EP1692687B1 - Transcoding between the indices of multipulse dictionaries used for coding in digital signal compression - Google Patents

Abstract

The invention relates to compressive transcoding between pulse coders using multipulse dictionaries in which each pulse occupies a position marked by an index. For each current pulse position supplied by a first coder, a neighborhood (Vge, Vde) is formed around that position. As a function of the pulse positions accepted by the second coder, pulse positions are selected in an ensemble constituted by a union of the neighborhoods. The second coder finally receives this selection (sj), involving a number of pulse positions smaller than the total number of pulse positions in the dictionary of the second coder.

Description

The present invention relates to the encoding / decoding of digital signals, particularly in applications for transmitting or storing multimedia signals such as audio signals (speech and / or sounds).

In the field of compression coding, many coders model a signal of L samples by pulses in a small number, much smaller than the total number of samples. This is, for example, the case of certain audio frequency coders such as the audio coder said "TDAC" described in particular in the published document US 2001 / 027,393 . In this coder, the modified discrete cosine transform coefficients normalized in each band are quantized by vector quantizers using algebraic dictionnaries nested in size, these algebraic codes generally comprise some non-zero components, the other components being equal to zero. This is also the case for the majority of synthetic analysis speech coders including "ACELP" (for "Algebraic Code Excited Linear Prediction") type MP-MLQ (for "MultiPulse Maximum Likelihood Quantization"), or other. Indeed, to model the innovation signal, these coders use a repertoire, composed of waveforms having very few non-zero components whose positions and amplitudes obey, moreover, predetermined rules.

Briefly hereinafter are described such synthetic analysis coders.

In these synthesis analysis coders, the synthesis model is used in coding to extract the parameters modeling the signals to be coded. These signals can be sampled at the telephone frequency (F _e = 8 kHz)
or at a higher frequency, for example at 16 kHz for wideband coding (bandwidth 50 Hz to 7 kHz). Depending on the application and Desired quality, the compression ratio varies from 1 to 16. These encoders operate at rates of 2 to 16 kbit / s in the telephone band, and at rates of 6 to 32 kbit / s in broadband.

The principle of a CELP-type digital coding / decoding device, which is the most widely used synthesis analysis coder / decoder for the coding / decoding of speech signals, is briefly described below. The speech signal is sampled and converted into a series of blocks of the samples called frames. In general, each frame is cut into smaller blocks of L samples, called subframes. Each block is synthesized by filtering a waveform extracted from a repertoire (also called dictionary), multiplied by a gain, through two filters varying in time. The excitation dictionary is a finite set of waveforms of L samples. The first filter is the long-term prediction filter. An analysis called "LTP" (for "Long Term Prediction") allows to evaluate the parameters of this long-term prediction filter that exploits the periodicity of voiced sounds (typically representing the frequency of the fundamental (frequency of vibration of the vocal cords) ). The second filter is the short-term prediction filter. The "LPC" ( Linear Prediction Coding) analysis methods make it possible to obtain these short-term prediction parameters, representative of the vocal tract transfer function and characteristics of the signal spectrum (typically representing the modulations due to the shape taken by the lips, at the position of the tongue and the larynx, or others).

The method used to determine the innovation sequence is the method of synthesis analysis. At the encoder level, a large number of excitation dictionary innovation sequences are filtered by the two LTP and LPC filters, and the selected waveform is the one producing the synthetic signal closest to the original signal according to a criterion. perceptual weighting, commonly known as the CELP criterion.

The use of multi-pulse dictionaries in such synthetic analysis encoders is briefly described below, it being understood, however, that CELP coders, such as CELP decoders, are well known to those skilled in the art.

• dans le premier utilisé pour les sous-trames paires, les formes d'ondes comportent 6 impulsions et,
• dans le deuxième utilisé pour les sous-trames impaires, elles comportent 5 impulsions.

The multi-rate encoder of ITU-T G.723.1 is a good example of a synthesis-based encoder using multi-pulse dictionaries. Here, the pulse positions are all distinct. The two encoder rates (6.3 kbit / s and 5.3 kbit / s) model the innovation signal by dictionary-derived waveforms that have only a limited number of non-zero pulses: 6 or 5 for broadband, 4 for low bit rate. These pulses are of amplitude +1 or -1. In its 6.3 kbit / s mode, the G.723.1 encoder alternately uses two dictionaries:

• in the first used for even subframes, the waveforms comprise 6 pulses and,
• in the second used for the odd subframes, they comprise 5 pulses.

For these two dictionaries, a single restriction is imposed on the positions of the pulses of any vector-code. These positions must all have the same parity, that is, they are all odd or even. In the 5.3 kbit / s mode dictionary, the positions of the 4 pulses are more constrained. In addition to the same parity constraint as the broadband mode dictionaries, each pulse has a limited choice of positions.

The 5.3 kbit / s mode multi-pulse dictionary belongs to the well-known family of ACELP dictionaries. The structure of an ACELP directory is based on the ISPP technique (for "Interleaved Single-Pulse Permutation") which consists in dividing all the L positions into K interleaved tracks, each of the N pulses being located in certain predefined tracks. In some applications, the L dimension of code words can be extended to L + N. Thus in the case of the encoder low bit rate directory according to ITU-T G.723.1, the block size of 60 samples was extended to 64 samples and the 32 even (odd) positions were divided into 4 tracks. interlaced length 8 not overlapping. So there are two groups of 4 tracks, one for each parity. Table 1 below shows for each pulse denoted i ₀ to i ₃ all of these 4 tracks for the even positions. <i> Table 1: Positions and magnitudes of the 5.3 kbit G.723.1 encoder ACELP dictionary </ i> / <i> s </ i> Impulse Sign positions i ₀ ± 1 0, 8, 16, 24, 32, 40, 48, 56 i ₁ ± 1 2, 10, 18, 26, 34, 42, 50, 58 i ₂ ± 1 4, 12, 20, 28, 36, 44, 52, (60) i ₃ ± 1 6, 14, 22, 30, 38, 46, 54, (62)

The ACELP innovation dictionaries are used in many synthetic analysis coders that are standardized (ITU-T G.723.1, ITU-T G.729, IS-641, 3GPP NB-AMR, 3GPP WB-AMR). Tables 2 to 4 below show some 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 shows the 17-bit ACELP dictionary and 4 non-zero pulses ± 1, used in the ITU-T G.729 8-bit / sec encoder, in the 7.4 kbit / s IS-641 encoder as well as in the 7.4 and 7.95 kbit / s modes of the 3GPP NB-AMR encoder. <i> Table 2: Positions and amplitudes of the ACELP dictionary pulses of ITU-T G. 729 coders at 8 kbit </ i> / <i> s, IS641 at 7.4 kbit </ i> / <i> s and 3GPP NB-AMR at 7.4 and 7.95 kbit </ i> / <i> s </ i> Impulse Sign positions i ₀ ± 1 0, 5, 10, 15, 20, 25, 30, 35 i ₁ ± 1 1, 6, 11, 16, 21, 26, 31, 36 i ₂ ± 1 2, 7, 12, 17, 22, 27, 32, 37 i ₃ ± 1 3, 8, 13, 18, 23, 28, 33, 38, 4, 9, 14, 19, 24, 29, 34, 39

In the 35-bit ACELP dictionary used in the 12.2 kbit / s mode of the 3GGPP NB-AMR encoder, each code vector contains 10 non-zero pulses of amplitude ± 1. The block of 40 samples is divided into 5 tracks of length 8, each containing 2 pulses. Note that the two pulses of the same track can overlap and result in a single pulse amplitude ± 2. This dictionary is presented in Table 3. <i> Table 3: Positions and amplitudes of the pulses of the ACELP dictionary of the 3GPP NB-AMR coder at 12.2 kbit </ i> / <i> s </ i> Impulse Sign positions i ₀ , i ₅ ± 1 0, 5, 10, 15, 20, 25, 30, 35 i ₁ , i ₆ ± 1 1, 6, 11, 16, 21, 26, 31, 36 i ₂ , i ₇ ± 1 2, 7, 12, 17, 22, 27, 32, 37 i ₃ , i ₈ ± 1 3, 8, 13, 18, 23, 28, 33, 38 i ₄ , i ₉ ± 1 4, 9, 14, 19, 24, 29, 34, 39

Finally, Table 4 presents the 11-bit ACELP dictionary and 2 non-zero pulses of ± 1 amplitude, used in the low-bit-rate (6.4 bps / sec) extension of the ITU-T G.729 encoder and in the 5.9 kbit / s mode of the 3GPP NB-AMR encoder. <i> Table 4: ACELP dictionary pulses and magnitudes of ITU-T G. 729 6.4 kbit </ i> / <i> s and 3GPP NB-AMR 5.9 kbit coders </ i> / <i> s </ i> Impulse Sign positions i ₀ ± 1 1, 3, 6, 8, 11, 13, 16, 18, 21, 23, 26, 28, 31, 33, 36, 38 i ₁ ± 1 0, 1, 2, 4, 5, 6, 7, 9, 10, 11, 12, 14, 15, 16, 17, 19, 20, 21, 22, 24, 25, 26, 27, 29, 30, 31,32,34,35,36,37,39

What is meant by "exploration" of multi-pulse dictionaries is described below.

As with any quantization operation, the search for the optimal modeling of a vector to be coded consists in choosing from the set (or in a subset) of the code vectors of the dictionary, the one that "resembles" it the most, that is to say the one that minimizes a measure of distance between him and this input vector. To do this, we proceed to a step called "exploration" of the dictionaries.

In the case of multi-pulse dictionaries, this step amounts to looking for the combination of pulses that optimizes the proximity between the signal to be modeled and the signal resulting from the choice of pulses. Depending on the size and / or structure of the dictionary, this exploration may be exhaustive or not (so more or less complex).

Since the dictionaries used in the aforementioned TDAC coder are type II permutation code unions, the algorithm for encoding a vector of standardized transform coefficients exploits this property to determine its nearest neighbor among all the code vectors while calculating only a limited number of distance criteria (with a use of so-called "absolute leaders").

In synthetic analysis coders, the exploration of multi-pulse dictionaries is not exhaustive except for small dictionaries. For higher-rate dictionaries, only a small percentage of the dictionary is explored. For example, the exploration of ACELP multi-pulse dictionaries is usually done in two steps. To simplify this search, a first step pre-selects for each possible pulse position its amplitude (and hence its sign as indicated above) by a simple quantization of a signal dependent on the input signal. The amplitudes of the pulses being fixed, the positions of the pulses are then sought by a synthesis analysis technique (according to the CELP criterion). Despite the exploitation of the IPPH structure and the limited number of pulses, an exhaustive search of combinations of positions is performed only for low rate dictionaries (typically less than or equal to 12 bits). This is the case, for example, for the 11-bit ACELP dictionary (table 4) used in the 6.4 kbit / s G.729 coder where the 512 combinations of 2 pulse positions are all tested to select the best one, which amounts to calculating the corresponding 512 CELP criteria.

For the higher rate dictionaries, various so-called focusing methods have been proposed. This is called "focused research".

Some of these known methods are used in the standard coders mentioned above. Their purpose is to reduce the number of combinations of positions to explore based on the properties of the signal to be modeled. For example, we can cite the so - called "depth-first tree" algorithm , used by many standardized ACELP coders. In this algorithm, certain positions are preferred, such as the local maxima of the tracks of a target signal depending on the input signal, the synthetic signal passed and the filter composed of the synthesis and perceptual weighting filters. There are several variants depending on the size of the dictionary used. To explore the ACELP dictionary at 35 bits and 10 pulses (table 3), the first pulse is placed at the same position as the overall maximum of the target signal. Then, four iterations are performed by circular permutation of the consecutive tracks. At each iteration, the position of the second pulse is set to the local maximum of one of the other 4 tracks, the positions of the other eight remaining pulses are searched sequentially in pairs in nested loops. At each iteration 256 (= 4 pairs × 8 × 8) different combinations are tested, which makes it possible to explore only 1024 combinations of positions of the 10 pulses among the 2 ²⁵ of the dictionary. Another variant is used in the IS641 encoder where a higher percentage of combinations of the 17-bit dictionary and 4 pulses (table 2) is explored. We tests 768 combinations among the 8192 (= 2 ¹³ ) combinations of pulse positions. In the 8 kbit / s G.729 encoder, this same ACELP dictionary is explored by a different method of focusing. The algorithm performs an iterative search by nesting four pulse search loops (one per pulse). The search is focused by making the entry in the inner loop (searching for the last pulse belonging to tracks 3 or 4) conditional on exceeding an adaptive threshold. This threshold also depends on the properties of the target signal (local and average maxima of the first 3 tracks). In addition, the maximum number of explorations of combinations of 4 pulses is fixed at 1440 (17.6% of the 8192 combinations).

(respectivement $2\times 2^6\times C_{30}^6$

) combinaisons de 5 (respectivement 6) impulsions ne sont pas explorées. Pour chaque grille, l'algorithme emploie une analyse connue, de type "multipulse", pour rechercher séquentiellement les positions et les amplitudes des impulsions. Comme pour les dictionnaires ACELP, des variantes existent pour restreindre le nombre de combinaisons testées.In the 6.3 kbit / s G.723.1 encoder, all $2\times 2^5\times {\mathrm{<figure-callout\; id="30"\; label="VS"\; filenames="imgf0002.png"\; state="\{\{state\}\}">VS</figure-callout>}}_{30}^5$

(respectively $2\times 2^6\times {\mathrm{<figure-callout\; id="30"\; label="VS"\; filenames="imgf0002.png"\; state="\{\{state\}\}">VS</figure-callout>}}_{30}^6$

) combinations of 5 (respectively 6) pulses are not explored. For each grid, the algorithm uses a known "multipulse" analysis to sequentially search the positions and amplitudes of the pulses. As with ACELP dictionaries, variants exist to restrict the number of combinations tested.

Such techniques, however, have the following problems.

The exploration, even suboptimal, of a multi-pulse dictionary is in many coders an expensive operation in computing time. For example, in G.723.1 encoders at 6.3kbit / s and G.729 at 8kbit / s, this search represents almost half of the total complexity of the encoder. For the NB-AMR, it represents one third of the total complexity of the encoder. For the TDAC encoder, it represents one quarter of the total complexity of the encoder.

