FR2830970A1 - Telephone channel transmission speech signal error sample processing has errors identified and preceding/succeeding valid frames found/samples formed following speech signal period and part blocks forming synthesised frame. - Google Patents

Abstract

The error correction system identifies an erroneous frame finding a first valid frame (t-1) preceding the erroneous frame and a second valid frame (t+1) following the frame. A succession of synthesised blocks are formed following the preceding and succeeding frames with a number of samples (TBL) calculated as a fundamental period of the speech signal in the valid frames (P0,P1). Part of the block samples are used to form a synthesised frame (tr).

Description

<Desc / Clms Page number 1>

 The invention relates to the field of the processing of speech signals received by samples.

The coding of a speech signal gives rise to a bit stream which is routed through a transmission channel. At the other end of the channel, a decoder renders the speech signal in the form of a succession of sample frames for its operation.

There are several types of encoders, among which: so-called temporal coders, perform the compression of the digitized signal samples, such as the MIC or ADPCM encoders described in particular in: W. R. 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, and X. Master. 117 kHz audio coding within 64 kbit / s, IEEE Journal on Selected Areas on Communications, Vol 6-2, February 1988, pp. 283-298, and parametric coders which analyze successive frames of samples of the signal at code to extract, at each of these frames, a number of parameters which are then coded and transmitted.

In this category, we know: vocoders, described in particular in:

<Desc / Clms Page number 2>

 T. E. Tremain, "The standard linear predictive coding algorithm: LPC 10". Speech technology, April 1982, PP. 40-49, IMBE coders, described in particular in: J. C. Hardwick and J. S. Lim. "The application of the IMBE speech coder to mobile communications". Proc. of ICASSP conference, 1991, PP. 249-252, or so-called transform encoders, described in particular in: K. H. 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.

Encoders are also known which complete the coding of the parameters representative of the parametric coders by the coding of a residual temporal waveform.

- et le codeur CELP, décrit notamment dans : M. R. Schroeder et B. S. Atal,"Code-Excited Linear Prediction (CELP) : High-Quality Speech at Very Low Bit Rates", Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), vol. 3, PP. 937-940, mars 1985. In this category, there are predictive coders and in particular the family of synthesis analysis coders such as: the RPE-LTP coder, described in particular in: K. Hellwig, P. Vary, D. Massaloux, JP Petit, C. Goland and M. Rosso, "Speech codec for the European mobile radio system", GLOBECOM conference, 1989, PP. 1065-1069,

- and the CELP coder, described in particular in: MR Schroeder and BS Atal, "Code-Excited Linear Prediction (CELP): High-Quality Speech at Very Low Bit Rates", Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), vol. 3, PP. 937-940, March 1985.

<Desc / Clms Page number 3>

Pour tous ces codeurs, les valeurs codées sont ensuite transformées en un train binaire qui sera transmis sur un canal de transmission. Selon la qualité de ce canal et le type de transport, des perturbations peuvent affecter le signal transmis et produire des erreurs sur le train binaire reçu par le décodeur. Des erreurs peuvent intervenir de manière isolée dans le train binaire ou encore se produire par rafales. Dans ce dernier cas, un paquet de données binaires, correspondant à une portion complète du signal, présente des erreurs ou n'est pas reçu. De tels problèmes se rencontrent par exemple pour les transmissions par trames sur les réseaux mobiles. Ils se rencontrent aussi dans les transmissions sur les réseaux par paquets et en particulier sur les réseaux de type interne.

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. Errors can occur in isolation in the bit stream or occur in bursts. In the latter case, a binary data packet, corresponding to a full portion of the signal, has errors or is not received. Such problems are encountered, for example, for frame transmissions over mobile networks. They are also found in transmission on packet networks and in particular on internal networks.

By the terms erroneous frame, a frame or a packet which has errors on reception, which is not received, or whose reception is too late to be processed, will then be designated by the terms erroneous frame.

When the transmission system or reception modules identify erroneous frames by detecting that the received data has errors (for example on mobile networks) or that a block of data has not been received ( case of packet transmission systems for example), error concealment procedures are then implemented. These procedures make it possible to extrapolate to the decoder the samples of the signal missing from the signals and

<Desc / Clms Page number 4>

 available data from the frames preceding the erased areas.

Such techniques have been implemented mainly in the case of parametric encoders, with techniques for synthesizing 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 treatments developed rely on the technique used for the encoder and decoder, and are in fact an extension of the decoder.

