Patent Grant
8612240
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 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 bit rate that may be used.
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 bitrate that may be used. 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; 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 onto a symbol code (symbol) in dependence on a context state; wherein the arithmetic decoder is configured to determine a numeric current context value describing the current context state in dependence on a plurality of previously decoded spectral values and also in dependence on whether a spectral value to be decoded is in a first predetermined frequency region or in a second predetermined frequency region.
According to another embodiment, an audio signal 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; an arithmetic encoder configured to encode spectral values, or a preprocessed version thereof, using a variable length codeword, wherein 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, wherein 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 context state, wherein the arithmetic encoder is configured to determine a numeric current context value describing the current context state in dependence on a plurality of previously encoded spectral values and also in dependence on whether a spectral value to be encoded is in a first predetermined frequency region or in a second predetermined frequency region.
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; and performing a frequency-domain-to-time-domain conversion, to provide a time-domain audio representation using the decoded spectral values, in order to acquire the decoded audio information; wherein a mapping rule describing a mapping of a code value onto a symbol code is selected in dependence on a context state; and wherein a numeric current context value describing the current context state is determined in dependence on a plurality of previously decoded spectral values and also in dependence on whether a spectral value to be decoded is in a first predetermined frequency region or in a second predetermined frequency region.
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 performing an energy-compacting time-domain-to-frequency-domain conversion, to provide 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 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 bit plane of a spectral value, is mapped onto a code value; wherein 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 is selected in dependence on a context state; wherein a numeric current context value describing the current context state is determined in dependence on a plurality of previously encoded spectral values and also in dependence on whether a spectral value to be encoded is in a first predetermined frequency region or in a second predetermined frequency region.
Another embodiment may have a computer program for performing the method for providing a decoded audio information on the basis of an encoded audio information, which method may have the steps of: providing a plurality of decoded spectral values on the basis of an arithmetically-encoded representation of the spectral values; and performing a frequency-domain-to-time-domain conversion, to provide a time-domain audio representation using the decoded spectral values, in order to acquire the decoded audio information; wherein a mapping rule describing a mapping of a code value onto a symbol code is selected in dependence on a context state; and wherein a numeric current context value describing the current context state is determined in dependence on a plurality of previously decoded spectral values and also in dependence on whether a spectral value to be decoded is in a first predetermined frequency region or in a second predetermined frequency region, when the computer program runs on a computer.
Another embodiment may have a computer program for performing the method for providing an encoded audio information on the basis of an input audio information, which method may have the steps of: performing an energy-compacting time-domain-to-frequency-domain conversion, to provide 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 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 bit plane of a spectral value, is mapped onto a code value; wherein 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 is selected in dependence on a context state; wherein a numeric current context value describing the current context state is determined in dependence on a plurality of previously encoded spectral values and also in dependence on whether a spectral value to be encoded is in a first predetermined frequency region or in a second predetermined frequency region, 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 code value (which may be extracted from a bitstream representing the encoded audio information) onto a symbol code (which may be a numeric value representing a decoded spectral value, or a most significant bitplane thereof) in dependence on a context state. The arithmetic decoder is configured to determine a numeric current context value describing the current context state in dependence on a plurality of previously decoded spectral values and also in dependence on whether a spectral value to be decoded is in a first predetermined frequency region or in a second predetermined frequency region.
It has been found that a consideration of the frequency region, in which a spectral value to be currently decoded lies, allows for a significant improvement of the quality of the context computation without significantly increasing the computational effort involved in the context computation. Moreover, by taking into consideration the fact that the statistical dependencies between previously decoded spectral values lying in a neighborhood of a spectral value to be decoded currently, vary over frequency, the context can be selected to allow for a high coding efficiency, both for decoding of spectral values associated with comparatively low frequencies and for decoding of spectral values associated with comparatively high frequencies. A good adaptation of the context to details of the statistical dependencies between the spectral value to be decoded currently and previously decoded spectral values (typically out of a direct or indirect neighborhood of the spectral value to be decoded currently) brings along the possibility to increase the coding efficiency while keeping the computational effort reasonably small. It has been found that the consideration of the frequency region is possible with very little effort, as a frequency index of the spectral value to be decoded currently is naturally known in the process of the arithmetic decoding. Thus, the selective adaptation of the context can be performed with little computational effort and still brings along an improvement of the coding efficiency.
In an advantageous embodiment, the arithmetic decoder is configured to selectively modify the numeric current context value in dependence on whether a spectral value to be decoded is in a first predetermined frequency region or in a second predetermined frequency region. A selective modification of the numeric current context value, in addition to a previous computation (or other determination) of the numeric current context value, allows a combination of a “normal” computation (or other determination) of the numeric current context value with a consideration of the frequency region in which the spectral values to be decoded currently lies. The “normal” computation of the numeric current context value may be handled separately from the region-dependent adaptation of the numeric current context value, which typically reduces the complexity of the algorithm and the computational effort. Also, it is easily possible to upgrade systems comprising a “normal” computation of the numeric current context value only using this concept.
In an advantageous embodiment, the arithmetic decoder is configured to determine the numeric current context value such that the numeric current context value is based on a combination of a plurality of previously decoded spectral values, or on a combination of a plurality of intermediate values derived from a plurality of previously decoded spectral values, and such that the numeric current context value is selectively increased over a value obtained on the basis of a combination of a plurality of previously decoded spectral values or on the basis of a combination of a plurality of intermediate values derived from a plurality of previously decoded spectral values, in dependence on whether a spectral value to be decoded is in a first predetermined frequency region or in a second predetermined frequency region. It has been found that a selective increase of the numeric current context value in dependence on the frequency region in which the spectral value to be decoded currently lies allows for an efficient evaluation of the numeric current context value while at the same time keeping the computation effort small.
In an advantageous embodiment, the arithmetic decoder is configured to distinguish at least between a first frequency region and a second frequency region in order to determine the numeric current context value, wherein the first frequency region comprises at least 15% of the spectral values associated with a given temporal portion (for example, a frame or a sub-frame) of the audio content, and wherein the first frequency region is a low-frequency region and comprises an associated spectral value having a lowest frequency (within the set of spectral values associated with the given (current) temporal portion of the audio content). It has been found that a good context adaptation can be achieved by commonly considering a lower part of a spectrum (comprising at least 15% of the spectral values) as a first frequency region, because the statistical dependencies between the spectral values do not comprise a strong variation over this low-frequency region. Accordingly, the number of different regions can be kept sufficiently small, which in turn helps to avoid the use of an excessive number of different mapping rules. However, in some embodiments it may be sufficient if the first frequency region comprises at least on spectral value, at least two spectral values or at least three spectral values, even though the choice of a more extended first spectral region is advantageous.
In an advantageous embodiment, the arithmetic decoder is configured to distinguish at least between a first frequency region and a second frequency region in order to determine the numeric current context value, wherein the second frequency region comprises at least 15% of the spectral values associated with a given temporal portion (for example, a frame or a sub-frame) of the audio content, and wherein the second frequency region is a high-frequency region and comprises an associated spectral value having a highest frequency (within the set of spectral values associated with the given (current) temporal portion of the audio content). It has been found that a good context adaptation can be achieved by commonly considering an upper part of a spectrum (comprising at least 15% of the spectral values) as a second frequency region, because the statistical dependencies between the spectral values do not comprise a strong variation over this high-frequency region. Accordingly, the number of different regions can be kept sufficiently small, which in turn helps to avoid the use of an excessive number of different mapping rules. However, in some embodiments it may be sufficient if the second frequency region comprises at least on spectral value, at least two spectral values or at least three spectral values, even though the choice of a more extended first spectral region is advantageous.
In an advantageous embodiment, the arithmetic decoder is configured to distinguish at least between a first frequency region, a second frequency region and a third frequency region, in order to determine the numeric current context value in dependence on a determination in which of the at least three frequency regions the spectral value to be decoded lies. In this case, each of the first frequency region, the second frequency region and the third frequency region comprises a plurality of associated spectral values. It has been found that for typical audio signals, it is recommendable to distinguish at least three different frequency regions, because there are typically at least three frequency regions in which there are different statistical dependencies between the spectral values. It has been found that it is recommendable (though not essential) to distinguish between three or more frequency regions even for narrow-band audio signals (for example, for audio signals having a frequency range between 300 Hz and 3 KHz). Also, for audio signals having a higher bandwidth, it has been found to be recommendable (though not essential) to distinguish three or more extended frequency regions (each having more than one spectral value associated therewith).
In an advantageous embodiment, at least one eighth of the spectral values of the (current) temporal portion of the audio information are associated with the first frequency region, and at least one fifth of the spectral values of the (current) temporal portion of the audio information are associated with the second frequency region, and at least one quarter of the spectral values of the (current) temporal portion of the audio information are associated with the third frequency region. It has been found that it is recommendable to have sufficiently large frequency regions, because such sufficiently large frequency regions bring along a good compromise between coding efficiency and computational complexity. Also, it has been found that the usage of very small frequency regions (for example, of frequency regions comprising only one spectral value associated therewith) is computationally inefficient and may even degrade the coding efficiency. Moreover, it should be noted that the choice of sufficiently large frequency regions (for example, of frequency regions comprising at least two spectral values associated therewith) is recommendable even when using only two frequency regions.