It will be understood in particular that this complexity becomes critical when several codings must be performed by the same processing unit, such as a gateway responsible for managing many parallel communications or a server distributing a large number of multimedia contents. The problem of complexity is further increased by the multiplicity of compression formats that circulate on networks.

Indeed, to offer mobility and continuity, modern and innovative multimedia communication services must be able to operate in a wide variety of conditions. The dynamism of the multimedia communication sector and the heterogeneity of networks, accesses and terminals have led to a proliferation of compression formats whose presence in communication channels requires several codings either in cascade (transcoding) or in parallel (multi-coding). -format or multi-mode encoding).

What is meant by "transcoding" is defined below . Transcoding becomes necessary when, in a transmission chain, a compressed signal frame sent by an encoder can not continue its path, in this format. Transcoding makes it possible to convert this frame into another format compatible with the rest of the transmission chain. The most basic solution (and the most common at the moment) is the end-to-end addition of a decoder and an encoder. The compressed frame comes in a first format, it is uncompressed. This decompressed signal is then recompressed in a second format accepted later in the communication chain. This cascading of a decoder and an encoder is called "tandem". This solution is very expensive in complexity (mainly because of the recoding) and it degrades the quality. Indeed, the second coding is done on a decoded signal, which is a degraded version of the original signal. In addition, a frame may encounter several tandems before reaching its destination. We can easily imagine the cost in calculation and the loss of quality. In addition, the delays related to each tandem operation builds up and can interfere with the interactivity of communications.

In addition, the complexity is also problematic in the context of a multi-format compression system where the same content is compressed in several formats. This is the case of content servers that broadcast the same content in several formats adapted to the conditions of access, networks and terminals of different customers. This multi-coding operation becomes extremely complex as the number of desired formats increases, which quickly saturates system resources.

Another case of multiple-coding in parallel is post-decision multi-mode compression. For each signal segment to be encoded, several compression modes are performed and the one that optimizes a given criterion or obtains the best compromise between bit rate and distortion is selected. Here again, the complexity of each of the modes of compression limits the number and / or leads to elaborate a selection a priori of a very limited number of modes.

The following are known approaches which have been implemented in an attempt to solve the problems raised described above.

New multimedia communication applications (such as audio and video) often require multiple codings either in cascade (transcoding) or in parallel (multi-coding and multi-mode coding with ex-post decision). The complexity barrier posed by all these codings remains a problem to be solved, in spite of the increase of the current processing powers. Most of these known multi-coding operations do not take into account the interactions between the formats and between the format of E and its contents. However, some intelligent transcoding techniques have been proposed that do not just decode but recode, but exploit the similarities between encoding formats and thus reduce the complexity while limiting the degradation provided.

Hereinafter described "intelligent" transcoding methods .

Within the same encoder family (CELP, parametric, transform, or other), all the encoders extract the same physical parameters from the signal. However, there is a wide variety of modeling and / or quantification of these parameters. Thus, from one coder to another, the same parameter can be coded in the same way, or, on the contrary, very differently.

Moreover, the encodings can be strictly identical. They can be identical in the modeling and the calculation of the parameter, but can be differentiated simply by the translation of the coding in the form of bits. Finally, they can be completely different as much by the modeling of the parameter as its quantification, or even by its frequency of analysis or sampling.

If the modeling and the calculation of a parameter are strictly identical, including in the binary translation, it is sufficient to copy the corresponding binary field of the bit stream generated by the first coder to that of the second. This very favorable case occurs for example when transcoding the G.729 standard to the IS-641 standard for adaptive excitation (LTP delays).

If, for the same parameter, the two coders are distinguished only by the binary translation of the calculated parameter, it suffices to decode the binary field of the first format, then to return to the binary domain using the coding method according to the second format . This conversion can also be performed by bijective correspondence tables. This is the case, for example when transcoding the fixed excitations of the G.729 standard to the AMR standard (7.4 and 7.95kbit / s).

In the two previous cases, the transcoding of the parameter is done while remaining at the level of the bits. Simple bit manipulation makes the parameter compatible with the second coding format. On the other hand, when a parameter extracted from the signal is modeled or quantified differently by two coding formats, the transition from one to the other is not so simple. Several methods have been proposed. They operate either at the parameter level, at the level of the excitation, or at the level of the decoded signal.

For transcoding in the parameter domain, staying at the parameter level is possible when both encoding formats compute a parameter in the same way but quantify it differently. The differences in quantification can be related to the chosen precision or to the chosen method (scalar, vector, predictive, or other). It is then enough to decode the parameter, then to quantify it by the method of the second coding format. This known method is currently applied, in particular for the transcoding of excitation gains. Often the decoded parameter must be modified before requantification. For example, if the encoders have parameter analysis frequencies or different frame / subframe lengths, it is common to interpolate / decimate the parameters. The interpolation can be done for example according to the method described in the published document US2003 / 033142 . Another possible modification is to round the parameter to the precision imposed on it by the second coding format. This case is especially for the height of the fundamental frequency (or "pitch").

If it is not possible to transcode a parameter while remaining in the parameter domain, one can go back to a higher level in the decoding. he This is the field of excitation, without going to the domain of the signal. This technique has been proposed for gains in the reference: " Improving transcoding capability of speech coders in clean and frame erasured channel environments "from Hong-Goo Kang, Kim Hong Kook, Cox, RV, in Speech Coding, 2000, Proceedings 2000, IEEE Workshop on Speech Coding, Pages 78-80 .

Finally, a last solution (the most complex and the least "intelligent") is to explicitly recalculate the parameter as the encoder would, but from a synthesized signal. This operation returns to a kind of partial tandem, only certain parameters being entirely recalculated. This method has been applied to various parameters such as fixed excitation, gains in the aforementioned IEEE reference, or pitch.

For transcoding pulses, even though several techniques have been developed to calculate parameters quickly and cheaply, few of these solutions today have an intelligent approach to compute the pulses of a format from the equivalent parameter under a other format, see for example the document WO 3058407 . In synthesis analysis coding, intelligent transcoding of pulse codes is applied only when modeling is identical (or near). On the other hand, if the modelizations are different, the partial tandem method is used. It should be noted that in order to limit the complexity of this operation, focused approaches exploiting the properties of the decoded signal or of a derived signal such as a target signal have been proposed. In the aforementioned document US 2001 / 027,393 , in the exemplary embodiment implementing an MDCT transform coder, a flow rate change procedure is presented which can be considered as a special case of intelligent transcoding. Indeed, this procedure makes it possible to requantize a vector of a first dictionary by a vector of a second dictionary. For this purpose, it distinguishes two cases according to whether or not the vector to be requantized in the second dictionary. If the quantified vector belongs to the new dictionary, the modeling is identical; otherwise, the partial decoding method is applied.

The present invention proposes, as indicated in claims 1 and 22, 23, by departing from all these known techniques, a multi-pulse transcoding based on a selection of a subset of combinations of pulse positions of a set of sets of pulses from a combination of pulse positions of another set of pulse sets, the two sets being distinguishable by the number of pulses they comprise as well as by the rules governing their positions and / or their amplitudes. This transcoding is very useful in particular for multiple coding in cascade (transcoding) or in parallel (multi-coding and multi-mode coding).

For this purpose, the present invention firstly proposes a transcoding method between a first encoder / decoder in compression and at least one second encoder / decoder in compression. These first and second encoders / decoders are of pulse type and use multi-pulse dictionaries in which each pulse has a position marked by an associated index.

1. a) le cas échéant, adaptation de paramètres de codage entre lesdits premier et second codeurs/décodeurs,
2. b) obtention, à partir du premier codeur/décodeur, d'un nombre choisi de positions d'impulsions et d'indices de positions respectivement associés,
3. c) pour chaque position d'impulsion courante d'indice donné, formation d'un groupe de positions d'impulsions comportant au moins la position d'impulsion courante et des positions d'impulsions d'indices associés immédiatement inférieurs et immédiatement supérieurs à l'indice donné,
4. d) sélection, en fonction de positions d'impulsions admises par le second codeur/décodeur, d'une partie au moins des positions d'impulsions dans un ensemble constitué par une union desdits groupes formés à l'étape c), et
5. e) transmission des positions des impulsions ainsi sélectionnées au second codeur/décodeur, pour un codage/décodage à partir desdites positions transmises.

The transcoding method within the meaning of the invention comprises the following steps:

1. a) where appropriate, adaptation of coding parameters between said first and second coders / decoders,
2. b) obtaining, from the first encoder / decoder, a selected number of respectively associated pulse positions and position indices,
3. c) for each current pulse position of given index, forming a group of pulse positions comprising at least the current pulse position and associated index pulse positions immediately below and immediately greater than 1 given index,
4. d) selecting, as a function of pulse positions allowed by the second coder / decoder, at least a portion of the pulse positions in a set consisting of a union of said groups formed in step c), and
5. e) transmitting the positions of the pulses thus selected to the second encoder / decoder, for coding / decoding from said transmitted positions.

Thus, the selection of step d) involves a number of possible pulse positions less than the total number of possible pulse positions of the dictionary of the second coder / decoder.

It will be understood in particular that in step e), in the case where the second encoder / decoder mentioned above is an encoder, the selected pulse positions are transmitted to this encoder, for search coding only among the transmitted positions. In the case where the aforementioned second encoder / decoder is a decoder, these selected pulse positions are transmitted for a decoding of these positions.

Preferably, step b) uses a partial decoding of the bitstream provided by the first coder / decoder in order to identify a first number of pulse positions used by the first coder / decoder, in a first coding format. Thus, the number chosen in step b) preferably corresponds to this first number of pulse positions.

In an advantageous embodiment, the above steps are implemented by a computer program product comprising program instructions for this purpose. In this respect, the present invention also aims at such a computer program product intended to be stored in a memory of a processing unit, in particular a computer or a mobile terminal, or on a removable memory medium intended for to cooperate with a drive of the processing unit.

The present invention also relates to a transcoding device between first and second encoders / decoders in compression, and then having a memory adapted to store instructions of a computer program product of the type described above.

• la figure 1a représente schématiquement le contexte de transcodage au sens de la présente invention, dans une configuration "en cascade",
• la figure 1b représente schématiquement le contexte de transcodage au sens de la présente invention, dans une configuration "en parallèle",
• la figure 2 représente schématiquement les différents cas prévus pour les traitements de transcodage à effectuer,
• la figure 2a représente schématiquement un traitement d'adaptation prévu pour le cas où les fréquences d'échantillonnage des premier E et second S codeurs sont différentes,
• la figure 2b représente schématiquement une variante du traitement de la figure 2a,
• la figure 3 résume les étapes du procédé de transcodage au sens de l'invention,
• la figure 4 représente schématiquement le cas de deux sous-trames respectivement des codeurs E et S, de durées différentes Le et L_s (avec L_e>L_s), mais de mêmes fréquences d'échantillonnage,
• la figure 4b représente à titre d'exemple un cas pratique de la figure 4 en illustrant la correspondance temporelle entre un codeur G.723.1 et un codeur G.729,
• la figure 5 illustre schématiquement le découpage de l'excitation du premier codeur E au rythme du second codeur S,
• la figure 6 illustre le cas où l'une des pseudo-sous-trames STE'O est vide, la figure 7 représente schématiquement un traitement d'adaptation prévu pour le cas où les durées de sous-trame des premier E et second S codeurs sont différentes.

Other features and advantages of the invention will appear on examining the detailed description below, and the attached drawings in which:

• the figure 1a schematically represents the transcoding context within the meaning of the present invention, in a "cascade" configuration ,
• the figure 1b schematically represents the context of transcoding in the sense of the present invention, in a "parallel" configuration ,
• the figure 2 schematically represents the different cases provided for the transcoding processing to be performed,
• the figure 2a schematically represents adaptation processing provided for the case where the sampling frequencies of the first E and second S encoders are different,
• the figure 2b schematically represents a variant of the treatment of figure 2a ,
• the figure 3 summarizes the steps of the transcoding method within the meaning of the invention,
• the figure 4 represents schematically the case of two sub-frames respectively of the encoders E and S, of different durations Le and L _s (with L _e > L _s ), but with the same sampling frequencies,
• the figure 4b represents as an example a practical case of the figure 4 illustrating the temporal correspondence between a G.723.1 encoder and a G.729 encoder,
• the figure 5 schematically illustrates the division of the excitation of the first encoder E to the rhythm of the second encoder S,
• the figure 6 illustrates the case where one of the pseudo-subframes STE'O is empty, the figure 7 schematically represents adaptation processing provided for the case where the subframe durations of the first E and second S encoders are different.

It is firstly indicated that the present invention is part of the modeling and coding of digital multimedia signals such as audio signals (speech and / or sounds) by multi-pulse dictionaries. It can be implemented in the context of cascading or parallel coding / multiple decoding or any other system using the modeling of a signal by a multi-pulse representation and which, from the knowledge of a first game of pulses belonging to a first set, must determine at least one set of pulses of a second set. For the sake of brevity, only the case of a passage from one set to another set is described but the invention is also applicable in the case of passage to n (n ≥ 2) sets. Furthermore, only the case of a "transcoding" between two coders is described below, but, of course, the transcoding between an encoder and a decoder can be deduced without major difficulty.

We therefore consider the case of two modelizations of a signal by sets of pulses corresponding to two coding systems. On the Figures 1a and 1b , there is shown a transcoding device D between a first encoder E, using a first encoding format COD1, and a second encoder S, using a second coding format COD2. The encoder E delivers a coded binary stream (or "train") s _CE (in the form of a succession of coded frames) to the transcoding device D, which comprises a partial decoding module 10 to recover the number Ne of positions d pulses used in the first coding format and the positions p _e of these pulses. As will be seen in detail below, the transcoding device in the sense of the invention proceeds to extract the neighborhoods of the right v ^e _d and left v ^e _g of each pulse position p _e and selects, in the union of these neighborhoods, pulse positions that will be recognized by the second encoder S. The module 11 of the transcoding device shown on the Figures 1a and 1b therefore performs these steps to deliver this selection of positions (noted S _j on these Figures 1a and 1b ) It will be understood in particular that from this selection S _j , there is constituted a sub-directory smaller than the size of the dictionary that usually uses the second encoder S, according to an advantage that provides the invention. By using this subdirectory, the encoding performed by the encoder S is of course faster because it is more restricted, without affecting the quality of the coding.