Most of the predictive coding algorithms propose techniques for recovering erased frames, described in particular in: - Recommendation GSM 06. 11, "Substitution and muting of lost frames for full rate speech traffic channels".

ETSI / TC SMG, ver. : 3.0. 1. February 1992. ITU-A T Recommendation G.723. Adoul, A. Kataoka, S.

Hayashi, T. Moriya, C. Lamblin, D. Massaloux, S. Proust, P. Kroon and Y. Shoham, "Design and description of CSACELP: a toll quality 8 kb / s speech coder". IEEE Trans. on Speech and Audio Processing, Vol. 6-2, March 1998, PP. 116-130,
T. Honkanen, J. Vainio, P. Kapanen, P. Haavisto, R.

Salami, C. Laflamme and J. P. Adoul, "GSM enhanced full

<Desc / Clms Page number 5>

 coded rate speech ", Proceedings of ICASSP conference, 1997, pp. 771-774, - R. V. Cox," An improved frame erasure concealment method for ITU-T Rec. G728 ", Delayed contribution D. 107 (WP 3/16), ITU-T, January 1998, - US-5574825, US-5550543, US-5615298, US-5717822 and published application EP0673015, and - CR Watkins , JH Chen, "Improving 16 kb / s G. 728 LDCELP Speech Coder for Frame Erasure Channels", Proceedings of ICASSP conference, 1995, pp. 241-244.

The decoder is informed of the occurrence of an erased (or erroneous) frame, for example in the case of mobile radio systems by transmitting the frame erase information from the channel decoder. The recovery devices by synthesis of erased frames are intended to extrapolate the parameters of the erased frame from the last previous frame, at least, considered valid. Some parameters manipulated or encoded by predictive coders have a strong inter-frame correlation (for example, LPC parameters that represent the spectral envelope, and long-term prediction parameters for voiced sounds, for example). Because of this correlation, it is much more advantageous to reuse the parameters of the last valid frame to synthesize the erased frame than to use erroneous or random parameters.

For the CELP coding algorithm, the parameters of the erased frame are generally obtained in the following manner.

<Desc / Clms Page number 6>

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 (case of the G723.1 encoder, as described in ITU T Annex A to recommendation G. 723 1 "Silence compression scheme for dual rate speech coder for multimedia communications transmitting at 5. 3 & 6. 3 kbit / sl1).

Voicing is also detected to determine the degree of harmonicity of the signal at the level of the erased frame, as described in particular in: R. Salami, C. Laflamme, J. P. Adoul, A. Kataoka, S.

Hayashi, T. Moriya, C. Lamblin, D. Massaloux, S. Proust, P. Kroon and Y. Shoham, "Design and description of CSACELP: a toll quality 8 kb / s speech coder", IEEE Trans. on Speech and Audio Processing, Vol. 6-2, March 1998, PP. 116-130.

Thus, in the case of an unvoiced signal, an excitation signal is generated randomly. In the technique described in the previous reference, a codeword is randomly drawn, with a gain of the past excitation slightly damped. Another technique in which random selection is made in past arousal is described in: JH Chen, Cox RV, Lin YC, Jayant N and MJ Melchner, "A low-delay CELP coder for the CCITT 16 kb / s speech standard coding ", IEEE Journal on Selected Areas on Communications, Vol. 10-5, June 1992, PP. 830-849.

<Desc / Clms Page number 7>

Other techniques make use of transmitted codes, even if they are totally erroneous, as in: T. Honkanen, J. Vainio, P. Kapanen, P. Haavisto, R.

Salami, C. Laflamme and J. P. Adoul, "GSM enhanced full rate speech coded", Proc. of ICASSP conference, 1997, PP. 771-774.

In the case of a voiced signal, the LTP delay is usually the delay calculated at the previous frame, possibly with a slight "jitter" (as in R.

Salami, C. Laflamme, JP Adoul, A. Kataoka, S. Hayashi, Moriya T., Lamblin C., Massaloux D., Proust S., Kroon P. 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). The LTP gain is taken very close to 1 or equal to 1. The excitation signal is limited to the long-term prediction made on the past excitation. However, these techniques using pre-loss signal modeling generally do not allow the model to evolve over one or more consecutive lost frames to a next valid frame.

In all the examples cited above, the methods for recovering erased frames are strongly related to the decoder and use modules of this decoder, such as the signal synthesis module. They also use intermediate signals available within this decoder, as the excitation signal passed and stored during the

<Desc / Clms Page number 8>

 processing valid frames preceding erased frames.

