EP1989705B1 - Method for limiting adaptive excitation gain in an audio decoder - Google Patents

Abstract

The invention concerns a decoder for an audio signal coded by an encoder comprising a long-term predictive filter. According to the invention, said decoder comprises: a block (211) for detecting losses of transmission frames, a module (222) for calculating values of an error indicating function, representing the accumulated error in decoding on the adaptive excitation following said transmission frame loss, an arbitrary value being assigned to said adaptive excitation for the lost frame, a module (213) for calculating an error indicating parameter based on said values of the error indicating function, a comparator (214) of said error indicating parameter with at least one given threshold, a discriminator (215) for determining based on of the result provided by the comparator (214) a value of at least one adaptive excitation gain to be used by the decoder. The invention is applicable to encoding and decoding digital signals such as audiofrequency signals.

Description

The present invention relates to a method of limiting adaptive excitation gain in a decoder of an audio signal. It also relates to a decoder of an audio signal encoded by means of an encoder comprising a long-term predictive filter.

The invention finds an advantageous application in the field of coding and decoding of digital signals such as audio-frequency signals.

The invention is particularly well suited to the transmission of speech and / or audio signals over packet networks, of the VoIP type, for example, to provide an acceptable quality during decoding after a loss of packets, in particular avoiding the saturation of the data. Long Term Prediction (LTP) filters used for decoding in the Code Exciting Linear Prediction (CELP) context.

An example of a CELP encoder is the G.729 system recommended in the ITU-T, designed for voiceband speech between 300 and 3400 Hz sampled at 8 kHz and transmitted at a fixed rate of 8 kbit / s with frames of 10 ms. The detailed operation of this encoder is specified in the article of A: Salami, C. Laflamme, JP Adoul, A. Kataoka, S. Hayashi, T. Moriya, C. Lamblin, D. Massaloux, S. Proust, P. Kroon and Y. Shoham. "Design and description of CS-ACELP: a toll quality 8 kb / s speech coder", IEEE Trans. on Speech and Audio Processing, Vol.6-2, March 1998, PP.116-130 .

On the Figure 1 (a) is represented a high level view of a G.729 encoder. This figure shows a high-pass filtering 101 of pre-processing for removing frequency signals below 50 Hz. The speech signal S (n) thus filtered is then analyzed by the block 102 to determine a filter ( z) Linear Prediction Coding (LPC), which is transmitted to the multiplexer 104 in the form of an index indexing the quantized vector (QV) in a dictionary.

The original signal S (n) filtered by the filter λ (z) , called excitation, is processed by the block 103 in order to extract the parameters mentioned in the table of the figure 2 . These parameters are then coded and then transmitted to the MUX multiplexer 104.

• dans une première étape, un filtrage de prédiction à long terme (LTP) est effectué par les blocs 106, 107, 110 Le filtre LTP du codeur G.729 est un filtre d'ordre égal à 1. La période P d'excitation adaptative, ou période de « pitch », exprimée en valeur entière P₀ complétée éventuellement par une valeur fractionnaire P₀ _fractionnaire, ainsi que le gain g_p d'excitation adaptative, ou gain de « pitch », sont déterminés par analyse par synthèse de façon à minimiser l'erreur entre le signal d'excitation cible issu du bloc 105 et le signal synthétisé donné par x(n) = g_p.x(n-P), n représentant un échantillon du signal,
• puis, dans une deuxième étape, la différence résiduelle entre ces deux signaux est modélisée, d'une part, par un code fixe c(n), ou code innovateur, extrait d'un dictionnaire innovateur ACELP 108 à quatre impulsions ± 1, et, d'autre part, par un gain g_c d'excitation fixe 109. Le code fixe c(n) et la gain g_c sont déterminés en minimisant en 111' l'erreur entre le signal résiduel issu de l'étage de LTP précédent et le signal g_c.c(n),
• enfin, dans une dernière étape, les paramètres résultant, à savoir la période P de pitch, le code fixe c(n) et les gains g_p de pitch et g_c d'excitation fixe, sont codés et transmis au multiplexeur 104.

