EP1316087B1 - Transmission error concealment in an audio signal - Google Patents

Abstract

A method of concealing transmission error in a digital audio signal, wherein a signal that has been decoded after transmission is received, the samples decoded while the transmitted data is valid are stored, at least one short-term prediction operator and one long-term prediction operator are estimated as a function of stored valid samples, and any missing or erroneous samples in the decoder signal are generated using the estimated operators. The energy of the synthesized signal that is thus generated is controlled by means of a gain that is computed and adapted sample by sample.

Description

1. TECHNICAL FIELD

The present invention relates to techniques for concealing consecutive transmission errors in transmission systems using any type of digital coding of the speech and / or sound signal.

• les codeurs dits temporels, qui effectuent la compression des échantillons de signal numérisé échantillon par échantillon (cas des codeurs MIC ou MICDA [DAUMER][MAITRE] par exemple)
• et les codeurs paramétriques qui analysent des trames successives d'échantillons du signal à coder pour extraire, à chacune de ces trames, un certain nombre de paramètres qui sont ensuite codés et transmis (cas des vocodeurs [TREMAIN], des codeurs IMBE [HARDWICK], ou des codeurs par transformée [BRANDENBURG]).

There are classically two main categories of coders:

• the so-called temporal encoders, which perform the compression of the sampled sample sample samples (in the case of MIC or ADPCM [DAUMER] [MASTER] coders for example)
• and the parametric coders which analyze successive frames of samples of the signal to be encoded in order to extract, at each of these frames, a certain number of parameters which are then coded and transmitted (in the case of vocoders [TREMAIN], IMBE coders [HARDWICK] , or transform coders [BRANDENBURG]).

There are intermediate categories that complement the encoding of parameters representative of parametric encoders by the coding of a residual time waveform. For simplicity, these encoders can be placed in the category of parametric encoders.

In this category we find predictive coders and in particular the family of synthetic analysis coders such as RPE-LTP ([HELLWIG]) or CELP ([ATAL]).

For all these coders, the coded values are then transformed into a bit stream which will be transmitted on a transmission channel. Depending on the quality of this channel and the type of transport, disturbances may affect the transmitted signal and produce errors on the bitstream received by the decoder. These errors may occur in isolation in the bit stream but occur very frequently in bursts. It is then a packet of bits corresponding to a complete portion of signal which is erroneous or not received. This type of problem occurs for example for transmissions on mobile networks. It is also found in transmission on packet networks and in particular on Internet-type networks.

When the transmission system or the modules in charge of reception can detect that the data received are highly erroneous (for example on mobile networks), or that a block of data has not been received (in the case of transmission systems for example), error concealment procedures are then implemented. These procedures make it possible to extrapolate to the decoder the samples of the missing signal from the signals and data available from the preceding frames and possibly following the erased zones.

Such techniques have been implemented mainly in the case of parametric coders (techniques for recovering erased frames). They make it possible to strongly limit the subjective degradation of the signal perceived at the decoder in the presence of erased frames. Most of the algorithms developed are based on the technique used for the encoder and decoder, and are in fact an extension of the decoder.
A general object of the invention is to improve, for any compression system of speech and sound, the subjective quality of the speech signal restored to the decoder when due to a poor quality of the transmission channel or as a result of the loss or non-receipt of a packet in a packet transmission system, a set of consecutive encoded data has been lost.

To this end, it proposes a technique making it possible to conceal the successive transmission errors (error packets) whatever the coding technique used, the proposed technique being able to be used for example in the case of time coders whose structure lends itself less easily. a priori to the concealment of error packets.

2. STATE OF THE PRIOR ART

Most predictive coding algorithms provide erased frame retrieval techniques ([GSM-FR], [REC G.723.1A], [SALAMI], [HONKANEN], [COX-2], [CHEN-2] ], [CHEN-3], [CHEN-4], [CHEN-5], [CHEN-6], [CHEN-7], [KROON-2], [WATKINS]). The decoder is informed of the occurrence of an erased frame in one way or another, for example in the case of mobile radio systems by the transmission of the frame erase information from the channel decoder. The purpose of the erased frame recovery devices is to extrapolate the parameters of the erased frame from the last (or more) previous frames considered valid. Some parameters manipulated or coded by the predictive coders have a strong inter-frame correlation (case of the short-term prediction parameters, also called "LPC" of "Linear Predictive Coding" (see [RABINER]) which represent the spectral envelope, and long-term prediction parameters for voiced sounds, for example). Because of this correlation, there is much it is more advantageous to reuse the parameters of the last valid frame to synthesize the erased frame than to use erroneous or random parameters.

• le filtre LPC est obtenu à partir des paramètres LPC de la dernière trame valide soit par recopie des paramètres ou avec introduction d'un certain amortissement (cf. codeur G723.1 [REC G.723.1A]).
• on détecte le voisement pour déterminer le degré d'harmonicité du signal au niveau de la trame effacée ([SALAMI], cette détection se intervenant de la façon suivante :
  • ■ dans le cas d'un signal non voisé :
    • un signal d'excitation est généré de manière aléatoire (tirage d'un mot de code et gain de l'excitation passée légèrement amorti [SALAMI], sélection aléatoire dans l'excitation passée [CHEN], usage des codes transmis éventuellement totalement erronés [HONKANEN],...)
  • ■ dans le cas d'un signal voisé :
    • le délai LTP est généralement le délai calculé à la trame précédente, éventuellement avec une légère "gigue" ([SALAMI]), le gain LTP étant pris très voisin de 1 ou égal à 1. Le signal d'excitation est limité à la prédiction à long terme effectuée à partir de l'excitation passée.

For the CELP (Code Excited Linear Prediction) coding algorithm, refer to [RABINER], the parameters of the erased frame are conventionally obtained as follows:

• the LPC filter is obtained from the LPC parameters of the last valid frame either by copying the parameters or by introducing a certain damping (cf encoder G723.1 [REC G.723.1A]).
• the voicing is detected to determine the degree of harmonicity of the signal at the level of the erased frame ([SALAMI], this detection intervening as follows:
  • ■ in the case of an unvoiced signal:
    • an excitation signal is generated in a random manner (drawing of a code word and gain of the past excitation slightly damped [SALAMI], random selection in the past excitation [CHEN], use of the transmitted codes possibly totally erroneous [ Honkanen], ...)
  • ■ in the case of a voiced signal:
    • the LTP delay is usually the delay calculated at the previous frame, possibly with a slight "jitter" ([SALAMI]), the LTP gain being taken very close to 1 or equal to 1. The excitation signal is limited to the prediction long-term made from past excitement.

In all the examples mentioned above, the procedures for concealing erased frames are strongly related to the decoder and use modules of this decoder, such as the signal synthesis module. They use also intermediate signals available within this decoder as the excitation signal passed and stored during the processing of valid frames preceding the erased frames.

Most of the methods used to conceal the errors produced by lost packets during the transport of coded data by time-type coders use waveform substitution techniques such as those presented in [GOODMAN], [ERDÖL], [AT & T]. Methods of this type reconstruct the signal by selecting portions of the decoded signal before the lost period and do not use synthesis models. Smoothing techniques are also used to avoid the artifacts produced by the concatenation of the different signals.

For transform coders, the techniques for reconstructing erased frames are also based on the coding structure used: the algorithms, such as [PICTEL, MAHIEUX-2], aim at regenerating the transformed coefficients lost from the values taken by these transformers. coefficients before deletion.

The method described in [PARIKH] can be applied to any type of signals; it is based on the construction of a sinusoidal model from the decoded valid signal preceding the erasure, to regenerate the part of the lost signal.

Finally, there is a family of techniques for concealing erased frames developed in conjunction with channel coding. These methods, such as those described in [FINGSCHEIDT], make use of information provided by the channel decoder, for example information concerning the degree of reliability received parameters. They are fundamentally different from the present invention which does not presuppose the existence of a channel encoder.