In an advantageous embodiment, the arithmetic decoder is configured to compute a sum comprising at least a first summand and a second summand, to obtain the numeric current context value as a result of the summation. In this case, the first summand is obtained by a combination of a plurality of intermediate values describing magnitudes of previously decoded spectral values, and the second summand describes to which frequency region, out of a plurality of frequency regions, a spectral value to be (currently) decoded is associated. Using such an approach, a separation between a context calculation based on a magnitude information about previously decoded spectral values and a context adaptation in dependence on the region to which the spectral value to be decoded currently is associated can be achieved. It has been found that the magnitudes of the previously decoded spectral values are an important indication about an environment of the spectral value to be decoded currently. However, it has also been found that the assessment of the statistical dependencies, which is based on an evaluation of the magnitudes of the previously decoded spectral values, can be improved by taking into consideration the frequency region to which the spectral value to be decoded currently is associated. However, it has been found that it is computationally sufficient to include the region information into the numeric current context value as a sum value, and that even such a simple mechanism brings along a good improvement of the numeric current context value.
In an advantageous embodiment, the arithmetic decoder is configured to modify one or more predetermined bit positions of a binary representation of the numeric current context value in dependence on a determination in which frequency region out of a plurality of different frequency regions the spectral value to be decoded lies. It has been found that the use of dedicated bit positions for the region information facilitates the selection of a mapping rule in dependence on the numeric current context value. For example, by using a predetermined bit position of the numeric current context value for a description of the frequency region to which the spectral value to be decoded currently is associated, the selection of a mapping rule can be simplified. For example, there are typically a number of context situations in which the same mapping rule may be used in the presence of a given neighborhood (in terms of spectral values) of the spectral value to be decoded currently, irrespective of the frequency region to which the spectral value to be decoded currently is associated. In such cases, the information regarding the frequency region, to which the spectral value to be decoded currently is associated, can be left unconsidered, which is facilitated by using a predetermined bit position for the encoding of the information. However, in other cases, i.e. for different environment constellations (in terms of spectral values) of the spectral value to be decoded currently, the information about the frequency region associated to the spectral values to be decoded currently can be exploited when selecting a mapping rule.
In an advantageous embodiment, the arithmetic decoder is configured to select a mapping rule in dependence on a numeric current context value, such that a plurality of different numeric current context values result in a selection of a same mapping rule. It has been found that the concept of taking into consideration the frequency region to which the spectral value to be decoded currently is associated may be combined with a concept in which the same mapping rule is associated with multiple different numeric current context values. It has been found that it is not necessary to consider the frequency, which is associated to the spectral value to be decoded currently, in all cases, but that it is recommendable to consider an information about the frequency region, to which the spectral value to be decoded currently is associated, at least in some cases.
In an advantageous embodiment, the arithmetic decoder configured to perform a two-stage selection of a mapping rule in dependence on the numeric current context value. In this case, the arithmetic decoder is configured to check, in a first selection step, whether the numeric current context value is equal to a significant state value described by an entry of a direct-hit table. The arithmetic decoder is also configured to determine, in a second selection step, which is only executed if the numeric current context value is different from the significant state values described by the entries of the direct-hit table, in which interval, out of a plurality of intervals, the numeric current context value lies. In this case, the arithmetic decoder is configured to select the mapping rule in dependence on a result of the first selection step and/or of the second selection step. The arithmetic decoder is also configured to select the mapping rule in dependence on whether a spectral value to be decoded is in a first frequency region or in a second frequency region. It has been found that a combination of the above-discussed concept for the computation of the numeric current context value with a two-step mapping rule selection brings along particular advantages. For example, using this concept, it is possible to define different “direct-hit” context configurations, to which a mapping rule is associated in the first selection step, for spectral values to be decoded and arranged in different frequency regions. Also, the second selection step, in which an interval-based selection of the mapping rule is performed, is well-suited for a handling of those situations (environments of previously decoded spectral values) in which it is not desired (or, at least, not necessary) to consider the frequency region to which the spectral value to be decoded currently is associated.
In an advantageous embodiment, the arithmetic decoder is configured to selectively modify one or more least-significant bit positions of a binary representation of the numeric current context value in dependence on a determination in which frequency region out of a plurality of different frequency regions the spectral value to be decoded lies. In this case, the arithmetic decoder is configured to determine, in the second selection step, in which interval out of a plurality of intervals the binary representation of the numeric current context value lies to select the mapping, such that some numeric current context values result in the selection of the same mapping rule independent from which frequency region the spectral value to be decoded lies in, and such that for some numeric current context values the mapping rule is selected in dependence on which frequency region the spectral value to be coded lies in. It has been found that the mechanism in which the frequency region is encoded in the least-significant bits of a binary representation of the numeric current context value is very well suited for an efficient cooperation with the two-step mapping rule selection.
An embodiment according to the invention creates an audio encoder for providing an encoded audio information on the basis of an 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 arithmetic encoder is configured to encode a spectral value, 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 (which may be included into a bitstream representing the input audio information in an encoded form). 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 the spectral value, onto a code value in dependence on a context state. The arithmetic encoder is configured to determine a numeric current context value describing the current context state in dependence on a plurality of previously encoded spectral values and also in dependence on whether a spectral value to be encoded is in a first predetermined frequency region or in a second predetermined frequency region.
This audio signal encoder is based on the same findings as the audio signal decoder discussed above. It has been found that the mechanism for the adaptation of the context, which has been shown to be efficient for the decoding of an audio content, should also be applied at the encoder side, in order to allow for a consistent system.
An embodiment according to the invention creates a method for providing decoded audio information on the basis of encoded audio information.
Yet another embodiment according to the invention creates a method for providing 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.
The methods and the computer program are based on the same findings as the above described audio decoder and the above described audio encoder.
Embodiments according to the present invention will subsequently be described taking reference to the enclosed figures, in which:
a shows a pseudo-program-code representation of an algorithm “arith_map_context( )” for mapping a context;
b and 5c show 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 an algorithm “get_pk(s)” for deriving a cumulative-frequencies-table index value “pki” from a state variable;
e shows a pseudo-program-code representation of an algorithm “arith_get_pk(s)” for deriving a cumulative-frequencies-table index value “pki” from a state value;
f shows a pseudo-program-code representation of an algorithm “get_pk(unsigned long s)” for deriving a cumulative-frequencies-table index value “pki” from a state value;
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 pseudo-program-code representation of an algorithm “arith_update_context( )” for updating the context;
i shows a legend of definitions and variables;
a shows as 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 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 a 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 the working draft 4 of the USAC draft standard;
b shows a schematic representation of a context for a state calculation, as it is used in embodiments according to the invention;
a shows an overview of the table as used in the arithmetic coding scheme according to the working draft 4 of the USAC draft standard;
b shows an overview of the table 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 schemes according to the present invention and according to the working draft 4 of the USAC draft standard;
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 the working draft 4 of the USAC draft standard;
a shows a table representation of average bitrates which are used by a unified-speech-and-audio-coding coder, using an arithmetic coder according to the working draft 3 of the USAC draft standard and an arithmetic decoder according to an embodiment of the present invention;
b shows a table representation of a bit reservoir control for a unified-speech-and-audio-coding coder, using the arithmetic coder according to the working draft 3 of the USAC draft standard and the arithmetic coder according to an embodiment of the present invention;
FIGS. 17(1) and 17(2) show a table representation of a content of a table “ari_s_hash[387]”;
FIGS. 19(1) and 19(2) show a table representation of a content of a table “ari_cf_m[64][9]”; and
FIGS. 20(1) and 20(2) show a table representation of a content of a table “ari_s_hash[387];
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 730 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 context state. The arithmetic encoder is configured to determine the current context state in dependence on a plurality of previously-encoded adjacent spectral values. For this purpose, the arithmetic encoder is configured to detect a group of a plurality of previously-encoded adjacent spectral values, which fulfill, individually or taken together, a predetermined condition regarding their magnitudes, and determine the current context state in dependence on a result of the detection.
As can be seen, the mapping of a spectral value or of a most-significant bit-plane of a spectral value onto a code value 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 and may comprise a group detector 752 to detect a group of a plurality of previously-encoded adjacent spectral values which fulfill, individually or taken together, the predetermined condition regarding their magnitudes. The state tracker 750 is also advantageously configured to determine the current context state in dependence on the result of said detection performed by the group detector 752. Accordingly, the state tracker 750 provides an information 754 describing the current context state. A mapping rule selector 760 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 760 provides the mapping rule information 742 to the spectral encoding 740.