The operation of the excitation coding block 103 is detailed in FIG. Figure 1 (b) . As can be seen in this figure, the excitation is coded in three steps:

• in a first step, a long-term prediction filtering (LTP) is performed by the blocks 106, 107, 110 The LTP filter of the G.729 encoder is a filter of order equal to 1. The period P of adaptive excitation , or "pitch" period, expressed as an integer value P ₀ supplemented optionally by a fractional fraction P ₀ -fractional value, as well as the gain g _p of adaptive excitation, or "pitch" gain, are determined by synthesis analysis of minimizing the error between the target excitation signal from block 105 and the synthesized signal given by x (n) = g _p .x (nP) , n representing a sample of the signal,
• then, in a second step, the residual difference between these two signals is modeled, on the one hand, by a fixed code c (n) , or innovative code, extracted from an innovatory dictionary ACELP 108 with four pulses ± 1, and on the other hand, by a fixed excitation gain g _c 109. The fixed code c (n) and the gain g _c are determined by minimizing the error between the residual signal from the LTP stage at 111 '. previous and the signal g _c .c (n),
• finally, in a last step, the resulting parameters, namely the pitch period P , the fixed code c (n) and the gains g _p of pitch and g _c of fixed excitation, are coded and transmitted to the multiplexer 104.

• une première contribution résultant du décodage 115 de la période P de pitch et du décodage 118 du gain g_p de pitch pour reconstituer en sortie des blocs 116, 117 le signal LTP d'excitation adaptative x(n) = g_p.x(n-P),
• une deuxième contribution résultant du décodage 113 de l'excitation fixe c(n) mise à l'échelle par le gain g_e décodé par le bloc 118 pour reconstituer l'excitation fixe g_e .c(n).
• ces deux contributions sont ensuite additionnées pour fournir l'excitation décodée x(n) = g_p.x(n-P) + g_e.c(n).

The Figure 1 (c) shows how a conventional G.729 decoder reconstructs the speech signal from the data received from the multiplexer 104 by the demultiplexer 112. The excitation is reconstituted by subframes of 5 ms by adding two contributions:

• a first contribution resulting from the decoding 115 of the pitch period P and the decoding 118 of the gain g _p of pitch to reconstitute at the output of the blocks 116, 117 the adaptive excitation LTP signal x (n) = g _p .x (nP ) ,
• a second contribution resulting from the decoding 113 of the fixed excitation c (n) scaled by the gain g _e decoded by the block 118 to reconstitute the fixed excitation g _e . c (n) .
• these two contributions are then added to provide the decoded excitation x (n) = g _p .x (nP) + g _e .c (n).

The excitation thus decoded is shaped by the synthesis filter 120 LPC 1 / Â (z) whose coefficients are decoded by the block 119 in the domain of the spectral line pairs (LSF) and interpolated by subframe 5. ms. In order to improve the quality and to mask certain coding artifacts, the reconstructed signal is then processed by an adaptive post-filter 121 and a post-processing high-pass filter 122. The decoder of the Figure 1 (c) therefore relies on the source-filter model to synthesize the signal.

In the case of excitation from the long-term prediction LTP filter, and in order to generate an excitation signal capable of rapidly following the signal attacks, the CELP type encoders generally allow the choice of a gain g _p pitch greater than 1. As a result, the decoder is locally unstable. However, this instability is controlled by the synthesis analysis model which permanently minimizes the difference between the LTP excitation signal and the original target signal.

During transmission or frame loss errors, this instability can lead to significant degradation due to the offset between the encoder and the decoder. Indeed, in these circumstances, the gain value g _p of pitch not received in a frame is generally replaced by the value of g _p in the previous frame, and although the variable nature of the speech signal consists of an alternation periods of voices with a pitch gain close to 1 and unvoiced with a pitch gain of less than 1 allows, in general, to limit the potential problems related to this local instability, it remains nonetheless true that, for certain signals, especially the voiced signals, transmission errors in periodic stationary zones can cause significant damage when, for example, the gain g _p of replacement is higher than the actual gain and the affected frame is followed by high gain frames, as happens during attacks. This situation can then quickly lead to a saturation of the LTP filter by cumulative effect linked to the recursive character of long-term predictive filtering.

A first solution to this problem is to limit the gp pitch gain to 1, but this constraint has the effect of degrading the performance of the CELP coders for attacks.