The methods used to conceal the errors produced by lost packets during the transport of coded data by time-type coders generally make use of waveform substitution techniques such as those presented in particular in: - 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, -N. Erdöl, C. Castelluccia, A. Zilouchian, "Recovery of Missing Speech Packets Using the ShortTime Energy and Zero-Crossing Measurements" IEEE Trans. on Speech and Audio Processing, Vol. 1-3, July 1993, PP. 295-303, and - AT & T (DA Kapilow, RV Cox) "A high quality lowcomplexity algorithm for frame erasure concealment (FEC) with G. 711", Delayed Contribution D. 249 (WP 3/16), ITU, May 1999.

These techniques reconstruct the signal by selecting portions of the decoded signal before the lost portion and do not use synthesis models. Smoothing techniques are also used to avoid the artifacts produced by the concatenation of the different signals.

Other techniques use the knowledge of the frame following the lost frame, which makes it possible to have a

<Desc / Clms Page number 9>

 a priori knowledge (previous frame) and a posterior knowledge (next frame) of the signal. It is therefore possible to use this information to reconstruct the missing portion of the signal, such as the Double Side Pattern Matching (DSPM) technique described in particular in: DJ Goodman, Lockhart GB, OJ Wasem, Wong WC, "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.

This is also the case of the "DSPS" (Double Side Periodic Substitution) techniques which implement a copy of fundamental periods of the speech signal (pitch replication). The fundamental period of the speech signal corresponds to the inverse of its fundamental frequency and is an important parameter since it represents a tonal pitch of the voice.

Such techniques are described in particular in: J. TANG, F. ITAKURA, "Double Side Periodic Substitution (DSPS) Method for Recovering Missing Speech", ISSPA, pp.

544-549, August 1987, and J. TANG, "Evaluation of Double Sided Periodic Substitution (DSPS) Method for Recovering Missing Speech in Packets and Voice Communications," IEEE Computers and Communications, pp. 454 to 458.1991.

However, the number of samples to be included in the frame to be synthesized does not correspond, in principle, to a

<Desc / Clms Page number 10>

 multiple of the number of samples in a segment representing a pitch period, which usually results in the presence of audible artifacts in the reconstructed signal. In addition, these "pitch replication" techniques only apply when the lost signal is located in a voiced area. In other cases (unvoiced area), the previous frame is simply copied.

(PictureTel Transform Coder)", Contribution ITU-T, SG15/WP2/Q6, 8-9 Octobre 1996 Baltimore meeting, TD7. Une autre technique est basée sur la construction d'un modèle sinusoïdal de signal décodé, lequel sert à régénérer la partie du signal perdue. Elle est décrite notamment dans : V. N. Parikh, J. H. Chen, G. Aguilar,"Frame Erasure Concealment Using Sinusoidal Analysis Synthesis and its Application to MDCT-Based Codecs", Proc. of ICASSP conference, 2000. For transform coders, the techniques for reconstructing erased frames also rely on the coding structure used. Some techniques aim at regenerating the lost transform coefficients from the values taken by these coefficients before the erasure. A technique of this type is described in particular in: - PictureTel Corporation, "Detailed Description of the PTC

(PictureTel Transform Coder) ", ITU-T Contribution, SG15 / WP2 / Q6, 8-9 October 1996 Baltimore meeting, TD7 Another technique is based on the construction of a sinusoidal decoded signal model, which serves to regenerate the part of the lost signal.It is described in particular in: 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.

Some synthetic concealment techniques for erroneous frames have been developed in conjunction with the

<Desc / Clms Page number 11>

 channel coding. These techniques make use of information provided by the channel decoder, for example information concerning the degree of reliability of the parameters received.

However, these are approaches which differ substantially from that of the present invention which does not presuppose the existence of a channel coder, as will be discussed below. An example of this type of technique is described in particular in: T. Fingscheidt, P. Vary, "Robust speech decoding: a universal approach to a bit error concealment", Proc. of ICASSP conference, 1997, pp. 1667-1670.

The techniques used to conceal erased frames in CELP encoders have been shown to be effective, particularly on speech signals. Other types of encoders, such as transform coders, which use extrapolation techniques related to their signal representation, generally give poorer results on speech signals.