In the example shown on the figure 1a , the transcoding device D furthermore comprises a module 12 for at least partial decoding of the coded stream S _CE delivered by the first coder E. The module 12 then supplies the second coder S with an at least partially decoded version s' ₀ of the signal d origin s ₀ . The second encoder S then delivers, based on this version s' ₀ , a coded bitstream s _CS .

In this configuration, the transcoding device D thus performs an encoding adaptation between the first encoder E and the second encoder S, advantageously favoring a faster (because more restricted) coding with the second encoder S. Of course, alternatively, the entity referenced S on the Figures 1a and 1b can be a decoder and, in this variant, the device D within the meaning of the invention performs a transcoding, properly speaking, between an encoder E and a decoder S, this decoding taking place quickly thanks to the information provided by the device D. Since the process is reversible, it will be understood that, more generally, the transcoding device D within the meaning of the present invention operates between a first encoder / decoder E and a second encoder / decoder S.

It is indicated that the arrangement of the encoder E, the transcoder D and the encoder S can respect a "cascade" configuration , as shown in FIG. figure 1a . In the variant shown on the figure 1b , this arrangement can comply with a configuration "in parallel". In this case, the two encoders E and S receive the original signal s ₀ and the two encoders E and S respectively deliver the coded streams S _CE and S _CS . Of course, the second encoder S no longer has to receive here the version s' ₀ of the figure 1a and the at least partial decoding module 12 of the transcoding device D is no longer necessary. It is furthermore indicated that, if the encoder E can provide an output compatible with the input of the module 11 (both with respect to its number of pulses and with respect to its pulse positions), the module 10 can simply be omitted or "short-circuited".

It is further indicated that the transcoding device D may simply be equipped with a memory storing the instructions for carrying out the above steps and a processor for processing these instructions.

The application of the invention is therefore as follows. The first encoder E has performed its coding operation on a given signal s ₀ (for example the original signal). The positions of the pulses chosen by the first coder E are thus available. This coder has determined these positions p _e by a technique which is specific to it during the coding process. The second encoder S must also perform its coding. In the case of transcoding, the second encoder S has only the bitstream generated by the first encoder and the invention is applicable here to "intelligent" transcoding as defined above. In the case of multiple coding in parallel, the second coder S also has the signal available to the first coder and the invention applies here to "intelligent multi-coding". Indeed, a system that wants to encode the same content in several formats can exploit the information of a first format to simplify the coding operations of other formats. The invention can also be applied to the particular case of multiple coding in parallel, which is the multi-mode coding with a posterior decision.

The present invention makes it possible to quickly determine the positions p _s (or, indiscriminately noted hereinafter, s _i ) pulses for another encoding format from positions p _e (or, noted indistinctly further below, e _i ) pulses of a first format. It makes it possible to considerably reduce the calculation complexity of this operation for the second encoder by limiting the number of possible positions. For this purpose, it uses the positions chosen by the first coder to define a restricted set of positions in all the possible positions of the second encoder, a restricted set in which the best set of positions for the pulses will be sought. This results in a significant gain in complexity while limiting the degradation of the signal compared to a conventional exhaustive or focused search.

It will thus be understood that the present invention limits the number of possible positions by defining a restricted set of positions from the positions of the first encoding format. It differs from existing solutions insofar as they use only the properties of the signal to be modeled to limit the number of possible positions, by favoring and / or eliminating positions.

Preferably, for each pulse of a set of a first set, two neighborhoods (a right and a left) of adjustable width more or less constrained are defined and a set of possible positions is extracted from it, in which at least one combination of pulses respecting the constraints of the second set.

Advantageously, the transcoding method makes it possible to optimize the complexity / quality compromise by adapting the number of pulse positions and / or the respective sizes (in terms of combinations of pulse positions) of the right and left neighborhoods for each pulse. . This adjustment can be done at the start of processing or at each subframe depending on the authorized complexity and / or the set of starting positions. The invention also makes it possible to adjust / limit the number of combinations of positions by advantageously favoring immediate neighborhoods.

As indicated above, the present invention also relates to a computer program product whose algorithm is designed in particular for the extraction of neighboring positions which facilitates the composition of the combinations of pulses of the second set.

As mentioned above, the heterogeneity of networks and content can bring into play a variety of coding formats. Encoders can be distinguished by many features. In particular, two of them substantially determine the mode of operation of the invention. This is the sampling frequency and the duration of a sub-frame. Hereinafter, the various possible cases are presented in correspondence of implementations of the invention according to these different cases.

• les nombres de positions d'impulsions N_e,N_s,
• les fréquences d'échantillonnage F_e,F_s respectives,
• et les durées de sous-trame L_e,L_s
  qu'utilisent respectivement les codeurs E et S (étape 21).
  Ainsi, on comprendra déjà que les étapes d'adaptation et de récupération des nombres de positions d'impulsions Ne, N_s peuvent avantageusement être interverties ou simplement menées simultanément.

The figure 2 synthesizes the different cases. At first, we obtain:

• the number of pulse positions N _e , N _s ,
• the sampling frequencies F _e , F _s ,
• and the subfield times L _e , L _s
  which E and S coders respectively use (step 21).
  Thus, it will be already understood that the adaptation and recovery steps of the number of pulse positions Ne, N _s can advantageously be reversed or simply conducted simultaneously.

In test 22, the sampling frequencies are compared. If the frequencies are equal, we compare, in test 23, the subframe times. Otherwise, the sampling frequencies are adapted, in step 32, according to a method described hereinafter. At the end of the test 23, if the subframe durations are equal, we compare, in the test 24, the numbers of pulse positions N _e and N _s used. respectively by the first and the second coding format. Otherwise, the subframe times are adapted in step 33 by a method which will also be described hereinafter. It will be understood that steps 22, 23, 32 and 33 together define step a) of adaptation of the coding parameters mentioned above. It is indicated that the steps 22 and 32 (adaptation of the sampling frequencies), on the one hand, and the steps 23 and 33 (adaptation of the subframe durations), on the other hand, can be reversed.

First, the case where the sampling frequencies are equal and the subframe durations are equal are described below.

This case is the most favorable. However, it is necessary to distinguish again the case where the first format uses more pulses than the second one (N _e ≥ N _s ), and the opposite case (N _e <N _s ), according to the result of the test 24.

* Case N _e ≥ N _s of Figure 2

The principle is the following. Considering the two encoders E and S, their repertoires respectively use N _e and N _s pulses to each subframe.

où $v_d^i$

et $v_g^i$

≥ 0 sont les tailles des voisinages droit et gauche de l'impulsion i. Les valeurs de $v_d^i$

et $v_g^i$

, choisies à l'étape 27 de la figure 2, sont plus ou moins grandes selon la complexité et la qualité désirées. Ces tailles peuvent être fixées arbitrairement au début du traitement, ou être choisies à chaque sous-trame S_e.The encoder E calculated the positions of its N _e pulses on the sub-frame s _e . We note below e _i (or, indistinctly, p _e ) these positions. The restricted set P _s of the preferred positions for the pulses of the repertoire of the encoder S then consists of Ne positions e _i and their neighborhoods. $$P_s=\underset{i=0}{\overset{{\mathrm{NOT}}_e-1}{\mathrm{U}}}\left\{\underset{i=0}{\overset{{\mathrm{NOT}}_e-1}{\mathrm{U}}}\left\{{\mathrm{e}}_{\mathrm{i}}\mathrm{+}\mathrm{k}\right\}\right\}$$

or $v_d^i$

and $v_{\mathrm{boy\; Wut}}^i$

≥ 0 are the sizes of the right and left neighborhoods of the pulse i. The values of $v_d^i$

and $v_{\mathrm{boy\; Wut}}^i$

chosen in step 27 of the figure 2 , are larger or smaller depending on the complexity and quality desired. These sizes can be set arbitrarily at the start of processing, or be chosen at each subfield S _e .

voisins de droite et ses $v_g^i$

voisins de gauche.At step 29 of the figure 2 , the set P _s then contains each position e _{i as} well as its $v_d^i$

neighbors right and his $v_{\mathrm{boy\; Wut}}^i$

neighbors on the left.

For each of the N _s pulses of the encoder directory S, it is then necessary to define the positions that this pulse has the right to take among those proposed by P _s .

For this, we will introduce rules governing the construction of the repertoire of S. It is assumed that the N _s pulses of S belong to pre-defined subsets of positions, a given number of pulses sharing the same subset of allowed positions. For example, the pulses of the 12.2 bit / s mode of the 3GPP NB-AMR encoder are distributed 2 by 2 in 5 different subsets, as shown in Table 3 given above. We denote N ' _s the number of subsets of different positions (N' _s ≤ N _s in this example since N ' _s = 5), and T _j (for j = 1 to N' _s ) the subsets of positions defining the repertoire of S.

From the set P _{s, N's} subsets S _j from the intersection of P _s with one of the sets T _j are formed at step 30 of the figure 2 , according to the relationship: $${\mathrm{S}}_{\mathrm{j}}\mathrm{=}{\mathrm{P}}_{\mathrm{s}}\cap T_{\mathrm{j}}$$

et $v_g^i$

doivent être d'une taille suffisante pour qu'aucune intersection ne soit vide. On doit permettre ainsi un réajustement des tailles de voisinage, si nécessaire, en fonction du jeu d'impulsions de départ. C'est l'objet du test 34 de la figure 2, avec une augmentation de la taille des voisinages (étape 35) et retour vers la définition de l'union P_s des groupes formés à l'étape c) (étape 29 sur la figure 2) si l'une des intersections est vide.Neighborhoods $v_d^i$

and $v_{\mathrm{boy\; Wut}}^i$

must be of sufficient size so that no intersection is empty. In this way, neighborhood sizes need to be readjusted, if necessary, according to the starting set of pulses. This is the object of test 34 of the figure 2 , with an increase in the size of the neighborhoods (step 35) and return to the definition of the union P _s of the groups formed in step c) (step 29 on the figure 2 ) if one of the intersections is empty.

On the contrary, if no intersection S _j is empty, it is the subdirectory constituted by these intersections S _j which is sent to the encoder S (end step 31).

Advantageously, the invention exploits. the directory structure. For example, if the directory of the encoder S is of type ACELP, it is the intersections of the positions of the tracks with P _s that are calculated. If the directory of the encoder E is also of the ACELP type, the procedure for extracting neighborhoods also exploits the structure in tracks and the two steps of extraction of the neighborhoods and composition of the restricted subsets of positions, are judiciously associated. In particular, it is interesting that the neighborhood extraction algorithm takes into account the composition of the combinations of pulses according to the constraints of the second set. As will be seen later, neighborhood extraction algorithms are elaborated to facilitate the composition of the combinations of pulses of the second set. An example of such an algorithm is illustrated by one of the embodiments given below (ACELP 2 pulses to ACELP 4 pulses).

The number of possible position combinations is thus limited and the size of the subset of the encoder directory S is generally much smaller than that of the original directory, which greatly reduces the complexity of the penultimate stage of the transcoding. It is specified here that the number of combinations of pulse positions defines the size of the aforementioned subset. It is furthermore specified that it is the number of pulse positions that is reduced in the sense of the invention, which leads to a reduction in the number of combinations of pulse positions and thus makes it possible to obtain a sub-number of pulse positions. restricted directory.

The referenced step 46 on the figure 3 then consists in launching the search for the best set of positions for the N _s pulses in this sub-directory of restricted size. The selection criterion is similar to that of the process of coding. To further reduce complexity, the exploration of this subdirectory can be accelerated using known focusing techniques described above.

The figure 3 summarizes the steps of the invention for the case where the encoder E uses at least as many pulses as the encoder S. However, it is indicated that, as has already been seen with reference to FIG. figure 2 if the number of positions N _s in the second format (the format of S) is greater than the number of positions Ne in the first format (the format of E), the treatment provided is distinguished by only a few advantageous variants which will be described later .

1. a) éventuelle d'adaptation des paramètres de codage (si nécessaire et représentée à cet effet par des traits pointillés sur la figure 3 dans le bloc 41):
   • récupération des positions e_i des impulsions du codeur E, et préférentiellement d'un nombre N_e de positions (étape 42 correspondant à l'étape b) précitée),
   • extraction des voisinages et formation des groupes de voisinages selon la relation : $$P_s=\underset{i=0}{\overset{N_e-1}{\mathrm{U}}}\left\{\underset{\mathrm{k}\mathrm{=}\mathrm{-}{\mathrm{v}}_{\mathrm{g}}^{\mathrm{i}}}{\overset{{\mathrm{v}}_{\mathrm{d}}^{\mathrm{i}}}{\mathrm{U}}}\left\{{\mathrm{e}}_{\mathrm{i}}\mathrm{+}\mathrm{k}\right\}\right\}$$
     
     (étape 43 correspondant à l'étape c) précitée)
   • composition des sous-ensembles restreints de positions {S_j= P_s∩T_j} formant la sélection de l'étape d) précitée et correspondant à l'étape 44 représentée sur la figure 3, et
   • et transmission de cette sélection au codeur S (étape 45 correspondant à l'étape e) précitée).

In short, these stages of the figures 3 can be summarized as follows. After a step

1. a) possible adaptation of the coding parameters (if necessary and represented for this purpose by dotted lines on the figure 3 in block 41):
   • recovery of the positions e _i of the pulses of the encoder E, and preferably of a number N _e of positions (step 42 corresponding to the step b) above),
   • extraction of neighborhoods and formation of neighborhood groups according to the relation: $$P_s=\underset{i=0}{\overset{{\mathrm{NOT}}_e-1}{\mathrm{U}}}\left\{\underset{\mathrm{k}\mathrm{=}\mathrm{-}{\mathrm{v}}_{\mathrm{boy\; Wut}}^{\mathrm{i}}}{\overset{{\mathrm{v}}_{\mathrm{d}}^{\mathrm{i}}}{\mathrm{U}}}\left\{{\mathrm{e}}_{\mathrm{i}}\mathrm{+}\mathrm{k}\right\}\right\}$$
     
     (step 43 corresponding to step c) mentioned above)
   • composition of the restricted subsets of positions {S _j = P _s ∩T _j } forming the selection of step d) above and corresponding to step 44 shown in FIG. figure 3 , and
   • and transmitting this selection to the encoder S (step 45 corresponding to the step e) mentioned above).

