Patent Application
20070124138
The present invention relates to coding and decoding digital signals, in particular in applications that transmit or store multimedia signals such as audio signals (speech and/or sound).
In the field of compression coding, many coders model a signal of L samples using a number of pulses very much less than the total number of samples. This is the case of certain audio-frequency coders, for example, such as the “TDAC” audio coder described in particular in the published document US-2001/027393, in which modified normalized discrete cosine transform coefficients in each band are quantized by vectorial quantifiers using algebraic dictionaries of interleaved size, these algebraic codes generally including a few components that are non-zero, the other components being equal to zero. This is also the case with most speech coders using analysis by synthesis, in particular coders of the Algebraic Code Excited Linear Prediction (ACELP), Multi-Pulse Maximum Likelihood Quantization (MP-MLQ) and other types. To model the innovation signal, these coders use a directory composed of waveforms having very few components that are non-zero, having positions and amplitudes that additionally obey predetermined rules.
Coders of the above kind using analysis by synthesis are briefly described below.
In coders using analysis by synthesis, a synthesis model is used on coding to extract parameters modeling the signals to be coded, which may be sampled at the telephone frequency (Fe=8 kilohertz (kHz)) or at a higher frequency, for example at 16 kHz for broadened band coding (passband from 50 hertz (Hz) to 7 kHz). Depending on the application and on the required quality, the compression rate varies from 1 to 16. These coders operate at bit rates from 2 kilobits per second (kbps) to 16 kbps in the telephone band and from 6 kbps to 32 kbps in the broadened band.
There follows a brief description of the CELP digital codec, which codec uses analysis by synthesis and is the one most widely used at present for coding/decoding speech signals. A speech signal is sampled and converted into a series of blocks of L′ samples called frames. As a general rule, each frame is divided into smaller blocks of L samples called subframes. Each block is synthesized by filtering a waveform extracted from a directory (also called a dictionary) multiplied by a gain via two filters varying in time. The excitation dictionary is a finite set of waveforms of L samples. The first filter is a long-term prediction (LTP) filter. An LTP analysis evaluates the parameters of this LTP filter, which exploits the periodic nature of voiced sounds (typically representing the frequency of the fundamental pitch (the vibration frequency of the vocal chords)). The second filter is a short-term prediction filter. Linear prediction coding (LPC) analysis methods are used to obtain short-term prediction parameters representing the transfer function of the vocal tract and characteristic of the spectrum of the signal (typically representing the modulation resulting from the shape assumed by the lips, the positions of the tongue and of the larynx, etc.).
The method used to determine the innovation sequence is the method known as analysis by synthesis. In the coder, a large number of innovation sequences from the excitation dictionary are filtered by the LTP and LPC filters and the waveform producing the synthetic signal closest to the original signal according to a perceptual weighting criterion, generally known as the CELP criterion, is selected.
The use of multipulse dictionaries in these analysis by synthesis coders is described briefly below, on the understanding that CELP coders and CELP decoders are well known to the person skilled in the art.
The multiple bit rate coder of the ITU-T G.723.1 Standard is a good example of a coder using analysis by synthesis that employs multipulse dictionaries. Here, the pulse positions are all separate. The two bit rates of the coder (6.3 kbps and 5.3 kbps) model the innovation signal by means of waveforms extracted from the dictionary that include only a small number of non-zero pulses: six or five for the high bit rate, four for the low bit rate. These pulses are of amplitude +1 or −1. In its 6.3 kbps mode, the G.723.1 coder uses two dictionaries alternately:
In both dictionaries, a single restriction is imposed on the positions of the pulses of any code-vector, which must all have the same parity, i.e. they must all be even or they must all be odd. In the 5.3 kbps mode dictionary, the positions of the four pulses are more severely constrained. Apart from the same parity constraint as the dictionaries of the high bit rate mode, there is a limited choice of positions for each pulse.
The 5.3 kbps mode multipulse dictionary belongs to the well-known family of ACELP dictionaries. The structure of an ACELP directory is based on the interleaved single-pulse permutation (ISPP) technique, which consists in dividing a set of L positions into K interleaved tracks, the N pulses being located in certain predefined tracks. In some applications, the dimension L of the code words can be expanded to L+N. Accordingly, in the case of the low bit rate mode directory of an ITU-T G.723.1 coder, the dimension of the block of 60 samples is expanded to 64 samples and the 32 even (or odd as the case may be) positions are divided into four non-overlapping interleaved tracks of length 8. There are therefore two groups of four tracks, one for each parity. Table 1 below sets out the four tracks for the even positions for each pulse i0 to i3.
The ACELP innovation dictionaries are used in many standardized coders employing analysis by synthesis (ITU-T G.723.1, ITU-T G.729, IS-641, 3GPP NB-AMR, 3GPP WB-AMR). Tables 2 to 4 below set out a few examples of these ACELP dictionaries for a block length of 40 samples. Note that the parity constraint is not used in these dictionaries. Table 2 covers the ACELP dictionary for 17 bits and four non-zero pulses of amplitude ±1, used in the 8 kbps mode ITU-T G.729 coder, the IS-641 7.4 kbps mode coder and the 7.4 and 7.95 kbps mode 3GPP NB-AMR coder.
Table 3 covers the ACELP dictionary for 35 bits used in the 12.2 kbps mode 3GPP NB-AMR coder, in which each code-vector contains 10 non-zero pulses of amplitude ±1. The block of 40 samples is divided into five tracks of length 8 each containing two pulses. Note that the two pulses of the same track can overlap and result in a single pulse of amplitude ±2.
Finally, Table 4 covers the ACELP dictionary for 11 bits and two non-zero pulses of amplitude ±1 used in the low bit rate (6.4 kbps) extension of the ITU-T G.729 coder and in the 5.9 kbps mode 3GPP NB-AMR coder.
What is meant by “exploring” multipulse dictionaries is explained below.
As with any quantizing operation, seeking the optimum modeling of a vector to be coded consists in selecting from the set (or a subset) of the code-vectors of the dictionary that which “resembles” it most closely, i.e. the one that minimizes the measured distance between it and that input vector. A step referred to as “exploring” the dictionaries is carried out for this purpose.
In the case of multipulse dictionaries, this amounts to seeking the combination of pulses that optimizes the proximity of the signal to be modeled and the signal resulting from the choice of pulses. Depending on the size and/or the structure of the dictionary, this exploration may be exhaustive or non-exhaustive (and therefore more or less complex).
Since the dictionaries used in the TDAC coder referred to above are unions of permutation codes of type II, the algorithm for coding a vector of normalized transform coefficients exploits this property to determine its nearest neighbor from all the code-vectors, calculating only a limited number of distance criteria (using so-called “absolute leader” vectors).