• La méthode décrite dans le brevet américain n° 5,960,386 peut se décomposer en plusieurs étages situés au codeur. Tout d'abord, une procédure de détection d'une possible instabilité utilisant le gain de pitch préalablement calculé et une moyenne des gains de pitch précédents. Ensuite, dans le cas où il n'y a pas de risque d'instabilité, le gain de pitch préalablement calculé est conservé. Dans le cas contraire, une procédure itérative de contrôle du gain de pitch permet d'adapter ce gain pour éliminer le risque d'instabilité.
• Dans les brevets américains n° 5,893,060 et 5,987,406 , une procédure de détection des instabilités au codeur est décrite. Cette procédure utilise les paramètres spectraux LSP pour déterminer la présence de résonances dans le spectre, calcule la durée de la résonance en nombre de trames, et évalue la possible instabilité en fonction de la valeur du gain de pitch. Dans le cas où une instabilité est détectée, la valeur du gain de pitch est saturée à un seuil et la recherche du vecteur de gain dans la quantification vectorielle des gains de pitch est modifiée pour que le vecteur choisi ait une valeur de gain de pitch inférieure à ce seuil.
• Dans l'article de R. Salami précité et le brevet américain n° 5'708757 est décrite une procédure de détection d'une possible saturation et du calcul de la valeur de gain de pitch associée, présente au codeur dans la norme G729, est décrite. Cette méthode, appelée "taming", prend en compte l'erreur maximum potentielle commise par le décodeur sur le calcul de l'excitation. Si cette erreur dépasse un certain seuil quand le gain de pitch est supérieur à 1, correspondant à un filtre instable, le gain est modifié pour prendre une valeur inférieure à 1 afin de rendre le filtre stable. L'idée est donc de détecter au codeur des zones où l'accumulation des erreurs de transmission précédentes peut causer une saturation du filtre à long terme localement instable, notamment lors de longues zones fortement voisées. Ces zones sont détectées en examinant la sortie d'un deuxième filtrage à long terme avec une excitation constante qui simule l'erreur maximum potentielle. Une technique identique est utilisée dans la norme ITU-T G.723.1. Ce codeur utilise un prédicteur à long terme d'ordre 5 pour lequel le gain de pitch est un vecteur de 5 coefficients appliqués sur 5 échantillons consécutifs du passé. Ces vecteurs de gain sont quantifiés par quantification vectorielle. Alors que la stabilité d'un filtre à long terme d'ordre 1, comme celui du codeur G.729, se vérifie très facilement en comparant le seul coefficient de gain avec la valeur 1, cette vérification est beaucoup plus compliquée pour un filtre à long terme d'ordre supérieur. En effet, la stabilité d'un filtre à long terme utilisant un jeu de gain dépend également de la nature du signal, par exemple du pitch. Donc, le même jeu de gain peut être stable dans une situation et instable dans une autre. C'est pourquoi il est difficile d'estimer la propagation d'une erreur, car la nature d'erreur potentielle ne peut pas être connue au codeur, et il n'est pas simple de détecter les zones potentiellement instables ni de déterminer l'atténuation à appliquer pour rétablir la stabilité du filtre. La solution mise en oeuvre dans la norme G.723.1 est de trouver, par apprentissage, pour chaque vecteur de gain possible du codeur un gain moyen équivalent d'ordre 1. Ces valeurs sont stockées dans un tableau. On utilise donc ce filtre équivalent d'ordre 1 pour estimer l'erreur maximum potentielle accumulée dans le filtre à long terme, et ainsi identifier les zones instables où il faut limiter le gain en cas d'une erreur accumulée importante et calculer le gain à appliquer pour rendre le filtre stable.

Other solutions propose to limit the gain g _p of pitch to a value less than or equal to 1 only when it is deemed necessary. In particular:

• The method described in US Patent No. 5,960,386 can be broken down into several stages located at the encoder. First, a procedure for detecting a possible instability using the previously calculated pitch gain and an average of the previous pitch gains. Then, in the case where there is no risk of instability, the previously calculated pitch gain is retained. In the opposite case, an iterative procedure for controlling the pitch gain makes it possible to adapt this gain to eliminate the risk of instability.
• In US Patents No. 5893060 and 5987406 , a procedure for detecting instabilities at the encoder is described. This procedure uses the spectral parameters LSP to determine the presence of resonances in the spectrum, calculates the duration of the resonance in number of frames, and evaluates the possible instability as a function of the value of the pitch gain. In the case where an instability is detected, the value of the pitch gain is saturated at a threshold and the search of the gain vector in the vector quantization of the pitch gains is modified so that the selected vector has a lower pitch gain value. at this threshold.
• In the article by R. Salami supra and US Patent No. 5'708757 A procedure for detecting a possible saturation and calculating the associated pitch gain value, present at the encoder in the G729 standard, is described. This method, called "taming", takes into account the maximum potential error committed by the decoder on the computation of the excitation. If this error exceeds a certain threshold when the pitch gain is greater than 1, corresponding to an unstable filter, the gain is modified to take a value less than 1 in order to make the filter stable. The idea is therefore to detect at the encoder areas where the accumulation of previous transmission errors can cause saturation of the long-term filter locally unstable, especially during long, highly voiced areas. These areas are detected by examining the output of a second long-term filter with constant excitation that simulates the potential maximum error. An identical technique is used in the ITU-T G.723.1 standard. This coder uses a long-term predictor of order 5 for which the pitch gain is a vector of 5 coefficients applied to 5 consecutive samples of the past. These gain vectors are quantified by vector quantization. While the stability of a first-order long-term filter, like that of the G.729 encoder, is very easy to verify by comparing the single gain coefficient with the value 1, this check is much more complicated for a filter with long-term higher order. Indeed, the stability of a long-term filter using a gain game also depends on the nature of the signal, for example the pitch. So the same winning game can be stable in one situation and unstable in another. This is why it is difficult to estimate the propagation of an error, because the nature of potential error can not be known to the coder, and it is not easy to detect potentially unstable areas or to determine the attenuation to apply to restore the stability of the filter. The solution implemented in the G.723.1 standard is to find, by learning, for each vector of possible gain of the encoder an equivalent average gain of order 1. These values are stored in a table. This equivalent filter of order 1 is then used to estimate the maximum potential error accumulated in the long-term filter, and thus to identify the unstable zones where the gain must be limited in the event of a large accumulated error and to calculate the gain at apply to make the filter stable.