It is indicated that after this step 45, the coder S then chooses a set of positions in the restricted directory obtained in step 44.

The method therefore continues with a step 46 of searching in this subdirectory received by the encoder S of an optimal set of positions (opt (S _j )) having the second number N _s of positions, as indicated above. This step 46 of searching for the optimal set of positions is implemented preferentially by a focused search to accelerate the exploration of the sub-directory. The processing continues naturally by the coding then carried out by the second coder S.

Procedures are now described for the case where the number of pulses Ne used by the first coding format is less than the number of pulses N _s used by the second coding format.

* case Ne <N _s of Figure 2

If the S format uses more pulses than the E format, the processing is similar to the treatment discussed above. However, it may happen that S-format pulses do not have positions in the restricted directory. In this case, in a first embodiment, all possible positions for these pulses are allowed. In a second, preferred embodiment, the size of the neighborhoods at V ' _d and V' _g is simply increased, at step 28 of FIG. figure 2 .

* case N _e <N _s <2N _e of figure 2

A special case must be emphasized here. If Ne is close to N _s , typically N _e <N _s <2N _e , then a preferred way of determining the positions is conceivable, although the above treatment is still quite applicable. We can still gain complexity by directly fixing the positions of the pulses of S from those of E. Indeed, the first Ne pulses of S are placed on the positions of those of E. The remaining N _s- N _e pulses are placed as close as possible (to the immediate vicinity) of the first impulses. So, we test at step 25 of the figure 2 if the numbers N _e and N _s are close (with N _e > N _s ) and, if so, proceed as described above for the choice of the pulse positions in step 26.

Of course, in both cases N _e <N _s and N _e <N _s <2N _e , if, despite these precautions, one of the intersections S _j is empty, the size of the neighborhoods V ⁺ _g , V is simply increased. ⁺ _d , in step 35, as described in the case where N _e ≥ N _s .
Finally, in any case, if no intersection S _j is empty, the subdirectory formed by the S _j is transmitted to the second encoder S (step 31).

The processing operations envisaged in the adaptation step a) are now described when the coding parameters of the first and second formats are not the same, in particular in terms of sampling frequencies and subframe durations.

We then distinguish the following cases.

* Equal subframe durations but different sampling frequencies

This situation corresponds to the case "n" for the test 22 and "o" for the test 23 of the figure 2 . The adaptation a) then relates to step 32 of the figure 2 .

We can not apply directly here the previous treatment because the two formats do not have the same division of time. Indeed, because of the different sampling rates, the two frames do not have the same number of samples over the same duration.

Rather than determining the positions of the pulses of the encoder S format without taking into account those of the format of the encoder E, as would a tandem, here two treatments are proposed according to two distinct embodiments. These These processes are of low complexity, by matching the positions of the two formats, which can then be reduced to the treatment described above (as if the sampling frequencies were equal).

The processing of the first embodiment uses a direct quantization of the time scale of the first format by that of the second format. This quantization operation, which can be tabulated or calculated by a formula, thus makes it possible to find, for each position of a subframe of the first format, its equivalent in a subframe of the second format and vice versa.

où F_e et F_s sont les fréquences d'échantillonnages respectives de E et S, et Le et L_s leurs longueurs de sous-trame,
└ ┘ dénotant la partie entière.For example, the correspondence between the positions p _e and p _s in the subframes of the two formats can be defined by the following formula: $$p_s=\lfloor \frac{F_s}{F_e}*p_e+0,5\rfloor ,0\le p_e<{\mathrm{The}}_e\mathrm{and}0\le p_s<{\mathrm{The}}_s$$

where F _e and F _s are the respective sampling frequencies of E and S, and the L _s and their sub-frame lengths,
└ ┘ denoting the entire part.

, d étant le plus grand commun diviseur de Le et L _s), les positions restantes se déduisant alors aisément.Depending on the characteristics of the processing unit, this correspondence may use the formula above or advantageously be tabulated for the values. One can also choose an intermediate solution that will tabulating the _e first values ( $l_e=\frac{{\mathrm{The}}_e}{d}$

, d being the greatest common divisor of Le and L _s ), the remaining positions being easily deduced then.

. It should be noted that several positions of the subframe of S can also be mapped to a position of a subframe of E. For example, by retaining the immediately lower and immediately higher positions of $\frac{F_s}{F_e}*p_e$

.

From the set of positions p _s corresponding to the positions p _e , the general treatment described above (neighborhood extraction, composition of the pulse combinations, selection of the optimal combination) is applied.

Table 5c: Table de correspondance temporelle restreinte du NB-AMR vers WB-AMR Positions NB-AMR 0 1 2 3 4 Positions WB-AMR 0 2 3 5 6 Table 5d: Table de correspondance temporelle restreinte du WB-AMR vers NB-AMR Positions WB-AMR 0 1 2 3 4 5 6 7 Positions NB-AMR 0 1 1 2 2 3 4 4 This case of equal subframe durations but of different sampling frequencies will be found on the tables 5a to 5d hereinafter, with reference to an exemplary embodiment where the encoder E is 3GPP NB-AMR type and the encoder S is of type WB-AMR. The NB-AMR encoder has a sub-frame of 40 samples for a sampling frequency of 8 kHz. The WB-AMR encoder 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 the reverse mapping. The restricted correspondence tables are given in tables 5c and 5d.

<i> Table 5c: Restricted time correspondence table from NB-AMR to WB-AMR </ i> NB-AMR Positions 0 1 2 3 4 WB-AMR positions 0 2 3 5 6 WB-AMR positions 0 1 2 3 4 5 6 7 NB-AMR Positions 0 1 1 2 2 3 4 4

• a1) de quantification directe d'échelle temporelle de la première fréquence vers la seconde fréquence (étape 51 de la figure 2a),
• a2) et de détermination, en fonction de cette quantification, de chaque position d'impulsion dans une sous-trame au second format de codage caractérisé par la seconde fréquence d'échantillonnage, à partir d'une position d'impulsion dans une sous-trame au premier format de codage caractérisé par la première fréquence d'échantillonnage (étape 52 de la figure 2a).

In short, referring to the figure 2a the following steps are foreseen:

• a1) of direct quantization of time scale from the first frequency to the second frequency (step 51 of the figure 2a )
• a2) and determining, based on this quantization, each pulse position in a subframe with the second encoding format characterized by the second sampling frequency, from a pulse position in a sub-frame. frame with the first coding format characterized by the first sampling frequency (step 52 of FIG. figure 2a ).

In general terms, the quantization step a1) is performed by calculation and / or tabulation from a function which, at a pulse position p _e in a first format subframe, matches a position p _s pulse in a subframe with the second format, and this function is substantially a linear combination involving a multiplier coefficient corresponding to the ratio of the second sampling frequency to the first sampling frequency.

Moreover, to switch inversely from a pulse position in a subframe to the second format p _s to a pulse position in a subframe of the first format p _e , of course an inverse function of this combination is applied. linear applied to a pulse position in a subframe in the second format p _s .

It will be understood that the transcoding method is completely reversible and adapts both in a transcoding direction (E-> S) and in the other (S-> E).

In a second embodiment of the adaptation of the sampling frequencies, a conventional principle of change of sampling frequency is used. We start from the sub-frame containing the pulses found by the first format. We oversample at the frequency equal to the least common multiple of the two sampling frequencies F _e and F _s . Then, after low-pass filtering, we sub-sample to return to the sampling frequency of the second format, that is to say F _s . We obtain a subframe at the frequency F _s containing the pulses of filtered E. Again, one can tabulate the result of the over-sampling / low-pass / subsampling operations for each possible position of a sub-frame of E. This processing can also be performed by "on-line" calculation . As in the first embodiment of the adaptation of the sampling frequencies, it will be possible to associate a single or several positions of S with a position of E, as explained hereinafter, and to apply the general treatment in the sense of the invention described above.

• a'1) sur-échantillonner une sous-trame au premier format de codage caractérisé par la première fréquence d'échantillonnage, à une fréquence F_pcm égale au plus petit commun multiple des première et seconde fréquences d'échantillonnage (étape 53 de la figure 2b), et
• a'2) appliquer à la sous-trame sur-échantillonnée un filtrage passe-bas (étape 54 de la figure 2b), suivi d'un sous-échantillonnage, pour atteindre une fréquence d'échantillonnage correspondant à la seconde fréquence d'échantillonnage (étape 55 de la figure 2b).

Le procédé se poursuit par l'obtention, préférentiellement par seuillage, d'un nombre de positions éventuellement variable, ces positions étant adaptées des impulsions de E (étape 56) comme dans le premier mode de réalisation ci-avant.As illustrated in the variant shown on the figure 2b the following steps are foreseen:

• a'1) oversampling a subframe with the first encoding format characterized by the first sampling frequency, at a frequency F _pcm equal to the least common multiple of the first and second sampling frequencies (step 53 of the figure 2b ), and
• a'2) applying to the oversampled subframe a low pass filtering (step 54 of the figure 2b ), followed by a subsampling, to reach a sampling frequency corresponding to the second sampling frequency (step 55 of the figure 2b ).

The method is continued by obtaining, preferably by thresholding, an optionally variable number of positions, these positions being adapted to the pulses of E (step 56) as in the first embodiment above.

* Equal sampling frequencies but different subframe durations

We now describe the treatment provided in the case where the sampling frequencies are equal but the subframe times are different. This situation corresponds to the case "n" for the test 23 but "o" for the test 22 of the figure 2 . The adaptation a) then relates to step 33 of the figure 2 .

As in the case above, it is not possible to directly apply the neighborhood extraction step as it is. First you have to make the two sub-frames compatible. Here, the subframes are different in size. Faced with this incompatibility, rather than calculating the positions of the pulses as the tandem does, a preferred embodiment proposes a low complexity solution for determining a restricted repertoire of position combinations for the second format pulses from the positions of the pulses. of the first format. However, since the subframe of S and that of E are not of the same size, it is not possible to establish a direct temporal correspondence between a subframe of S and a subframe of E. As the show it figure 4 (In which the E and S frames are respectively designated ST _E and ST _S ), the boundaries of the subframes of the two formats are not aligned and over time these subframes shift relative to one another. to the other.

In a preferred embodiment, it is proposed to cut the excitation of E into pseudo-subframes the size of those of S and at the rate of S. On the figure 5 the pseudo subframes denoted ST _E 'are represented. In practice, this also means performing a temporal correspondence between the positions in the two formats by taking into account the difference in size of the subframes to align the positions relative to an origin common to E and S. The determination of this common origin is described in detail below.

A position p ^o _e (respectively p ^o _s ) of the first format (respectively the second format) with respect to this origin coincides with the position p _e (respectively p _s ) of the subframe i _e (respectively j _s ) of E (S respectively) relative to this subframe. We thus have: $${p^o}_e=p_e+i_e{\mathrm{The}}_e{{\mathrm{and}p}^o}_s=p_s+j_s{\mathrm{The}}_s\mathrm{with}0\le p_e<{\mathrm{The}}_e\mathrm{and}0\le p_s<{\mathrm{The}}_s$$

avec 0 ≤ p_e < L_e et 0 ≤ p_s < L_s At a position p _e of the sub-frame i _e of the format of E corresponds the position p _s of the sub-frame j _s of the format of S, where p _s and j _s are respectively the remainder and the quotient of the Euclidean division by L _s of the position p ^o _e of p _e with respect to an origin 0 common to E and S: $$j_s=\lfloor {}^{\left(p_e+i_e{\mathrm{The}}_e\right)}{/}_{{\mathrm{The}}_s}\rfloor {\mathrm{and}p}_s\equiv \left(p_e+i_e{\mathrm{The}}_e\right)\left[{\mathrm{The}}_s\right]$$

with 0 ≤ p _e < L _e and 0 ≤ p _s < L _s

└ ┘ denoting the integer part, ≡ denoting the modulo, the index of a subframe of E (respectively S) being given with respect to the common origin O.

Thus, the positions p _e located in a sub-frame j _s are used to determine, according to the general processing described above, a restricted set of positions for pulses of S in the sub-frame j _s . However, when Le> L _s , it may happen that a subframe of S contains no pulse. In the example of the figure 6 , the pulses of the sub-frame STE0 are represented by vertical lines. The format of E can very well concentrate the pulses of STE0 at the end of the subframe so that the pseudo subframe STE'0 does not contain any impulses. All the pulses placed by E are found in STE'1 when cutting. In this case, a conventional focused search is preferentially applied to the pseudo STE'O subframe.

Preferred embodiments are now described for determining a time origin 0 common to both formats. This common reference is the position (number 0) from which the positions of the pulses in the following subframes are numbered. This position 0 can be defined in different ways, depending on the system using the transcoding method within the meaning of the present invention. For example, for a transcoding module included in an equipment of a transmission system, it will be natural to take as origin the first position of the first frame received after the start of the equipment.

However, the disadvantage of this choice is that the positions take larger and larger values and it may become necessary to limit them. For this, it is sufficient to update the position of the common origin whenever possible. Thus, if the respective lengths Le and L _s of the subframes of E and S are constant over time, the position of the common origin is updated each time the subframes of E and S are aligned. This happens periodically, the period (in samples) being equal to the least common multiple of L _e and L _s .

One can also consider the case where Le and / or L _s are not constant in time. It is no longer possible to find a common multiple at the two subframe lengths, now denoted L _e ( n ) and L _s ( n ) , where n is the subframe number. In this case, the values L _e (n) and L _s (n) should be summed up as they are and compare the two obtained sums with each subframe: $$T_e\left(k\right)={\displaystyle \underset{\mathrm{not}=1}{\overset{k}{\Sigma }}}{\mathrm{The}}_e\left(\mathrm{not}\right)\mathrm{and}{T}_s\left(\mathit{k\; \text{'}}\right)={\displaystyle \underset{\mathrm{not}=1}{\overset{\mathit{k\; \text{'}}}{\Sigma }}}{\mathrm{The}}_s\left(\mathrm{not}\right)$$

Whenever we have T _e (k) = T _s (k '), the common origin is updated (and taken at the position k × L _e or else at k' × L _s ). As for the two sums T _e and T _s , they are preferentially reset.