In coders employing analysis by synthesis, the exploration of the multipulse dictionaries is not exhaustive except in the case of small dictionaries. Only a small percentage of dictionaries of higher bit rate is explored. For example, multipulse ACELP dictionaries are generally explored in two stages. To simplify this search, a first stage preselects the amplitude (and therefore the sign, see above) of each possible pulse position by simply quantizing a signal depending on the input signal. Since the amplitudes of the pulses are fixed, it is the positions of the pulses that are then searched for using an analysis by synthesis technique (conforming to the CELP criterion). Despite using the ISPP structure, and despite the small number of pulses, an exhaustive search of the combinations of positions is effected only for the low bit rate dictionaries (typically less than or equal to 12 bits). This applies to the 11-bit ACELP dictionary used in the 6.4 kbps mode G.729 coder (see Table 4), for example, in which the 512 combinations of positions of two pulses are all tested to select the best one, which amounts to calculating the corresponding 512 CELP criteria.
Various focusing methods have been proposed for dictionaries of higher bit rate. The expression “focused search” is then used.
Some of those prior art methods are used in the standardized coders mentioned above. Their aim is to reduce the number of combinations of positions to be explored on the basis of the properties of the signal to be modeled. One example is the “depth-first tree” algorithm used by many standardized ACELP coders, in which preference is given to certain positions, such as the local maxima of the tracks of a target signal depending on the input signal, the past synthetic signal, and a filter composed of synthesis and perceptual weighting filters. There are several variants of this, depending on the size of the dictionary used. To explore the ACELP dictionary for 35 bits and 10 pulses (see Table 3), the first pulse is placed at the same position as the global maximum of the target-signal. This is followed by four iterations by circular permutation of the consecutive tracks. On each iteration, the position of the second pulse is fixed at the local maximum of one of the other four tracks, and the positions of the remaining other eight pulses are searched for sequentially in pairs in interleaved loops. 256 (8×8×4 pairs) different combinations are tested on each iteration, which means that only 1024 combinations of positions of the 10 pulses among the 225 of the dictionary can be explored. A different variant is used in the IS641 coder, in which a higher percentage of combinations of the dictionary for 17 bits and four pulses (see Table 2) is explored. 768 combinations of the 8192 (=213) combinations of pulse positions are tested. In the 8 kbps G.729 coder, the same ACELP dictionary is explored by a different focusing method. The algorithm effects an iterative search by interleaving four pulse search loops (one per pulse). The search is focused by making entry into the interior loop (search for the last pulse belonging to tracks 3 or 4) conditional on exceeding an adaptive threshold that also depends on the properties of the target-signal (local maximum values and mean values of the first three tracks). Moreover, the maximum number of explorations of combinations of four pulses is fixed at 1440 (which represents 17.6% of the 8192 combinations).
In the 6.3 kbps mode G.723.1 coder, not all the 2×25×C305 (or 2×26×C306) combinations of five (or six) pulses are explored. For each chart, the algorithm employs a known “multipulse” analysis to search sequentially for the positions and the amplitudes of the pulses. As with the ACELP dictionaries, there are variants that restrict the number of combinations tested.
The above techniques suffer from the following problems, however.
The exploration of a multipulse dictionary, even a sub-optimum exploration thereof, constitutes in many coders a costly operation in terms of calculation time. For example, in the 6.3 kbps mode G.723.1 and 8 kbps mode G.729 coders, the search represents close to half the total complexity of the coder. For the NB-AMR coder, it represents one third of the total complexity. For the TDAC coder, it represents one quarter of the total complexity.
It is clear in particular that this complexity becomes critical if a plurality of coding operations have to be carried out by the same processor unit, such as a gateway managing many calls in parallel or a server distributing many multimedia contents. The complexity problem is accentuated by the multiplicity of compression formats circulating on the networks.
To offer mobility and continuity, modern and innovative multimedia communications services must be able to operate under a wide variety of conditions. The dynamism of the multimedia communications sector and the heterogeneous nature of the networks, access points and terminals have generated a plethora of compression formats whose presence in communications systems necessitates multiple coding either in cascade (transcoding) or in parallel (multiformat coding or multimode coding).
The meaning of the term “transcoding” is explained below. Transcoding becomes necessary if, in a transmission system, a compressed signal frame sent by a coder can no longer proceed in the same format. Transcoding converts the frame to another format compatible with the remainder of the transmission system. The most elementary solution (and therefore that in most widespread use at present) is to place a decoder and a coder back to back. The compressed frame arrives with a first format and is decompressed. The decompressed signal is then compressed with a second format accepted by the remainder of the communications system. Such a cascade of a decoder and a coder is referred to as 37 tandem”. That solution is very costly in terms of complexity (essentially because of the recoding) and degrades quality because the second coding is effected on a decoded signal, which is a degraded version of the original signal. Moreover, a frame may encounter several tandems before reaching its destination. The calculation cost and the loss of quality are not difficult to imagine. Moreover, the delays linked to each tandem operation are cumulative and can compromise the interactivity of calls.
What is more, complexity also causes problems in a multiformat compression system in which the same content is compressed to more than one format. This is the case of content servers that broadcast the same content in a plurality of formats adapted to the access conditions, networks and terminals of different customers. This multicoding operation becomes extremely complex as the number of formats required increases, which rapidly saturates the resources of the system.
Another case of multiple coding in parallel is a posteriori decision multimode compression. A plurality of compression modes are applied to each segment of the signal to be coded, and that which optimizes a given criterion or achieves the best bit rate/distortion trade-off is selected. Once again, the complexity of each of the compression modes limits the number thereof and/or leads to an a priori selection of a very small number of modes.
Prior art approaches to solving the above problems are described below.
New multimedia communications applications (such as audio and video applications) often necessitate a plurality of coding operations either in cascade (transcoding) or in parallel (multicoding and a posteriori decision multimode coding). The problem of the complexity barrier resulting from all these coding operations remains to be solved, despite the increase in current processing powers. Most prior art multiple coding operations do not take account of interactions between formats and between the format of the coder E and its content. Nevertheless, a few intelligent transcoding techniques have been proposed that are not satisfied merely by decoding and then recoding, but instead exploit the similarities between coding formats so that complexity can be reduced whilst limiting the resulting degradation.
So-called “intelligent” transcoding methods are described below.
All the coders in the same family of coders (CELP, parametric, transform, etc.) extract the same physical parameters from the signal. There is nevertheless great variety in terms of modeling and/or quantizing those parameters. Thus the same parameter may be coded in the same way or very differently from one coder to another.
Moreover, the coding may be strictly identical, or it may be identical in terms of modeling and calculation of the parameter, but differ simply in how the coding is translated into the form of bits. Finally, the coding may be completely different in terms of modeling and quantizing the parameter, or even in terms of its analysis or sampling frequency.
If modeling and parameter calculation are strictly identical, including translation to bit form, it suffices to copy the corresponding bit field from the bit stream generated by the first coder to that of the second. This highly favorable situation arises on transcoding from the G.729 standard to the IS-641 standard for adaptive excitation (LTP delays), for example.
If, for the same parameter, the two coders differ only in terms of the translation of the calculated parameter into bit form, it suffices to decode the bit field of the first format and then to return it to the binary domain using the coding method of the second format. This conversion may also be effected by means of one-to-one correspondence tables. This is the situation when transcoding fixed excitations from the G.729 standard to the AMR standard (7.4 kbps and 7.95 kbps modes), for example.