One prior art that can be considered to be closest to the present invention is that described in [COMBESCURE], which proposed a method for hiding erased frames equivalent to that used in CELP encoders for a transform coder. The drawbacks of the proposed method were the introduction of audible spectral distortions ("synthetic" voice, spurious resonances, etc.), due in particular to the use of poorly controlled long-term synthesis filters (unique harmonic component in voiced sounds). generation of the excitation signal limited to the use of portions of the past residual signal). In addition, the energy control was carried out in [COMBESCURE] at the excitation signal, the energy target of this signal was kept constant for the duration of the erasure, which also generated annoying artifacts. The same remarks apply to the document US5884010 .

3. PRESENTATION OF THE INVENTION

The invention as defined in claims 1, 17 and 18 allows for the concealment of erased frames without marked distortion at higher error rates and / or for longer erased intervals.

It proposes in particular a method for concealing a transmission error in an audio-digital signal according to which a decoded signal is received after transmission, the decoded samples are stored when the transmitted data are valid, and at least one short-term prediction operator is estimated. term and at least one long-term prediction operator based valid samples stored and any missing or erroneous samples are generated in the decoded signal using the operators thus estimated.

According to a first particularly advantageous aspect of the invention, the energy of the synthesis signal thus generated is controlled by means of a gain calculated and adapted sample by sample.

This contributes in particular to improving the performance of the technique on deletion zones of a longer duration.

In particular, the gain for the control of the synthesis signal is advantageously calculated as a function of at least one of the following parameters: energy values previously stored for the samples corresponding to valid data, fundamental period for the voiced sounds, or any parameter characterizing the frequency spectrum.

Advantageously also, the gain applied to the synthesis signal decreases progressively as a function of the time during which the synthesis samples are generated.

Also preferably, stationary and non-stationary sounds are discriminated in the valid data and adaptation laws of this gain (eg decreasing speed) are used, different on the one hand for the generated samples. as a result of valid data corresponding to stationary sounds and secondly for the samples generated as a result of valid data corresponding to nonstationary sounds.

According to another independent aspect of the invention, the contents of the memories used for the decoding process are updated according to the generated synthesis samples.

In this way, on the one hand, the possible desynchronization of the encoder and the decoder (see paragraph 5.1.4 below) is limited, and sudden discontinuities between the erased zone reconstructed according to the invention and the samples following this are avoided. zoned.

In particular, at least part of the synthesized samples is implemented with a coding analogous to that implemented at the transmitter, optionally followed by a (possibly partial) decoding operation, the data obtained being used to regenerate the decoder memories.

In particular, this possibly partial coding-decoding operation can be advantageously used to regenerate the first erased frame because it makes it possible to exploit the contents of the decoder memories before the cutoff, when these memories contain information not provided by the last valid samples decoded (eg in the case of overlap-transform transform coders, see paragraph 5.2.2.2.1 point 10).

According to yet another aspect of the invention, the input of the short-term prediction operator is generated by an excitation signal which, in a voiced area, is the sum of a harmonic component and a weakly harmonic component or non-harmonic, and in voiced area limited to the non-harmonic component.

In particular, the harmonic component is advantageously obtained by implementing a filtering by means of the long-term prediction operator applied to a residual signal calculated by implementing inverse short-term filtering on the stored samples.

The other component can be determined using a long-term prediction operator to which disturbances (eg perturbation of gain, or period), pseudo-random, are applied.

Particularly preferably, for the generation of a voiced excitation signal, the harmonic component represents the low frequencies of the spectrum, while the other component the high frequency portion.

In yet another aspect, the long-term prediction operator is determined from the stored valid frame samples, with a number of samples used for this estimate varying between a minimum value and a value equal to at least two times the fundamental period estimated for voiced sound.

Moreover, the residual signal is advantageously modified by nonlinear type of processing to eliminate amplitude peaks.

Also, according to another advantageous aspect, the voice activity is detected by estimating noise parameters when the signal is considered as non-active, and parameters of the synthesized signal are adjusted to those of the estimated noise.

Even more preferably, the spectral envelope of the noise of the decoded samples is estimated valid and generates a synthesized signal evolving towards a signal having the same spectral envelope.

The invention also proposes a method for processing sound signals, characterized in that it implements a discrimination between the speech and the musical sounds and when musical sounds are detected, a method of the aforementioned type is implemented. without estimating a long-term prediction operator, the excitation signal being limited to a non-harmonic component obtained for example by generating a uniform white noise.

The invention further relates to a transmission error concealment device in an audio-digital signal which receives as input a decoded signal transmitted to it by a decoder and which generates missing or erroneous samples in this decoded signal, characterized in that it comprises processing means capable of implementing the aforementioned method.

It also relates to a transmission system comprising at least one encoder, at least one transmission channel, a module able to detect that transmitted data has been lost or is highly erroneous, at least one decoder and an error concealment device which receives the decoded signal, characterized in that this error concealment device is a device of the aforementioned type.

4. PRESENTATION OF FIGURES

• la figure 1 est un schéma synoptique illustrant un système de transmission conforme à un mode de réalisation possible de l'invention ;
• la figure 2 et la figure 3 sont des schémas synoptiques illustrant une mise en oeuvre conforme à un mode possible de l'invention ;
• les figures 4 à 6 illustrent schématiquement les fenêtres utilisées avec le procédé de dissimulation d'erreurs conforme à un mode de mise en oeuvre possible de l'invention ;
• les figures 7 et 8 sont des représentations schématiques illustrant un mode de mise en oeuvre possible de l'invention dans le cas de signaux musicaux.

Other features and advantages of the invention will become apparent from the description which follows, which is purely illustrative and not limiting and should be read in conjunction with the attached drawings in which:

• Figure 1 is a block diagram illustrating a transmission system according to a possible embodiment of the invention;
• FIG. 2 and FIG. 3 are block diagrams illustrating an implementation according to a possible embodiment of the invention;
• Figures 4 to 6 schematically illustrate the windows used with the error concealment method according to a possible embodiment of the invention;
• Figures 7 and 8 are schematic representations illustrating a possible embodiment of the invention in the case of musical signals.

5. DESCRIPTION OF ONE OR MORE POSSIBLE EMBODIMENTS OF THE INVENTION 5.1 Principle of a possible embodiment

FIG. 1 presents a device for coding and decoding the digital audio signal, comprising an encoder 1, a transmission channel 2, a module 3 making it possible to detect that data transmitted has been lost or is highly erroneous, a decoder 4, and a module 5 for concealing errors or lost packets according to a possible embodiment of the invention.

Note that this module 5, in addition to the indication of erased data, receives the decoded signal in valid period and transmits to the decoder signals used for its update.

1. 1. la mémorisation des échantillons décodés lorsque les données transmises sont valides (traitement 6);
2. 2. durant un bloc de données effacées, la synthèse des échantillons correspondant aux données perdues (traitement 7) ;
3. 3. lorsque la transmission est rétablie, le lissage entre les échantillons de synthèse produits pendant la période effacée et les échantillons décodés (traitement 8);
4. 4. la mise à jour des mémoires du décodeur (traitement 9) (mise à jour qui s'effectue soit pendant la génération des échantillons effacés, soit au moment du rétablissement de la transmission).

More precisely, the processing implemented by the module 5 is based on:

1. 1. storage of the decoded samples when the transmitted data are valid (processing 6);
2. 2. during a block of erased data, the synthesis of the samples corresponding to the lost data (treatment 7);
3. 3. when the transmission is restored, the smoothing between the synthetic samples produced during the erased period and the decoded samples (treatment 8);
4. 4. the updating of the decoder memories (processing 9) (updating which is carried out either during the generation of the erased samples, or at the time of the restoration of the transmission).

5.1.1 In valid period

After decoding the valid data, the memory of the decoded samples, containing a sufficient number of samples for the regeneration of any erased periods in the following, is updated. Typically, one stores on the order of 20 to 40 ms of signal. The energy of the valid frames is also calculated and the energies corresponding to the last valid processed frames (typically of the order of 5 s) are stored in memory.

5.1.2 During a block of erased data.

The following operations, illustrated in FIG. 3, are carried out:

1. Estimation of the current spectral envelope:

This spectral envelope is calculated as an LPC [RABINER] [KLEIJN] filter. The analysis is carried out by conventional methods ([KLEIJN]) after winding the samples stored in valid period. In particular, an LPC analysis (step 10) is used to obtain the parameters of a filter A (z), the inverse of which is used for the LPC filtering (step 11). Since the coefficients thus calculated do not have to be transmitted, a high order can be used for this analysis, which makes it possible to obtain good performances on the musical signals.