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, a detection is performed in order to detect whether there is a group of a plurality of previously-encoded adjacent spectral values which fulfill, individually or taken together, a predetermined condition regarding their magnitudes. The result of this detection is applied in the selection of the current context state, i.e. in the selection of a mapping rule. By detecting whether there is a group of a plurality of spectral values which are particularly small or particularly large, it is possible to recognize special features within the frequency-domain audio representation, which may be a time-frequency representation. Special features such as, for example, a group of a plurality of particularly small or particularly large spectral values, indicate that a specific context state should be used as this specific context state may provide a particularly good coding efficiency. Thus, the detection of the group of adjacent spectral values which fulfill the predetermined condition, which is typically used in combination with an alternative context evaluation based on a combination of a plurality of previously-coded spectral values, provides a mechanism which allows for an efficient selection of an appropriate context if the input audio information takes some special states (e.g., comprises a large masked frequency range).
Accordingly, an efficient encoding can be achieved while keeping the context calculation sufficiently simple.
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 the mapping in dependence on a mapping rule, which may be described by a mapping rule information 828a.
The arithmetic decoder 820 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 826a). The arithmetic decoder 820 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 826 may be used, which receives an information describing the previously-decoded spectral values. The arithmetic decoder is also configured to detect a group of a plurality of previously-decoded adjacent spectral values, which fulfill, individually or taken together, a predetermined condition regarding their magnitudes, and to determine the current context state (described, for example, by the context state information 826a) in dependence on a result of the detection.
The detection of the group of a plurality of previously-decoded adjacent spectral values which fulfill the predetermined condition regarding their magnitudes may, for example, be performed by a group detector, which is part of the state tracker 826. Accordingly, a current context state information 826a is obtained. The selection of the mapping rule may be performed by a mapping rule selector 828, which derives a mapping rule information 828a from the current context state information 826a, and which provides the mapping rule information 828a to the spectral value determinator 824.
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 an 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 detecting a group of a plurality of previously-decoded adjacent spectral values which fulfill, individually or taken together, a predetermined condition regarding their magnitudes, it is possible to adapt the mapping rule to special conditions (or patterns) of previously-decoded spectral values. For example, a specific mapping rule may be selected if a group of a plurality of comparatively small previously-decoded adjacent spectral values is identified, or if a group of a plurality of comparatively large previously-decoded adjacent spectral values is identified. It has been found that the presence of a group of comparatively large spectral values or of a group of comparatively small spectral values may be considered as a significant indication that a dedicated mapping rule, specifically adapted to such a condition, should be used. Accordingly, a context computation can be facilitated (or accelerated) by exploiting the detection of such a group of a plurality of spectral values. Also, characteristics of an audio content can be considered that could not be considered as easily without applying the above-mentioned concept. For example, the detection of a group of a plurality of spectral values which fulfill, individually or taken together, a predetermined condition regarding their magnitudes, can be performed on the basis of a different set of spectral values, when compared to the set of spectral values used for a normal context computation.
Further details will be described below.
3. 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 advantageously 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, 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.
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 advantageously 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”. 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 64 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 to the codeword determinator. Thus, the codeword determinator 180 may use the selected cumulative-frequencies-table for the provision of the codeword acod_m[pki][m] of the most-significant hit-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.
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) is encoded to obtain an arithmetic codeword “acod_m[pki][m]” of a most-significant bit-plane value. One or more less-significant bit-planes (each of the less-significant bit-planes comprising, for example, one, two or three bits) 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, 64 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.
4. Audio Decoder
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, and a codeword “acod_r” representing a content of a less-significant bit-plane of the spectral value a 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 rescaled 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 rescaled frequency-domain audio representation 242 and to provide, on the basis thereof, a pre-processed version 252 of the inversely-quantized and rescaled 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 rescaled frequency-domain audio representation 242 (or, alternatively, the inversely-quantized and rescaled 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, rescaled and pre-processed. Accordingly, an inversely-quantized, rescaled 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, rescaled 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 64 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 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 the 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 64 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 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 64 different cumulative-frequencies-tables in dependence on a state index 298, which is obtained by observing the previously-computed decoded spectral values.
5. Overview Over 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 put 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 are inversed.
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 to further reduce the redundancy of the quantized spectrum, which is obtained, for example, by an energy-compacting time-domain to a 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. The noiseless coding is fed by (original or encoded representations of) quantized spectral values and uses context-dependent cumulative-frequencies-tables derived, for example, from a plurality of previously-decoded neighboring spectral values. Here, the 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 in dependence on the respective probabilities. The binary code is generated by mapping a probability interval, where the set of symbol lies, to a codeword.
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, seven previously-decoded neighboring spectral values
Here, the 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.
6. Decoding Process
6.1 Decoding Process Overview
In the following, an overview of the process of decoding 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. The initialization 310 of the context comprises a derivation of the current context from a previous context using the function “arith_map_context (1g)”. The derivation of the current context from a previous context may 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 314, which context update is performed by a function “Arith_update_context(a,i,1g)” which is described below. The spectral value decoding 312 and the context update 314 are repeated 1g times, wherein 1g indicates the number of spectral values to be decoded (e.g. for an audio frame). The spectral value decoding 312 comprises a context-value calculation 312a, a most-significant bit-plane decoding 312b, and a less-significant bit-plane addition 312c.
The state value computation 312a comprises the computation of a first state value s using the function “arith_get_context(i, 1g, arith_reset_flag, N/2)” which function returns the first state value s. The state value computation 312a also comprises a computation of a level value “lev0” and of a level value “lev”, which level values “lev0”, “lev” are obtained by shifting the first state value s to the right by 24 bits. The state value computation 312a also comprises a computation of a second state value t according to the formula shown in
The most-significant bit-plane decoding 312b comprises an iterative execution of a decoding algorithm 312ba, wherein a variable j is initialized to 0 before a first execution of the algorithm 312ba.
The algorithm 312ba comprises a computation of a state index “pki” (which also serves as a cumulative-frequencies-table index) in dependence on the second state value t, and also in dependence on the level values “lev” and lev0, using a function “arith_get_pk( )”, which is discussed below. The algorithm 312ba also comprises the selection of a cumulative-frequencies-table in dependence on the state index pki, wherein a variable “cum_freq” may be set to a starting address of one out of 64 cumulative-frequencies-tables in dependence on the state index pki. Also, a variable “cfl” may be initialized to a length of the selected cumulative-frequencies-table, which is, for example, equal to the number of symbols in the alphabet, i.e. the number of different values which can be decoded. The lengths of all the cumulative-frequencies-tables from “arith_cf_m[pki=0][9]” to “arith_cf_m[pki=63][9]” available for the decoding of the most-significant bit-plane value m is 9, as eight different most-significant bit-plane values and an escape symbol can be decoded. 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 therefore skipped. Accordingly, execution of the process is continued with the setting of the spectral value a to be equal to the most-significant bit-plane value m (instruction “a=m”). In contrast, if the decoded most-significant bit-plane value m is identical to the arithmetic escape symbol “ARITH_ESCAPE”, the level value “lev” is increased by one. As mentioned, the algorithm 312ba is then repeated until the decoded most-significant bit-plane value m is different from the arithmetic escape symbol.
As soon as most-significant bit-plane decoding is completed, i.e. a most-significant bit-plane value m different from the arithmetic escape symbol has been decoded, the spectral value variable “a” is set to be equal to the most-significant bit-plane value m. Subsequently, the less-significant bit-planes are obtained, for example, as shown at reference numeral 312c in
6.2 Decoding Order According to
In the following, the decoding order of the spectral values will be described.
Spectral coefficients are noiselessly coded and transmitted (e.g. in the bitstream) starting from the lowest-frequency coefficient and progressing to the highest-frequency coefficient.
Coefficients from an advanced audio coding (for example obtained using a modified-discrete-cosine-transform, as discussed in ISO/IEC 14496, part3, subpart 4) are stored in an array called “x_ac_quant[g][win][sfb][bin]”, and the order of transmission of the noiseless-coding-codeword (e.g. acod_m, acod_r) is such that when they are decoded in the order received and stored in the array, “bin” (the frequency index) is the most rapidly incrementing index and “g” is the most slowly incrementing index.
Spectral coefficients associated with a lower frequency are encoded before spectral coefficients associated with a higher frequency.
Coefficients from the transform-coded-excitation (tcx) are stored directly in an 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 slowest incrementing index. 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 are associated to adjacent and increasing frequencies of the transform-coded-excitation.
Spectral coefficients associated to a lower frequency are encoded before spectral coefficients associated with a higher frequency.
Notably, the audio decoder 200 may be configured to apply the decoded frequency-domain audio 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 an 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 200, 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 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).
6.3. 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( )”, which is shown in
The variable “1g” describes a number of spectral coefficients to decode in the frame. The variable “previous—1g” describes a previous number of spectral lines of a previous frame.
A mapping of the context may be performed in accordance with the algorithm “arith_map_context( )”. It should be noted here that the function “arith_map_context( )” sets the entries q[0][i] of the current context array q to the values qs[i] of the past context array qs, if the number of spectral values associated with the current (e.g. frequency-domain-encoded) audio frame is identical to the number of spectral values associated with the previous audio frame for i=0 to i=1g−1.