• La décision de modifier le gain g_p associé à la prédiction à long terme étant réalisée au codeur a priori, il n'est pas possible de contrôler complètement l'état du décodeur et son comportement après une perte de trame, lesquels sont par hypothèse ignorés du codeur. Aussi, les techniques existantes peuvent continuer à générer des dégradations audio au décodage lors d'erreurs de transmission, ceci malgré la décision prise par le codeur de modifier le gain.
• La limitation à 1 du gain g_p de pitch associée aux techniques décrites plus haut peut entraîner une légère dégradation de la qualité par exemple sur les attaques qui génèrent normalement des gains supérieurs à 1. Le choix du seuil de déclenchement est en effet un compromis entre qualité et sécurité. Un seuil bas déclencherait la limitation trop souvent, entraînant une dégradation inutile, surtout en cas d'absence d'erreurs de transmission. Inversement, un seuil plus élevé ne garantirait pas une protection suffisante en cas de taux d'erreur élevés.

However, the solutions proposed by these known techniques to avoid the risk of saturation LTP filters in case of loss or transmission errors pose the following problems:

• Since the decision to modify the gain g _p associated with the long-term prediction is carried out at the coder a priori, it is not possible to control completely the state of the decoder and its behavior after a loss of frame, which are supposedly ignored by the encoder. Also, existing techniques can continue to generate audio degradations during decoding during transmission errors, despite the decision made by the coder to change the gain.
• The limitation to 1 of the gain g _p of pitch associated with the techniques described above can lead to a slight deterioration of the quality, for example on the attacks which normally generate gains greater than 1. The choice of the triggering threshold is indeed a compromise between quality and safety. A low threshold would trigger the limitation too often, causing unnecessary degradation, especially in the absence of transmission errors. Conversely, a higher threshold would not guarantee sufficient protection in case of high error rates.

Also, the technical problem to be solved by the object of the present invention is to propose a method for limiting adaptive excitation gain in a decoder of an audio signal encoded by means of an encoder comprising a long-term predictive filter. , following a loss of transmission frame between said encoder and said decoder, which would limit the adaptive excitation g _p gain, or pitch gain, only in the case where an instability LTP filter is actually noted, and to ensure the best possible compromise between the quality of the decoding and its robustness vis-à-vis the frame loss.

The solution to the technical problem posed consists, in the present invention, in that said method comprises the steps corresponding to claim 1.

In general, the term "frame loss" is used here to mean the non-reception of a frame as well as the transmission errors in a frame.

According to one embodiment, said arbitrary value is equal to a value of the adaptive excitation gain determined during said lost frame by an error concealment algorithm.

As an example of an error concealment algorithm, said arbitrary value is equal to the value of the adaptive excitation gain for the non-lost frame preceding said lost frame.

In another example, said arbitrary value is defined from a voicing detection of the previous frame. For a voiced frame, said arbitrary value is equal to 1, otherwise the arbitrary value is equal to 0. In the latter case, the excitation is composed of a random noise.

As will be seen in detail below, the method according to the invention has the advantage of modifying the gain g _p of pitch only when a possible instability of the LTP filter is detected at the decoder itself and not at the encoder as in known techniques. In addition, the method of the invention takes into account both the actual state of the decoder and the exact information on the transmission errors reached.

The method, object of the invention, can be used autonomously, that is to say in coding structures that do not provide for limiting the pitch gain at the coder.