It also appears that the so-called "pitch replication" techniques described above (which use an amplitude modulated copy of a fundamental period of the speech signal), as well as the techniques based on the use of a modeling of the signal (LPC, LTP, or other), are those that give the most promising results.

<Desc / Clms Page number 12>

However, with signal modeling techniques prior to loss, it is difficult to evolve the model on one or more consecutive lost frames.

Pitch replication techniques generally introduce artifacts that result in audible frequency hopping, energy contrast, and / or ribbing edge effects of segments representing fundamental periods (or pitch periods). ).

The present invention improves the situation.

To this end, it proposes a method for synthesizing a speech signal represented by successive frames of samples, in which: a) at least one erroneous frame is identified, as well as at least one first valid frame preceding the erroneous frame; and a second valid frame succeeding the erroneous frame, b) forming a succession of synthesized blocks according to the valid frames preceding and succeeding the erroneous frame, each block comprising a number of samples calculated according to a fundamental period of the signal of speech in at least one of the valid frames, and c) constructing a synthesis frame using at least a portion of the block samples.

According to one of the advantages provided by the present invention, the treatment in the sense of the present invention

<Desc / Clms Page number 13>

 does not depend on the type of digital coding initially performed on the speech signal.

According to another advantage provided by the present invention, the valid signal, both before and after the erroneous frame, is used to construct the synthesis frame.

According to an advantage that follows from the present invention, it is possible to change, from the previous valid frame to the next valid frame, models associated with the synthesis blocks, the respective durations of these blocks corresponding to at least one period fundamental of the speech signal in the synthesis frame.

According to an advantage that follows from the present invention, it is possible to synthesize more than one substitution frame, knowing the signal before and after the erroneous frame or frames.

In a preferred embodiment, the blocks formed are speech signal synthesis blocks.

According to an advantageous characteristic of the invention, provision is made for: obtaining values of first and second parameters of a model applied to first and second portions respectively of the first frame and the second frame, and

<Desc / Clms Page number 14>

 - The formation of the samples of a block by applying to this block the aforementioned model, with intermediate parameters between the first and second parameters.

According to another advantageous characteristic of the invention, provision is made for obtaining respective information of a degree of voicing of the speech signal in the first frame and in the second frame. In this embodiment, it is advantageously taken into account that the preceding and following frames are substantially voiced or not to form the samples of the synthesis blocks.

Advantageously, samples are added or deleted in the blocks so as to obtain a total number of samples in the blocks which corresponds substantially to the number of samples to be provided in the synthesis frame.

Preferably, the binary values of the samples preceding and succeeding a deleted sample are weighted according to those of the deleted sample, and / or the binary values of an added sample as a function of those of the samples preceding and succeeding the added sample. .

Preferably, smoothing is carried out between the respective binary values of the last sample of at least one block and the first sample of at least the next block, and / or

<Desc / Clms Page number 15>

 between the respective bit values of the last sample of at least one valid frame preceding the erroneous frame and at least the first sample of the synthesis frame, and / or between the respective bit values of the last least sample of the synthesis frame and the first sample at least one valid frame succeeding the erroneous frame.

The present invention also relates to a synthesizer device of a speech signal represented by successive frames of samples, which comprises means for implementing the method above.

In a particular embodiment, the device further comprises connection means intended to be connected to a decoder of a coded data stream of speech signals, so that the device receives decoded successive frames.

In a variant, the device comprises means for receiving an encoded data stream of speech signals, as well as means for at least partial decoding of this stream.

Other features and advantages of the invention will appear on examining the following detailed description, and the accompanying drawings, in which: FIG. 1 illustrates a system for transmitting a speech signal, via a transmission channel;

<Desc / Clms Page number 16>

- la figure 2 représente schématiquement des modules d'un dispositif synthétiseur au sens de la présente invention ; - la figure 3 illustre les étapes d'un procédé d'identification d'une trame valide précédant une ou plusieurs trames erronées et d'une trame valide succédant une ou plusieurs trames erronées ; - la figure 4A représente schématiquement un signal de parole échantillonné et comportant une trame erronée Te ; - la figure 4B représente une succession de blocs de synthèse, dont une partie au moins des échantillons est destinée à être utilisée pour remplacer la trame erronée Te ; - la figure 4C représente le signal de parole synthétisé, comportant une trame reconstruite tr, en remplacement de la trame erronée Te ; - la figure 5 illustre les étapes d'un procédé d'estimation de période fondamentale du signal de parole pour une trame reconstruite tr, en fonction du degré de voisement des trames valides précédentes et suivantes ; - la figure 6 représente les étapes d'un procédé de synthèse des blocs, suivies d'un lissage des premiers et derniers échantillons de ces blocs ; - la figure 7 représente schématiquement des parties du signal destinées à être lissées, en particulier au début et à la fin de la trame reconstruite tr ; - la figure 8A représente un signal de parole, dont une trame centrale a été reconstruite, avec une trame voisée précédant la partie reconstruite et une trame voisée succédant la trame reconstruite ; - la figure 8B représente le signal de parole d'origine, qui correspond à la figure 8A ;