However, 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 present invention, such that reference is made to the pseudo program code of
6.4 State Value Computation According to
In the following, the state value computation 312a will be described in more detail.
It should be noted that the first state value s (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 value 420 may, nevertheless, be evaluated for a detection of a plurality of previously-decoded adjacent spectral values which fulfill, individually or taken together, a predetermined condition regarding their magnitudes.
Taking reference now to
It should be noted that the function “arith_get_context( )” receives, as input variables an index i of the spectral value to decode. The index i is typically a frequency index. An input variable 1g describes a (total) number of expected quantized coefficients (for a current audio frame). A variable N describes a number of lines of the transformation. A flag “arith_reset_flag” indicates whether the context should be reset. The function “arith_get_context” provides, as an output value, a variable “t”, which represents a concatenated state index s and a predicted bit-plane level lev0.
The function “arith_get_context( )” uses integer variables a0, c0, c1, c2, c3, c4, c5, c6, lev0, and “region”.
The function “arith_get_context( )” comprises as main functional blocks, a first arithmetic reset processing 510, a detection 512 of a group of a plurality of previously-decoded adjacent zero spectral values, a first variable setting 514, a second variable setting 516, a level adaptation 518, a region value setting 520, a level adaptation 522, a level limitation 524, an arithmetic reset processing 526, a third variable setting 528, a fourth variable setting 530, a fifth variable setting 532, a level adaptation 534, and a selective return value computation 536.
In the first arithmetic reset processing 510, it is checked whether the arithmetic reset flag “arith_reset_flag” is set, while the index of the spectral value to decode is equal to zero. In this case, a context value of zero is returned, and the function is aborted.
In the detection 512 of a group of a plurality of previously-decoded zero spectral values, which is only performed if the arithmetic reset flag is inactive and the index i of the spectral value to decode is different from zero, a variable named “flag” is initialized to 1, as shown at reference numeral 512a, and a region of spectral value that is to be evaluated is determined, as shown at reference numeral 512b. Subsequently, the region of spectral values, which is determined as shown at reference number 512b, is evaluated as shown at reference numeral 512c. If it is found that there is a sufficient region of previously-decoded zero spectral values, a context value of 1 is returned, as shown at reference numeral 512d. For example, an upper frequency index boundary “lim_max” is set to i+6, unless index i of the spectral value to be decoded is close to a maximum frequency index 1g−1, in which case a special setting of the upper frequency index boundary is made, as shown at reference numeral 512b. Moreover, a lower frequency index boundary “lim_min” is set to −5, unless the index i of the spectral value to decode is close to zero (i+lim_min<0), in which case a special computation of the lower frequency index boundary lim_min is performed, as shown at reference numeral 512b. When evaluating the region of spectral values determined in step 512b, an evaluation is first performed for negative frequency indices k between the lower frequency index boundary lim_min and zero. For frequency indices k between lim_min and zero, it is verified whether at least one out of the context values q[0][k].c and q[1][k].c is equal to zero. If, however, both of the context values q[0][k].c and q[1][k].c are different from zero for any frequency indices k between lim_min and zero, it is concluded that there is no sufficient group of zero spectral values and the evaluation 512c is aborted. Subsequently, context values q[0][k].c for frequency indices between zero and lim_max are evaluated. If it found that any of the context values q[0][k].c for any of the frequency indices between zero and lim_max is different from zero, it is concluded that there is no sufficient group of previously-decoded zero spectral values, and the evaluation 512c is aborted. If, however, it is found that for every frequency indices k between lim_min and zero, there is at least one context value q[0][k].c or q[1][k].c which is equal to zero and if there is a zero context value q[0][k].c for every frequency index k between zero and lim_max, it is concluded that there is a sufficient group of previously-decoded zero spectral values. Accordingly, a context value of 1 is returned in this case to indicate this condition, without any further calculation. In other words, calculations 514, 516, 518, 520, 522, 524, 526, 528, 530, 532, 534, 536 are skipped, if a sufficient group of a plurality of context values q[0][k].c, q[1][k].c having a value of zero is identified. In other words, the returned context value, which describes the context state (s), is determined independent from the previously decoded spectral values in response to the detection that the predetermined condition is fulfilled.
Otherwise, i.e. if there is no sufficient group of context values [q][0][k].c, [q][1][k].c, which are zero at least some of the computations 514, 516, 518, 520, 522, 524, 526, 528, 530, 532, 534, 536 are executed.
In the first variable setting 514, which is selectively executed if (and only if) index i of the spectral value to be decoded is less than 1, the variable a0 is initialized to take the context value q[1][i−1], and the variable c0 is initialized to take the absolute value of the variable a0. The variable “lev0” is initialized to take the value of zero. Subsequently, the variables “lev0” and c0 are increased if the variable a0 comprises a comparatively large absolute value, i.e. is smaller than −4, or larger or equal to 4. The increase of the variables “lev0” and c0 is performed iteratively, until the value of the variable a0 is brought into a range between −4 and 3 by a shift-to-the-right operation (step 514b).
Subsequently, the variables c0 and “lev0” are limited to maximum values of 7 and 3, respectively (step 514c).
If the index i of the spectral value to be decoded is equal to 1 and the arithmetic reset flag (“arith_reset_flag”) is active, a context value is returned, which is computed merely on the basis of the variables c0 and lev0 (step 514d). Accordingly, only a single previously-decoded spectral value having the same time index as the spectral value to decode and having a frequency index which is smaller, by 1, than the frequency index i of the spectral value to be decoded, is considered for the context computation (step 514d). Otherwise, i.e. if there is no arithmetic reset functionality, the variable c4 is initialized (step 514e).
To conclude, in the first variable setting 514, the variables c0 and “lev0” are initialized in dependence on a previously-decoded spectral value, decoded for the same frame as the spectral value to be currently decoded and for a preceding spectral bin i−1. The variable c4 is initialized in dependence on a previously-decoded spectral value, decoded for a previous audio frame (having time index t−1) and having a frequency which is lower (e.g., by one frequency bin) than the frequency associated with the spectral value to be currently decoded.
The second variable setting 516 which is selectively executed if (and only if) the frequency index of the spectral value to be currently decoded is larger than 1, comprises an initialization of the variables c1 and c6 and an update of the variable lev0. The variable c1 is updated in dependence on a context value q[1][i−2].c associated with a previously-decoded spectral value of the current audio frame, a frequency of which is smaller (e.g. by two frequency bins) than a frequency of a spectral value currently to be decoded. Similarly, variable c6 is initialized in dependence on a context value q[0][i−2].c, which describes a previously-decoded spectral value of a previous frame (having time index t−1), an associated frequency of which is smaller (e.g. by two frequency bins) than a frequency associated with the spectral value to currently be decoded. In addition, the level variable “lev0” is set to a level value q[1][i−2].1 associated with a previously-decoded spectral value of the current frame, an associated frequency of which is smaller (e.g. by two frequency bins) than a frequency associated with the spectral value to currently be decoded, if q[1][i−2].1 is larger than lev0.
The level adaptation 518 and the region value setting 520 are selectively executed, if (and only if) the index i of the spectral value to be decoded is larger than 2. In the level adaptation 518, the level variable “lev0” is increased to a value of q[1][i−3].1 if the level value q[1][i−3].1 which is associated to a previously-decoded spectral value of the current frame, an associated frequency of which is smaller (e.g. by three frequency bins) than the frequency associated with the spectral value to currently be decoded, is larger than the level value lev0.
In the region value setting 520, a variable “region” is set in dependence on an evaluation, in which spectral region, out of a plurality of spectral regions, the spectral value to currently be decoded is arranged. For example, if it is found that the spectral value to be currently decoded is associated to a frequency bin (having frequency bin index i) which is in the first (lower most) quarter of the frequency bins (0≦i<N/4), the region variable “region” is set to zero. Otherwise, if the spectral value currently to be decoded is associated to a frequency bin which is in a second quarter of the frequency bins associated to the current frame (N/4≦i<N/2), the region variable is set to a value of 1. Otherwise, i.e. if the spectral value currently to be decoded is associated to a frequency bin which is in the second (upper) half of the frequency bins (N/2≦i<N), the region variable is set to 2. Thus, a region variable is set in dependence on an evaluation to which frequency region the spectral value currently to be decoded is associated. Two or more frequency regions may be distinguished.
An additional level adaptation 522 is executed if (and only if) the spectral value currently to be decoded comprises a spectral index which is larger than 3. In this case, the level variable “lev0” is increased (set to the value q[1][i−4].1) if the level value q[i][i−4].1, which is associated to a previously-decoded spectral value of the current frame, which is associated to a frequency which is smaller, for example, by four frequency bins, than a frequency associated to the spectral value currently to be decoded is larger than the current level “lev0” (step 522). The level variable “lev0” is limited to a maximum value of 3 (step 524).
If an arithmetic reset condition is detected and the index i of the spectral value currently to be decoded is larger than 1, the state value is returned in dependence on the variables c0, c1, lev0, as well as in dependence on the region variable “region” (step 526). Accordingly, previously-decoded spectral values of any previous frames are left out of consideration if an arithmetic reset condition is given.