In the above two situations, transcoding the parameter remains at the bit level. Simple bit manipulation renders the parameter compatible with the second coding format. On the other hand, if a parameter extracted from the signal is modeled or quantized differently by two coding formats, passing from one to the other is not such a simple matter. Several methods have been proposed. They operate at the parameter level, the excitation level, or the decoded signal level.
For transcoding in the parameter domain, remaining at the parameter level is possible if the two coding formats calculate a parameter in the same way but quantize it differently. Quantizing differences may be related to the accuracy or the method selected (scalar, vectorial, predictive, etc.). It then suffices to decode the parameter and then to quantize it using the method of the second coding format. That prior art method is used at present for transcoding excitation gains in particular. The decoded parameter must often be modified before it is requantized. For example, if the coders have different parameter analysis frequencies or different frame/subframe lengths, it is standard practice to interpolate/decimate the parameters. Interpolation may be effected by the method described in the published document US2003/033142, for example. Another modification option is to round off the parameter to the accuracy imposed on it by the second coding format. This situation is encountered for the most part for the height of the fundamental frequency (“pitch”).
If it is not possible to transcode a parameter within the parameter domain, decoding can go to a higher level. This is the excitation domain, without going so far as the signal domain. This technique has been proposed for gains in the document “Improving transcoding capability of speech coders in clean and frame erasured channel environments”, Hong-Goo Kang, Hong Kook Kim, Cox, R. V., Speech Coding, 2000, Proceedings 2000, IEEE Workshop on Speech Coding, Pages 78-80.
Finally, a last solution (the most complex and the least “intelligent”) consists in recalculating the parameter explicitly, as the coder would, but based on a synthesized signal. This operation amounts to a kind of partial tandem, with only some parameters being entirely recalculated. This method has been applied to diverse parameters such as the fixed excitation, the gains in the IEEE reference cited above, or the pitch.
For transcoding pulses, although several techniques have been developed to calculate the parameters quickly and at lower cost, few solutions available today use an intelligent approach to calculating the pulses of one format from the equivalent parameter in another format. In coding using analysis by synthesis, intelligent transcoding of pulse codes is applied only if the modeling is identical (or close). In contrast, if the modeling is different, the partial tandem method is used. Note that to limit the complexity of this operation, focused approaches have been proposed that exploit the properties of the decoded signal or a derived signal such as a target-signal. In the document US-2001/027393 cited above, in an embodiment utilizing an MDCT transform coder, there is described a bit rate change procedure that may be considered a special case of intelligent transcoding. That procedure requantizes a vector from a first dictionary using a vector from a second dictionary. To this end it distinguishes between two situations depending on whether the vector to be requantized belongs to the second dictionary or not. If the quantized vector belongs to the new dictionary, the modeling is identical; if not, the partial decoding method is applied.
Setting itself apart from all the above prior art techniques, the present invention proposes a method of multipulse transcoding based on selecting a subset of combinations of pulse positions of an ensemble of sets of pulses from a combination of pulse positions of another ensemble of sets of pulses, the two ensembles being distinguished by the numbers of pulses that they include and by rules governing their positions and/or their amplitudes. This form of transcoding is very beneficial for multiple coding in cascade (transcoding) or in parallel (multicoding and multimode coding) in particular.
To this end, the present invention firstly proposes a method of transcoding between a first compression codec and a second compression codec. The first and second codecs are of pulse type and use multipulse dictionaries in which each pulse has a position marked by an associated index.
The transcoding method of the invention includes the following steps:
a) where appropriate, adapting coding parameters between said first and second codecs;
b) obtaining from the first codec a selected number of pulse positions and respective position indices associated therewith;
c) for each current pulse position of given index, forming a group of pulse positions including at least the current pulse position and the pulse positions with associated indices immediately below and immediately above the given index;
d) selecting as a function of pulse positions accepted by the second codec at least some of the pulse positions in an ensemble constituted by a union of said groups formed in step c); and
e) sending the selected pulse positions to the second codec for coding/decoding from the positions sent.
The selection step d) therefore involves a number of pulse positions that is less than the total number of pulse positions in the dictionary of the second codec.
It is clear in particular that if, in the step e), the second above-mentioned codec is a coder, the selected pulse positions are transmitted to that coder for coding by searching only the positions transmitted. If the second above-mentioned codec is a decoder, the selected pulse positions are transmitted for the positions to be decoded.
The step b) preferably uses partial decoding of the bit stream supplied by the first codec to identify a first number of pulse positions that the first codec uses in a first coding format. The number chosen in the step b) therefore preferably corresponds to this first number of pulse positions.
In an advantageous embodiment, the above steps are executed by a software product including program instructions to that effect. In this regard, the present invention is also directed to a software product of the above kind adapted to be stored in a memory of a processor unit, in particular of a computer or a mobile terminal, or on a removable memory medium adapted to cooperate with a reader of the processor unit.
The present invention is also directed to a device for transcoding between first and second compression codecs, in which case it includes a memory adapted to store instructions of a software product of the type described above.
Other features and advantages of the invention become apparent on reading the following detailed description and examining the appended drawings, in which:
a is a diagram of a transcoding context in the terms of the present invention in a “cascade” configuration;
b is a diagram of a transcoding context in the terms of the present invention in a “parallel” configuration;
a is a diagram of an adaptation process for use when the sampling frequencies of the first coder E and the second coder S are different;
b is a diagram of a variant of the
b represents a practical implementation of
Note first that the present invention relates to modeling and coding digital multimedia signals such as audio (speech and/or sound) signals using multipulse dictionaries. It may be implemented in the context of multiple coding/decoding in cascade or in parallel or of any other system modeling a signal by means of a multipulse representation and which, based on the knowledge of a first set of pulses belonging to a first ensemble, has to determine at least one set of pulses of a second ensemble. For conciseness, only the passage from a first ensemble to another ensemble is described, but the invention applies equally to passage to n ensembles (n≧2). Moreover, only the situation of “transcoding” between two coders is described below, but transcoding between a coder and a decoder can of course be deduced from this without major difficulty.
Consider the case therefore of modeling a signal by sets of pulses corresponding to two coding systems.
In the example represented in
In this configuration, the transcoder D therefore effects coding adaptation between the first coder E and the second coder S, advantageously favoring faster (because more restricted) coding by the second coder S. Of course, as an alternative to this, the entity referenced S in
Note that the arrangement of the coder E, the transcoder D and the coder S may conform to a “cascade” configuration as represented in
Note further that the transcoder D may simply be equipped with a memory for storing instructions for implementing the foregoing steps and a processor for processing those instructions.
The invention is therefore applied as follows. The first coder E has effected its coding operation on a given signal s0 (for example the original signal). The positions of the pulses selected by the first coder E are therefore available. That coder determined these positions Pe using a technique of its own during the coding process. The second coder S must also perform its coding. In the case of transcoding, the second coder S has only the bit stream generated by the first coder and the invention is here applicable to “intelligent” transcoding as defined above. In the case of multiple coding in parallel, the second coder S also has the signal that the first coder has and here the invention applies to “intelligent multicoding”. A system that requires to code the same content in a plurality of formats can exploit the information of a first format to simplify coding the other formats. The invention can also be applied to the particular situation of multiple coding in parallel constituting a posteriori decision multimode coding.