FIG. 2 schematically represents modules of a synthesizer device within the meaning of the present invention; FIG. 3 illustrates the steps of a method of identification of a valid frame preceding one or more erroneous frames and of a valid frame succeeding one or more erroneous frames; FIG. 4A schematically represents a sampled speech signal comprising an erroneous frame Te; FIG. 4B represents a succession of synthesis blocks, of which at least part of the samples is intended to be used to replace the erroneous frame Te; FIG. 4C represents the synthesized speech signal comprising a reconstructed frame tr, replacing the erroneous frame Te; FIG. 5 illustrates the steps of a method for estimating the fundamental period of the speech signal for a reconstructed frame tr, as a function of the degree of voicing of the preceding and following valid frames; FIG. 6 represents the steps of a method for synthesizing the blocks, followed by a smoothing of the first and last samples of these blocks; FIG. 7 schematically represents parts of the signal intended to be smoothed, in particular at the beginning and at the end of the reconstructed tr-frame; FIG. 8A represents a speech signal, a central frame of which has been reconstructed, with a voiced frame preceding the reconstructed part and a voiced frame succeeding the reconstructed frame; FIG. 8B represents the original speech signal, which corresponds to FIG. 8A;

<Desc / Clms Page number 17>

 FIG. 9A represents a speech signal, of which a central frame has been reconstructed, with a voiced frame preceding the reconstructed part and an unvoiced frame succeeding the reconstructed frame; FIG. 9B represents the original speech signal, which corresponds to FIG. 9A; FIG. 10A represents a speech signal, of which two central frames have been reconstructed, with an unvoiced frame preceding the reconstructed part and an unvoiced frame succeeding the reconstructed part; and Figure 10B shows the original speech signal, which corresponds to Figure 10A.

The drawings and the description below contain, for the most part, elements of a certain character. They can therefore not only serve to better understand the invention, but also contribute to its definition, if any.

Referring first to Figure 1 in which a digital speech signal S is, at first, encoded by a CO encoder which converts the speech signal S into a bit stream of encoded data FC. This bit stream FC is then transported by a transmission channel CL.

It may also be a GSM-type network, in which case the speech signal is represented by a succession of consecutive frames, or an Internet-type network, in which case the signal S is represented by a succession bit packets. Subsequently, the term "frame" will be used to designate a frame transmitted by GSM network,

<Desc / Clms Page number 18>

 or a packet transmitted over the Internet, or other types of quanta of samples.

A block for receiving and decoding a bit stream FC'issu of the transmission channel CL is described below. The bitstream FC 'is not always identical to the bitstream FC, since some frames or packets could be erroneous during transmission via the CL channel, or may have been lost, or received too late to be used.

In what follows, the terms "erroneous frame", a frame or an erroneous packet, or received late, or which was lost during transmission.

By way of non-limiting example, the reception and decoding block of FIG. 1 corresponds to a G711-type digital coding / decoding Internet Protocol (IP) audio conferencing terminal, which is however equipped with a synthesizer device 1 at sense of the present invention. The G711 is a telephone band coder (300-3400 Hz) with a sampling frequency of 8 kHz at 64 kbit / s. In the example, the coded frames have a duration of 30 milliseconds (which corresponds to 240 samples, with the sampling frequency of 8 kHz).

On reception, a module 3 for detecting erroneous data receives the bit stream FC'issu of the transmission channel CL. In particular, Module 3 detects

<Desc / Clms Page number 19>

 Erroneous frames Te and transmits an erroneous ITE indication signal to a synthesizer device 1, for error concealment. The valid frames Tv (non-erroneous frames) are transmitted to a decoding device 2. The synthesizer device 1 further receives at least portions of the decoded signal SD from the decoding device 2 and uses these signal portions to synthesize a reconstruction signal SR. The reconstruction signal SR is, if necessary, added to the decoded signal (at the output of the decoding device 2), so as to deliver a speech signal Ready to be used. It should be noted that the signal S'n is not always identical to the signal S, since portions of the signal could be synthesized.