In the third variable setting 528, the variable c2 is set to the context value q[0][i].c, which is associated to a previously-decoded spectral value of the previous audio frame (having time index t−1), which previously-decoded spectral value is associated with the same frequency as the spectral value currently to be decoded.
In the fourth variable setting 530, the variable c3 is set to the context value q[0][i+1].c, which is associated to a previously-decoded spectral value of the previous audio frame having a frequency index i+1, unless the spectral value currently to be decoded is associated with the highest possible frequency index 1g−1.
In the fifth variable setting 532, the variable c5 is set to the context value q[0][i+2].c, which is associated with a previously-decoded spectral value of the previous audio frame having frequency index i+2, unless the frequency index i of the spectral value currently to be decoded is too close to the maximum frequency index value (i.e. takes the frequency index value 1g−2 or 1g−1).
An additional adaptation of the level variable “lev0” is performed if the frequency index i is equal to zero (i.e. if the spectral value currently to be decoded is the lowermost spectral value). In this case, the level variable “lev0” is increased from zero to 1, if the variable c2 or c3 takes a value of 3, which indicates that a previously-decoded spectral value of a previous audio frame, which is associated with the same frequency or even a higher frequency, when compared to the frequency associated with the spectral value currently to be encoded, takes a comparatively large value.
In the selective return value computation 536, the return value is computed in dependence on whether the index i of the spectral values currently to be decoded takes the value zero, 1, or a larger value. The return value is computed in dependence on the variables c2, c3, c5 and lev0, as indicated at reference numeral 536a, if index i takes the value of zero. The return value is computed in dependence on the variables c0, c2, c3, c4, c5, and “lev” as shown at reference numeral 536b, if index i takes the value of 1. The return value is computed in dependence on the variable c0, c2, c3, c4, c1, c5, c6, “region”, and lev0, if the index i takes a value which is different from zero or I (reference numeral 536c).
To summarize the above, the context value computation “arith_get_context( )” comprises a detection 512 of a group of a plurality of previously-decoded zero spectral values (or at least, sufficiently small spectral values). If a sufficient group of previously-decoded zero spectral values is found, the presence of a special context is indicated by setting the return value to 1. Otherwise, the context value computation is performed. It can generally be said that in the context value computation, the index value i is evaluated in order to decide how many previously-decoded spectral values should be evaluated. For example, a number of evaluated previously-decoded spectral values is reduced if a frequency index i of the spectral value currently to be decoded is close to a lower boundary (e.g. zero), or close to an upper boundary (e.g. 1g−1). In addition, even if the frequency index i of the spectral value currently to be decoded is sufficiently far away from a minimum value, different spectral regions are distinguished by the region value setting 520. Accordingly, different statistical properties of different spectral regions (e.g. first, low frequency spectral region, second, medium frequency spectral region, and third, high frequency spectral region) are taken into consideration. The context value, which is calculated as a return value, is dependent on the variable “region”, such that the returned context value is dependent on whether a spectral value currently to be decoded is in a first predetermined frequency region or in a second predetermined frequency region (or in any other predetermined frequency region).
6.5 Mapping Rule Selection
In the following, the selection of a mapping rule, for example, a cumulative-frequencies-table, which describes a mapping of a code value onto a symbol code, will be described. The selection of the mapping rule is made in dependence on the context state, which is described by the state value s or t.
6.5.1 Mapping Rule Selection Using the Algorithm According to
In the following, the selection of a mapping rule using the function “get_pk” according to
It should also be noted that a function “get_pk” according to
The function “get_pk” receives, as an input variable, a state value s, which may be obtained by a combination of the variable “t” according to
The function “get_pk” comprises a first table evaluation 540, and a second table evaluation 544. The first table evaluation 540 comprises a variable initialization 541 in which the variables i_min, i_max, and i are initialized, as shown at reference numeral 541. The first table evaluation 540 also comprises an iterative table search 542, in the course of which a determination is made as to whether there is an entry of the table “ari_s_hash” which matches the state value s. If such a match is identified during the iterative table search 542, the function get_pk is aborted, wherein a return value of the function is determined by the entry of the table “ari_s_hash” which matches the state value s, as will be explained in more detail. If, however, no perfect match between the state value s and an entry of the table “ari_s_hash” is found during the course of the iterative table search 542, a boundary entry check 543 is performed.
Turning now to the details of the first table evaluation 540, it can be seen that a search interval is defined by the variables i_min and i_max. The iterative table search 542 is repeated as long as the interval defined by the variables i_min and i_max is sufficiently large, which may be true if the condition i_max−i_min>1 is fulfilled. Subsequently, the variable i is set, at least approximately, to designate the middle of the interval (i=i_min+(i_max−i_min)/2). Subsequently, a variable j is set to a value which is determined by the array “ari_s_hash” at an array position designated by the variable i (reference numeral 542). It should be noted here that each entry of the table “ari_s_hash” describes both, a state value, which is associated to the table entry, and a mapping rule index value which is associated to the table entry. The state value, which is associated to the table entry, is described by the more-significant bits (bits 8-31) of the table entry, while the mapping rule index values are described by the lower bits (e.g. bits 0-7) of said table entry. The lower boundary i_min or the upper boundary i_max are adapted in dependence on whether the state value s is smaller than a state value described by the most-significant 24 bits of the entry “ari_s_hash[i]” of the table “ari_s_hash” referenced by the variable i. For example, if the state value s is smaller than the state value described by the most-significant 24 bits of the entry “ari_s_hash[i]”, the upper boundary i_max of the table interval is set to the value i. Accordingly, the table interval for the next iteration of the iterative table search 542 is restricted to the lower half of the table interval (from i_min to i_max) used for the present iteration of the iterative table search 542. If, in contrast, the state value s is larger than the state values described by the most-significant 24 bits of the table entry “ari_s_hash[i]”, then the lower boundary i_min of the table interval for the next iteration of the iterative table search 542 is set to value i, such that the upper half of the current table interval (between i_min and i_max) is used as the table interval for the next iterative table search. If, however, it is found that the state value s is identical to the state value described by the most-significant 24 bits of the table entry “ari_s_hash[i]”, the mapping rule index value described by the least-significant 8-bits of the table entry “ari_s_hash[i]” is returned by the function “get_pk”, and the function is aborted.
The iterative table search 542 is repeated until the table interval defined by the variables i_min and i_max is sufficiently small.
A boundary entry check 543 is (optionally) executed to supplement the iterative table search 542. If the index variable i is equal to index variable i_max after the completion of the iterative table search 542, a final check is made whether the state value s is equal to a state value described by the most-significant 24 bits of a table entry “ari_s_hash[i_min]”, and a mapping rule index value described by the least-significant 8 bits of the entry “ari_s_hash[i_min]” is returned, in this case, as a result of the function “get_pk”. In contrast, if the index variable i is different from the index variable i_max, then a check is performed as to whether a state value s is equal to a state value described by the most-significant 24 bits of the table entry “ari_s_hash[i_max]”, and a mapping rule index value described by the least-significant 8 bits of said table entry “ari_s_hash[i_max]” is returned as a return value of the function “get_pk” in this case.
However, it should be noted that the boundary entry check 543 may be considered as optional in its entirety.
Subsequent to the first table evaluation 540, the second table evaluation 544 is performed, unless a “direct hit” has occurred during the first table evaluation 540, in that the state value s is identical to one of the state values described by the entries of the table “ari_s_hash” (or, more precisely, by the 24 most-significant bits thereof).
The second table evaluation 544 comprises a variable initialization 545, in which the index variables i_min, i and i_max are initialized, as shown at reference numeral 545. The second table evaluation 544 also comprises an iterative table search 546, in the course of which the table “ari_gs_hash” is searched for an entry which represents a state value identical to the state value s. Finally, the second table search 544 comprises a return value determination 547.
The iterative table search 546 is repeated as long as the table interval defined by the index variables i_min and i_max is large enough (e.g. as long as i_max−i_min>1). In the iteration of the iterative table search 546, the variable i is set to the center of the table interval defined by i_min and i_max (step 546a). Subsequently, an entry j of the table “ari_gs_hash” is obtained at a table location determined by the index variable i (546b). In other words, the table entry “ari_gs_hash[i]” is a table entry at the center of the current table interval defined by the table indices i_min and i_max. Subsequently, the table interval for the next iteration of the iterative table search 546 is determined. For this purpose, the index value i_max describing the upper boundary of the table interval is set to the value i, if the state value s is smaller than a state value described by the most-significant 24 bits of the table entry “j=ari_gs_hash[i]” (546c). In other words, the lower half of the current table interval is selected as the new table interval for the next iteration of the iterative table search 546 (step 546c). Otherwise, if the state value s is larger than a state value described by the most-significant 24 bits of the table entry “j=ari_gs_hash[i]”, the index value i_min is set to the value i. Accordingly, the upper half of the current table interval is selected as the new table interval for the next iteration of the iterative table search 546 (step 546d). If, however, it is found that the state value s is identical to a state value described by the uppermost 24 bits of the table entry “j=ari_gs_hash[i]”, the index variable i_max is set to the value i+1 or to the value 224 (if i+1 is larger than 224), and the iterative table search 546 is aborted. However, if the state value s is different from the state value described by the 24 most-significant bits of “j=ari_gs_hash[i]”, the iterative table search 546 is repeated with the newly set table interval defined by the updated index values i_min and i_max, unless the table interval is too small (i_max−i_min≦1). Thus, the interval size of the table interval (defined by i_min and i_max) is iteratively reduced until a “direct hit” is detected (s==(j>>8)) or the interval reaches a minimum allowable size (i_max−i_min≦1). Finally, following an abortion of the iterative table search 546, a table entry “j=ari_gs_hash[i_max]” is determined and a mapping rule index value, which is described by the 8 least-significant bits of said table entry “j=ari_gs_hash[i_max]” is returned as the return value of the function “get_pk”. Accordingly, the mapping rule index value is determined in dependence on the upper boundary i_max of the table interval (defined by i_min and i_max) after the completion or abortion of the iterative table search 546.