• a20) définition d'une origine commune 0 aux sous-trames des premier et second formats (étape 70),
• a21) découpage des sous-trames successives du premier format de codage caractérisé par une première durée de sous-trame, pour former des pseudo-sous-trames de durées L'_e correspondantes à la seconde durée de sous-trame (étape 71),
• a22) mise à jour de l'origine commune O (étape 79),
• a23) et détermination de correspondance entre les positions d'impulsions dans les pseudo-sous-trames p'_e et dans les sous-trames au second format (étape 80).

In short and more generally, by calling the first (respectively second) subframe duration the subframe duration of the first (respectively second) encoding format, the adaptation steps performed when the subframe times are different are summarized on the figure 7 and are preferentially the following:

• a20) defining a common origin 0 to the subframes of the first and second formats (step 70),
• a21) cutting successive subframes of the first coding format characterized by a first subframe duration, to form pseudo-subframes of durations L ' _e corresponding to the second subframe duration (step 71),
• a22) updating the common origin O (step 79),
• a23) and determining correspondence between the pulse positions in the pseudo-subframes p ' _e and in the subframes in the second format (step 80).

• les première et seconde durées sont fixes dans le temps (sortie "o" du test 72), et
• les première et seconde durées varient dans le temps (sortie "n" du test 72).

Preferably, to determine the common origin O, the test 72 is discriminated against. figure 7 the following cases:

• the first and second durations are fixed in time (output "o" of the test 72), and
• the first and second durations vary in time (output "n" of the test 72).

In the first case, the temporal position of the common origin is periodically updated (step 74) at each instant when respective subframe boundaries of first duration St (L _e ) and second duration St (L _s ). are aligned in time (test 73 performed on these borders).

• a221) on effectue successivement les deux sommations respectives des sous-trames au premier format T_e(k) et des sous-trames au second format T_s(k') (étape 76),
• a222) on détecte une occurrence d'une égalité entre lesdites deux sommes, définissant un instant de remise à jour de ladite origine commune (test 77),
• a223) on réinitialise les deux sommes précitées (étape 78), après ladite occurrence, pour une future détection d'une prochaine origine commune.
  Maintenant, dans le cas où les durées de sous-trame et les fréquences d'échantillonnage sont différentes, il suffit de combiner judicieusement les algorithmes de correspondance entre les positions de E et S décrites dans les deux cas précédents.

In the second case, preferentially:

• a221) the respective two summations of the subframes in the first format T _e (k) and the subframes in the second format T _s (k ') are carried out successively (step 76),
• a222) detecting an occurrence of an equality between said two sums, defining a time of updating of said common origin (test 77),
• a223) the two aforementioned amounts (step 78) are reset, after said occurrence, for a future detection of a next common origin.
  Now, in the case where the subframe times and the sampling frequencies are different, it is sufficient to judiciously combine the correspondence algorithms between the positions of E and S described in the two previous cases.

* EXAMPLES OF REALIZATION

Three embodiments of transcoding within the meaning of the invention will now be described. These exemplary embodiments describe the implementation of the treatments provided in the cases presented above in standard speech analysis coders. The first two modes illustrate the favorable case where the sampling frequencies, like the durations of the subframes, are identical. The last example illustrates the case where the duration of the sub-frames are different.

* Example of realization n ° 1

The first embodiment is applied to intelligent transcoding between the MP-MLQ model of G.723.1 at 6.3 kbit / s and the ACELP model at 4 pulses of G.723.1 at 5.3 kbit / s.

Intelligent transcoding of the G.723.1's high bitrate to low bit rate brings together a 6 and 5 pulse MP-MLQ model with an ACELP model at 4 pulses. The exemplary embodiment presented here makes it possible to determine the positions of the four pulses of the ACELP from the positions of the pulses of the MP-MLQ.

The operation of the G.723.1 encoder is described below.

The ITU-T G.723.1 multi-rate encoder and its multi-pulse directories have been presented above. It is specified only that a frame of G.723.1 has 240 samples at 8 kHz, and is divided into four subframes of 60 samples. The same restriction is imposed on the pulse positions of any vector-code of each of the three multi-pulse dictionaries. These positions must all have the same parity (all pairs or all odd). The sub-frame of 60 (+4) positions is thus cut into two gates of 32 positions. The even grid has the positions numbered [0, 2, 4, ..., 58, (60,62)]. The odd grid has the positions [1, 3, 5, ..., 59, (61,63)]. For each flow, the exploration of the directory, even if not exhaustive, remains complex as indicated previously.

The selection of a subset of the 5.3 kbps G.723.1 ACELP directory from an item in a 6.3 kbps G.723.1 MP-MLQ directory is now described.

The innovation signal of a sub-frame is modeled by an element of the 5.3 kbps G.723.1 ACELP directory that knows the 6.3 kbps G.723.1 MP-MLQ directory element. / s determined during a first coding. There are therefore Ne positions (Ne = 5 or 6) pulses chosen by the G.723.1 coder at 6.3 kbit / s.

For example, it can be assumed that the positions extracted from the bit stream of the G.723.1 encoder at 6.3kbit / s for a subframe whose excitation is modeled by Ne = 5 pulses are: $${\mathrm{e}}_{\mathrm{0}}\mathrm{=}\mathrm{0}\mathrm{;}{\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{1}}\mathrm{=}\mathrm{8}\mathrm{;}{\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{2}}\mathrm{=}\mathrm{28}\mathrm{;}{\mathrm{<figure-callout\; id="3"\; label="e"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{3}}\mathrm{=}\mathrm{38}\mathrm{;}{\mathrm{<figure-callout\; id="4"\; label="e"\; filenames="imgf0004.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{4}}\mathrm{=}\mathrm{46}\mathrm{;}$$

It is recalled that no adaptation of sampling frequencies or subframe times is to be done here. After this step of recovering the positions e _i , a next step then consists in directly extracting the right and left neighborhoods of these 5 pulses. The right and left neighborhoods are taken here equal to 2. The set P _s of the selected positions is: $$P_s=\left\{-2,-1,0,1,2\right\}\cup \left\{6,,7,,8,,9,,10\right\}\cup \left\{26,,27,,28,,29,,30\right\}\cup \left\{36,,37,,38,,39,,40\right\}\cup \left\{44,,45,,46,,47,,48\right\}$$

The third step consists of composing the restricted set of possible positions for each pulse (here a track) of the G.723.1 ACELP directory at 5.3 kbit / s by taking N _s = 4 intersections of P _s with the 4 sets of positions of the odd (odd) tracks allowed by the latter directory (as shown in Table 1).

$$S_3=P_s\cap \left\{6,,14,22,\cdots ,54,\left(62\right)\right\};$$

d'où : $$S_0=\left\{0,,8,,40,,48\right\};S_1=\left\{2,,10,,26\right\};S_2=\left\{28,,36,,44\right\};S_3=\left\{6,,30,,38,,46\right\};$$

For the even parity: $${\mathrm{<figure-callout\; id="0"\; label="S"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">S</figure-callout>}}_0=P_s\cap \left\{0,,8,,16,...,56\right\};{\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">S</figure-callout>}}_1=P_s\cap \left\{2,,10,,18,...,58\right\};{\mathrm{<figure-callout\; id="2"\; label="S"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">S</figure-callout>}}_2=P_s\cap \left\{4,,12,,20,\cdots ,52,\left(60\right)\right\};$$

$${\mathrm{<figure-callout\; id="3"\; label="S"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">S</figure-callout>}}_3=P_s\cap \left\{6,,14,22,\cdots ,54,\left(62\right)\right\};$$

from where : $${\mathrm{<figure-callout\; id="0"\; label="S"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">S</figure-callout>}}_0=\left\{0,,8,,40,,48\right\};{\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">S</figure-callout>}}_1=\left\{2,,10,,26\right\};{\mathrm{<figure-callout\; id="2"\; label="S"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">S</figure-callout>}}_2=\left\{28,,36,,44\right\};{\mathrm{<figure-callout\; id="3"\; label="S"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">S</figure-callout>}}_3=\left\{6,,30,,38,,46\right\};$$

$$S_3=P_s\cap \left\{7,,15,22,\cdots ,55,\left(63\right)\right\};$$

d'où : $$S_0=\left\{1,,9\right\};S_1=\left\{27\right\};S_2=\left\{29,,37,,45\right\};S_3=\left\{7,,39,,47\right\};$$

For odd parity: $${\mathrm{<figure-callout\; id="0"\; label="S"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">S</figure-callout>}}_0=P_s\cap \left\{1,9,,\cdots ,57\right\};{\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">S</figure-callout>}}_1=P_s\cap \left\{3,,11,\cdots ,59\right\};{\mathrm{<figure-callout\; id="2"\; label="S"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">S</figure-callout>}}_2=P_s\cap \left\{5,,13,\cdots ,53,\left(61\right)\right\};$$

$${\mathrm{<figure-callout\; id="3"\; label="S"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">S</figure-callout>}}_3=P_s\cap \left\{7,,15,22,\cdots ,55,\left(63\right)\right\};$$

from where : $${\mathrm{<figure-callout\; id="0"\; label="S"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">S</figure-callout>}}_0=\left\{1,,9\right\};{\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">S</figure-callout>}}_1=\left\{27\right\};{\mathrm{<figure-callout\; id="2"\; label="S"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">S</figure-callout>}}_2=\left\{29,,37,,45\right\};{\mathrm{<figure-callout\; id="3"\; label="S"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">S</figure-callout>}}_3=\left\{7,,39,,47\right\};$$

The combination of these selected positions constitutes the new restricted directory in which the search will take place. For this last step, the procedure for selecting the optimal positions set is based on the criterion CELP as the G.723.1 does in the 5.3 kbit / s mode, the exploration can be exhaustive or, preferably, focused.

The number of combinations of positions in the restricted repertoire is equal to 180 (= 4 * 3 * 3 * 4 + 2 * 1 * 3 * 3) instead of 8192 (= 2 * 8 * 8 * 8 * 8) combinations of ACELP directory positions of G.723.1 at 5.3 kbit / s.

It is indicated that the number of combinations can be further restricted by considering only the parity chosen in the 6.3 kbit / s mode (in the example cited the even parity). In this case, the number of combinations in the restricted directory is 144.

It can happen (depending on the size of the considered neighborhoods) that for one of the four pulses, the set P _s contains no position for a track of the model ACELP (case where one of the set Si is empty). Thus, for neighborhoods of size 2, when the positions of the Ne pulses are all on the same track, P _s contains only positions of this track and adjacent tracks. In this case, depending on the desired quality / complexity compromise, it is possible either to replace the set S _i by T _i (which amounts to not restricting all the positions of this track), or to increase the neighborhood right (or left) pulses. For example, if all pulses in the 6.3 kbit / s code are on track 2, with right and left neighborhoods equal to 2, track 0 will have no positions regardless of parity. It is then sufficient to increase by 2 the size of the left and / or right neighborhood to assign positions to this track 0.

To illustrate this realization, we start from the following example: $${\mathrm{<figure-callout\; id="0"\; label="e"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{0}}\mathrm{=}\mathrm{4}\mathrm{;}{\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{1}}\mathrm{=}\mathrm{12}\mathrm{;}{\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{2}}\mathrm{=}\mathrm{20}\mathrm{;}{\mathrm{<figure-callout\; id="3"\; label="e"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{3}}\mathrm{=}\mathrm{36}\mathrm{;}{\mathrm{<figure-callout\; id="4"\; label="e"\; filenames="imgf0004.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{4}}\mathrm{=}\mathrm{52}\mathrm{;}$$

The set P _s of the selected positions is: $$P_s=\left\{2,,3,,4,,5,,6\right\}\cup \left\{10,,11,,12,,13,,14\right\}\cup \left\{18,,19,,20,,21,,22\right\}\cup \left\{34,,35,,36,,37,,38\right\}\cup \left\{50,,51,,52,,53,,54\right\}$$

Assuming that we want to keep the same parity, the initial distribution of these positions for the 4 pulses is: $${\mathrm{<figure-callout\; id="0"\; label="S"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathit{0}}=Ø;{\mathit{<figure-callout\; id="1"\; label="S"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathit{1}}=\left\{2,,10,,18,,34,,50\right\};{\mathit{<figure-callout\; id="2"\; label="S"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathit{2}}=\left\{4,,12,,20,,36,,52\right\};{\mathit{<figure-callout\; id="3"\; label="S"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathit{3}}=\left\{6,,14,,22,,38,,54\right\};$$

$${\mathit{S}}_{\mathit{2}}=\left\{4,,12,,20,,36,,52\right\};{\mathit{S}}_{\mathit{3}}=\left\{6,,14,,22,,38,,54\right\}$$

(avec donc S ₀ ≠ ⌀).By increasing the left neighborhood of the pulses by 2, we obtain: $${\mathit{<figure-callout\; id="0"\; label="S"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathit{0}}=\left\{0,,8,,16,,32,,48\right\};{\mathit{<figure-callout\; id="1"\; label="S"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathit{1}}=\left\{2,,10,,18,,34,,50\right\};$$

$${\mathit{<figure-callout\; id="2"\; label="S"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathit{2}}=\left\{4,,12,,20,,36,,52\right\};{\mathit{<figure-callout\; id="3"\; label="S"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathit{3}}=\left\{6,,14,,22,,38,,54\right\}$$

(so with S ₀ ≠ ⌀).

* Execution example n ° 2

The following second example illustrates the application of the invention to intelligent transcoding between ACELP models of the same length. In particular, this second exemplary embodiment is applied to intelligent transcoding between the 4-pulse ACELP of the G.729 at 8 kbit / s and the ACELP at 2 pulses of the G.729 at 6.4 bit / s.

Intelligent transcoding between the 6.4 kbit / s and 8 kbit / s modes of the G.729 encoder brings together a two-pulse and a four-pulse ACELP repertoire. The example presented here makes it possible to determine the positions of 4 pulses (8 kbit / s) from the positions of 2 pulses (6.4 kbit / s) and vice versa.