The present invention can be used to determine quickly the positions ps (interchangeably denoted si below) of the pulses for another coding format from positions pe (interchangeably denoted ei below) of the pulses of a first format. It considerably reduces the calculation complexity of this operation for the second coder by limiting the number of possible positions. To this end, it uses the positions selected by the first coder to define a restricted set of positions from all possible positions of the second coder, in which restricted set the best set of positions for the pulses is searched for. This results in a significant increase in complexity whilst limiting degradation of the signal relative to a standard exhaustive or focused search.
It is therefore clear that the present invention limits the number of possible positions by defining a restricted set of positions based on positions from the first coding format. It differs from existing solutions in that they use only the properties of the signal to be modeled to limit the number of possible positions, by giving preference to and/or eliminating positions.
For each pulse of a set of a first ensemble, two neighbors (one on the right and one on the left) of variable width and of greater or lesser constraint are preferably defined and an ensemble of possible positions extracted therefrom within which at least one combination of pulses complying with the constraints of the second ensemble will be preselected.
The transcoding method has the advantage of optimizing the complexity/quality trade-off by adapting the number of pulse positions and/or the respective sizes (in terms of combinations of pulse positions) of the right-hand and left-hand neighborhoods for each pulse, either at the beginning of the processing or for each subframe as a function of the authorized complexity and/or the set of starting positions. The invention also adjusts/limits the number of combinations of positions by advantageously favoring the immediate neighborhoods.
As indicated above, the present invention is also directed to a software product the algorithm whereof is designed in particular to extract neighbor positions that facilitate composing the combinations of pulses of the second ensemble.
As indicated above, the heterogeneous nature of the networks and the contents may call highly varied coding formats into play. Coders may be distinguished by numerous characteristics, of which two in particular, the sampling frequency and the duration of a subframe, substantially determine the mode of operation of the invention. The options are described below in corresponding relationship to embodiments of the invention suited to these situations.
The sampling frequencies are compared in a test 22. If the frequencies are equal, the subframe durations are compared in a test 23.If not, the sampling frequencies are adapted in a step 32 by a method described below. Following the test 23, if the subframe durations are equal, the numbers Ne and Ns of pulse positions used by the first and second coding formats, respectively, are compared in a test 24. If not, the subframe durations are adapted in a step 33 using a method that is also described below. It is clear that the steps 22, 23, 32 and 33 together define the above step a) of adapting the coding parameters. Note that the steps 22 and 32 (sampling frequency adaptation), on the one hand, and the steps 23 and 33 (subframe duration adaptation), on the other hand, may be interchanged.
There is first described below a situation in which the sampling frequencies are equal and the subframe durations are equal.
This is the most favorable situation, but it is nevertheless necessary to distinguish the situation in which the first format uses more pulses than the second (Ne≧Ns) and the contrary situation (Ne<Ns), according to the result of the test 24.
Ne≧Ns in
The principle is as follows. The directories of the two coders E and S use Ne and Ns pulses in each subframe, respectively.
The coder E calculates the positions of its Ne pulses over the subframe se. These positions are interchangeably denoted ei and pe below. The restricted ensemble Ps of privileged positions for the pulses of the directory of the coder S is then made up of Ne positions ei and their neighborhoods:
where vdi and vgi≧0 are the sizes of the right-hand and left-hand neighborhoods of the pulse i. The values of vdi and vgi, which are chosen in the step 27 in
In step 29 in
It is then necessary to define for each of the Ns pulses from the directory of the coder S the positions which that pulse is authorized to assume among those proposed by Ps.
To this end, rules governing the construction of the directory of S are introduced. It is assumed that the Ns pulses of S belong to predefined subsets of positions, a given number of pulses sharing the same sub-set of authorized positions. For example, the 10 pulses of the 12.2 kbps mode 3GPP NB-AMR coder are distributed two by two into five different subsets, as shown in Table 3 above. N′s denotes the number of subsets of different positions (N′s≦Ns in this example since N′s=5) and Tj (for j=1 to N′s) denotes the subsets of positions defining the directory of S.
Starting from the ensemble Ps, the N′s, subsets Sj resulting from the intersection of Ps with one of the ensembles Tj are constituted in step 30 in
Sj=Ps∩Tj
The neighborhoods vdi and vg i must be of sufficient size for no intersection to be empty. It is therefore necessary to allow adjustment of the neighborhood sizes, if necessary, as a function of the starting set of pulses. This is the purpose of the test 34 in
The invention advantageously exploits the structure of the directories. For example, if the directory of the coder S is of the ACELP type, it is the intersections of the positions of the tracks with Ps that are calculated. If the directory of the coder E is also of the ACELP type, the neighborhood extraction procedure also exploits the track structure and the steps of extracting the neighborhoods and composing restricted subsets of positions are judiciously combined. In particular, it is beneficial for the neighborhood extraction algorithm to take account of the composition of the combinations of pulses in accordance with the constraints of the second ensemble. As will emerge later, neighborhood extraction algorithms are produced to facilitate the composition of combinations of pulses of the second ensemble. One of the embodiments described later (from ACELP with two pulses to ACELP with four pulses) is an example of this kind of algorithm.
The number of possible combinations of positions is therefore small and the size of the subset of the directory of the coder S is generally very much less than that of the original directory, which greatly reduces the complexity of the penultimate transcoding step. The number of combinations of pulse positions defines the size of the aforementioned subset. It is the number of pulse positions the invention reduces, which leads to a reduction in the number of combinations of pulse positions and thus makes it possible to obtain a subdirectory of restricted size.
Step 46 in
In outline, the
The next step is therefore a step 46 of searching the subdirectory received by the coder S for a set (opt(Sj)) of optimum positions including the second number Ns of positions, as indicated above. To accelerate the exploration of the subdirectory, this step 46 of searching for the optimum set of positions is preferably implemented by means of a focused search. Processing continues naturally with the coding that is effected thereafter by the second coder S.
There are described next the forms of processing provided for the situation in which the number Ne of pulses used by the first coding format is lower than the number Ns of pulses used by the second coding format.
Ne<Ns in
If the format of S uses more pulses than the format of E, the process is similar to that explained above. However, pulses of the format of S may not have positions in the restricted directory. In this case, in a first embodiment, all possible positions are authorized for those pulses. In a second and preferred embodiment the sizes of the neighborhoods V′d and V′g are simply increased in step 28 in
Ne<Ns<2Ne in
A special case must be emphasized here. If Ne is close to Ns, typically if Ne<Ns<2Ne, then a preferred way to determine the positions may be envisaged, even though the above form of processing remains entirely applicable. A further reduction in complexity may be obtained by directly fixing the positions of the pulses of S on the basis of those of E. The Ne first pulses of S are placed at the positions of those of E. The remaining Ns−Ne pulses are placed as close as possible the first Ne pulses (in their immediate neighborhood). Step 25 in
Of course, in both cases, Ne<Ns and Ne<Ns<2Ne, if one of the intersections Sj is empty despite the above precautions, the size of the neighborhoods V+g, V+d, is simply increased in step 35, as described in the situation where Ne≧Ns.