However, and advantageously, the invention provides that said adaptive excitation gain is supplied to said decoder by an encoder equipped with a gain limitation device. The method according to the invention can therefore also be used in combination with a known taming technique, installed at the encoder. The advantages of the two techniques are then cumulated: the prior technique makes it possible to limit the long sequences of pitch gains greater than 1. Indeed, such sequences cause a large propagation of the error, forcing the method of the invention to modify the signal over long periods. However, a too low trigger threshold of the "taming" technique a priori degrades the signal. The invention thus makes it possible to reduce the number of triggers of the "taming" technique a priori by increasing the threshold, because even if this technique does not detect the risk of explosion, the posterior method according to the invention detects it and cure it.

où:

• N est l'ordre du filtre prédictif à long terme, généralement impair,
• les gains g_it sont égaux aux gains d'excitation adaptative g_i dudit filtre prédictif à long terme pour les trames reçues ou aux gains d'excitation adaptative g _{i_FEC} (FEC pour « Frame Erasure Concealment») dudit filtre prédictif à long terme dans la trame précédente pour les trames perdues,
• e_t(n) vaut 0 pour les trames reçues et 1 pour les trames perdues.
• P est la période d'excitation adaptative.

According to the invention, said error indication function is of the form: $${\mathit{x}}_t\left(\mathit{not}\right)\mathit{=}{\mathit{e}}_t\left(\mathit{not}\right)+{\displaystyle \underset{i}{\Sigma }}{\mathit{boy\; Wut}}_{\mathit{it}}\mathrm{.}{\mathit{x}}_{\mathit{t}}\left(\mathit{not}\mathit{-}\mathit{P}\mathit{+}\mathit{i}\right)\mathit{i}\mathit{\in }\left[\mathit{-}\left(\mathit{NOT}\mathit{-}\mathit{1}\right)\mathit{/}\mathit{2}\mathit{,}\left(\mathit{NOT}\mathit{-}\mathit{1}\right)\mathit{/}\mathit{2}\right]$$

or:

• N is the order of the long-term, generally odd, predictive filter
• the gains g _it are equal to the adaptive excitation gains g _{i of} said long-term predictive filter for the received frames or to the adaptive excitation gains g _{i _ FEC} (FEC for "Frame Erasure Concealment") of said long-term predictive filter in the previous frame for lost frames,
• e _t (n) is 0 for received frames and 1 for lost frames.
• P is the period of adaptive excitation.

Of course, in the simplest case, the order N of the LTP filter can be taken as 1.

In a first embodiment of the method according to the invention, the gain g _p of adaptive excitation of a long-term predictive filter of order 1 is limited to the value 1 if said parameter of indication of error is greater than said given threshold.

Likewise, the invention provides that a correction factor is applied to the gains g _i of adaptive excitation of a long-term predictive filter of order greater than 1 if said error indication parameter is greater than said given threshold. .

In a second embodiment, said at least one adaptive excitation gain is limited by a linear function of said given threshold if said error indication parameter is greater than said threshold. This provision advantage makes it possible to make the gain limitation more progressive and to avoid a sudden threshold effect.

The invention also relates to a program comprising instructions recorded on a computer-readable medium for carrying out the steps of the method according to the invention, when said program is executed on a computer.

The invention finally relates to a decoder of an audio signal encoded by means of an encoder comprising the features of claim 11.

• La figure 1 (a) est un schéma de haut niveau d'un codeur G.729.
• La figure 1 (b) est un schéma détaillé du bloc de codage de l'excitation du codeur de la figure 1 (a).
• La figure 1 (c) est un schéma du décodeur associé au codeur de la figure 1 (a).
• La figure 2 est un tableau donnant les divers paramètres de codage du codeur de la figure 1 (a)
• La figure 3 est un schéma d'un décodeur conforme à l'invention.

The following description with reference to the accompanying drawings, given by way of non-limiting examples, will make it clear what the invention consists of and how it can be achieved.

• The Figure 1 (a) is a high-level schema of a G.729 encoder.
• The Figure 1 (b) is a detailed diagram of the coding block of the encoder excitation of the Figure 1 (a) .
• The Figure 1 (c) is a diagram of the decoder associated with the coder of the Figure 1 (a) .
• The figure 2 is a table giving the various coding parameters of the coder of the Figure 1 (a)
• The figure 3 is a diagram of a decoder according to the invention.

The invention will now be described in detail in the context of a G.729 decoder and long-term prediction filtering LTP of order N = 1. The case of an LTP filter of any order N will be treated at the end of the present description.

où :

• g_p est le gain de l'excitation adaptative ou gain de pitch,
• P est la valeur du pitch ou longueur de la période. Le codeur G.729 utilise une résolution fractionnelle par pas de 1/3 pour les petites valeurs de pitch (P < 85) pour mieux modéliser les sons voisés aigus. L'excitation adaptative avec un pitch fractionnel est obtenue par interpolation avec sur-échantillonnage,
• g_c est le gain de l'excitation fixe,
• c(n) est le mot de code fixe, ou innovateur.