Briefly, the operation of the ITU-T G.729 coder is described. This encoder can operate at three rates: 6.4; 8 and 11.8 kbit / s. We therefore consider here the first two flows. A G.729 frame has 80 samples at 8 kHz. This frame is divided into two subframes of 40 samples. For each subframe, the G.729 models the innovation signal by pulses according to the ACELP model. It uses four in the 8 kbit / s mode and two in the 6.4 kbit / s mode. Tables 2 and 4 above give the positions that pulses can take for these two rates. At 6.4 kbit / s, a exhaustive search of all combinations (512) positions is performed. At 8 kbit / s, a focused search is preferably used.

General treatment within the meaning of the invention is still used here. However, and advantageously, here we take advantage of the ACELP structure common to both directories. The mapping of the sets of positions thus exploits a subframe cut of 40 samples in 5 tracks of 8 positions, given in table 6 below. <i> Table 6: Distribution of positions in five tracks in ACELP </ i> dictionaries of <i> G.729 </ i> Tracks positions P ₀ 0 5 10 15 20 25 30, 35 P ₁ 1 6 11 16 21 26 31 36 P ₂ 2 7 12 17 22 27 32 37 P ₃ 3 8 13 18 23 28 33 38 P ₄ 4 9 14 19 24, 29 34, 39

In both directories, the pulse positions share these tracks, as shown in Table 7 below.

All pulses are characterized by their track and their rank in this track. The 8 kbit / s 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 kbit / s mode puts its first pulse on the tracks P ₁ or P ₃ , and its second pulse on the tracks P ₀ , P ₁ , P ₂ or P ₄ . <i> Table 7: Distribution of pulses from G.729 ACELP directories at 8 and 6.4 kbit </ i> / <i> s in the five tracks. </ i> Fashion pulses Tracks 6.4 kbit / s i ₀ P ₁ , P ₃ i ₁ P ₀ , P ₁ , P ₂ , P ₄ 8 kbit / s i ₀ P ₀ i ₁ P ₁ i ₂ P ₂ i ₃ P ₃ , P ₄

The interlacing of tracks (ISSP structure) is used in this embodiment to facilitate the extraction of neighborhoods and the composition of restricted subsets of positions. Thus, to move from one track to another, simply move one unit to the right or to the left. For example, if you are at the 5 ^th position of track 2 (of absolute position 22), a shift of 1 to the right (+1) moves to the 5 ^th position of track 3 (of position absolute 23) and a shift to the left (-1) moves to the 5 ^th position of track 1 (absolute position 21).

• voisinage droit : P_i ⇒ P _(i+d)≡5
• voisinage gauche : P_i ⇒ P _(i-d)≡5

Au niveau du rang m dans la piste :

• * voisinage droit : $${\mathit{P}}_{\mathit{i}}\Rightarrow {\mathit{P}}_{(\mathit{i}+\mathit{d})\equiv 5}$$
  
  $${\mathit{P}}_{\mathit{i}}\Rightarrow {\mathit{P}}_{(\mathit{i}\mathit{-}\mathit{d})\equiv 5}$$
  
• * voisinage gauche : $$\mathrm{si}(\mathrm{i}-\mathrm{d})\ge 0:{\mathit{m}}_{\mathit{i}}\Rightarrow {\mathit{m}}_{\mathit{i}}$$
  
  $$\mathrm{sinon}:{\mathit{m}}_{\mathit{i}}\Rightarrow {\mathit{m}}_{\mathit{i}}-1$$
  

More generally, an offset of ± d of a position is reflected here by the following effects.
At the level of the tracks P _i :

• right neighbor: P _i ⇒ P _{( i + d ) ≡5}
• left neighborhood: P _i ⇒ P _{( id ) ≡ 5}

At the level of the row m in the track:

• * right neighbor: $${\mathit{P}}_{\mathit{i}}\Rightarrow {\mathit{P}}_{(\mathit{i}+\mathit{d})\equiv 5}$$
  
  $${\mathit{P}}_{\mathit{i}}\Rightarrow {\mathit{P}}_{(\mathit{i}\mathit{-}\mathit{d})\equiv 5}$$
  
• * left neighborhood: $$\mathrm{if}(\mathrm{i}-\mathrm{d})\ge 0:{\mathit{m}}_{\mathit{i}}\Rightarrow {\mathit{m}}_{\mathit{i}}$$
  
  $$\mathrm{if\; not}:{\mathit{m}}_{\mathit{i}}\Rightarrow {\mathit{m}}_{\mathit{i}}-1$$
  

The selection of a subset of the 4-pulse ACELP repertoire of the 8 kbit / s G.729 encoder is now described from a 2-pulse ACELP directory element of the G.729 to 6.4 encoder. kbit / s.

A subframe of G.729 is considered to be in the 6.4 kbit / s mode. Two pulses are placed by this encoder but it is necessary to determine the positions of the other pulses that must place the G.729 at 8 kbit / s. To radically restrict complexity, only one position per pulse is chosen and only one combination of positions is retained. Advantageously, the selection step is therefore immediate. Two of the four G.729 pulses at 8 kbit / s are selected at the same positions as the 6.4 kbit / s mode, and then the two remaining pulses are placed in close proximity to 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 9 bits) of the two positions, the two corresponding tracks are also determined. From these two tracks (possibly identical), the last three steps of neighborhood extraction, composition of restricted subsets and selection of a combination of pulses are then judiciously associated. There are several cases according to the tracks P _i ( i = 0 to 4) on which are located the two pulses 6.4 kbit / s mode.

The positions of the mode pulses is noted e _k 6.4 kbit / s and that of the S _k mode 8 kbit / s. Table 8 below presents for each case the positions chosen. The columns marked "P _{j + d} = P _i 'specify the neighborhood law at the slopes and leading to the track P _i. Recall that level P _i tracks.
* for the right neighborhood: P _i ⇒ P _{( i + d ) ≡5}
* for the left neighborhood: P _i ⇒ P _{( id )} ≡ 5 <i> Table 8: Choosing the G.729 8-Kit Restricted Directory </ i> / <i> s from the two pulses in the 6.4 kbit G.729 ACELP directory </ i> / <i > s. </ i> e ₀ (Track) e ₁ (Track) s ₀ s ₁ s ₂ s ₃ Position P _{l + d} = P ₀ Position P _{l + d} = P ₁ Position P _{l + d} = P ₂ Position P _{l + d} = P ₃ / P ₄ P ₁ e ₀ = e ₁ P ₁ e ₁ -1 P ₁ -1 E ₁ P ₁ e ₁ +1 P ₁ +1 e ₁ +2 P ₁ +2 e ₀ ≠ e ₁ e ₀ -1 P ₁ -1 E ₀ P ₁ e ₁ +1 P ₁ +1 e ₁ +2 P ₁ +2 P ₁ Po e ₁ Po eo P ₁ e ₀ +1 P ₁ +1 e ₁ -1 ⁽¹⁾ P ₀ ⁽¹⁾ -1 P ₁ P ₂ e ₀ -1 P ₁ -1 eo P ₁ e ₁ P ₂ e ₁ +1 P ₂ +1 P ₁ P ₄ e ₁ +1 ⁽²⁾ P ₄ ⁽²⁾ +1 E ₀ P ₁ e ₀ +1 P ₁ +1 e ₁ P ₄ P ₃ P ₀ e ₁ P ₀ E ₁ +1 P ₀ +1 e ₀ -1 P ₃ -1 e ₀ P ₃ P ₃ P ₁ e ₁ -1 P ₁ -1 E ₁ P ₁ e ₀ -1 P ₃ -1 e ₀ P ₃ P ₃ P ₂ e ₀ +2 ⁽³⁾ P ₃ ⁽³⁾ +2 E ₁ -1 P ₂ -1 e ₁ P ₂ e ₀ P ₃ P ₃ P ₄ e ₁ +1 ⁽⁴⁾ P ₄ ⁽⁴⁾ +1 E ₀ -2 P ₃ -2 e ₀ -1 P ₃ -1 e ₁ P ₄

• Cas (1) : si e₁ = 0, alors on ne peut pas prendre s₃=e₁-1. On choisira S₃= e₀+2.
• Cas (2) : si e₁ = 39, alors on ne peut pas prendre s₀=e₁+1. On choisit s₀= e₀ -1.
• Cas (3): si e₁ = 38, alors on ne peut pas prendre s₀=e₀+2. On choisit s₀= e₁ -2.
• Cas (4): si e₁ = 39, alors on ne peut pas prendre s₀=e₁+1. On choisit s₀= e₀ -3.

Preferentially, it is therefore sought to balance the distribution of the 4 positions with respect to the two starting positions, but it indicates that another choice can be made. Four cases (indicated by an exponent in parentheses in table 8) can, however, cause edge effects problems:

• Case (1): if e ₁ = 0, then we can not take s ₃ = e ₁ -1. We choose S ₃ = e ₀ + 2.
• Case (2): if e ₁ = 39, then we can not take s ₀ = e ₁ +1. We choose s ₀ = e ₀ -1.
• Case (3): if e ₁ = 38, then we can not take s ₀ = e ₀ +2. We choose s ₀ = e ₁ -2.
• Case (4): if e ₁ = 39, then we can not take s ₀ = e ₁ +1. We choose s ₀ = e ₀ -3.

To further reduce the complexity, the sign of each pulse s _k can be taken as equal to that of the pulse e _j from which it is deduced.

The selection of a subset of the ACELP directory at 2 pulses of the G.729 at 6.4 kbit / s from an element of a 4-pulse G.729 ACELP repertoire at 8 kbit / s is now described. s.

For a G.729 subframe in the 8 kbit / s mode, the first step is the recovery of the positions of the four pulses generated by the 8 kbit / s mode. The decoding of the bit index (over 13 bits) of the 4 positions makes it possible to obtain 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 as well. than his rank in this track. Each position e _i (0 i i <4) is characterized by the pair (p _i , m _i ) where p _i is the index of its track and m _i its rank in this track. We have: $${\mathrm{e}}_{\mathrm{i}}\mathrm{=}\mathrm{5}{\mathrm{m}}_{\mathrm{i}}\mathrm{+}{\mathrm{p}}_{\mathrm{i}}\mathrm{,}\mathrm{with}\mathrm{0}\mathrm{\le }{\mathrm{m}}_{\mathrm{i}}\mathrm{<}\mathrm{8}{\mathrm{and\; P}}_{\mathrm{i}}\mathrm{=}\mathrm{i\; for\; i}\mathrm{<}\mathrm{3}{\mathrm{and\; <figure-callout\; id="3"\; label="P"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">P</figure-callout>}}_{\mathrm{3}}\mathrm{=}\mathrm{3}\mathrm{or}\mathrm{4.}$$

As already mentioned, the extraction of neighborhoods and the composition of restricted subsets are combined and advantageously exploit the ISSP structure common to both directories. By exploiting the property of neighboring positions induced by the interlacing of the tracks, the five intersections T ' _j of the set P _s of neighborhoods of the 4 positions with the 5 tracks P _{j are constructed} . $${\mathrm{T\text{'}}}_{\mathrm{j}}\mathrm{=}{\mathrm{P}}_{\mathrm{s}}\mathrm{\cap }{\mathrm{P}}_{\mathrm{j}}$$

Thus, a right (respectively left) neighbor 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) , otherwise (case p = 4) at T ' ₀ provided that m <7 (respectively at T' ₄ (case i = 0) provided that m> 0). The restriction on the right neighbor for a position of the fourth pulse belonging to the fourth track (respectively left for a position of the first track) makes it possible to ensure that the neighboring position is not outside the sub-frame.

Thus, using the notation modulo 5 (≡5), a right (respectively left) neighbor of +1 (respectively -1) of the impulse (p, m) belongs to T ' _{(p + 1) ≡5} (respectively at T ' _{(p-1) ≡5} ). Remember that edge effects must be taken into account. By generalizing to a neighborhood size d, a right neighbor of + d (respectively left of -d) of the pulse (p, m) belongs to T ' _{(p + d) ≡5} (respectively to T' _{(pd) ≡5} ). The rank of the neighbor at ± d is equal to m if p + d≤4 (or pd≥0), otherwise the rank m is incremented for a right neighbor and decremented for a left neighbor. Taking into account edge effects therefore amounts to ensuring that: m <7 if p + d> 4 and that m> 0 if pd <0.
From this distribution of the neighbors in the tracks, it is easy to determine the restricted subsets S ₀ and S ₁ of the positions of the two pulses: $${\mathrm{<figure-callout\; id="0"\; label="S"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathrm{0}}\mathrm{=}{\mathrm{T\text{'}}}_{\mathrm{1}}\mathrm{\cup }{\mathrm{T\text{'}}}_{\mathrm{3}}{\mathrm{and\; <figure-callout\; id="1"\; label="S"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathrm{1}}\mathrm{=}{\mathrm{T\text{'}}}_{\mathrm{0}}\mathrm{\cup }{\mathrm{T\text{'}}}_{\mathrm{1}}\mathrm{\cup }{\mathrm{T\text{'}}}_{\mathrm{2}}\mathrm{\cup }{\mathrm{T\text{'}}}_{\mathrm{4}}$$

The fourth and last step is to perform the search for the optimal torque in the two subsets obtained. The search algorithm (like the normalized one using the track structure) and the track storage of the pulses further simplify the search algorithm. In practice, it is therefore unnecessary to explicitly form the restricted subsets S ₀ and S ₁ because sets T ' _j can be used alone.