The above-described table evaluations 540, 544, which both use iterative table search 542, 546, allow for the examination of tables “ari_s_hash” and “ari_gs_hash” for the presence of a given significant state with very high computational efficiency. In particular, a number of table access operations can be kept reasonably small, even in a worst case. It has been found that a numeric ordering of the table “ari_s_hash” and “ari_gs_hash” allows for the acceleration of the search for an appropriate hash value. In addition, a table size can be kept small as the inclusion of escape symbols in tables “ari_s_hash” and “ari_gs_hash” is not required. Thus, an efficient context hashing mechanism is established even though there are a large number of different states: In a first stage (first table evaluation 540), a search for a direct hit is conducted (s==(j>>8)).
In the second stage (second table evaluation 544) ranges of the state value s can be mapped onto mapping rule index values. Thus, a well-balanced handling of particularly significant states, for which there is an associated entry in the table “ari_s_hash”, and less-significant states, for which there is a range-based handling, can be performed. Accordingly, the function “get_pk” constitutes an efficient implementation of a mapping rule selection.
For any further details, reference is made to the pseudo program code of
6.5.2 Mapping Rule Selection Using the Algorithm According to
In the following, another algorithm for a selection of the mapping rule will be described taking reference to
It should be noted that the function “arith_get_pk” according to
It should also be noted that the function “arith_get_pk” may, for example, evaluate the table ari_s_hash according to
The function “arith_get_pk” according to
If a direct hit is not identified within the first table evaluation 550, a second table evaluation 560 is executed. In the course of the second table evaluation, a linear scan with entry indices i increasing linearly from zero to a maximum value of 224 is performed. During the second table evaluation, an entry “ari_gs_hash[i]” of the table “ari_gs_hash” for table i is read, and the table entry “j=ari_gs_hash[i]” is evaluated in that it is determined whether the state value represented by the 24 most-significant bits of the table entry j is larger than the state value s. If this is the case, a mapping rule index value described by the 8 least-significant bits of said table entry j is returned as the return value of the function “arith_get_pk”, and the execution of the function “arith_get_pk” is aborted. If, however, the state value s is not smaller than the state value described by the 24 most-significant bits of the current table entry j=ari_gs_hash[i], the scan through the entries of the table ari_gs_hash is continued by increasing the table index i. If, however, the state value s is larger than or equal to any of the state values described by the entries of the table ari_gs hash, a mapping rule index value “pki” defined by the 8 least-significant bits of the last entry of the table ari_gs_hash is returned as the return value of the function “arith_get_pk”.
To summarize, the function “arith_get_pk” according to
6.5.3 Mapping Rule Selection Using the Algorithm According to
The function “get_pk” according to
It should be noted that the function “get_pk” according to
6.6. Function “arith_decode( )” According to
In the following, the functionality of the function “arith_decode( )” will be discussed in detail taking reference to
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. Also, the function “arith_decode( )” uses the input variable “cfl”, which indicates the length of the selected cumulative-frequencies-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, 20 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 1048575.
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 524286, nothing is done, and the interval renormalization continues with an interval-size-increase operation 570fb. If, however, the variable “high” is not smaller than 524286 and the variable “low” is greater than or equal to 524286, the variables “values”, “low” and “high” are all reduced by 524286, 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 524286, and that the variable “low” is not greater than or equal to 524286, and that the variable “low” is greater than or equal to 262143 and that the variable “high” is smaller than 786429, the variables “value”, “low” and “high” are all reduced by 262143, 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 may be used 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 that may be used 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
6.7 Escape Mechanism
While the decoded most-significant bit-plane value m (which is returned as a symbol value by the function “arith_decode( )” is the escape symbol “ARITH_ESCAPE”, an additional most-significant bit-plane value m is decoded and the variable “lev” is incremented by 1. Accordingly, an information is obtained about the numeric significance of the most-significant bit-plane value m as well as on the number of less-significant bit-planes to be decoded.
If an escape symbol “ARITH_ESCAPE” is decoded, the level variable “lev” is increased by 1. Accordingly, the state value which is input to the function “arith_get_pk” is also modified in that a value represented by the uppermost bits (bits 24 and up) is increased for the next iterations of the algorithm 312ba.
6.8 Context Update According to
Once the spectral value is completely decoded (i.e. all of the least-significant bit-planes have been added, the context tables q and qs are updated by calling the function “arith_update_context(a,i,1g))”. In the following, details regarding the function “arith_update_context(a,i,1g)” will be described taking reference to
The function “arith_update_context( )” receives, as input variables, the decoded quantized spectral coefficient a, the index i of the spectral value to be decoded (or of the decoded spectral value) and the number 1g of spectral values (or coefficients) associated with the current audio frame.
In a step 580, the currently decoded quantized spectral value (or coefficient) a is copied into the context table or context array q. Accordingly, the entry q[1][i] of the context table q is set to a. Also, the variable “a0” is set to the value of “a”.
In a step 582, the level value q[1][i].1 of the context table q is determined. By default, the level value q[1][i].1 of the context table q is set to zero. However, if the absolute value of the currently coded spectral value a is larger than 4, the level value q[1][i].1 is incremented. With each increment, the variable “a” is shifted to the right by one bit. The increment of the level value q[1][i].1 is repeated until the absolute value of the variable a0 is smaller than, or equal to, 4.
In a step 584, a 2-bit context value q[1][i].c of the context table q is set. The 2-bit context value q[1][i].c is set to the value of zero if the currently decoded spectral value a is equal to zero. Otherwise, if the absolute value of the decoded spectral value a is smaller than, or equal to, 1, the 2-bit context value q[1][i].c is set to 1. Otherwise, if the absolute value of the currently decoded spectral value a is smaller than, or equal to, 3, the 2-bit context value q[1][i].c is set to 2. Otherwise, i.e. if the absolute value of the currently decoded spectral value a is larger than 3, the 2-bit context value q[1][i].c is set to 3. Accordingly, the 2-bit context value q[1][i].c is obtained by a very coarse quantization of the currently decoded spectral coefficient a.
In a subsequent step 586, which is only performed if the index i of the currently decoded spectral value is equal to the number 1g of coefficients (spectral values) in the frame, that is, if the last spectral value of the frame has been decoded) and the core mode is a linear-prediction-domain core mode (which is indicated by “core_mode==1”), the entries q[1][j].c are copied into the context table qs[k]. The copying is performed as shown at reference numeral 586, such that the number 1g of spectral values in the current frame is taken into consideration for the copying of the entries q[1][j].c to the context table qs[k]. In addition, the variable “previous—1g” takes the value 1024.
Alternatively, however, the entries q[1][j].c of the context table q are copied into the context table qs[j] if the index i of the currently decoded spectral coefficient reaches the value of 1g and the core mode is a frequency-domain core mode (indicated by “core_mode==0”).
In this case, the variable “previous—1g” is set to the minimum between the value of 1024 and the number 1g of spectral values in the frame.
6.9 Summary of the 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 quantized spectral coefficients a are noiselessly coded and transmitted, starting from the lowest frequency coefficient and progressing to the highest frequency coefficient.
The coefficients from the advanced-audio coding (AAC) are stored in the array “x_ac_quant[g][win][sfb][bin]”, and 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. Index bin designates frequency bins. The index “sfb” designates scale factor bands. The index “win” designates windows. The index “g” designates audio frames.
The coefficients from the transform-coded-excitation are stored directly in an 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.
First, a mapping is done between the saved past context stored in the context table or array “qs” and the context of the current frame q (stored in the context table or array q). The past context “qs” is stored onto 2-bits per frequency line (or per frequency bin).
The mapping between the saved past context stored in the context table “qs” and the context of the current frame stored in the context table “q” is performed using the function “arith_map_context( )”, a pseudo-program-code representation of which is shown in
The noiseless decoder outputs signed quantized spectral coefficients “a”.