Finally, in all cases, if none of the intersections Sj is empty, the subdirectory formed by the intersections Sj is forwarded to the second coder S (step 31).
There are described next the forms of processing used in the adaptation step a) if the coding parameters of the first and second formats are not the same, in particular their sampling frequencies and subframe durations.
The following situations are then distinguished.
Equal Subframe Durations but Different Sampling Frequencies
This situation corresponds to “n” for the test 22 and “y” for the test 23 in
The previous processing cannot be applied directly here because the two formats do not have the same time subdivision. Because the sampling frequencies are different, the two frames do not have the same number of samples over the same duration.
Rather that determining the positions of the pulses of the format of the coder S without taking account of those of the format of the coder E, as a tandem would do, two different forms of processing constituting two different embodiments are proposed here. They limit complexity by establishing a correspondence between the positions of the two formats, after which the processing reverts to the processing described above (as if the sampling frequencies were equal).
The processing of the first embodiment uses direct quantization of the time scale of the first format by that of the second format. This quantizing operation, which may be tabulated or computed from a formula, finds for each position of a subframe of the first format its equivalent in a subframe of the second format, and vice-versa.
For example, the correspondence between the positions pe and ps in the subframes of the two formats may be defined by the following equation:
in which Fe and Fs are the sampling frequencies of E and S, respectively,
Le and Ls are their subframe lengths, and └┘ denotes the integer part.
Depending on the characteristics of the processor unit, this correspondence could use the above formula or advantageously be tabulated for the Le values. An intermediate solution may also be selected by tabulating only the first le values
being the highest common factor of Le and Ls), the remaining positions then being easily deduced.
Note that it is also possible to make a plurality of positions of the subframe of S correspond to a position of a subframe of E. For example, retaining the positions immediately below and immediately above
The general processing described above is applied starting from the ensemble of positions ps corresponding to the positions pe, (extraction of neighborhoods, composition of combinations of pulses, selection of the optimum combination).
This situation of equal subframe durations but different sampling frequencies is found in Tables 5a to 5d below, referring to an embodiment in which the coder E is of the 3GPP NB-AMR type and the coder S is of the WB-AMR type. The NB-AMR coder has a subframe of 40 samples for a sampling frequency of 8 kHz. The WB-AMR coder uses 64 samples per subframe at 12.8 kHz. In both cases, the subframe has a duration of 5 ms. Table 5a gives the correspondence of the positions in a NB-AMR subframe to a WB-AMR subframe and Table 5b gives the converse correspondence. Tables 5c and 5d are the restricted correspondence tables.
Briefly, the following steps apply (see
In general terms, the quantization step a1) is effected by calculation and/or tabulation from a function which makes correspond to a pulse position pe in a subframe with the first format a pulse position ps in a subframe with the second format; that function actually takes the form of a linear combination involving a multiplier coefficient corresponding to the ratio of the second sampling frequency to the first sampling frequency.
Moreover, to go in the opposite direction from a pulse position in a subframe with the second format ps to a pulse position in a subframe with the first format pe, there is of course applied an inverse function of this linear combination applied to a pulse position in a subframe with the second format ps.
Clearly the transcoding process is completely reversible and is as equally adapted to one transcoding direction (E->S) as to the other (S->E).
A second embodiment of sampling frequency adaptation uses a conventional change of sampling frequency principle. Starting from the subframe containing the pulses found by the first format, oversampling is applied at the frequency equal to the lowest common multiple of the two sampling frequencies Fe and Fs. Then, after low-pass filtering, undersampling is applied to revert to the sampling frequency of the second format, i.e. Fs. There is obtained a subframe at the frequency Fs containing the filtered pulses from E. Once again, the result of the oversampling/LP filtering/undersampling operations can be tabulated for each possible position of a subframe of E. This processing can also be effected by “on line” calculation. As in the first embodiment of sampling frequency adaptation, one or more positions of S may be associated with a position of E, as explained below, and the general processing in the sense of the above-described invention applied.
As indicated in the variant represented in
The process continues by obtaining, preferably by a thresholding method, a number of positions, possibly a variable number of positions, adapted from the pulses of E (step 56), as in the above first embodiment.
Equal Sampling Frequencies but Different Subframe Durations
The processing carried out in the situation where the sampling frequencies are equal but the subframe durations are different is described next. This situation corresponds to “n” for the test 23 but “o” for the test 22 of
As in the above situation, the neighborhood extraction step as such cannot be applied directly. It is first necessary to make the two subframes compatible. Here the subframes differ in size. Faced with this incompatibility, rather than calculate the positions of the pulses like the tandem does, a preferred embodiment offers a solution of low complexity that determines a restricted directory of combinations of positions for the pulses of the second format from the positions of the pulses of the first format. However, the subframe of S and that of E not being the same size, it is not possible to establish a direct temporal correspondence between a subframe of S and a subframe of E. As shown in
In a preferred embodiment, it is proposed to divide the excitation of E into pseudosubframes the size of those of S and at the timing rate of S. The pseudosubframes are denoted STE′ in
A position poe (respectively pos) of the first format (respectively the second format) relative to that origin coincides with the position pe (respectively ps) of the subframe ie (respectively js) of E (respectively S) relative to that subframe. Thus:
poe=pe+ieLe and po=ps+jsLs with 0≦pe<Le and 0≦ps<Ls
To a position pe of the subframe ie of the format of E there corresponds the position ps of the subframe js of the format of S, ps and js being respectively the remainder and the quotient of the Euclidian division by Ls of the position poe of pe relative to an origin O common to E and S:
js=└(pe+ieLe)/Ls┘ and ps≡(pe+ieLe)[Ls]
with 0<pe≦Le and 0≦ps<Ls └┘ denoting the integer part, ≡denoting the modulus, the index of a subframe of E (respectively S) being given relative to the common origin O.
Accordingly, the positions pe in a subframe js are used to determine a restricted ensemble of positions for pulses of S in the subframe js by means of the general process described above. However, if Le>Ls, a subframe of S may not contain any pulse. In the
Preferred embodiments for the determination of a time origin O common to the two formats are described next. That common reference constitutes the position (number 0) from which the positions of the pulses are numbered in the subsequent subframes. This position 0 can be defined in various ways, depending on the system utilizing the transcoding method of the present invention. For example, for a transcoder module included in a transmission system equipment, it will be natural to take for the origin the first position of the first frame received after the equipment is started up.
However, the disadvantage of that choice is that the positions take increasingly large values and it may become necessary to limit them. For this it suffices to update the position of the common origin whenever possible. Accordingly, if the respective lengths Le and Ls, of the subframes of E and S are constant over time, the position of the common origin is reset each time that the boundaries of the subframes of E and S are aligned. This occurs periodically, the period (expressed in samples) being equal to the lowest common multiple of Le and Ls.