2. Detection of voiced sounds and calculation of LTP parameters:

A method for detecting voiced sounds (processing 12 of FIG. 3: V / NV detection, for "voiced / unvoiced") is used on the last stored data. For example, the normalized correlation ([KLEIJN]) or the criterion presented in the following exemplary embodiment can be used for this purpose.

When the signal is declared voiced, the parameters enabling the generation of a long-term synthesis filter, also called LTP filter ([KLEIJN]) are calculated (FIG. 3: LTP analysis, the inverse filter is defined by B (Z). Calculated LTP). Such a filter is generally represented by a period corresponding to the fundamental period and a gain. The accuracy of this filter can be improved by the use of fractional pitch or a multi-coefficient structure [KROON].

When the signal is declared unvoiced, a particular value is assigned to the LTP synthesis filter (see section 4).

It is particularly interesting in this estimation of the LTP synthesis filter to restrict the analyzed area at the end of the period preceding the erasure. The length of the analysis window varies between a minimum value and a value related to the fundamental period of the signal.

3. Calculation of a residual signal:

A residual signal is calculated by LPC inverse filtering (processing 10) of the last stored samples. This signal is then used to generate an excitation signal of the LPC synthesis filter 11 (see below).

4. Summary of missing samples:

The synthesis of the replacement samples is carried out by introducing an excitation signal (calculated at 13 from the signal at the output of the inverse LPC filter) in the LPC synthesis filter 11 (1 / A (z)) calculated in 1. This excitation signal is generated in two different ways depending on whether the signal is voiced or unvoiced:

4.1 In the voiced area:

The excitation signal is the sum of two signals, one strongly harmonic component and the other less harmonic or not at all.

The strongly harmonic component is obtained by LTP filtering (processing module 14) using the parameters calculated in 2, of the residual signal mentioned in 3.

The second component can also be obtained by LTP filtering but made non-periodic by random modifications of the parameters, by generating a pseudo-random signal.

It is particularly interesting to limit the bandwidth of the first component to the low frequencies of the spectrum. Similarly, it will be interesting to limit the second component to higher frequencies.

4.2 In unvoiced area:

When the signal is unvoiced, a non-harmonic excitation signal is generated. It is interesting to use a method of generation similar to that used for the voiced sounds, with variations of parameters (period, gain, signs) making it possible to make it harmless.

4.3 Control of the amplitude of the residual signal:

When the signal is unvoiced, or weakly voiced, the residual signal used for generating the excitation is processed to eliminate the amplitude peaks significantly above the average.

5. Control of the energy of the synthesis signal

The energy of the synthesis signal is controlled using a calculated and adapted sample gain per sample. In the case where the erasure period is relatively long, it is necessary to gradually lower the energy of the synthesis signal. The gain adaptation law is calculated according to different parameters: stored energy values before the deletion (see in 1), fundamental period, and local stationarity of the signal at the time of the cut.

If the system includes a module that discriminates between stationary (such as music) and non-stationary sounds (such as speech), different adaptation laws may also be used.

In the case of transform coders with overlap-add, the first half of the memory of the last correctly received frame contains fairly accurate information about the first half of the first lost frame (its weight in addition-overlap is greater than that of the current frame). This information can also be used for calculating adaptive gain.

6. Evolution of the synthesis procedure over time:

In the case of relatively long erasure period, it is also possible to change the synthesis parameters. If the system is coupled to a voice activity detection device with estimation of the noise parameters (such as [REC-G.723.1A], [SALAMI-2], [BENYASSINE]), it is particularly interesting to make the parameters for generating the signal to be reconstructed towards those of the estimated noise: in particular at the level of the spectral envelope (interpolation of the LPC filter with that of the estimated noise, the interpolation coefficients evolving over time until the filter is obtained noise) and energy (level gradually evolving towards that of noise, for example by windowing).

5.1.3 At the reestablishment of the transmission

In the restoration of transmission, it is particularly important to avoid sudden breaks between the erased period that has been reconstructed according to the techniques defined in the preceding paragraphs and the periods that follow, during which all the information transmitted is available. to decode the signal. The present invention performs time domain weighting with interpolation between replacement samples prior to restoration of communication and valid decoded samples following the erased period. This operation is a priori independent of the type of encoder used.

In the case of transform coders with addition-overlap, this operation is common with the update of the memories described in the following paragraph (see embodiment example).

5.1.4 Updating the decoder memories

When the decoding of the valid samples resumes after an erased period, there may be degradation when the decoder uses data normally generated at the previous frames and stored. It is important to cleanly update these memories to avoid these artifacts.

This is particularly important for coding structures using recursive methods, which for a sample or a sequence of samples, use information obtained after decoding previous samples. For example, predictions ([KLEIJN]) make it possible to extract signal redundancy. This information is normally available to both the coder, who must have done this for these samples preceding a local decoding form, and to the remote decoder present at the reception. As soon as the transmission channel is disturbed and the remote decoder no longer has the same information as the local decoder present on the transmission, there is desynchronization between the encoder and the decoder. In the case of highly recursive coding systems, this desynchronization can cause audible impairments that can persist for a long time or even increase over time if there are instabilities in the structure. In this case, it is therefore important to try to resynchronize the encoder and the decoder, ie to make an estimation of the decoder memories as close as possible to those of the encoder. However resynchronization techniques depend on the coding structure used. One will be presented whose principle is general in the present patent, but whose complexity is potentially important.

One possible method consists in introducing into the decoder on reception a coding module of the same type as that present at the transmission, making it possible to perform the coding-decoding of the samples of the signal produced by the techniques mentioned in the preceding paragraph during the periods deleted. In this way the memories necessary to decode the next samples are completed with data a priori close (subject to a certain stationarity during the erased period) of those which have been lost. In the event that this assumption of stationarity is not respected, after a long erased period for example, we do not have any information sufficient to do better.

In fact it is generally not necessary to perform the complete coding of these samples, it is limited to the modules necessary to update the memories.

This update can be done at the time of the production of the replacement samples, which distributes the complexity over the entire erasure zone, but is cumulative with the synthesis procedure described above.

When the coding structure permits, the above procedure can also be limited to an intermediate zone at the beginning of the valid data period after an erased period, the updating procedure then being combined with the decoding operation. .

5.2. Description of particular embodiments

Specific examples of possible implementation are given below. The case of transform coders of TDAC or TCDM type ([MAHIEUX]) is in particular addressed.

5.2.1 Description of the device

TDAC-type digital transform coding / decoding system.

Wide band encoder (50-7000 Hz) at 24 kb / s or 32 kb / s.

20 ms frame (320 samples).

40 ms windows (640 samples) with 20 ms overlap overlays. A bit frame contains the encoded parameters obtained by the TDAC transformation on a window. After decoding these parameters, doing the inverse TDAC transformation, we obtain an output frame of 20 ms which is the sum of the second one. half of the previous window and the first half of the current window. In Figure 4, it was marked in bold the two parts of windows used for the reconstruction of the frame n (in time). Thus, a lost binary frame disrupts the reconstruction of two consecutive frames (the current one and the next one, figure 5). On the other hand, by correctly replacing the lost parameters, it is possible to recover the parts of the information coming from the preceding and following bit frame (FIG. 6), for the reconstruction of these two frames.

5.2.2 Implementation

All the operations described below are implemented at the reception, in accordance with FIGS. 1 and 2, either within the module for concealing the erased frames which communicates with the decoder, or in the decoder itself (update of the memories decoder).

5.2.2.1 In valid period

Corresponding to paragraph 5.1.2, the memory of the decoded samples is updated. This memory is used for the LPC and LTP analyzes of the signal passed in the case of erasure of a bit frame. In the example presented here, the LPC analysis is done over a 20 ms signal period (320 samples). In general, LTP analysis requires more samples to memorize. In our example, to be able to perform the LTP analysis correctly, the number of stored samples is twice the maximum value of the pitch. For example, if the maximum MaxPitch pitch is 320 samples (50 Hz, 20 ms), the last 640 samples will be stored (40 ms of the signal). We also calculate the energy of the frames valid and stored in a circular buffer of 5 s length. When an erased frame is detected, the energy of the last valid frame is compared to the maximum and the minimum of this circular buffer to know its relative energy.