At first, the state of the context is calculated based on the previously-decoded spectral coefficients surrounding the quantized spectral coefficients to decode. The state of the context s corresponds to the 24 first bits of the value returned by the function “arith_get_context( )”. The bits beyond the 24th bit of the returned value correspond to the predicted bit-plane-level lev0. The variable “lev” is initialized to lev0. A pseudo program code representation of the function “arith_get_context” is shown in
Once the state s and the predicted level “lev” are known, the most-significant 2-bits wise plane m is decoded using the function “arith_decode( )”, fed with the appropriated cumulative-frequencies-table corresponding to the probability model corresponding to the context state.
The correspondence is made by the function “arith_get_pk( )”.
A pseudo-program-code representation of the function “arith_get_pk( )” is shown in
A pseudo program code of another function “get_pk” which may take the place 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 the function “arith_get_pk( )” (or, alternatively, by the function “get_pk( )”).
The arithmetic coder is an integer implementation using the method of tag generation with scaling (see, e.g., K. Sayood “Introduction to Data Compression” third edition, 2006, Elsevier Inc.). The pseudo-C-code shown in
When the decoded value m is the escape symbol, “ARITH_ESCAPE”, another value m is decoded and the variable “lev” is incremented by 1. Once the value m is not the escape symbol, “ARITH_ESCAPE”, the remaining bit-planes are then decoded from the most-significant to the least-significant level, by calling “lev” times the function “arith_decode( )” with the cumulative-frequencies-table “arith_cf_r[ ]”. Said cumulative-frequencies-table “arith_cf_r[ ] may, for example, describe an even probability distribution.
The decoded bit planes r permit the refining of the previously-decoded value m in the following manner:
Once the spectral quantized coefficient a is completely decoded, the context tables q, or the stored context qs, is updated by the function “arith_update_context( )”, for the next quantized spectral coefficients to decode.
A pseudo program code representation of the function “arith_update_context( )” is shown in
In addition, a legend of the definitions is shown in
7. Mapping Tables
In an embodiment according to the invention, particularly advantageous tables “ari_s_hash” and “ari_gs_hash” and “ari_cf_m” are used for the execution of the function “get_pk”, which has been discussed with reference to
7.1. Table “ari_s_hash[387]” According to
A content of a particularly advantageous implementation of the table “ari_s_hash”, which is used by the function “get_pk” which was described with reference to
It should further be noted that the most-significant 24 bits of the table entries of the table “ari_s_hash” represent state values, while the least-significant 8-bits represent mapping rule index values pki.
Thus, the entries of the table “ari_s_hash” describe a “direct hit” mapping of a state value onto a mapping rule index value “pki”.
7.2 Table “ari_gs_hash” According to
A content of a particularly advantageous embodiment of the table “ari_gs_hash” is shown in the table of
It should be noted that the entries of the table “ari_gs_hash” are listed in an ascending order of the table index i for table index values i between zero and 224. The term “0x” indicates that the table entries are described in a hexadecimal format. Accordingly, the first table entry “0X00000401” corresponds to table entry “ari_gs_hash[0]” having table index 0 and the last table entry “0Xffffff3f” corresponds to table entry “ari_gs_hash[224]” having table index 224.
It should also be noted that the table entries are ordered in a numerically ascending manner, such that the table entries are well-suited for the second table evaluation 544 of the function “get_pk”. The most-significant 24 bits of the table entries of the table “ari_gs_hash” describe boundaries between ranges of state values, and the 8 least-significant bits of the entries describe mapping rule index values “pki” associated with the ranges of state values defined by the 24 most-significant bits.
7.3 Table “ari_cf_m” According to
As can be seen from
Within a line (e.g. a line 1910 or a line 1912 or a line 1964), a leftmost value describes a first entry of a cumulative-frequencies-table and a rightmost value describes the last entry of a cumulative-frequencies-table.
Accordingly, each line 1910, 1912, 1964 of the table representation of
7.4 Table “ari_s_hash” According to
The table “ari_s_hash” according to
The “0x” indicates that the table entries are represented in a hexadecimal form. The 24 most-significant bits of the entries of the table “ari_s_hash” describe significant states, and the 8 least-significant bits of the entries of the table “ari_s_hash” describe mapping rule index values.
Accordingly, the entries of the table “ari_s_hash” describe a mapping of significant states onto mapping rule index values “pki”.
8. 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 computation complexity, memory requirements, and coding efficiency.
Generally speaking, the embodiments according to the invention create an improved spectral noiseless coding.
The present description describes embodiments for the CE on improved spectral noiseless coding of spectral coefficients. The proposed scheme is based on the “original” context-based arithmetic coding scheme, as described in the working draft 4 of the USAC draft standard, but significantly reduces memory requirements (RAM, ROM), while maintaining a noiseless coding performance. A lossless transcoding of WD3 (i.e. of the output of an audio encoder providing a bitstream in accordance with the working draft 3 of the USAC draft standard) was proven to be possible. The scheme described herein is, in general, scalable, allowing further alternative tradeoffs between memory requirements and encoding performance. Embodiments according to the invention aim at replacing the spectral noiseless coding scheme as used in the working draft 4 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 4 (WD4) of the USAC draft standard. Spectral coefficients previous in frequency or in time model a context. This context is used for the selection of cumulative-frequencies-tables for the arithmetic coder (encoder or decoder). Compared to the embodiment according to WD4, the context modeling is further improved and the tables holding the symbol probabilities were retrained. The number of different probability models was increased from 32 to 64.
Embodiments according to the invention reduce the table sizes (data ROM demand) to 900 words of length 32-bits or 3600 bytes. In contrast, embodiments according to WD4 of the USAC draft standard may use 16894.5 words or 76578 bytes. The static RAM demand is reduced, in some embodiments according to the invention, from 666 words (2664 bytes) to 72 (288 bytes) per core coder channel. At the same time, it fully preserves the coding performance and can even reach a gain of approximately 1.04% to 1.39%, compared to the overall data rate over all 9 operating points. All working draft 3 (WD3) bitstreams can be transcoded in a lossless manner without affecting the bit reservoir constraints.
The proposed scheme according to the embodiments of the invention is scalable: flexible tradeoffs between memory demand and coding performance are possible. By increasing the table sizes to the coding gain can be further increased.
In the following, a brief discussion of the coding concept according to WD4 of the USAC draft standard will be provided to facilitate the understanding of the advantages of the concept described herein. In USAC WD4, 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. According to WD4, a maximum number of 16 spectral coefficients are used as context, 12 of which are previous in time. Both, spectral coefficients used for the context and to be decoded, are grouped as 4-tuples (i.e. four spectral coefficients neighbored in frequency, see
For the complete WD4 noiseless coding scheme, a memory demand (ROM) of 16894.5 words (67578 bytes) may be used. Additionally, 666 words (2664 byte) of static ROM per core-coder channel may be used to store the states for the next frame.
The table representation of
A total memory demand of a complete USAC WD4 decoder is estimated to be 37000 words (148000 byte) for data ROM without a 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 tables and the large individual tables exceed typical cache sizes as provided by fixed point chips for low-budget portable devices, which is in a typical range of 8-32 kByte (e.g. ARM9e, TIC64xx, 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.
In the following, the proposed new scheme will briefly be described.
To overcome the problems mentioned above, an improved noiseless coding scheme is proposed to replace the scheme as in WD4 of the USAC draft standard. As a context based arithmetic coding scheme, it is based on the scheme of WD4 of the USAC draft standard, but features a modified scheme for the derivation of cumulative-frequencies-tables from the context. Further on, context derivation and symbol coding is performed on granularity of a single spectral coefficient (opposed to 4-tuples, as in WD4 of the USAC draft standard). In total, 7 spectral coefficients are used for the context (at least in some cases). By reduction in mapping, one of in total 64 probability models or cumulative frequency tables (in WD4: 32) is selected.
b shows a graphical representation of a context for the state calculation, as used in the proposed scheme (wherein a context used for the zero region detection is not shown in
In the following, a brief discussion will be provided regarding the reduction of the memory demand, which can be achieved by using the proposed coding scheme. The proposed new scheme exhibits a total ROM demand of 900 words (3600 Bytes) (see the table of
Compared to the ROM demand of the noiseless coding scheme in WD4 of the USAC draft standard, the ROM demand is reduced by 15994.5 words (64978 Bytes)(see also
Further on, the amount of information needed for the context derivation in the next frame (static RAM) is also reduced. According to WD4, the complete set of coefficients (maximally 1152) with a resolution of typically 16-bits additional to a group index per 4-tuple of 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, which is used in embodiments according to the invention, reduces the persistent information to only 2-bits per spectral coefficient, which sums up to 72 words (288 Bytes) in total per core-coder channel. The demand on static memory can be reduced by 594 words (2376 Bytes).