The situation may also be envisaged in which Le and/or Ls are not constant in time. It is no longer possible to find a multiple common to the two subframe lengths, at present denoted Le(n) and Ls(n), where n represents the subframe number. In this case, it is necessary to sum the values Le(n) and Ls(n) on the fly and to compare the two sums obtained in each subframe:
Each time that Te(k)=Ts(k′), the common origin is updated (and taken at the position k×Le or k′×Ls). The two sums Te and Ts are preferably reset.
Briefly, and more generally, calling the first (respectively second) subframe duration the subframe duration of the first (respectively second) coding format, the adaptation steps executed when the subframe durations are different are summarized in
To determine the common origin O, the following cases are preferably discriminated in the test 72 in
In the former case, the time position of the common origin is updated periodically (step 74), each time that the boundaries of the respective subframes of first duration St(Le) and second duration St(Ls) are aligned in time (test 73 applied to those boundaries).
In the second case, it is preferable if:
Now, in the situation in which the subframe durations and sampling frequencies are different, it suffices to combine judiciously the algorithms of the correspondences between the positions of E and S described for the above two situations.
Three embodiments of transcoding in accordance with the invention are described next. These embodiments describe the application of the processing provided in the situations described above in standard speech coders using analysis by synthesis. The first two embodiments illustrate the favorable situation in which the sampling frequencies and the subframe durations are identical. The final embodiment illustrates the situation in which the subframe durations are different.
The first embodiment applies to intelligent transcoding between the 6.3 kbps mode G.723.1 MP-MLQ model and the 5.3 kbps mode G.723.1 ACELP model with four pulses.
Intelligent transcoding from the high bit rate to the low bit rate of G.723.1 employs an MP-MLQ model with six and five pulses with an ACELP model with four pulses. The embodiment described here determines the positions of the four ACELP pulses from the positions of the MP-MLQ pulses.
The operation of the G.723.1 coder is summarized below.
The ITU-T G.723.1 multiple bit rate coder and its multipulse directories have been described above. Suffice to say that a G.723.1 frame contains 240 samples at 8 kHz and is divided into four subframes each of 60 samples. The same restriction is imposed on the positions of the pulses of any code-vector of each of the three multipulse dictionaries. These positions must all have the same parity (they must all be even or all be odd). The subframe of 60(+4) positions is therefore divided into two grids each of 32 positions. The even grid includes the positions numbered [0, 2, 4, . . . , 58, (60,62)]. The odd grid includes the positions [1, 3, 5, . . . , 59, (61,63)]. For each bit rate, exploration of the directory, although not exhaustive, remains complex, as indicated above.
The selection of a subset of the 5.3 kbps mode G.723.1 ACELP directory from an element of a 6.3 kbps mode G.723.1 MP-MLQ directory is described next.
The aim is to model the innovation signal of a subframe by means of an element from the 5.3 kbps mode G.723.1 ACELP directory knowing the element of the 6.3 kbps mode MP-MLQ G.723.1 directory determined during a first coding operation. The Ne positions (Ne=5 or 6) of the pulses selected by the 6.3 kbps mode G.723.1 coder are therefore available.
For example, it may be assumed that the positions extracted from the bit stream of the 6.3 kbps mode G.723.1 coder for a subframe whose excitation is modeled by Ne=5 pulses are as follows:
e0=0; e1=8; e2=28; e3=38; e4=46;
Remember that no adaptation of sampling frequency or subframe duration is required here. After this step of recovering the positions ei, a subsequent step then consists in extracting the right-hand and left-hand neighborhoods of those five pulses directly. The right-hand and left-hand neighborhoods are here taken to be equal to two. The ensemble Ps of positions selected is:
Ps={−2,−1,0,1,2}∪{6,7,8,9,10}∪{26,27,28,29,30}∪{36,37,38,39,40}∪{44,45,46,47,48}
The third step consists in composing the restricted ensemble of possible positions for each pulse (here one track) of the ACELP directory of the 5.3 kbps mode G.723.1 coder by taking Ns=4 intersections of Ps with the four ensembles of positions of the even tracks (respectively odd tracks) authorized by said directory (as represented in Table 1).
For even parity:
S0=Ps∩{8,16, . . . ,56}; S1=Ps∩{2,10,18, . . . ,58}; S2=Ps∩{4,12,20, . . . ,52,(60)}; S3=Ps∩{6,14,22, . . . ,54,(62)};
whence : S0={0,8,40,48}; S1={2,10,26,}; S2={28,36,44}; S3={6,30,38,46};
For odd parity:
S0=Ps109 {1,9, . . . ,57}; S1=Ps∩{3,11, . . . ,59}; S2=Ps∩{5,13, . . . ,53,(61)}; S3=Ps∩{7,15, . . . ,55,(63)};
whence : S0={1,9}; S1={27}; S2={29,37,45}; S3={7,39,47};
The combination of these selected positions constitutes the new restricted directory in which the search will be effected. For this step, the procedure for selecting the set of optimum positions is based on the CELP criterion, as in the 5.3 kbps mode G.723.1 coder. The exploration may be exhaustive but is preferably focused.
The number of combinations of positions in the restricted directory is equal to 180 (=4*3*3*4+2*1*3*3) instead of 8192 (=2*8*8*8*8) combinations of positions of the ACELP directory of the 5.3 kbps mode G.723.1 coder.
The number of combinations may be further restricted by considering only the parity chosen for the 6.3 kbps mode (in the present example that is the even parity). In this case, the number of combinations in the restricted directory is equal to 144.
Depending on the size of the neighborhoods concerned, for one of the four pulses the ensemble Ps may not contain any position for a track of the ACELP model (situation in which one of the ensembles Si is empty). Accordingly, for neighborhoods of size 2, when the positions of the Ne pulses are all on the same track, Ps contains only positions of that track and adjacent tracks. In this case, depending on the required quality/complexity trade-off, it is possible either to replace the ensemble Si with Ti (which amounts to not restricting the ensemble of positions of that track) or to increase the right-hand (or left-hand) neighborhood of the pulses. For example, if all the pulses of the 6.3 kbps mode coder are on track 2, with right-hand and left-hand neighborhoods equal to two, then track 0 will have no positions regardless of the parity. It then suffices to increase by 2 the size of the left-hand and/or right-hand neighborhood to assign positions to that track 0.
To illustrate this embodiment, consider the following example:
e0=4; e1=12; e2=20; e3=36; e4=52;
The ensemble Ps of selected positions is as follows:
Ps={2,3,4,5,6 }∩{10,11,12,13,14}∩{18,19,20,21,22}∩{34,35,36,37,38}∩{50,51,52,53,54}
Assuming that it is wished to retain the same parity, the initial division of these positions for the four pulses is as follows:
The following second embodiment illustrates the application of the invention to intelligent transcoding between ACELP models of the same length. In particular, this second embodiment is applied to intelligent transcoding between the ACELP model with four pulses of 8 kbps mode G.729 and the ACELP with two pulses of 6.4 kbps mode G.729.