5.2.2.2 During an erased data block

When a binary frame is lost, two different cases are distinguished:

5.2.2.2.1 First bit frame lost after a valid period

First, an analysis of the stored signal is made to estimate the parameters of the model used to synthesize the regenerated signal. This model then allows us to synthesize 40 ms of signal, which corresponds to the window of 40 ms lost. By doing the TDAC transformation followed by the inverse TDAC transformation on this synthesized signal (without coding - decoding of the parameters), the output signal of 20 ms is obtained. With these inverse TDAC-TDAC operations, the information from the previous correctly received window is exploited (see FIG. 6). At the same time, the memories of the decoder are updated. Thus, the next bit frame, if it is well received, can be decoded normally, and the decoded frames will be automatically synchronized (Figure 6).

1. 1. Fenêtrage du signal mémorisé. Par exemple, on peut utiliser une fenêtre asymétrique de Hamming de 20 ms.
2. 2. Calcul de la fonction d'autocorrélation sur le signal fenêtré.
3. 3. Détermination des coefficients du filtre LPC. Pour cela, classiquement on utilise l'algorithme itératif de Levinson-Durbin. L'ordre d'analyse peut être élevé, surtout lorsque le codeur est utilisé pour coder des séquences de musique.
4. 4. Détection de voisement et analyse à long terme du signal mémorisé pour la modélisation de l'éventuelle périodicité du signal (sons voisés). Dans la réalisation présentée, les inventeurs ont limité l'estimation de la période fondamentale Tp aux valeurs entières, et calculé une estimation du degré de voisement sous la forme du coefficient de corrélation MaxCorr (voir ci-dessous) évalué à la période sélectionnée. Soit Tm = max(T, Fs/200), où Fs est la fréquence d'échantillonnage, donc Fs/200 échantillons correspondent à une durée de 5 ms. Pour mieux modéliser l'évolution du signal à la fin de la trame précédente, on calcule les coefficients de corrélation Corr(T) correspondant à un retard T en n'utilisant que 2*Tm échantillons à la fin du signal mémorisé : $$\mathit{Corr}\left(T\right)=\frac{2{\displaystyle \sum _{i=\mathit{Lmem}-2T_m+T}^{\mathit{Lmem}-1}}{m}_i{m}_{i-T}}{{\displaystyle \sum _{i=\mathit{Lmem}-2T_m}^{\mathit{Lmem}-1}}{m_i}^2+{\displaystyle \sum _{i=\mathit{Lmem}-2T_m+T}^{\mathit{Lmem}-1-\mathrm{<figure-callout\; id="2"\; label="T\; m\; i"\; filenames="imgf0001.png"\; state="\{\{state\}\}">T</figure-callout>}}}{m_i}^{}{{}_{}}^2}$$
   
   
   où m ₀···m _Lmem-1 est la mémoire du signal décodé précédemment. De cette formule, on voit que la longueur de cette mémoire L_mem doit être au moins 2 fois la valeur maximale de la période fondamentale (encore appelée "pitch") MaxPitch.
   On a également fixé la valeur minimale de la période fondamentale MinPitch correspondant à une fréquence de 600 Hz (26 échantillons à Fs = 16 kHz).
   On calcule Corr(T) pour T = 2,
   
   , MaxPitch. Si T' est le plus petit retard tel que Corr(T')<0 (on élimine ainsi les corrélations à très court terme), alors on cherche MaxCorr, maximum de Corr(T) pour T'<T<=MaxPitch. Soit Tp la période correspondant à MaxCorr ( Corr(Tp) = MaxCorr). On cherche également MaxCorrMP, maximum de Corr(T) pour T'<T<=0.75*MinPitch,. Si Tp<MinPitch ou MaxCorrMP > 0.7*MaxCorr et si l'énergie de la dernière trame valide est relativement faible, on décide que la trame est non voisée, car en utilisant la prédiction LTP on risquerait d'obtenir une résonance dans les hautes fréquences très gênante. Le pitch choisi est Tp=MaxPitch/2, et le coefficient de corrélation MaxCorr fixé à une valeur faible (0.25).
   On considère également la trame comme non-voisée lorsque plus de 80% de son énergie se concentre dans les derniers MinPitch échantillons. Il s'agit donc d'un démarrage de la parole, mais le nombre d'échantillons n'est pas suffisant pour estimer la période fondamentale éventuelle, il vaut mieux le traiter comme trame non voisée, et même diminuer plus rapidement l'énergie du signal synthétisé (pour signaler cela, on met DiminFlag=1).
   Dans le cas où MaxCorr > 0.6, on vérifie que l'on n'a pas trouvé un multiple (4, 3 ou 2 fois) de la période fondamentale. Pour cela, on cherche le maximum local de la corrélation autour de T_p/4, Tp/3 et Tp/2. Notons T₁ la position de ce maximum, et MaxCorrL = Corr(T₁). Si T₁ > MinPitch et MaxCorrL > 0.75* MaxCorr, on choisit T₁ comme nouvelle période fondamentale.
   Si Tp est inférieur à MaxPitch/2, on peut vérifier s'il s'agit vraiment d'une trame voisée en cherchant le maximum local de la corrélation autour de 2*TP (TPP) et en vérifiant si Corr(T_PP)>0.4. Si Corr(T_PP)<0.4 et si l'énergie du signal diminue, on met DiminFlag=1 et on diminue la valeur de MaxCorr, sinon on cherche le maximum local suivant entre le T_P actuel et MaxPitch.
   Un autre critère de voisement consiste à vérifier si au moins dans 2/3 des cas le signal retardé par la période fondamentale a le même signe que le signal non retardé.
   On vérifie cela sur une longueur égale au maximum entre 5ms et 2*T_P.
   On vérifie également si l'énergie du signal a tendance à diminuer ou non. Si oui, on met DiminFlag=1 et on fait décroître la valeur de MaxCorr en fonction de degré de diminution.
   La décision de voisement tient compte également de l'énergie du signal: si l'énergie est forte, on augmente la valeur de MaxCorr, ainsi il est plus probable que la trame soit décidée voisée. Par contre, si l'énergie est très faible, on diminue la valeur de MaxCorr.
   Finalement, on prend la décision de voisement en fonction de la valeur de MaxCorr: la trame est non voisée si et seulement si MaxCorr < 0.4. La période fondamentale T_p d'une trame non voisée est bornée, elle doit être inférieure ou égale à MaxPitch/2.
5. 5. Calcul du signal résiduel par filtrage inverse LPC des derniers échantillons mémorisés. Ce signal résiduel est stocké dans la mémoire ResMem.
6. 6. Egalisation de l'énergie du signal résiduel. Dans le cas d'un signal non voisé ou faiblement voisé (MaxCorr < 0.7), l'énergie du signal résiduel stocké dans ResMem peut changer brusquement d'une partie à l'autre. La répétition de cette excitation entraîne une perturbation périodique très désagréable dans le signal synthétisé. Pour éviter cela, on s'assure qu'aucun pic d'amplitude important ne se présente dans l'excitation d'une trame faiblement voisée. Comme l'excitation est construite à partir des derniers Tp échantillons du signal résiduel, on traite ce vecteur de Tp échantillons. La méthode utilisée dans notre exemple est la suivante :
   • ■ On calcule la moyenne MeanAmpl des valeurs absolues des derniers Tp échantillons du signal résiduel.
   • ■ Si le vecteur d'échantillons à traiter contient n passages à zéro, on le coupe en n+1 sous-vecteurs, le signe du signal dans chaque sous-vecteur étant donc invariant.
   • ■ On cherche l'amplitude maximale MaxAmplSv de chaque sous-vecteur. Si MaxAmplSv>1.5*MeanAmpl , on multiplie le sous-vecteur par 1.5*MeanAmpl/MaxAmplSv.
7. 7. Préparation du signal d'excitation d'une longueur de 640 échantillons correspondant à la longueur de la fenêtre TDAC. On distingue 2 cas selon le voisement :
   • 
     Le signal d'excitation est la somme de deux signaux, une composante fortement harmonique limitée en bande aux basses fréquences du spectre excb et une autre moins harmonique limitée aux plus hautes fréquences exch.
   