In the following example, the 8 kbit / s mode of the G.729 has set its four pulses to the following positions: $${\mathrm{<figure-callout\; id="0"\; label="e"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{0}}\mathrm{=}\mathrm{5}\mathrm{;}{\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{1}}\mathrm{=}\mathrm{21}\mathrm{;}{\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{2}}\mathrm{=}\mathrm{22}\mathrm{;}{\mathrm{<figure-callout\; id="3"\; label="e"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathrm{3}}\mathrm{=}\mathrm{34}\mathrm{;}$$

• e₀: (0, 1) donne : (4,0) à gauche et (1, 1) à droite
• e₁ : (1,4) donne : (0,4) à gauche et (2,4) à droite
• e₂: (2,4) donne : (1,4) à gauche et (3,4) à droite
• e₃ :(4,6) donne : (3, 6) à gauche et (0,7) à droite

These 4 positions are characterized by the 4 pairs (p _i , m _i ) = (0,1), (1, 4), (2,4) (4,6).
If we take a fixed neighborhood equal to 1, we construct the 5 intersections T ' _j as follows:

• e ₀ : (0, 1) gives: (4.0) on the left and (1, 1) on the right
• e ₁ : (1,4) gives: (0,4) on the left and (2,4) on the right
• e ₂ : (2,4) gives: (1,4) on the left and (3,4) on the right
• e ₃ : (4,6) gives: (3, 6) on the left and (0,7) on the right

So we have: $${\mathrm{T\text{'}}}_0=\left\{\left(0,,1\right),,\left(0,,4\right),,\left(0,,7\right)\right\};{\mathrm{T\text{'}}}_1=\left\{\left(1,,4\right),\left(1,,1\right))\right\};{\mathrm{T\text{'}}}_2=\left\{\left(3,,4\right),,\left(3,,6\right)\right\};{\mathrm{T\text{'}}}_4=\left\{\left(4,,6\right),,\left(4,,0\right)\right\}$$

Returning to the notation by positions: $${\mathrm{T\text{'}}}_0=\left\{5,,20,,35\right\};{\mathrm{T\text{'}}}_1=\left\{21,6)\right\};{\mathrm{T\text{'}}}_2=\left\{22\right\};{\mathrm{T\text{'}}}_3=\left\{23,,33\right\};{\mathrm{T\text{'}}}_4=\left\{34,,4\right\}$$

In the last step, an algorithm similar to that of the 6.4 kbit / s G.729 performs the search for the best pulse pair. This algorithm is here less complex because the number of combinations of positions to explore is very small. In the example, there are only 4 (= Cardinal (T ' ₁ ) + Cardinal (T' ₃ )) times 8 (= Cardinal (T ' ₀ ) + Cardinal (T' ₁ ) + Cardinal (T ') ₂ ) + Cardinal (T ' ₄ )) combinations to be tested, ie 32 combinations instead of 512.

For a neighborhood of size 1, less than 8% of combinations of positions is to explore on average without exceeding 10% (50 combinations). For a neighborhood size 2, less than 17% of combinations of positions is to explore on average and at most 25% of the combinations is to explore. For a neighborhood size 2, the complexity of the treatment proposed in the invention (by adding the cost of the search in the directory restricted to the cost of the extraction of neighborhoods associated with the composition of the intersections) represents less than 30% of an exhaustive search for an equivalent quality.

* Execution example n ° 3

The last example illustrates the transitions between the 8 kbit / s G.729 ACELP model and the 6.3 kbit / s G.723.1 MP-MLQ model.

Intelligent pulse transcoding between G.723.1 (6.3 kbit / s mode) and G.729 (8 kbit / s mode) has two major difficulties. First, the size of the frames is different (40 samples for G.729 versus 60 samples for G.723.1). The second difficulty is related to the different structure of the dictionaries, the ACELP type for the G.729 and the MP-MLQ type for the G.723.1. The example of realization presented here shows how the invention overcomes these two difficulties in order to transcode the pulses at a lower cost while preserving the quality of the transcoding.

Firstly, a temporal correspondence is made between the positions in the two formats taking into account the difference in size of the subframes for aligning the positions relative to an origin common to E and S. The lengths of the subframes of the G: 729 and G.723.1 having as least common multiple 120, the temporal correspondence is performed in blocks of 120 samples or two subframes of G.723.1 for three subframes of G.729, as shown by the example of the figure 4b . Alternatively, it may be preferred to work on blocks of complete frames. In this case, blocks of 240 samples are chosen, ie a G.723.1 frame (4 subframes) for three G.729 frames (6 subframes).

The selection of a subset of the MP-MLQ directory of G.723.1 at 6.3 kbit / s is now described from elements of the 4-pulse G.729 ACELP directory at 8 kbit / s. The first step consists of recovering the positions of the pulses in blocks of 3 subframes (index i _e , 0 i i _e ≤ 2) of G.729. We denote p _e ( i _e ) a position of the subframe i _e of this block.

$${\mathrm{si}\mathit{i}}_{\mathit{e}}=2,{\mathrm{alors}\mathit{j}}_{\mathit{s}}=1{\mathrm{et}\mathit{p}}_{\mathit{s}}={\mathit{p}}_{\mathit{e}}+20$$

$${\mathrm{si}\mathit{i}}_{\mathit{e}}=1,{\mathrm{alors\; si\; p}}_{\mathit{e}}<20{\mathit{j}}_{\mathit{s}}=0{\mathrm{et}\mathit{p}}_{\mathit{s}}={\mathit{p}}_{\mathit{e}}+20+40,\mathrm{sinon}\left({\mathit{p}}_{\mathit{e}}\ge 20\right):{\mathit{j}}_{\mathit{s}}=1$$

et $${\mathit{p}}_{\mathit{s}}\approx {\mathit{p}}_{\mathit{e}}-20$$

Before extracting the neighborhoods, these 12 positions p _e ( i _e ) are converted into 12 positions denoted p _s ( j _s ), distributed in two subframes (of index j _s , 0 ≤ j _s ≤ ₁ ) of G.723.1. We can use the general relation above (involving the modulo of the subframe length) to realize the adaptation of the subframe durations. However, it is preferred here to distinguish three cases simply according to the index i _e : $${\mathrm{if}\mathit{i}}_{\mathit{e}}=0,{\mathrm{so}\mathit{j}}_{\mathit{s}}=0{\mathrm{and}\mathit{p}}_{\mathit{s}}={\mathit{p}}_{\mathit{e}}$$

$${\mathrm{if}\mathit{i}}_{\mathit{e}}=2,{\mathrm{so}\mathit{j}}_{\mathit{s}}=1{\mathrm{and}\mathit{p}}_{\mathit{s}}={\mathit{p}}_{\mathit{e}}+20$$

$${\mathrm{if}\mathit{i}}_{\mathit{e}}=1,{\mathrm{so\; if\; p}}_{\mathit{e}}<20{\mathit{j}}_{\mathit{s}}=0{\mathrm{and}\mathit{p}}_{\mathit{s}}={\mathit{p}}_{\mathit{e}}+20+40,\mathrm{if\; not}\left({\mathit{p}}_{\mathit{e}}\ge 20\right):{\mathit{j}}_{\mathit{s}}=1$$

and $${\mathit{p}}_{\mathit{s}}\approx {\mathit{p}}_{\mathit{e}}-20$$

This is neither division nor modulo operation.

The 4 positions retrieved in the block sub-frame STE0 are directly assigned to the sub-frame STS0 with the same position, those of the block sub-frame STE2 are directly assigned to the sub-frame STS1 with an increment of +20 in position, the sub-frame positions STE1 less than 20 are assigned to the sub-frame STS0 with an increment of +40, the others being allocated to the sub-frame STS1 with a decrement of -20.

The neighborhoods are then extracted from these 12 positions. It should be noted that the right (respectively left) neighborhoods of the positions of the subframe STS0 (respectively STS1) can be allowed to leave their subframe, these neighboring positions then being in the subframe STS1 (respectively STSO ).

The temporal matching and neighborhood extraction step can be reversed. In this case, the right (respectively left) neighborhoods of the positions of the STE0 subframe (respectively STE2) can be allowed to leave their subframe, these neighboring positions then being in the subframe STE1. Similarly, the right (respectively left) neighborhoods of the positions in STE1 can lead to neighboring positions in STE2 (respectively STEO).

Once the set of restricted positions for each sub-frame STS constituted, the last step is to explore for each sub-frame STS its restricted directory thus formed to select the Np (6 or 5) pulses of the same parity. This procedure may be derived from the standard algorithm or may be inspired by other focusing procedures.

$$\mathit{STE}\mathit{1}\mathit{=}{\mathit{e}}_{\mathit{10}}\mathit{=}\mathit{15}\mathit{;}{\mathit{e}}_{\mathit{11}}\mathit{=}\mathit{31}\mathit{;}{\mathit{e}}_{\mathit{12}}\mathit{=}\mathit{22}\mathit{;}{\mathit{e}}_{\mathit{13}}\mathit{=}\mathit{4}\mathit{;}$$

$$\mathit{STE}\mathit{2}\mathit{=}{\mathit{e}}_{\mathit{20}}\mathit{=}\mathit{0}\mathit{;}{\mathit{e}}_{\mathit{21}}\mathit{=}\mathit{1}\mathit{;}{\mathit{e}}_{\mathit{22}}\mathit{=}\mathit{37}\mathit{;}{\mathit{e}}_{\mathit{23}}\mathit{=}\mathit{24},$$

après application de l'étape ci-avant de correspondance temporelle, l'attribution de ces 12 positions aux sous-trames STS0 et STS1 est : $$\mathit{STS}\mathit{0}\mathit{:}{\mathit{s}}_{\mathit{00}}\mathit{=}\mathit{5}\mathit{;}{\mathit{s}}_{\mathit{01}}\mathit{=}\mathit{1}\mathit{;}{\mathit{s}}_{\mathit{02}}\mathit{=}\mathit{32}\mathit{;}{\mathit{s}}_{\mathit{03}}\mathit{=}\mathit{39}\mathit{;}\left({\mathit{s}}_{\mathit{0}\mathit{k}}\mathit{=}{\mathit{e}}_{\mathit{0}\mathit{k}}\right)$$

$$\mathit{STS}\mathit{0}\mathit{=}{{\mathit{s}}^{\mathit{\prime }}}_{\mathit{10}}\mathit{=}\mathit{55}\mathit{;}{{\mathit{s}}^{\mathit{\prime }}}_{\mathit{13}}\mathit{=}\mathit{44}\mathit{;}\left({{\mathit{s}}^{\mathit{\prime }}}_{\mathit{0}\mathit{k}}\mathit{=}{\mathit{e}}_{\mathit{1}\mathit{k}}\mathit{+}\mathit{40}\mathit{,}{\mathit{si\; e}}_{\mathit{1}\mathit{k}}\mathit{<}\mathit{20}\right)$$

$$\mathit{STS}\mathit{1}\mathit{=}{{\mathit{s}}^{\mathit{\prime }}}_{\mathit{11}}\mathit{=}\mathit{11}\mathit{;}{{\mathit{s}}^{\mathit{\prime }}}_{\mathit{12}}\mathit{=}\mathit{2}\mathit{;}\left({{\mathit{s}}^{\mathit{\prime }}}_{\mathit{1}\mathit{k}}\mathit{=}{\mathit{e}}_{\mathit{1}\mathit{k}}\mathit{-}\mathit{20}\mathit{,}{\mathit{si\; e}}_{\mathit{1}\mathit{k}}\mathit{\ge }\mathit{20}\right)$$

$$\mathit{StS}\mathit{1}\mathit{:}{\mathit{s}}_{\mathit{20}}\mathit{=}\mathit{20}\mathit{;}{\mathit{s}}_{\mathit{21}}\mathit{=}\mathit{21}\mathit{;}{\mathit{s}}_{\mathit{22}}\mathit{=}\mathit{57}\mathit{;}{\mathit{s}}_{\mathit{23}}\mathit{=}\mathit{44}\mathit{;}\left({\mathit{s}}_{\mathit{0}\mathit{k}}\mathit{=}{\mathit{e}}_{\mathit{2}\mathit{k}}\mathit{+}\mathit{20}\right)$$

To illustrate this exemplary embodiment, three subframes of G.729 are considered which will make it possible to build the sub-directories of two subframes of G.723.1. Assuming that the G.729 gives the following positions: $$\mathit{ETS}\mathit{0}\mathit{=}{\mathit{e}}_{\mathit{00}}\mathit{=}\mathit{5}\mathit{;}{\mathit{e}}_{\mathit{01}}\mathit{=}\mathit{1}\mathit{;}{\mathit{e}}_{\mathit{02}}\mathit{=}\mathit{32}\mathit{;}{\mathit{e}}_{\mathit{03}}\mathit{=}\mathit{39}\mathit{;}$$

$$\mathit{ETS}\mathit{1}\mathit{=}{\mathit{<figure-callout\; id="10"\; label="e"\; filenames="imgf0001.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{10}}\mathit{=}\mathit{15}\mathit{;}{\mathit{<figure-callout\; id="11"\; label="e"\; filenames="imgf0001.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{11}}\mathit{=}\mathit{31}\mathit{;}{\mathit{<figure-callout\; id="12"\; label="e"\; filenames="imgf0001.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{12}}\mathit{=}\mathit{22}\mathit{;}{\mathit{e}}_{\mathit{13}}\mathit{=}\mathit{4}\mathit{;}$$

$$\mathit{ETS}\mathit{2}\mathit{=}{\mathit{e}}_{\mathit{20}}\mathit{=}\mathit{0}\mathit{;}{\mathit{<figure-callout\; id="21"\; label="e"\; filenames="imgf0002.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{21}}\mathit{=}\mathit{1}\mathit{;}{\mathit{<figure-callout\; id="22"\; label="e"\; filenames="imgf0002.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{22}}\mathit{=}\mathit{37}\mathit{;}{\mathit{<figure-callout\; id="23"\; label="e"\; filenames="imgf0002.png,imgf0003.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{23}}\mathit{=}\mathit{24},$$

after applying the above step of temporal correspondence, the allocation of these 12 positions to the subframes STS0 and STS1 is: $$\mathit{STS}\mathit{0}\mathit{:}{\mathit{s}}_{\mathit{00}}\mathit{=}\mathit{5}\mathit{;}{\mathit{s}}_{\mathit{01}}\mathit{=}\mathit{1}\mathit{;}{\mathit{s}}_{\mathit{02}}\mathit{=}\mathit{32}\mathit{;}{\mathit{s}}_{\mathit{03}}\mathit{=}\mathit{39}\mathit{;}\left({\mathit{s}}_{\mathit{0}\mathit{k}}\mathit{=}{\mathit{<figure-callout\; id="0"\; label="e"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{0}\mathit{k}}\right)$$

$$\mathit{STS}\mathit{0}\mathit{=}{{\mathit{s}}^{\mathit{\text{'}}}}_{\mathit{10}}\mathit{=}\mathit{55}\mathit{;}{{\mathit{s}}^{\mathit{\text{'}}}}_{\mathit{13}}\mathit{=}\mathit{44}\mathit{;}\left({{\mathit{s}}^{\mathit{\text{'}}}}_{\mathit{0}\mathit{k}}\mathit{=}{\mathit{<figure-callout\; id="1"\; label="e"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{1}\mathit{k}}\mathit{+}\mathit{40}\mathit{,}{\mathit{if\; <figure-callout\; id="1"\; label="e"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{1}\mathit{k}}\mathit{<}\mathit{20}\right)$$

$$\mathit{STS}\mathit{1}\mathit{=}{{\mathit{s}}^{\mathit{\text{'}}}}_{\mathit{11}}\mathit{=}\mathit{11}\mathit{;}{{\mathit{s}}^{\mathit{\text{'}}}}_{\mathit{12}}\mathit{=}\mathit{2}\mathit{;}\left({{\mathit{s}}^{\mathit{\text{'}}}}_{\mathit{1}\mathit{k}}\mathit{=}{\mathit{<figure-callout\; id="1"\; label="e"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{1}\mathit{k}}\mathit{-}\mathit{20}\mathit{,}{\mathit{if\; <figure-callout\; id="1"\; label="e"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{1}\mathit{k}}\mathit{\ge }\mathit{20}\right)$$

$$\mathit{StS}\mathit{1}\mathit{:}{\mathit{s}}_{\mathit{20}}\mathit{=}\mathit{20}\mathit{;}{\mathit{s}}_{\mathit{21}}\mathit{=}\mathit{21}\mathit{;}{\mathit{s}}_{\mathit{22}}\mathit{=}\mathit{57}\mathit{;}{\mathit{s}}_{\mathit{23}}\mathit{=}\mathit{44}\mathit{;}\left({\mathit{s}}_{\mathit{0}\mathit{k}}\mathit{=}{\mathit{<figure-callout\; id="2"\; label="e"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">e</figure-callout>}}_{\mathit{2}\mathit{k}}\mathit{+}\mathit{20}\right)$$