Referring to FIG. 2, the synthesizer device 1 comprises a memory 10 intended to store a selected number of valid frames coming from the decoder 2. This memory 10 cooperates with a block synthesizer 11 which receives the erroneous frame indication signal. ITE, to start the synthesis of BL blocks. The synthesizer device 1 further comprises a module 12 for smoothing the synthesis samples produced to replace the samples of the erroneous frame. The synthesizer device 1 provides, at the output of the module 12, the reconstructed signal SR.

Thus, after decoding the valid data, there are stored in the memory 10 decoded samples in sufficient number to synthesize erased or erroneous frames, thereafter. Preferably, the memory 10 stores two

<Desc / Clms Page number 20>

trames valides précédant la trame à reconstruire (soit 60 millisecondes de signal avant la trame synthétisée à jouer à l'instant t), ainsi qu'une trame valide qui succède la trame à synthétiser (soit 30 millisecondes du signal après la trame synthétisée à jouer à l'instant t). Le contenu temporaire de la mémoire 10 comprend donc 960 échantillons (480 échantillons avant l'instant t et 240 échantillons après l'instant t pour une fréquence d'échantillonnage de 8 kHz). Dans ce mode de réalisation, il est pris une trame de retard, soit donc 30 millisecondes, pour avoir des informations a posteriori sur le signal futur. Les échantillons de la dernière trame reçue sont placés en fin du contenu temporaire de la mémoire 10 et les échantillons des trames les plus anciennes sont décalés, selon le fonctionnement classique d'une mémoire à registre de décalage.

valid frames preceding the frame to be reconstructed (ie 60 milliseconds of signal before the synthesized frame to play at time t), as well as a valid frame that succeeds the frame to be synthesized (ie 30 milliseconds of the signal after the synthesized frame to play at time t). The temporary content of the memory 10 therefore comprises 960 samples (480 samples before the instant t and 240 samples after the instant t for a sampling frequency of 8 kHz). In this embodiment, a delay frame, ie 30 milliseconds, is taken to have a posteriori information on the future signal. The samples of the last received frame are placed at the end of the temporary content of the memory 10 and the samples of the oldest frames are shifted according to the conventional operation of an offset register memory.

Thus, with reference to FIG. 2, the synthesizer device 1 thus receives two pieces of information: the first piece of information corresponds to the samples of the frames received previously, and the second piece of information corresponds to the number of successive frames lost or erroneous, to be synthesized.

In the example described, the number of successive frames that the device is able to synthesize is limited to two successive erroneous frames. Of course, by choosing a memory 10 with sufficient storage capacity, it is possible to synthesize more than two successive erroneous frames, depending on the desired quality of the restored signal.

<Desc / Clms Page number 21>

Referring to FIG. 3, the synthesizer device 1 receives an erroneous frame indication relative to the frame Ti, in step 30. Following the tests 31 and 33, if the following two frames are also erroneous, the three consecutive frames Ti, Ti + 1, Ti + 2 are replaced preferentially by a comfort noise BCO, in step 35. On the other hand, if, following the test 31, it turns out that the following frame Ti + i is valid, we designate the previous frame Shot and the following frame Ti + 1 as valid frames which will be used as a basis, in step 32, to reconstruct the erroneous frame Ti. On the other hand, if the two consecutive frames Ti and Ti + 1 are erroneous, while the following frame Ti + 2 is valid, the latter frame Ti + 2 and the frame Ti-1 which precedes the first erroneous frame Ti, are chosen at step 34 as valid base frames for the reconstruction of the two erroneous frames Ti and Ti + 1.

As a variant of the replacement of the three successive erroneous frames Ti, Ti + 1 and Ti + 2 by a comfort noise BCO, provision is made for a reset of the temporary content of the memory 10 (zeroing of the sample binary values). However, the replacement of erroneous frames of long duration, by a comfort noise, can be better perceived in terms of quality of communication by the recipient of the speech signal.

A particular embodiment is described below in which the number of samples that is assigned to each

<Desc / Clms Page number 22>

 synthesis block is substantially constant from one synthesis block to another.