   La composante fortement harmonique est obtenue par filtrage LTP d'ordre 3 du signal résiduel : $$\mathrm{excb}\left(\mathrm{i}\right)\mathrm{=}\mathrm{0.15}\mathrm{*}\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{Tp}\mathrm{-}\mathrm{1}\right)\mathrm{+}0.7*\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{Tp}\right)\mathrm{+}\mathrm{0.15}\mathrm{*}\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{<figure-callout\; id="1"\; label="Tp\; +"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">Tp</figure-callout>}+1\right)$$
   
   
   Les coefficients [0.15, 0.7, 0.15] correspondent à un filtre FIR passe-bas de 3 dB d'atténuation à Fs/4.
   La seconde composante est obtenue également par un filtrage LTP rendu non périodique par la modification aléatoire de sa période fondamentale Tph. Tph est choisie comme la partie entière d'une valeur réelle aléatoire Tpa. La valeur initiale de Tpa est égale à Tp puis elle est modifiée échantillon par échantillon en l'additionnant une valeur aléatoire dans [-0.5, 0.5]. De plus, ce filtrage LTP est combiné avec un filtrage IIR passe haut : $$\mathrm{exch}\left(\mathrm{i}\right)\mathrm{=}\mathrm{-}\mathrm{0.0635}\mathrm{*}\left(\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{Tph}\mathrm{-}\mathrm{1}\right)\mathrm{+}\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{<figure-callout\; id="1"\; label="Tph\; +"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">Tph</figure-callout>}\mathrm{+}\mathrm{}\mathrm{1}\right)\right)\mathrm{+}\mathrm{0.1182}\mathrm{*}\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{Tph}\right)\mathrm{-}\mathrm{0.9926}\mathrm{*}\mathrm{exch}\left(\mathrm{i}\mathrm{-}\mathrm{1}\right)\mathrm{-}\mathrm{0.7679}\mathrm{*}\mathrm{exch}\left(\mathrm{i}\mathrm{-}\mathrm{2}\right)$$
   
   
   L'excitation voisée est alors la somme de ces 2 composante : $$\mathrm{Exc}\left(\mathrm{i}\right)\mathrm{=}\mathrm{excb}\left(\mathrm{i}\right)\mathrm{+}\mathrm{exch}\left(\mathrm{i}\right)$$
   
   • ■ Dans le cas d'une trame non voisée, le signal d'excitation exc est obtenu également par filtrage LTP d'ordre 3 avec les coefficients [0.15, 0.7, 0.15] mais il est rendu non périodique par augmentation de la période fondamentale d'une valeur égale à 1 tous les 10 échantillons, et inversion du signe avec une probabilité de 0.2.
8. 8. Synthèse des échantillons de remplacement en introduisant le signal d'excitation exc dans le filtre LPC calculé en 3.
9. 9. Contrôle du niveau de l'énergie du signal de synthèse. L'énergie tend progressivement vers un niveau fixé par avance dès la première trame de remplacement synthétisée. Ce niveau peut être défini, par exemple, comme l'énergie de la trame de sortie la plus faible trouvée durant les 5 dernières secondes précédant l'effacement. Nous avons défini deux lois d'adaptation du gain qui sont choisies en fonction du drapeau DiminFlag calculé en 4. La vitesse de diminution de l'énergie dépend également de la période fondamentale. Il existe une troisième loi d'adaptation plus radicale qui est utilisée quand on détecte que le début du signal généré ne correspond pas bien au signal originel, comme expliqué ultérieurement (voir point 11).
10. 10. Transformation TDAC sur le signal synthétisé en 8, comme expliqué au début de ce chapitre. Les coefficients TDAC obtenus remplacent les coefficients TDAC perdus. Ensuite, en faisant la transformation inverse TDAC, on obtient la trame de sortie. Ces opérations ont trois buts :
    • ■ Dans le cas de la première fenêtre perdue, de cette façon on exploite l'information de la fenêtre précédente correctement reçue qui contient la moitié des données nécessaires pour reconstruire la première trame perturbée (figure 6).
    • ■ On met à jour la mémoire du décodeur pour le décodage de la trame suivante (synchronisation du codeur et du décodeur, voir paragraphe 5.1.4).
    • ■ On assure automatiquement la transition continue (sans rupture) du signal de sortie lorsque la première trame binaire correctement reçue arrive après une période effacée que l'on a reconstruite selon les techniques présentées ci-dessus (voir paragraphe 5.1.3).
11. 11. La technique d'addition-recouvrement permet de vérifier si le signal voisé synthétisé correspond bien au signal d'origine ou non car pour la première moitié de la première trame perdue le poids de la mémoire de dernière fenêtre correctement reçue est plus important (figure 6). Donc en prenant la corrélation entre la première moitié de la première trame synthétisée et la première moitié de la trame obtenue après les opérations TDAC
    
    TDAC inverse, on peut estimer la similitude entre la trame perdue et la trame de remplacement. Une corrélation faible (< 0.65) signale que le signal originel est assez différent de celui obtenu par la méthode de remplacement, et il vaut mieux diminuer l'énergie de ce dernier rapidement vers le niveau minimal.

The operations to be performed are as follows:

1. 1. Winding of the memorized signal. For example, an asymmetric Hamming window of 20 ms can be used.
2. 2. Calculation of the autocorrelation function on the windowed signal.
3. 3. Determination of the coefficients of the LPC filter. For this, conventionally we use the iterative algorithm of Levinson-Durbin. The order of analysis can be high, especially when the encoder is used to encode music sequences.
4. 4. Voicing detection and long-term analysis of the memorized signal for modeling the possible periodicity of the signal (voiced sounds). In the embodiment presented, the inventors limited the estimation of the fundamental period Tp to integer values, and calculated an estimate of the degree of voicing in the form of the correlation coefficient MaxCorr (see below) evaluated at the selected period. Let Tm = max (T, Fs / 200), where Fs is the sampling frequency, so Fs / 200 samples correspond to a duration of 5 ms. To better model the evolution of the signal at the end of the previous frame, the correlation coefficients Corr (T) corresponding to a delay T are calculated using only 2 * Tm samples at the end of the memorized signal: $$\mathit{Corr}\left(T\right)=\frac{2{\displaystyle \underset{i=\mathit{LMEM}-2T_m+T}{\overset{\mathit{LMEM}-1}{\Sigma }}}{m}_i{m}_{i-T}}{{\displaystyle \underset{i=\mathit{LMEM}-2T_m}{\overset{\mathit{LMEM}-1}{\Sigma }}}{m_i}^2+{\displaystyle \underset{i=\mathit{LMEM}-2T_m+T}{\overset{\mathit{LMEM}-1-\mathrm{<figure-callout\; id="2"\; label="T\; m\; i"\; filenames="imgf0001.png"\; state="\{\{state\}\}">T\; </figure-callout>}}{\Sigma }}}{m_i}^{}{{}_{}}^2}$$
   
   
   where m ₀ ··· m _{Lmem -1} is the memory of the previously decoded signal. From this formula, we see that the length of this memory L _mem must be at least 2 times the maximum value of the fundamental period (also called "pitch") MaxPitch .
   We also set the minimum value of the fundamental MinPitch period corresponding to a frequency of 600 Hz (26 samples at Fs = 16 kHz).
   Corr (T) is calculated for T = 2,
   