It is recalled that the excitation signal x _c (n) from block 103 coding the excitation of the Figure 1 (a) and explained to the Figure 1 (b) is the sum of the adaptive excitation g _p .x _c (nP) and the fixed excitation g _p .c (n) : $${\mathit{x}}_{\mathit{e}}\left(\mathit{not}\right)\mathit{=}{\mathit{boy\; Wut}}_{\mathit{p}}\mathit{,}{\mathit{x}}_{\mathit{e}}\left(\mathit{not}\mathit{-}\mathit{P}\right)\mathit{+}{\mathit{boy\; Wut}}_{\mathit{vs}}\mathit{.}\mathit{vs}\left(\mathit{not}\right)$$

or :

• g _p is the gain of the adaptive excitation or pitch gain,
• P is the value of the pitch or length of the period. The G.729 encoder uses fractional resolution in 1/3 increments for small pitch values ( P <85) to better model high-pitched voices. Adaptive excitation with a fractional pitch is obtained by interpolation with oversampling,
• g _c is the gain of the fixed excitation,
• c (n) is the fixed code word, or innovator.

Adaptive excitation depends solely on the past excitation and makes it possible to efficiently model the periodic signals, especially voiced signals, where the excitation itself is repeated almost periodically. The fixed part c (n) brings the innovation in the total excitation to model the difference between the periods, that is to say to correct the error between the adaptive excitation and the prediction residue.

As seen above, this excitation signal is optimized to the encoder using the technique of synthesis analysis. The synthesis filtering of this excitation is thus performed with the quantized filter to check the result that will be obtained at the decoder. This explains why it is possible to use a locally unstable long-term filtering, that is with a value of g _p greater than 1, to model a signal attack because the increase in energy due to this instability is controlled. On the other hand, this control is disturbed by the possible frame losses.

At the decoder, in the case of a lost frame, or erroneous, the error concealment algorithm uses an estimated excitation signal from the past excitation signal. Typically, we only reuse the filtering long-term LTP by keeping the last value of the pitch correctly decoded

. A disturbance is thus injected into the decoder excitation signal, denoted x _d (n). For the following valid frames, even if it is possible to correctly decode all the excitation parameters g _p , _p , g _c and c (n) , the excitation obtained will not be exact because the excitation is past x _d (nP) is disturbed. The error injected during the lost frame can therefore propagate later on many frames because of the recursion of the long-term filtering in the voiced periods, in particular when g _p is close to 1. By cons, when g _p has a low or zero value during several voiceless zones, the effect of the perturbation weakens or vanishes because the weight of the innovating code c (n) is greater than the weight of the past.

It is therefore essential to be able to estimate the magnitude of the error accumulated in the adaptive part, due to transmission errors. For this purpose, it is proposed to amend the figure 3 the decoder shown on the Figure 1 (c) .

We can see on the figure 3 that, in parallel with the long-term filtering LTP, the decoder comprises a processing line of the excitation signal from the demultiplexer 112 constituted by the blocks 211 to 215. This processing line of the decoder thus described also serves to illustrate the main steps of the adaptive excitation gain limiting method according to the invention.

où e_t(n) est égal à :

• 1 pour les trames non reçues ou erronées afin de modéliser l'erreur injectée dans la boucle adaptative,
• 0 pour les trames valides, quand l'erreur se propage uniquement à cause de la récursivité du filtre à long terme.
  g_t est égal à :
• g_{p_FEC} , valeur du gain de pitch de la trame précédente pour les trames non reçues,
• g_p pour les trames valides.

The block 211 is intended to detect whether a frame is correctly received or not. This detection block is followed by a module 212 which performs a similar operation to LTP long-term filtering. More precisely, the module 212 calculates an error indication function x _t (n) whose values are representative of the accumulated error on decoding on the adaptive excitation as a result of a loss of transmission. In one embodiment, this function is given by: $${\mathit{x}}_{\mathit{t}}\left(\mathit{not}\right)\mathit{=}{\mathit{boy\; Wut}}_{\mathit{t}}\mathit{.}{\mathit{x}}_{\mathit{t}}\left(\mathit{not}\mathit{-}\mathit{p}\right)\mathit{+}{\mathit{e}}_{\mathit{t}}\left(\mathit{not}\right)$$

where e _t (n) is equal to:

• 1 for frames not received or erroneous in order to model the error injected into the adaptive loop,
• 0 for valid frames, when the error propagates only because of the long-term filter recursion.
  g _t is equal to:
• g _{p_FEC} , value of the pitch gain of the previous frame for the frames not received,
• g _p for valid frames.

Then, a module 213 calculates from the values of the function x _t (n) provided by the module 212, an error indication parameter S _t . For a valid frame, a comparator 214 checks whether the parameter S _t does not exceed a certain threshold S ₀ . In case of overshoot and if the decoded pitch gain g _p is greater than 1, the value of g _p is limited, because in this case there is a risk of saturation of the LTP filter.

