless coding for subband audio coding”, May 31, 2005, XP008071322. |
| Neuendorf, Max et al.: “A Novel Scheme for Low Bitrate Unified Speech and Audio Coding—MPEG RMO”, May 1, 2009, XP040508995. |
| Quackenbush, et al., “Revised Report on Complexity of MPEG-2 AAC Tools”, ISO/IEC JTC1/SC29/WG11 N2005, MPEG98, Feb. 1998, San José. |
| Sayood, K., “Introduction to Data Compression”, Third Edition, 2006, Elsevier Inc. |
| Neuendorf, Max et al., “Detailed Technical Description of Reference Model 0 of the CfP on Unified Speech and Audio Coding (USAC)”, ISO/IEC JTC1/SC29/WG11, MPEG2008/M15867, Busan, South Korea, Oct. 2008, 100 pp. |
| Wubbolt, Oliver , “Spectral Noiseless Coding CE: Thomson Proposal”, ISO/IEC JTC1/SC29/WG11, MPEG2009/M16953, Xian, China, Oct. 2009, 20 pp. |
| Yang, D et al., “High-Fidelity Multichannel Audio Coding”, EURASIP Book Series on Signal Processing and Communications. Hindawi Publishing Corporation., 2006, 12 Pages. |
| Yu, , “MPEG-4 Scalable to Lossless Audio Coding”, 117th AES Convention, Oct. 31, 2004, XP040372512, 1-14. |
| Number | Date | Country | |
|---|---|---|---|
| 20130013323 A1 | Jan 2013 | US |
| Number | Date | Country | |
|---|---|---|---|
| 61294357 | Jan 2010 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/EP2011/050273 | Jan 2011 | US |
| Child | 13547664 | US |