Intelligent transcoding between the 6.4 kbps and 8 kbps modes of the G.729 coder utilizes one ACELP directory with two pulses and a second one with four pulses. The embodiment described here determines the positions of four pulses (8 kbps) from the positions of two pulses (6.4 kbps) and vice-versa.
The operation of the ITU-T G.729 encoder is described briefly. This coder can operate at three bit rates: 6.4, 8 and 11.8 kbps. The first two bit rates are considered here. A G.729 frame contains 80 samples at 8 kHz and is divided into two subframes each of 40 samples. For each subframe, G.729 models the innovation signal by means of pulses conforming to the ACELP model. It uses four pulses for the 8 kbps mode and two pulses for the 6.4 kbps mode. Tables 2 and 4 above give the positions that the pulses can adopt for those two bit rates. At 6.4 kbps, an exhaustive search of all (512) combinations of positions is effected. At 8 kbps, a focused search is preferably used.
The general processing in accordance with the invention is used again here. However, the ACELP structure common to the two directories is advantageously exploited here. Establishing the correspondence between the sets of positions therefore exploits a division of the subframe of 40 samples into five tracks each of eight positions, as set out in Table 6 below.
In the two directories, the positions of the pulses share these tracks, as shown in Table 7 below.
All the pulses are characterized by the their track and their rank in that track. The 8 kbps mode places a pulse on each of the first three tracks and the last pulse on one of the last two tracks. The 6.4 kbps mode places its first pulse on track P1 or P3 and its second pulse on track P0, P1, P2 or P4.
This embodiment exploits interleaving of the tracks (ISSP structure) to facilitate extracting the neighborhoods and composing the restricted subensembles of positions. Accordingly, to move from one track to another, it suffices to shift one unit to the right or to the left. For example, at the 5th position of track 2 (absolute position 22), a shift of one unit to the right (+1) goes to the 5th position on track 3 (absolute position 23) and a shift of one unit to the left (−1) goes to the 5th position of track 1 (absolute position 21).
More generally, a position shift of ±d is reflected here in the following effects.
At the level of the tracks Pi:
right-hand neighborhood: Pi=P(i+d)≡5
left-hand neighborhood: Pi=P(i−d)≡5
At the level of the rank m in the track:
The selection of a subensemble of the ACELP directory with four pulses of the 8 kbps mode G.729 coder from an element of an ACELP directory with two pulses of the 6.4 kbps mode G.729 coder is described next.
A 6.4 kbps mode G.729 subframe is considered. Two pulses are placed by the coder, but it is necessary to determine the positions of the other pulses that the 8 kbps mode G.729 must place. To restrict complexity radically, only one position per pulse is selected and only one combination of positions is retained. This has the advantage that the selection step is therefore immediate. Two of the four pulses of the 8 kbps mode G.729 are selected at the same positions as those of the 6.4 kbps mode, after which the remaining two pulses are placed in the immediate neighborhood of the first two. As indicated above, the track structure is exploited. In the first step of recovering the two positions by decoding the binary index (on nine bits) of the two positions, the corresponding two tracks are also determined. From those two tracks (which may be identical), the last three steps of extracting the neighborhoods, composing the restricted subensembles and selecting a combination of pulses are then judiciously associated. Different cases are then distinguished according to the tracks Pi (i=0 to 4) containing the two 6.4 kbps mode pulses.
The positions of the 6.4 kbps mode pulses are denoted ek and those of the 8 kbps mode pulses are denoted Sk. Table 8 below gives the selected positions in each case. The columns labeled “Pj+d=Pi” specify the neighborhood law at the level of the tracks and terminating at the track Pi. At the level of the tracks Pi:
for the left-hand neighborhood: Pi P(i−d)≡5
The aim is therefore preferably to balance the distribution of the four positions relative to the two starting positions, although a different choice may be made. Four situations (indicated by an exponent in parentheses in Table 8) may nevertheless give rise to edge effect problems:
The selection of a subensemble of the 6.4 kbps mode G.729 ACELP directory with two pulses from an element of an 8 kbps mode G.729 ACELP directory with four pulses is described next.
For an 8 kbps mode G.729 subframe, the first step is to recover the positions of the four pulses generated by the 8 kbps mode. Decoding the binary index (on 13 bits) of these four positions yields their rank in their respective track for the first three positions (tracks 0 to 2) and the track (3 or 4) of the fourth pulse together with its rank in that track. Each position ei (0≦i<4) is characterized by the pair (pi,mi) in which pi is the index of its track and mi is its rank in that track. We have:
ei=5mi+pi
with 0≦mi<8 and pi=i for I<3 and p3=3 or 4.
As already mentioned, neighborhood extraction and restricted subensemble composition are combined and advantageously exploit the ISSP structure common to the two directories. The five intersections T′j of the ensemble Ps of the neighborhoods of the four positions with the five tracks Pj are constructed by exploiting the adjacent position property induced by interleaving the tracks:
T′j=Ps∩Pj
Accordingly, a right-hand (respectively left-hand) neighborhood of +1 (respectively −1) of the pulse (p,m) belongs to T′p+1 if p<4 (respectively to T′p−1 if p>0), if not (p=4) to T′0 on condition that m<7 (respectively to T′4 (I=0) on condition that m>0). The restriction on the right-hand neighbor for a position of the fourth pulse belonging to the fourth track (respectively left-hand neighbor for a position of the first track) ensure that adjacent position is not outside the sub-frame.
Accordingly, using the modulo 5 notation (≡5), a right-hand (respectively left-hand) neighbor of +1 (respectively −1) of the pulse (p,m) belongs to T′(p+1)≡5 (respectively to T′(p−1)≡5). Note that it is necessary to take account of edge effects. Generalizing to a neighborhood size d, a right-hand neighbor of +d (respectively a left-hand neighbor of −d) of the pulse (p,m) belongs to T′(p+d)≡5 (respectively T′(p−d)≡5). The rank of the neighbor of ±d is equal to m if p+d≧4 (or p−d≧0), otherwise the rank m is incremented for a right-hand neighbor and decremented for a left-hand neighbor. Taking account of edge effects therefore amounts to ensuring that m<7 if p+d>4 and m>0 if p−d≧0.
Starting from this distribution of the neighbors in the five tracks, it is a simple matter to determine the subensembles S0 and S1 of the positions of the two pulses:
S0=T′1∪T′3 and S1=T′0∪T′1∪T′2∪T′4
The fourth and final step consists in searching for the optimum pair in the two subensembles obtained. The search algorithm (like the standardized algorithm exploiting the track structure) and the track by track storage of pulses once again simplify the search algorithm. In practice, it is therefore of no utility to construct the restricted subensembles S0 and S1 explicitly, as the ensembles T′j can be used alone.
In the following example, the four 8 kbps mode G.729 pulses have been placed at the following positions:
e0=5; e1=21; e2=22; e3=34.
Taking a fixed neighborhood equal to 1, the five intersections T′j are constructed as follows:
Thus we have:
Reverting to the position notation:
In the final step, an algorithm similar to that of the G.729 6.4 kbps mode effects the search for the best pair of pulses. That algorithm is much less complex here as the number of combinations of positions to be explored is very small. In the example, there number of combinations to be tested is only 4 (Cardinal(T′1)+Cardinal(T′3)) multiplied by 8 (Cardinal(T′0) +Cardinal(T′1)+Cardinal(T′2)+Cardinal(T′4)), i.e. 32 combinations instead of 512.