The error indication parameter S _t may be the sum of the values of the function x _t (n) , or the maximum value, the average or the sum of the squares of these values.

The comparator 214 is followed by a discriminator 215 able to determine the value g ' _t of the pitch gain to be applied to the block 117 for the current frame, namely the decoded pitch value g _p or a limited value.

S étant un coefficient arbitraire permettant d'ajuster la pente de la variation de g'_t avec S_t .In the case where the parameter S _t exceeds the threshold S ₀ and the decoded pitch gain g _p is greater than 1, the gain g ' _t can be systematically limited to 1 for example, regardless of the magnitude of the overshoot. But we can also provide a more progressive limitation which consists in defining the gain g ' _t as a linear function of the parameter S _t of the form: $${\mathrm{boy\; Wut}}_t={\mathrm{boy\; Wut}}_{\mathrm{p}}+\left({\mathrm{boy\; Wut}}_p-1\right)\left({\mathrm{S}}_{\mathit{0}}-{\mathrm{S}}_t\right)/\mathrm{S}$$

S being an arbitrary coefficient for adjusting the slope of the variation of g ' _t with S _t .

It is also possible to provide a limitation of the gain with respect to two successive thresholds, with a linear limitation between the two thresholds and a limitation to 1 beyond the second, as illustrated in the following example.

As a practical example, for a valid frame, the LTP, P and g _p parameters are transmitted for each 5 ms subframe containing 40 samples. The treatment to avoid saturation of the LTP filter, which is the subject of the invention, is also carried out at the rate of the subframes. The error indication parameter S _t , for example the sum of the function x _t (n) , is calculated for each subframe. The value of this parameter is limited to 120, which corresponds to an average value of 3: $$\mathit{St}\mathit{=}\mathit{min}\left({\displaystyle \underset{\mathit{i}\mathit{=}\mathit{0}}{\overset{\mathit{39}}{\mathit{\Sigma }}}}\mathit{xt}\left(\mathit{not}\right)\mathit{,}\mathit{120}\right)$$

If the pitch gain of the current subframe is greater than 1 and the value of S _t is greater than a threshold of 80, corresponding to an average value of the samples x _t (n) greater than 2, which shows that the Cumulative error is important, we reduce the value of the pitch gain according to the following equation: $${\mathit{boy\; Wut}}_{\mathit{t}}\mathit{=}\mathit{1}\mathit{+}\left({\mathit{boy\; Wut}}_{\mathit{t}}\mathit{-}\mathit{1}\right)\mathit{.}\left(\mathit{120}\mathit{-}{\mathit{S}}_{\mathit{t}}\right)\mathit{/}\mathit{40}$$

For the maximum value of S _t ( S _t = 120) the new pitch gain will be g ' _t = 1, for the other values of S _t 80 < S _t <120, 1>g' _t > g _t .

When the value of the pitch gain is modified by the method described above, the memory of the signal x _t (n) is updated with the new value g ' _t .

On the contrary, if the pitch gain of the current sub-frame is less than 1 or the value of S _t is less than 80, corresponding to a cumulative error in the long-term weak synthesis filter, the value is not modified. the decoded pitch gain and g ' _t = g _t .

Finally, to generate the excitation of the synthesis filter, instead of the decoded pitch gain, we use g ' _t : $${\mathit{x}}_{\mathit{d}}\left(\mathit{not}\right)\mathit{=}{\mathit{boy\; Wut}}_{\mathit{t}}\mathit{.}{\mathit{x}}_{\mathit{d}}\left(\mathit{not}\mathit{-}\mathit{P}\right)\mathit{+}{\mathit{boy\; Wut}}_{\mathit{vs}}\left(\mathit{not}\right)\mathit{.}\mathit{vs}\left(\mathit{not}\right)$$

In the exemplary embodiment presented here, the long-term filter of the encoder is a filter of order 1. However, if the encoder uses a long-term filter LTP of higher order N, as for example for the encoder G. 723.1, the pseudo-LTP filter used to define the error indication function may be the equivalent filter of order 1 or more advantageously a filter identical to that used in the encoder, in particular of the same order. To identify the unstable areas where the gain must be limited in the event of a large cumulative error during the valid frames and to determine the necessary attenuation, the equivalent filter of order 1 is always used.

In the case where the parameter S _t exceeds the threshold S ₀ and if the equivalent gain g _c is greater than 1, the gain g ' _t can be calculated in the same way as for a filter of order 1. It then applies the corrective factor g ' _{t /} g _e to the gains g _i of the higher order filter.