   , MaxPitch . If T 'is the smallest delay such that Corr (T') <0 (thus eliminating very short term correlations), then we look for MaxCorr , maximum of Corr (T) for T '<T <= MaxPitch. Let Tp be the period corresponding to MaxCorr (Corr (Tp) = MaxCorr). One also seeks MaxCorrMP, maximum of Corr (T) for T '<T <= 0.75 * MinPitch ,. If Tp <MinPitch or MaxCorrMP> 0.7 * MaxCorr and if the energy of the last valid frame is relatively weak, it is decided that the frame is unvoiced, because by using the LTP prediction it would be likely to obtain a resonance in the high frequencies very embarrassing. The pitch chosen is Tp = MaxPitch / 2, and the correlation coefficient MaxCorr set to a low value (0.25).
   The frame is also considered unvoiced when more than 80% of its energy is concentrated in the last MinPitch samples. It is thus a start of the speech, but the number of samples is not sufficient to estimate the possible fundamental period, it is better to treat it as an unvoiced frame, and even to decrease more quickly the energy of the signal synthesized (to signal this, we put DiminFlag = 1 ).
   In the case where MaxCorr> 0.6, we check that we did not find a multiple (4, 3 or 2 times) of the fundamental period. For this, we look for the local maximum of the correlation around T _p / 4, Tp / 3 and Tp / 2. Let T ₁ denote the position of this maximum, and MaxCorrL = Corr (T ₁ ). If T ₁ > MinPitch and MaxCorrL> 0.75 * MaxCorr, we choose T ₁ as the new fundamental period.
   If Tp is less than MaxPitch / 2, we can check if it is really a voiced frame looking for the local maximum of the correlation around 2 * TP (TPP) and checking if Corr (T _PP )> 0.4. If Corr (T _PP ) <0.4 and if the signal energy decreases, we set DiminFlag = 1 and decrease the value of MaxCorr, otherwise we look for the next local maximum between the current T _P and MaxPitch.
   Another voicing criterion consists of checking whether at least in 2/3 of the cases the signal delayed by the fundamental period has the same sign as the undelayed signal.
   This is verified for a length equal to the maximum between 5ms and 2 * T _P.
   It is also checked whether the signal energy tends to decrease or not. If yes, we set DiminFlag = 1 and we decrease the value of MaxCorr according to the degree of decrease.
   The decision of voicing also takes into account the energy of the signal: if the energy is strong, one increases the value of MaxCorr, thus it is more probable that the frame is decided voiced. On the other hand, if the energy is very weak, one decreases the value of MaxCorr.
   Finally, we make the voicing decision based on the value of MaxCorr: the frame is unvoiced if and only if MaxCorr <0.4. The fundamental period T _p of an unvoiced frame is bounded, it must be less than or equal to MaxPitch / 2.
5. 5. Calculation of the residual signal by LPC inverse filtering of the last stored samples. This residual signal is stored in the ResMem memory.
6. 6. Equalization of the energy of the residual signal. In the case of an unvoiced or weakly voiced signal (MaxCorr <0.7), the energy of the residual signal stored in ResMem can change abruptly from one party to another. The repetition of this excitation causes a very unpleasant periodic disturbance in the synthesized signal. To avoid this, it is ensured that no peak of significant amplitude occurs in the excitation of a weakly voiced frame. Since the excitation is constructed from the last Tp samples of the residual signal, this vector of Tp samples is processed. The method used in our example is as follows:
   • ■ MeanAmpl mean of the absolute values of the last Tp samples of the residual signal.
   • ■ If the sample vector to be processed contains n zero crossings, we cut it in n + 1 sub-vectors, the sign of the signal in each sub-vector being therefore invariant.
   • ■ We search for the maximum amplitude MaxAmplSv of each sub-vector. If MaxAmplSv> 1.5 * MeanAmpl, multiply the sub-vector by 1.5 * MeanAmpl / MaxAmplSv.
7. 7. Preparation of the excitation signal with a length of 640 samples corresponding to the length of the window TDAC. We distinguish 2 cases according to the voicing:
   • 
     The excitation signal is the sum of two signals, a strongly harmonic component limited in band at the low frequencies of the excb spectrum and another less harmonic limited to the higher frequencies exch.
   
   The strongly harmonic component is obtained by third order LTP filtering of the residual signal: $$\mathrm{ExCB}\left(\mathrm{i}\right)\mathrm{=}\mathrm{0.15}\mathrm{*}\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{Tp}\mathrm{-}\mathrm{1}\right)\mathrm{+}0.7*\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{Tp}\right)\mathrm{+}\mathrm{0.15}\mathrm{*}\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{<figure-callout\; id="1"\; label="Tp\; +"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">Tp\; </figure-callout>}+1\right)$$
   
   
   The coefficients [0.15, 0.7, 0.15] correspond to a low-pass FIR filter of 3 dB attenuation at Fs / 4.
   The second component is also obtained by LTP filtering made non-periodic by the random modification of its fundamental period Tph. Tph is chosen as the integer part of a random real value Tpa. The initial value of Tpa is equal to Tp then it is modified sample by sample by adding a random value in [-0.5, 0.5]. In addition, this LTP filtering is combined with high pass IIR filtering: $$\mathrm{exch}\left(\mathrm{i}\right)\mathrm{=}\mathrm{-}\mathrm{0.0635}\mathrm{*}\left(\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{tph}\mathrm{-}\mathrm{1}\right)\mathrm{+}\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{<figure-callout\; id="1"\; label="tph\; +"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">tph\; </figure-callout>}\mathrm{+}\mathrm{}\mathrm{1}\right)\right)\mathrm{+}\mathrm{0.1182}\mathrm{*}\mathrm{exc}\left(\mathrm{i}\mathrm{-}\mathrm{tph}\right)\mathrm{-}\mathrm{0.9926}\mathrm{*}\mathrm{exch}\left(\mathrm{i}\mathrm{-}\mathrm{1}\right)\mathrm{-}\mathrm{0.7679}\mathrm{*}\mathrm{exch}\left(\mathrm{i}\mathrm{-}\mathrm{2}\right)$$
   
   
   The voiced excitation is then the sum of these 2 components: $$\mathrm{HE}\left(\mathrm{i}\right)\mathrm{=}\mathrm{ExCB}\left(\mathrm{i}\right)\mathrm{+}\mathrm{exch}\left(\mathrm{i}\right)$$
   
   • ■ In the case of an unvoiced frame, the excitation signal exc is also obtained by LTP filtering of order 3 with the coefficients [0.15, 0.7, 0.15] but it is made non-periodic by increasing the fundamental period d a value equal to 1 every 10 samples, and inversion of the sign with a probability of 0.2.
8. 8. Synthesis of the replacement samples by introducing the exc excitation signal into the LPC filter calculated in 3.
9. 9. Control of the energy level of the synthesis signal. The energy tends gradually to a level set in advance from the first synthesized replacement frame. This level can be defined, for example, as the energy of the lowest output frame found during the last 5 seconds preceding erasure. We have defined two gain adaptation laws which are chosen according to the DiminFlag flag calculated in 4. The speed of energy decrease also depends on the fundamental period. There is a third, more radical adaptation law that is used when we detect that the The beginning of the generated signal does not correspond well to the original signal, as explained later (see point 11).
10. 10. Transformation TDAC on the signal synthesized in 8, as explained at the beginning of this chapter. The TDAC coefficients obtained replace the TDAC coefficients lost. Then, doing the reverse TDAC transformation, we get the output frame. These operations have three goals:
    • ■ In the case of the first window lost, this way we use the information from the previous window correctly received which contains half of the data needed to reconstruct the first disturbed frame (Figure 6).
    • ■ The decoder memory is updated for the decoding of the next frame (synchronization of the encoder and the decoder, see section 5.1.4).
    • ■ The continuous transition (without breaking) of the output signal is automatically performed when the first correctly received bit frame arrives after an erased period which has been reconstructed according to the techniques presented above (see paragraph 5.1.3).
11. 11. The addition-overlay technique makes it possible to check whether the synthesized voiced signal corresponds to the original signal or not because for the first half of the first lost frame the weight of the last received window memory is greater ( Figure 6). So taking the correlation between the first half of the first synthesized frame and the first half of the frame obtained after TDAC operations
    
    Reverse TDAC, we can estimate the similarity between the lost frame and the replacement frame. A weak correlation (<0.65) indicates that the original signal is quite different from that obtained by the replacement method, and it is better to reduce the energy of the latter rapidly towards the minimum level.