Thus we have the sets of positions {1, 5, 32, 39, 44, 55} for subframe STS0 and {2, 11, 20, 21, 44, 57} for subframe STS1.

$$P_{s1}=\left\{1,,2,,3\right\}\cup \left\{10,,11,,12\right\}\cup \left\{20,,21,,22\right\}\cup \left\{21,,22,,23\right\}\cup \left\{43,,44,,45\right\}\cup \left\{56,,57,,58\right\}$$

We must now extract the neighborhoods. By taking for example a neighborhood fixed at 1, we obtain: $$P_{s0}=\left\{0,,1,,2\right\}\cup \left\{4,,5,,6\right\}\cup \left\{31,,32,,33\right\}\cup \left\{43,,39,,40\right\}\cup \left\{43,,44,,45\right\}\cup \left\{54,,55,,56\right\}$$

$$P_{s1}=\left\{1,,2,,3\right\}\cup \left\{10,,11,,12\right\}\cup \left\{20,,21,,22\right\}\cup \left\{21,,22,,23\right\}\cup \left\{43,,44,,45\right\}\cup \left\{56,,57,,58\right\}$$

• P_s0 : {0,2,4,6,32,40,44,54,56} et {1,5,31,33,39,43,45,55}
• P_s1 : {2,10,12,20,22,44,56} et {1,3,11.21,23,43,45,57}

The MP-MLQ does not impose any constraints on the impulses, apart from their parity. On a sub-frame, they must all be of the same parity. We must here split P _s0 and P _s1 into two subsets, with:

• P _s0 : {0,2,4,6,32,40,44,54,56} and {1,5,31,33,39,43,45,55}
• P _s1 : {2,10,12,20,22,44,56} and {1,3,11.21,23,43,45,57}

This subdirectory is finally transmitted to the selection algorithm which determines the N _p best positions in the sense of the CELP criterion for the subframes STS0 and STS1 of G.723.1. This considerably reduces the number of combinations tested. Indeed, for example in the sub-frame STS0, there remain 9 even positions and 8 odd positions instead of 30 and 30.

• augmenter la taille du voisinage pour les sous-trames concernées jusqu'à obtenir une taille suffisante pour P_s (taille≥N_p ),
• ou sélectionner les N premières impulsions et autoriser pour les N_p-N impulsions restantes une recherche parmi les 30-N positions restantes de la grille, comme décrit ci-avant.

However, certain precautions are taken in cases where the positions chosen by G.729 are such that the extraction of neighborhoods gives a number N of possible positions less than the number of positions of G.723.1 (N <N _p ). This is particularly the case when the G.729 positions follow each other (eg {0,1,2,3}). Two possibilities are then foreseen:

• increasing the size of the neighborhood for the subframes concerned until a sufficient size is obtained for P _s (size ≥N _p ),
• or select the first N pulses and allow for the remaining N _p -N pulses a search among the remaining 30-N positions of the grid, as described above.

The inverse process of selecting a subset of the ACELP repertoire of four 8 kbit / s G.729 pulses from elements in a 6.3 kbit G.723.1 MP-MLQ directory is now described. s.

Overall, the treatment is similar. Two subframes of G.723.1 correspond to 3 frames of G.729. Here too, we extract the positions of G.723.1 that we translate in the time scale of G.729. We can advantageously translate these positions in the form "track - rank in the track" to take advantage of the structure ACELP as previously to extract neighborhoods and search for optimal positions.

The same arrangements as previously are planned to prevent cases where the extraction of neighborhoods would give positions in insufficient number (here less than 4 positions).

Thus, the present invention makes it possible to determine at a lower cost the positions of a set of pulses from a first set of pulses, the two sets of pulses belonging to two multi-pulse repertoires. These two directories can be distinguished by their size, the length and the number of pulses of their codewords as well as by the rules governing the positions and / or amplitudes of the pulses. We favor the neighborhoods of the positions of the impulses of the (or) set (s) chosen (s) in the first directory to determine those of a game in the second directory. The invention also makes it possible to exploit the structure of the starting and / or arrival directories in order to further reduce the complexity. Through the first example above showing the transition from an MP-MLQ model to an ACELP model, it will be understood that the invention is easily applicable to two multi-pulse models having different structural constraints. Through the second exemplary embodiment having the passage between two models having a different number of pulses but based on the same ACELP structure, it will be understood that the invention advantageously allows the exploitation of the directory structure to reduce the complexity of transcoding. Through the third example showing the passage between an MP-MLQ model and an ACELP model, it will be understood that the invention can be applied even for coders of different subframe lengths or sampling frequencies. The invention makes it possible to adjust the quality / complexity compromise and, in particular, to greatly reduce the computation complexity for a minimal degradation compared to a conventional search for a multi-pulse model.

Claims (23)

1. Method of transcoding between a first compression coder/decoder and at least one second compression coder/decoder, said first and second coders/decoders being of pulse type and using multipulse dictionaries in which each pulse has a position marked by an associated index, characterized in that it comprises the following steps:

   a) where appropriate, adapting coding parameters between said first and second coders/decoders;

   b) obtaining from the first coder/decoder a selected number (N_e) of pulse positions and respective position indices (e_i) 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 (T_j) accepted by the second coder/decoder at least some of the pulse positions in a set (P_s) constituted by a union of said groups formed in step c); and

   e) sending the thus selected pulse positions to the second coder/decoder for coding/decoding from said positions sent,
   said selection of step d) then involving a number of possible pulse positions less than the total number of possible pulse positions in the dictionary of the second coder/decoder.
2. Method according to Claim 1, in which the first coder/decoder (E) uses a first number of pulses in a first coding format, characterized in that said number (N_e) selected in step b) corresponds to said first number of pulse positions.
3. Method according to Claim 2, in which:

   - the first coder/decoder (E) uses a first number (N_e) of pulse positions in a first coding format;

   - the second coder/decoder (S) uses a second number (N_s) of pulse positions in a second coding format,

   characterized in that the method further includes a step of distinguishing at least the cases in which:

   - the first number (N_e) is greater than or equal to the second number (N_s) ;

   - the first number (N_e) is less than the second number (N_s).
4. Method according to Claim 3, in which the first number (N_e) is greater than or equal to the second number (N_s) (N_e≥N_s), characterized in that each group formed in step c) includes right-hand neighbour pulse positions (vⁱ _r) and left-hand neighbour pulse positions (vⁱ ₁) of said current pulse position of given index, and in that the respective numbers of left-hand and right-hand neighbour pulse positions are selected as a function of a complexity/transcoding quality trade-off.
5. Method according to Claim 4, characterized in that there is constructed in step d) a subdirectory of combinations of pulse positions resulting from intersections (S_j) of:

   - the set (P_s) constituted by a union of said groups formed in step c); and

   - pulse positions (T_j) accepted by the second coder/decoder,
   so that said subdirectory has a size less than the number of combinations of pulse positions (T_j) accepted by the second coder/decoder.
6. Method according to Claim 5, characterized in that, after step e), said subdirectory is searched for an optimum set of positions including said second number (N_s) of positions at the level of the second coder (S) .
7. Method according to Claim 6, characterized in that the step of searching for the optimum set of positions is effected by means of a focused search to accelerate the exploration of said subdirectory.
8. Method according to one of the preceding claims, in which said first coder/decoder is designed to deliver a succession of coded frames, characterized in that the respective numbers of pulse positions in the groups formed in step c) are selected successively from one frame to the other.
9. Method according to Claim 3, in which the first number (N_e) is less than the second number (N_s)(N_e<N_s),
   characterized in that a further test is effected to determine if the pulse positions provided in the second number of pulse positions (N_s) are included in the pulse positions of the groups formed in step c), and, in the event of a negative result to said test, the number of pulse positions in the groups formed in step c) is increased.
10. Method according to Claim 3, characterized in that it further distinguishes the case in which the second number N_s is between the first number N_e and twice the first number N_e (N_e<N_s<2N_e), and if so:

    c1) the N_e pulse positions are selected from the outset; and

    c2) there is further selected a complementary number of pulse positions N_s-N_e defined in the immediate neighbourhood of the pulse positions selected in step c1).
11. Method according to one of the preceding claims, in which said first coder/decoder is designed to operate with a given first sampling frequency and from a given first subframe duration, characterized in that said coding parameters for which said adaptation is carried out in step a) include at least a subframe duration and a sampling frequency, whereas the second coder/decoder operates with a second sampling frequency and a second subframe duration, and in that the following four cases are distinguished in step a):

    - the first and second durations are equal and the first and second frequencies are equal;

    - the first and second durations are equal and the first and second frequencies are different;

    - the first and second durations are different and the first and second frequencies are equal;

    - the first and second durations are different and the first and second frequencies are different.
12. Method according to Claim 11, in which the first and second durations are equal and the first and second sampling frequencies are different, characterized in that the method includes steps of:

    a1) direct time scale quantization from the first frequency to the second frequency; and

    a2) determination, as a function of said quantization, of each pulse position in a subframe with the second coding format characterized by the second sampling frequency, from a pulse position in a subframe with the first coding format characterized by the first sampling frequency.
13. Method according to Claim 12, characterized in that the quantization step a1) is effected by calculation and/or tabulation on the basis of a function which at a pulse position in a subframe with the first format (p_e) establishes the correspondence of a pulse position in a subframe with the second format (p_s), said function substantially taking the form of a linear combination involving a multiplier coefficient corresponding to the ratio of the second sampling frequency to the first sampling frequency.
14. Method according to Claim 13, characterized in that, to pass conversely from a pulse position in a subframe with the second format (p_s) to a pulse position in a subframe with the first format (p_e), there is applied a function inverse to said linear combination applied to a pulse position in a subframe with the second format (p_s).
15. Method according to Claim 11, in which the first and second durations are equal and the first and second sampling frequencies are different, characterized in that the method comprises the following steps:

    a'1) oversampling a subframe with the first coding format characterized by the first sampling frequency, at a frequency equal to the lowest common multiple of the first and second sampling frequencies; and

    a'2) applying to the oversampled subframe low-pass filtering followed by undersampling to attain a sampling frequency corresponding to the second sampling frequency.
16. Method according to Claim 15, characterized in that the method continues by obtaining, by thresholding, a number of positions which can be variable where appropriate.
17. Method according to Claim 12, characterized in that it further includes a step of establishing the correspondence for each position (p_e) of a pulse of a subframe with the first coding format characterized by the first sampling frequency of a group of pulse positions (p_s) in a subframe with the second coding format characterized by the second sampling frequency, each group including a number of positions that is a function of the ratio between the second sampling frequency and the first sampling frequency (F_s/F_e).
18. Method according to Claim 11, in which the first and second subframe durations are different, characterized in that it includes the steps of:

    a20) defining an origin common (O) to the subframes of the first and second formats;

    a21) dividing successive subframes of the first coding format characterized by a first subframe duration, to form pseudosubframes of durations corresponding to the subframe duration of the second format;

    a22) updating said common origin; and

    a23) determining the correspondence between the pulse positions in the pseudosubframes and in the subframes with the second format.
19. Method according to Claim 18, characterized in that it also distinguishes the following cases:

    - the first and second durations are fixed in time; and

    - the first and second durations vary in time.
20. Method according to Claim 19, in which the first and second durations are fixed in time, characterized in that the temporal position of said common origin is periodically updated whenever boundaries of respective subframes of first and second duration are aligned in time.
21. Method according to Claim 19, in which the first and second durations vary in time, characterized in that:

    a221) two respective summations of the durations of subframes with the first format and the durations of subframes with the second format are effected successively;

    a222) an occurrence of equality between said two sums is detected, defining a time of updating said common origin; and

    a223) said two sums are reset, after said occurrence, for future detection of a next common origin.
22. Software product intended to be stored in a memory of a processing unit, in particular a computer or a mobile terminal, or on a removable memory medium intended to cooperate with a reader of the processing unit, characterized in that it includes instructions for implementing the method of transcoding according to one of the preceding claims.
23. Device for transcoding between a first compression coder/decoder and at least one second compression coder/decoder, said first and second coders/decoders being of pulse type and using multipulse dictionaries in which each pulse has a position marked by an associated index, characterized in that it comprises a memory comprising the instructions of a software product according to Claim 22.

Images