Claims (11)

1. Method for limiting the adaptive excitation gain in a decoder of an audio signal coded by a coder comprising a long-term predictive filter, following transmission frame loss between said coder and said decoder, characterized in that said method comprises the steps consisting, in the decoder, in:

   - detecting whether a frame is correctly received or not,

   - establishing an error indication function of the form: $${\mathit{x}}_t\left(\mathit{n}\right)\mathit{=}{\mathit{e}}_t\left(\mathit{n}\right)+{\displaystyle \sum _i}{\mathit{g}}_{\mathit{it}}\mathrm{.}{\mathit{x}}_{\mathit{t}}\left(\mathit{n}\mathit{-}\mathit{P}\mathit{+}\mathit{i}\right)\mathit{i}\mathit{\in }\left[\mathit{-}\left(\mathit{N}\mathit{-}\mathit{1}\right)\mathit{/}\mathit{2}\mathit{,}\left(\mathit{N}\mathit{-}\mathit{1}\right)\mathit{/}\mathit{2}\right]$$

   

   
   in which:

   - N is the order of the long-term predictive filter,

   - the gains g_it are equal to the adaptive excitation gains of said long-term predictive filter for the frames received or to the adaptive excitation gains of said long-term predictive filter in the preceding frame for the frames lost,

   - e_t(n) has the value 0 for the frames received and 1 for the frames lost.

   - P is the adaptive excitation period decoded for the frames received or most recently correctly decoded for the frames lost,

   - calculating, during the decoding, the values of said error indication function,

   - calculating an error indication parameter on the basis of said values of the error indication function,

   - comparing said error indication parameter to at least one given threshold,

   - applying a limitation to at least one adaptive excitation gain in case of a positive comparison if a gain equivalent to said at least one adaptive excitation gain is greater than a given value.
2. Method according to Claim 1, characterized in that said equivalent gain is the adaptive excitation gain g_p of a 1st order long-term predictive filter.
3. Method according to Claim 1, characterized in that said equivalent gain is the equivalent gain g_e of a long-term predictive filter of an order greater than 1.
4. Method according to any one of Claims 1 to 3, characterized in that said error indication parameter is a parameter representative of the energy of said error indication function.
5. Method according to Claim 4, characterized in that said representative parameter is given by the sum of the values of the error indication function.
6. Method according to any one of Claims 1 to 5, characterized in that the adaptive excitation gain g_p of a first order long-term predictive filter is limited to the value 1 if said error indication parameter is greater than said given threshold.
7. Method according to any one of Claims 1 to 5, characterized in that a corrective factor is applied to the adaptive excitation gains g_i of a long-term predictive filter of an order greater than 1 if said error indication parameter is greater than said given threshold.
8. Method according to any one of Claims 1 to 5, characterized in that said at least one adaptive excitation gain is limited by a linear function of said given threshold if said error indication parameter is greater than said threshold.
9. Method according to any one of Claims 1 to 8, characterized in that said adaptive excitation gain is supplied to said decoder by a coder equipped with a gain limiting device.
10. Program comprising instructions stored on a computer-readable medium for implementing the steps of the method according to Claims 1 to 9, when said program is executed on a computer.
11. Decoder of an audio signal coded by a coder comprising a long-term predictive filter, characterized in that said decoder comprises:

    - a block (211) for detecting transmission frame losses,

    - a module (222) for calculating values of an error indication function, the error indication function being of the form: $${\mathit{x}}_t\left(\mathit{n}\right)\mathit{=}{\mathit{e}}_t\left(\mathit{n}\right)+{\displaystyle \sum _i}{\mathit{g}}_{\mathit{it}}\mathrm{.}{\mathit{x}}_{\mathit{t}}\left(\mathit{n}\mathit{-}\mathit{P}\mathit{+}\mathit{i}\right)\mathit{i}\mathit{\in }\left[\mathit{-}\left(\mathit{N}\mathit{-}\mathit{1}\right)\mathit{/}\mathit{2}\mathit{,}\left(\mathit{N}\mathit{-}\mathit{1}\right)\mathit{/}\mathit{2}\right]$$

    

    
    in which:

    - N is the order of the long-term predictive filter,

    - the gains g _it are equal to the adaptive excitation gains of said long-term predictive filter for the frames received or to the adaptive excitation gains of said long-term predictive filter in the preceding frame for the frames lost,

    - e_t(n) has the value 0 for the frames received and 1 for the frames lost,

    - P is the adaptive excitation period decoded for the frames received or most recently correctly decoded for the frames lost,

    - a module (213) for calculating an error indication parameter on the basis of said values of the error indication function,

    - a comparator (214) for comparing said error indication parameter to at least one given threshold,

    - a discriminator (215) suitable for determining, according to the result supplied by the comparator (214), a value of at least one adaptive excitation gain to be used by the decoder.

Images