5.2.2.2.2 Lost frames following the first frame of an erased area

• ■ Dans la partie synthèse (points 7 et 8), on génère uniquement 320 nouveaux échantillons, car la fenêtre de la transformation TDAC couvre les derniers 320 échantillons générés lors de la trame effacée précédente et ces nouveaux 320 échantillons.
• ■ Dans le cas où la période d'effacement serait relativement longue, il est important de faire évoluer les paramètres de synthèse vers les paramètres d'un bruit blanc ou vers ceux du bruit de fond (voir point 5 dans paragraphe 3.2.2.2). Comme le système présenté dans cet exemple ne comprend pas de VAD /CNG, nous avons, par exemple, la possibilité de faire une ou plusieurs des modifications suivantes :

• ■ Interpolation progressive du filtre LPC avec un filtre plat pour rendre le signal synthétisé moins coloré.
• ■ Augmentation progressive de la valeur du pitch.
• ■ En mode voisé, on bascule en mode non-voisé après un certain temps (par exemple quand l'énergie minimale est atteinte).

In the preceding paragraph, the points 1-6 concern the analysis of the decoded signal preceding the first erased frame and allowing the construction of a synthesis model (LPC and possibly LTP) of this signal. For the following erased frames, the analysis is not repeated, the replacement of the lost signal is based on the parameters (LPC coefficients, pitch, MaxCorr, ResMem) calculated during the first erased frame. Therefore, only the operations corresponding to the synthesis of the signal and the synchronization of the decoder are carried out, with the following modifications with respect to the first erased frame:

• ■ In the summary section (points 7 and 8), only 320 new samples are generated because the TDAC transformation window covers the last 320 samples generated during the previous cleared frame and these new 320 samples.
• ■ In the case where the erasure period is relatively long, it is important to change the synthesis parameters to the parameters of a white noise or to those of the background noise (see point 5 in paragraph 3.2.2.2). As the system presented in this example does not include VAD / CNG, we have, for example, the ability to make one or more of the following changes:

• ■ Progressive interpolation of the LPC filter with a flat filter to make the synthesized signal less colorful.
• ■ Gradual increase of the pitch value.
• ■ In the voiced mode, you switch to unvoiced mode after a certain time (for example when the minimum energy is reached).

5.3 Specific processing for musical signals. If the system comprises a module for speech / music discrimination, then after selecting a music synthesis mode, it is possible to implement a specific processing of the musical signals. In FIG. 7, the music synthesis module has been referenced by 15, that of speech synthesis by 16 and the speech / music switch by 17.

Such a processing implements, for example, for the music synthesis module the following steps, illustrated in FIG. 8:

1. Estimation of the current spectral envelope:

This spectral envelope is calculated as an LPC [RABINER] [KLEIJN] filter. The analysis is carried out by conventional methods ([KLEIJN]). After windowing of the samples stored in valid period, an LPC analysis is used to calculate an LPC A (Z) filter (step 19). For this analysis, a high order (> 100) is used to obtain good performances on musical signals.

2. Summary of missing samples:

The synthesis of the replacement samples is carried out by introducing an excitation signal into the LPC synthesis filter (1 / A (z)) calculated in step 19. This excitation signal - calculated in a step 20 - is a white noise whose amplitude is chosen to obtain a signal having the same energy as that of the last N samples stored in valid period. In FIG. 8, the filtering step is referenced 21.

Example of the control of the amplitude of the residual signal:

If the excitation is a uniform white noise multiplied by a gain, this gain G can be calculated as follows:

Estimation of the gain of the LPC filter:

The Durbin algorithm gives the energy of the residual signal. Also knowing the energy of the signal to modeled one estimates the gain G _LPC of the filter LPC as the ratio of these two energies.

Calculation of the target energy:

The target energy is estimated as equal to the energy of the last N samples stored in valid period (N is typically <the length of the signal used for the LPC analysis).

The energy of the synthesized signal is the product of the white noise energy by G ² and G _LPC . We choose G so that this energy is equal to the target energy.

3. Control of the energy of the synthesis signal

As for speech signals, except that the rate of decrease of the energy of the synthesis signal and much slower, and that it does not depend on fundamental period (non-existent):

The energy of the synthesis signal is controlled using a calculated and adapted sample gain per sample. In the case where the erasure period is relatively long, it is necessary to gradually lower the energy of the synthesis signal. The gain adaptation law can be calculated according to various parameters such as energy values stored before the erasure, and local stationarity of the signal at the time of the cutoff.

6. Evolution of the synthesis procedure over time: As for speech signals:

In the case of relatively long erasure periods, it is also possible to change the synthesis parameters. If the system is coupled to a voice activity detection device or musical signals with estimation of noise parameters (such as [REC-G.723.1A], [SALAMI-2], [BENYASSINE]), it will be of particular interest to make the parameters of generation of the signal to be reconstructed tend towards those of the estimated noise: in particular at the level of the spectral envelope (interpolation of the LPC filter with that of the estimated noise, the interpolation coefficients evolving over time until to obtain the noise filter) and energy (level gradually evolving towards that of the noise, for example by windowing).

6. GENERAL REMARK

As will be understood, the technique which has just been described has the advantage of being usable with any type of encoder; in particular, it makes it possible to remedy the problems of lost bit packets for time or transform coders, on speech and music signals with good performances: in fact in the present technique, the only signals stored during the periods when the data transmitted are valid are the samples from the decoder, information that is available regardless of the coding structure used.