For a neighborhood of size 1, less than 8% of the combinations of positions are to be explored on average, without exceeding 10% (50 combinations). For a neighborhood of size 2, less than 17% of combinations of positions are to be explored on average and at most 25% of the combinations are to be explored. For a neighborhood of size 2, the complexity of the processing proposed by the invention (lumping together the cost of searching the restricted directory and the cost of extracting the neighborhoods associated with the composition of the intersections) represents less than 30% of an exhaustive search for an equivalent quality.
The final embodiment illustrates passing between the 8 kbps mode G.729 ACELP model and the 6.3 kbps mode G.723.1 MP-MLQ model.
Intelligent transcoding of the pulses between G.723.1 (6.3 kbps mode) and G.729 (8 kbps mode) entails two major difficulties. Firstly, the size of the frames is different (40 samples for G.729 as against 60 samples for G.723.1). The second difficulty is linked to the different structures of the dictionaries (ACELP type for G.729 and MP-MLQ type for G.723.1). The embodiment described here shows how the invention eliminates these two problems in order to transcode the pulses at reduced cost whilst preserving transcoding quality.
First of all a temporal correspondence is set up between the positions in the two formats, taking account of the size difference of the subframes to align the positions relative to an origin common to E and S. The G.729 and G.723.1 subframe lengths having a lowest common multiple of 120, the temporal correspondence is set up by blocks of 120 samples, i.e. two G.723.1 subframes for every three G.729 subframes, as shown in the
There is described next the selection of a subensemble of the 6.3 kbps mode G.723.1 MP-MLQ directory from elements of the 8 kbps mode G.729 ACELP directory with four pulses. The first step consists in recovering the positions of the pulses by blocks of three G.729 subframes (with index ie, 0≦ie≦2). The position of that block in the subframe ie is denoted pe(ie).
Before neighborhood extraction, the 12 positions pe(ie) are converted into 12 positions ps(js) divided into two G.723.1 subframes (of index js0≦js≦1). The above general equation may be used (involving the modulus of the subframe length) to perform the adaptation of the subframe durations. However, it is preferred here merely to distinguish three situations according to the value of the index ie:
The four positions recovered in the subframe STE0 of the block are directly assigned to the subframe STS0 with the same position, those of the subframe STE2 of the block are directly assigned to the subframe STS1 with a position increment of +20, the positions of the subframe STE1 below 20 are assigned to the subframe STS0 with an increment of +40, and the others are assigned to the subframe STS1 with an increment of −20.
The neighborhoods of those 12 positions are then extracted. Note that the right-hand (respectively left-hand) neighborhoods of the positions of the subframe STS0 (respectively STS1) to be extracted from their subframe can be authorized, these neighbor positions being then in the subframe STS1 (respectively STS0).
The temporal correspondence and neighborhood extraction steps can be interchanged. In this case, the right-hand (respectively left-hand) neighborhoods of the positions of the subframe STE0 (respectively STE2) to be extracted from their subframe can be authorized, those neighbor positions then being in the subframe STE1. Similarly, the right-hand (respectively left-hand) neighborhoods of the positions in STE1 can lead to neighbor positions in STE2 (respectively STE0).
Once the ensemble of restricted positions for each subframe STS has been constituted, the final step consists in exploring the restricted directory constituted in this way for each subframe STS to select the Np (=6 or 5) pulses with the same parity. This procedure can be derived from the standardized algorithm or take its inspiration from other focusing procedures.
To illustrate this embodiment, consider three G.729 subframes that can be used to construct the subdirectories of two G.723.1 subframes. Assume that G.729 yields the following positions:
Thus we have the sets of positions {1, 5, 32, 39, 44, 55} for the subframe STSO and {2, 11, 20, 21, 44, 57} for the subframe STS1.
At this stage it is necessary to extract the neighborhoods. Taking a neighborhood fixed at 1, for example, we obtain:
Ps0={0,1,2}∪{4,5,6}∪{31,32,33}∪{38,39,40}∪{43,44,45}∪{54,55,56}
Ps1={1,2,3}∪{10,11,12}∪{20,21,22}∪{21,22,23}∪{43,44,45}∪{56,57,58}
MP-MLQ imposes no constraint on the pulses, apart from their parity. Over a subframe, they must all have the same parity. It is therefore necessary here to split Ps0 and Ps1 into two subensembles, as follows:
Finally, this subdirectory is transmitted to the selection algorithm that determines the Np best positions in the sense of the CELP criterion for the G.723.1 subframes FTS0 et STS1 .This considerably reduces the number of combinations to be tested. For example, there remain in the subframe STS0 nine even positions and eight odd positions, rather than 30 and 30.
Certain precautions are nevertheless required in situations in which the positions selected by G.729 are such that the extraction of the neighborhoods yields a number N of possible positions lower than the G.723.1 number of positions (N <Np). This is the case in particular if the G.729 positions are all in sequence (for example: {0,1,2,3}). There are then two options:
The opposite processing operation, consisting in selecting a subensemble of the 8 kbps mode G.729 ACELP directory with four pulses from elements of a 6.3 kbps mode G.723.1 MP-MLQ directory, is described next.
Overall, the process is similar. Two G.723.1 subframes correspond to three G.729 frames. Once again, the G.723.1 positions are extracted and translated into the G.729 time frame. These positions could advantageously be translated in the form “track−rank in the track” in order to benefit as before from the ACELP structure to extract the neighborhoods and search for the optimum positions.
The same arrangements as before are adopted to prevent situations in which neighborhood extraction would yield an insufficient number of positions (here fewer than four positions).
Thus the present invention determines at lower cost the positions of a set of pulses from a first set of pulses, the two sets of pulses belonging to two multipulse directories. Those two directories may be distinguished by their size, the length and the number of pulses of their code words, and the rules governing the positions and/or amplitudes of the pulses. Preference is given to the neighborhoods of the positions of the pulses of the selected set(s) in the first directory to determine those of a set in the second directory. The invention further exploits the structure of the starting and/or destination directories to reduce complexity further. From the first embodiment described above entailing changing from an MP-MLQ model to a ACELP model, it will be clear that the invention is easy to apply to two multipulse models having different structural constraints. From the second embodiment, entailing passing between two models having different numbers of pulses based on the same ACELP structure, it will be clear that the invention advantageously exploits the structure of the directories to reduce transcoding complexity. From the third embodiment, entailing passing between an MP-MLQ model and an ACELP model, it will be clear that the invention may even be applied to coders with different subframe lengths or sampling frequencies. The invention adjusts the quality/complexity trade-off and in particular greatly reduces the calculation complexity for a minimum deterioration compared to a conventional search of a multipulse model.
| Number | Date | Country | Kind |
|---|---|---|---|
| 0314489 | Dec 2003 | FR | national |
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/FR04/03008 | 11/24/2004 | WO | 10/6/2006 |