In the following, some details regarding a possible increase of coding efficiency will be described. The coding efficiency of embodiments according to the new proposal was compared against the reference quality bitstreams according to WD3 of the USAC draft standard. The comparison was performed by means of a transcoder, based on a reference software decoder. For details regarding the comparison of the noiseless coding according to WD3 of the USAC draft standard and the proposed coding scheme, reference is made to
Although the memory demand is drastically reduced in embodiments according to the invention when compared to embodiments according to WD3 or WD4 of the USAC draft standard, the coding efficiency is not only maintained, but slightly increased. The coding efficiency is on average increased by 1.04% to 1.39%. For details, reference is made to the table of
By measurement of the bit reservoir fill level, it was shown that the proposed noiseless coding is able to losslessly transcode the WD3 bitstream for every operating point. For details, reference is made to the table of
Details on average bitrates per operating mode, minimum, maximum and average bitrates on a frame basis and a best/worst case performance on a frame basis can be found in the tables of
In addition, it should be noted that embodiments according to the present invention provide a good scalability. By adapting the table size, a tradeoff between memory requirements, computational complexity and coding efficiency can be adjusted in accordance with the requirements.
9. Bitstream Syntax
9.1. 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 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, a “linear-prediction domain” coded signal and a “frequency-domain” coded signal are scalar quantized and then noiselessly coded by an adaptively context dependent arithmetic coding. The quantized coefficients are transmitted from the lowest-frequency to the highest-frequency. Each individual quantized coefficient is split into the most significant 2-bits-wise plane m, and the remaining less-significant bit-planes r. The value m is coded according to the coefficient's neighborhood. The remaining less-significant bit-planes r are entropy-encoded, without considering the context. The values m and r form the symbols of the arithmetic coder.
A detailed arithmetic decoding procedure is described herein.
9.2. 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_mode0”, “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
The structure of the arithmetically-encoded data block will be described taking reference to
The context for the encoding of the current set 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 discussed above taking reference to
If, however, one or more less-significant bit-planes may be used (in addition to the most-significant bit plane) for a proper representation of the spectral value, 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) may be used. If one or more less-significant bit planes may be used, 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 into consideration the context adaptation caused by the inclusion of the arithmetic escape codewords), and wherein m designates the most-significant bit-plane value of the spectral value to be encoded or decoded.
As discussed above, the presence of any less-significant-bit planes results in the presence of one or more codewords “acod_r[r]”, each of which represents one bit of the least-significant bit plane. The one or more codewords “acod_r[r]” are encoded in accordance with a corresponding cumulative-frequencies-table, which is constant and context-independent.
In addition, it should be noted that the context is updated after the encoding of each spectral value, as shown at reference numeral 668, such that the context is typically different for encoding of two subsequent spectral values.
h shows a legend of definitions and help elements defining the syntax of the arithmetically-encoded data block.
To summarize the above, a bitstream format has been described, which may be provided by the audio coder 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 that is defined in the bitstream syntax and may be used by the arithmetic decoder.
10. Further Embodiments According to
In the following, some further simplified embodiments according to the invention will be described.
The arithmetic encoder 2130 comprises a mapping rule selection 2132 and a context value determination 2136. The arithmetic encoder is configured to select a mapping rule describing a mapping of a spectral value 2124, or of a most significant bit plane of a spectral value 2124, onto a code value (which may represent a variable length codeword) in dependence on a numeric current context value describing a context state. The arithmetic decoder is configured to determine a numeric current context value 2134, which is used for the mapping rule selection 2132, in dependence on a plurality of previously encoded spectral values and also in dependence on whether a spectral value to be encoded is in a first predetermined frequency region or in a second predetermined frequency region. Accordingly, the mapping 2131 is adapted to the specific characteristics of the different frequency regions.
The arithmetic decoder 2220 comprises a mapping 2225, which is used to map a code value (for example, a code value extracted from a bit stream representing the encoded audio information) onto a symbol code (which symbol code may describe, for example, a decoded spectral value or a most significant bit plane of the decoded spectral value). The arithmetic decoder further comprises a mapping rule selection 2226, which provides a mapping rule selection information 2227 to be mapping 2225. The arithmetic decoder 2220 also comprises a context value determination 2228, which provides a numeric current context value 2229 to the mapping rule selection 2226. The arithmetic decoder 2220 is configured to select a mapping rule describing a mapping of a code value (for example, a code value extracted from a bit stream representing the encoded audio information) onto a symbol code (for example, a numeric value representing the decoded spectral value or a numeric value representing a most significant bit plane of the decoded spectral value) in dependence on a context state. The arithmetic decoder is configured to determine a numeric current context value describing the current context state in dependence on a plurality of previously decoded spectral values and also in dependence on whether a spectral value to be decoded is in a first predetermined frequency region or in a second predetermined frequency region.
Accordingly, different characteristics of different frequency regions are considered in the mapping 2225, which typically brings along increased coding efficiency without significantly increasing the computational effort.
11. 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.
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.
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 advantageously 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.
While the foregoing has been particularly shown and described with reference to particular embodiments above, it will be understood by those skilled in the art that various other changes in the forms and details may be made without departing from the sprit and cope thereof. It is to be understood that various changes may be made in adapting to different embodiments without departing from the broader concept disclosed herein and comprehended by the claims that follow.
12. Conclusion
To conclude, it can be noted that embodiments according to the invention create an improved spectral noiseless coding scheme. Embodiments according to the new proposal allows for the significant reduction of the memory demand from 16894.5 words to 900 words (ROM) and from 666 words to 72 (static RAM per core-coder channel). This allows for the reduction of the data ROM demand of the complete system by approximately 43% in one embodiment. Simultaneously, the coding performance is not only fully maintained, but on average even increased. A lossless transcoding of WD3 (or of a bitstream provided in accordance with WD3 of the USAC draft standard) was proven to be possible. Accordingly, an embodiment according to the invention is obtained by adopting the noiseless decoding described herein into the upcoming working draft of the USAC draft standard.
To summarize, in an embodiment the proposed new noiseless coding may engender the modifications in the MPEG USAC working draft with respect to the syntax of the bitstream element “arith_data( )” as shown in
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/EP2010/065726, filed Oct. 19, 2010, which is incorporated herein by reference in its entirety, and additionally claims priority from U.S. Application No. 61/253,459, filed Oct. 20, 2009, which is also incorporated herein by reference in its entirety.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5222189 | Fielder | Jun 1993 | A |
| 5388181 | Anderson et al. | Feb 1995 | 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 | Herre 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 | 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 |
| 20020016161 | Dellien et al. | Feb 2002 | 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 |
| 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 |
| 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 |
| 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 |
| 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 |
|---|
| “Subpart 4: General Audio Coding (GA)—AAC, TwinVQ, BSAC”, ISO/IEC 14496-3:2005, Dec. 2005, pp. 1-344., Dec. 2005, pp. 1-344. |
| Lu, M. et al., “Dual-mode switching used for unified speech and audio codec”, Int'l Conference on Audio Language and Image Processing 2010 (ICALIP), Nov. 23-25, 2010, pp. 700-704. |
| Meine, et al., “Improved Quantization and Lossless Coding for Subband Audio Coding”, 118th AES Convention, vol. 1-4, XP040507276, May 2005, pp. 1-9., May 31, 2005, 1-9. |
| Neuendorf, et al., “Detailed Technical Description of Reference Model 0 of the CfP on Unified Speech and Audio Coding (USAC)”, Int'l Organisation for Standardisation ISO/IEC JTC1/SC29/WG11 Coding of Moving Pictures and Audio, MPEG2008/M15867, Busan, South Korea, Oct. 2008, 95 pages. |
| Neuendorf, et al., “Unified Speech and Audio Coding Scheme for High Quality at Low Bitrates”, IEEE Int'l Conference on Acoustics, Speech and Signal Processing, Apr. 19-24, 2009, 4 pages. |
| Neuendorf, Max et al., “A Novel Scheme for Low Bitrate Unified Speech and Audio Coding—MPEG RMO”, AES 126th Convention, Paper 7713, Munich, Germany. XP040508995, May 2009, 13 Pages. |
| Oger, M. et al., “Transform Audio Coding with Arithmetic-Coding Scalar Quantization and Model-Based Bit Allocation”, IEEE Int'l Conference on Acoustics, Speech and Signal Processing 2007 (ICASSP 2007); vol. 4, Apr. 15-20, 2007, pp. IV-545-IV-548. |
| Quackenbush, et al., “Revised Report on Complexity of MPEG-2 AAC Tools”, Quackenbush, et al., “Revised Report on Complexity of MPEG-2 AAC Tools”, ISO/IEC JTC1/SC29/WG11 N2957, 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”, Sayood, K., “Introduction to Data Compression”, Third Edition, Chapter 4 “Arithmetic Coding,” 2006, Elsevier Inc., pp. 81-97. |
| Shin, Sang-Wook et al., “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, IEEE, Piscataway, NJ, USA, Mar. 31-Apr. 4, 2008, pp. 185-188. |
| 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. |
| 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., Sep. 2009, pp. 525-530. |
| 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. |
| Number | Date | Country | |
|---|---|---|---|
| 20120278086 A1 | Nov 2012 | US |
| Number | Date | Country | |
|---|---|---|---|
| 61253459 | Oct 2009 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/EP2010/065726 | Oct 2010 | US |
| Child | 13450699 | US |