7. BIBLIOGRAPHIC REFERENCES

• [AT & T] AT & T (DA Kapilow, RV Cox) "A high quality low-complexity algorithm for frame erasure concealment (FEC) with G.711, Delayed Contribution D.249 (WP 3/16), ITU, May 1999.
• [ATAL] BS Atal and MR Schroeder. "Predictive coding of speech signal and subjective error criteria". IEEE Trans. on Acoustics, Speech and Signal Processing, 27: 247-254, June 1979 .
• [BENYASSINE] A. Benyassin, E. Shlomot and HY Su. "ITU-T recommendation G.729 Annex B: A silent compression scheme for use with G.729 for V.70 digital simultaneous voice and data applications". IEEE Communication Magazine, September 97, PP. 56-63 .
• [BRANDENBURG] KH Brandenburg and M. Bossi. "Overview of MPEG audio: current and future standards for low-bit-rate audio coding". Journal of Audio Eng. Soc., Vol.45-1 / 2, January / February 1997, PP.4-21 .
• [Chen] JH Chen, Cox RV, YC Lin, N. Jayant and MJ Melchner. "A low-delay CELP coder for the CCITT standard 16 kb / s speech coding". IEEE Journal on Selected Areas on Communications, Vol.10-5, June 1992, PP.830-849 .
• [CHEN-2] JH Chen, CR Watkins. "Linear prediction coefficient generation during frame erasure gold packet loss "Patent US5574825 , EP0673018 .
• [CHEN-3] JH Chen, CR Watkins. "Linear prediction coefficient generation during frame erasure gold packet loss "Patent 884010 .
• [CHEN-4] JH Chen, CR Watkins. "Frame erase gold packet loss compensation method "Patent US5550543 , EP0707308 .
• [CHEN-5] JH Chen. "Excitement signal synthesis during frame erasure gold packet loss "Patent US5615298 , EP0673017 .
• [CHEN-6] JH Chen. "Computational complexity reduction during the erasure of packet loss "Patent US5717822 .
• [CHEN-7] JH Chen. "Computational complexity reduction during the frame erasure or packet loss "Patent US940212435 , EP0673015 .
• [COX] RV Cox. "Three new speech coders from the ITU covers a range of applications". IEEE Communication Magazine, September 97, PP. 40-47 .
• [COX-2] RV Cox. "An improved frame erasure concealment method for ITU-T Rec G728". Delayed contribution D.107 (WP 3/16), ITU-T, January 1998 .
• [COMBESCURE] P.Combescure, J. Schnitzler, K. Ficher, R. Kirchherr, C. Lamblin, A. Guyader, D. Massaloux, C. Quinquis, J. Stegmann, P. Vary. "At 16.24,32 kbps Wideband Speech Codec Based on ATCELP". Proc. of ICASSP conference, 1998 .
• [DAUMER] WR Daumer, P. Mermelstein, X. Master and I. Tokizawa. "Overview of the ADPCM coding algorithm". Proc. of GLOBECOM 1984, PP.23.1.1-23.1.4 .
• [ERDÖL]. N. Erdöl, C. Castelluccia, A. Zilouchian "Recovery of Missing Speech Packets Using the Short-Time Energy and Zero-Crossing Measurements" IEEE Trans. on Speech and Audio Processing, Vol.1-3, July 1993, PP.295-303 .
• [FINGSCHEIDT] T. Fingscheidt, P. Vary, "Robust speech decoding: a universal approach to a bit error concealment", Proc. of ICASSP conference, 1997, pp. 1667-1670 .
• [Goodman] DJ Goodman, GB Lockhart, OJ Wasem, WC Wong. "Waveform Substitution Techniques for Recovering Missing Speech Segments in Packet Voice Communications". IEEE Trans. On Acoustics, Speech and Signal Processing, Vol. ASSP-34, December 1986, PP. 1440-1448 .
• [GSM-FR] Recommendation GSM 06.11. " ETSI / TC SMG, Ver .: 3.0.1., February 1992, Substitution and muting of lost frames for full rate speech. .
• [HARDWICK] JC Hardwick and JS Lim. "The application of the IMBE speech coder to mobile communications". Proc. of ICASSP conference, 1991, PP.249-252 .
• [Hellwig] K. Hellwig, P. Vary, D. Massaloux, JP Petit, C. Galand and M. Rosso. "Speech codec for the European mobile radio system". GLOBECOM conference, 1989, PP. 1065-1069 .
• [Honkanen] T. Honkanen, J. Vainio, P. Kapanen, P. Haavisto, R. Salami, C. Laflamme and JP Adoul. "GSM enhanced full rate speech codec". Proc. of ICASSP conference, 1997, PP.771-774 .
• [KROON] P. Kroon, BS Atal. "On the use of pitch predictors with high temporal resolution". IEEE Trans. on Signal Processing, Vol.39-3, March 1991, PP.733-735 .
• [KROON-2] P. Kroon. "Linear prediction coefficient generation during frame erasure gold packet loss "Patent US5450449 , EP0673016 .
• [MAHIEUX] Y. Mahieux, JP Small. "High quality audio coding at 64 kbit / s". IEEE Trans. on Com., Vol.42-11, Nov. 1994, PP.3010-3019 .
• [MAHIEUX-2] Y. Mahieux, "Concealment of transmission errors "patent 92 06720 filed June 3, 1992 .
• [MASTER] X. Master. "7 kHz audio coding within 64 kbit / s". IEEE Journal on Selected Areas on Communications, Vol.6-2, February 1988, PP.283-298 .
• [PARIKH] VN Parikh, JH Chen, G. Aguilar. "Frame Erasure Concealment Using Sinusoidal Analysis-Synthesis and Its Application to MDCT-Based Codecs". Proc. of ICASSP conference, 2000 .
• [PICTEL] PictureTel Corporation, "Detailed Description of the PTC (PictureTel Transform Coder), ITU-T Contribution, SG15 / WP2 / Q6, October 8-9, 1996 Baltimore meeting, TD7
• [RABINER] LR Rabiner, RW Schafer. "Digital processing of speech signals". Bell Laboratories Inc., 1978 .
• [REC G.723.1A] ITU-T Annex A to Recommendation G.723.1 "Silence compression scheme for dual rate speech coding for multimedia communications transmitting at 5.3 and 6.3 kbit / s"
• [SALAMI] R. 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 .
• [SALAMI-2] R. Salami, C. Laflamme, JP Adoul. "ITU-T G.729 Annex A: Reduced complexity 8 kb / s CS-ACELP codec for digital simultaneous voice and data". IEEE Communication Magazine, September 97, PP. 56-63 .
• [TREMAIN] TE Tremain. "The standard linear predictive government coding algorithm: LPC 10". Speech technology, April 1982, PP.40-49 .
• [Watkins] CR Watkins, JH Chen. "Improving 16kbps G.728 LD-CELP Speech Coder for Frame Erasure Channels". Proc. of ICASSP conference, 1995, PP.241-244 .

Claims (18)

1. Method of concealing transmission error in a digital audio signal in which upon detecting (3) samples that are missing or erroneous in a signal, synthesis samples (5) are generated with the aid of at least one short-term prediction operator and at least for the voiced sounds a long-term prediction operator estimated as a function of decoded samples of a past decoded signal, said decoded samples being stored (6) previously when the transmitted data of said past signal are valid, characterized in that the energy of the synthesis signal thus generated is controlled with the aid of a gain that is calculated and adapted sample by sample according to an adaptation law dependent on at least one parameter of said stored decoded samples.
2. Method according to Claim 1, characterized in that the gain for controlling the synthesis signal is calculated as a function of at least one of the following parameters: energy values previously stored for the samples corresponding to valid data, pitch period for the voiced sounds, or any parameter characterizing the frequency spectrum.
3. Method according to one of the preceding claims, characterized in that the gain applied to the synthesis signal decreases progressively as a function of the duration for which the synthesis samples are generated.
4. Method according to one of the preceding claims, characterized in that the steady sounds and the non-steady sounds are discriminated in the valid data, and gain adaptation laws are implemented for controlling the synthesis signal that differ on the one hand for the samples generated following valid data corresponding to steady sounds and on the other hand for the samplers generated following valid data corresponding to non-steady sounds.
5. Method according to one of the preceding claims, characterized in that the content of memories that are used for the decoding processing is updated as a function of the synthesis samples generated.
6. Method according to Claim 5, characterized in that a coding analogous to that implemented at the transmitter is implemented at least partially on the synthesized samples, optionally followed by an at least partial decoding operation, the data obtained serving to regenerate the memories of the decoder.
7. Method according to Claim 6, characterized in that the first erased frame is regenerated by means of this coding-decoding operation, utilizing the content of the memories of the decoder prior to the interruption, when said memories contain information that can be utilized in this operation.
8. Method according to one of the preceding claims, characterized in that an excitation signal is generated as input to the short-term prediction operator, which signal, in a voiced zone, is the sum of a harmonic component and a weakly harmonic or non-harmonic component, and in a non-voiced zone, limited to a non-harmonic component.
9. Method according to Claim 8, characterized in that the harmonic component is obtained by implementing a filtering by means of the long-term prediction operator applied to a residual signal calculated by implementing an inverse short-term filtering on the stored samples.
10. Method according to Claim 9, characterized in that the other component is determined with the aid of a long-term prediction operator to which pseudo-ransom disturbances are applied.
11. Method according to one of Claims 8 to 10, characterized in that in order to generate a voiced excitation signal, the harmonic component is limited to the low frequencies of the spectrum, while the other component is limited to the high frequencies.
12. Method according to one of the preceding claims, characterized in that the long-term prediction operator is determined from the stored valid frame samples, with a number of samples used for this estimation varying between a minimum value and a value that is equal to at least twice the pitch period estimated for the voiced sound.
13. Method according to one of the preceding claims, characterized in that the residual signal is processed in a non-linear manner in order to eliminate amplitude peaks.
14. Method according to one of the preceding claims, characterized in that voice activity is detected by estimating noise parameters and in that parameters of the synthesized signal are made to tend towards those of the estimated noise.
15. Method according to Claim 14, characterized in that the spectral envelope of the noise of the valid decoded samples is estimated and a synthesized signal that evolves towards a signal possessing the same spectral envelope is generated.
16. Method of processing sound signals, characterized in that a discrimination is implemented between the voiced sounds and the musical sounds, and when ;musical sounds are detected, a method is implemented according to one of the preceding claims without estimating a long-term prediction operator.
17. Device for concealing transmission error in a digital audio signal which receives as input a decoded signal transmitted to it by a decoder and which generates samples that are missing or erroneous in this decoded signal, characterized in that it comprises processing means suitable for implementing the method according to one of the preceding claims.
18. Transmission system comprising at least one coder, at least one transmission channel, a module suitable for detecting that transmitted data have been lost or are highly erroneous, at least one decoder and a device for concealing errors which receives the decoded signal, characterized in that this device for concealing errors is a device according to Claim 17.

Images