In the example of FIG. 5, a first estimate of a fundamental period P of the speech signal in the frame to be synthesized is calculated, as a function of the valid frames preceding (t-1) and succeeding (t + 1) the frame wrong. As will be seen below, the number of samples per block depends on this first estimate P.

In a preferred embodiment, the first estimate of the fundamental period P is made as a function of the voicing of the valid frames preceding (t-1) and succeeding (t + 1) the erroneous frame.

The processing starts with a detection 50 of VO voiced or voiceless voices NV in the last data stored in the memory 10. To perform this detection, it is advantageous to use a normalized correlation, for example of the type described in: WB KLEIJN, KK PALIWAL, "Speech Coding and Synthesis", ELSEVIER, 1995.

Preferably, the technique of detection of voicing in the frames and determining their fundamental period (or "Ptchs" PO and PI) which is used here is that described in the reference: H. Kobayashi, T. Shimamura, "A weighted autocorrelation method for pitch extraction of noisy speech ", Proc; of ICASSP conference, 2000,

<Desc / Clms Page number 23>

 which implements the joint use of the normalized autocorrelation function and the AMDF (Average Magnitude Difference Functior) function, in order to improve the estimation of the pitch in the presence of noisy speech signals. Preferably, the estimation of the pitch PO and PI is constrained to maximum and minimum values which respectively correspond to a fundamental frequency of the speech signal of 400 Hz (20 samples for a sampling frequency of 8 kHz) and of 66 Hz (120 samples for a sampling frequency of 8 kHz).

Of course, any other technique for obtaining the voicing period in a speech signal frame can be used here.

Depending on the degree of frame voicing (t-1) and (t + 1), decisions (which follow tests 51, 52 and 55) are made as to which frame synthesis technique to use. In the example described, four cases are distinguished: if the frames (t1) and (t + 1) are voiced (output 0 of the test 52), a first estimate of the fundamental period P of the signal is calculated (step 53) in the frame to be synthesized, based on the fundamental periods PO of the valid frame preceding the erroneous frame and PI of the valid frame succeeding the erroneous frame. The function making it possible to obtain the period P can be the average between the respective pitches PO and PI,

<Desc / Clms Page number 24>

 or the maximum between these two pitches, or the minimum between these two pitches, or any combination of PO and PI; if the frame (t-1) is voiced, while the frame (t + 1) is unvoiced (output N of test 52), the frame to be synthesized probably corresponds to a word end (speech / noise transition) and the estimate of P preferably corresponds to the period PO of the voiced valid frame (step 54); if the frame (t-1) is unvoiced, while the frame (t + 1) is voiced (output 0 of the test 55), the frame to be synthesized probably corresponds to a start of a word (noise / speech transition) and the estimation of P preferably corresponds to the pitch PI of the voiced valid frame (step 57); and if the frames (t-1) and (t + 1) are unvoiced (output N of test 55), the frame to be synthesized probably corresponds to a noise period and, preferably, the synthesis of this frame is done by a copy of the previous frame (t-1) (step 56). As a variant, the synthesis of the erroneous frame can be done by copying the next frame (t + 1).

In what follows, a preferred embodiment is described for estimating the NBL number of blocks to be synthesized, as well as the TBL number of samples to be provided in each block, here supposed constant from one block to another.

Referring to FIG. 6, after detection and estimation, at step 60, PO and / or PI periods of the valid frames surrounding the erroneous frame, a first estimate of the period P in the frame of the frame is calculated.

<Desc / Clms Page number 25>

synthèse, en fonction des périodes PO et/ou PI comme décrit ci-avant.

synthesis, according to the periods PO and / or PI as described above.

The number of NBL blocks to be synthesized is a function of this first estimate of the period P and of the number NTR of samples lost in the erroneous frame (obtained in step 62).

Cependant, les périodes PO et PI (si les trames valides sont toutes deux voisées) ne sont pas forcément identiques et il faut définir un nombre d'échantillons par bloc TBL qui soit adapté. Avantageusement, la valeur du nombre de blocs NBL, en relation avec le nombre d'échantillons par bloc TBL comme on le verra plus loin, est optimisée par minimisation de la valeur absolue de la quantité suivante :

En fonction de cette valeur minimum de AN calculée à l'étape 72 et prise pour la valeur i, on incrémente le nombre de blocs NBL de cette valeur de i (NBL = NBL+i à l'étape 74). Finalement, la valeur de NBL correspond à l'entier le plus proche du rapport NTR/P. La valeur AN The number of NBL blocks is estimated at step 61, depending on the ratio:

However, the periods PO and PI (if the valid frames are both voiced) are not necessarily identical and it is necessary to define a number of samples per block TBL that is adapted. Advantageously, the value of the number of NBL blocks, in relation to the number of samples per TBL block, as will be seen below, is optimized by minimizing the absolute value of the following quantity:

Based on this minimum value of AN calculated in step 72 and taken for the value i, the number of NBL blocks of this value of i is incremented (NBL = NBL + i in step 74). Finally, the value of NBL corresponds to the integer closest to the ratio NTR / P. AN value

<Desc / Clms Page number 26>

 thus estimated corresponds to the number of samples to be added to the blocks to form the synthesis frame (if the aforementioned difference AN is positive), or else to delete to form the synthesis frame (if the difference AN is negative).

dont la valeur est arrondie à l'entier le plus proche. In step 63 of FIG. 6, the number of samples per TBL block to be synthesized is then estimated. This number is given by the report:

whose value is rounded to the nearest integer.

Finally, the duration of the TBL blocks (given by the ratio of the number of samples per TBL block thus estimated on the sampling frequency) is representative of a fundamental period of the speech signal in the synthesis frame tr.

In what follows, in a preferred embodiment, the synthesis of the blocks, whose NBL number and the number of samples per block TBL have been calculated beforehand, is described.

précédée d'une trame (t-l) valide et voisée, de pitch PO, et succédée d'une trame (t+l) valide et voisée, de pitch PI. Après avoir détecté et estimé les pitchs PO et PI, les échantillons d'un segment SEG (PO) de durée correspondant à In the example shown in FIG. 4A, the received and decoded speech signal has an erroneous frame Te,

preceded by a frame (tl) valid and voiced pitch PO, and succeeded by a frame (t + 1) valid and voiced pitch PI. After detecting and estimating the pitch PO and PI, the samples of a segment SEG (PO) of duration corresponding to

<Desc / Clms Page number 27>

 the period PO in the frame (t-1), as well as the samples of a segment SEG (P1) of duration corresponding to the period PI in the frame (t + 1), are stored in memory 10, as vectors Xo and Xi.

Preferably, these segments SEG (PO) and SEG (P1) respectively comprise the last samples and first samples of the previous (t-1) and following (t + 1) valid frames. A modeling of these signal elements is then carried out, for example by using a spectral transform (Fourier transform, discrete cosine transform, discrete sinus transform, or other), or by applying an LPC model (of the type used in FIG. CELP coding). The models that result from this modeling are noted below Mod (PO) and Mod (pal).

- si les deux trames valides précédente et suivante sont voisées, on sélectionne à l'étape 53 un segment dans la trame valide précédente, dont la durée correspond à une période PO de la trame valide précédente, ainsi qu'un segment dans la trame valide suivante, dont la durée correspond à la période PI dans la trame valide suivante ; - si seule la trame précédente est voisée, on sélectionne, à l'étape 54, un segment dans la trame valide précédente dont la durée correspond à la période PO de la trame valide précédente, ainsi qu'un segment dans la trame valide suivante (non voisée) dont la durée correspond à la période PO de la trame valide précédente (voisée) ; More particularly, referring again to FIG. 5, three cases are preferentially distinguished:

if the two preceding and following valid frames are voiced, in step 53 a segment is selected in the previous valid frame whose duration corresponds to a period PO of the previous valid frame, as well as a segment in the valid frame next, whose duration is the PI period in the next valid frame; if only the previous frame is voiced, a segment in the previous valid frame whose duration corresponds to the period PO of the previous valid frame and a segment in the following valid frame are selected in step 54 ( unvoiced) whose duration corresponds to the period PO of the previous valid frame (voiced);

<Desc / Clms Page number 28>

- de même, si seule la trame valide suivante est voisée, on sélectionne, à l'étape 57, un segment comprenant les derniers échantillons de la trame valide précédente (non voisée) dont la durée correspond à la période PI dans la trame suivante (voisée), ainsi qu'un segment comportant les premiers échantillons de la trame valide suivante et dont la durée correspond à la période PI de la trame valide suivante.

- Similarly, if only the following valid frame is voiced, in step 57, a segment is selected comprising the last samples of the previous valid frame (unvoiced) whose duration corresponds to the period PI in the following frame ( voiced), as well as a segment comprising the first samples of the next valid frame and whose duration corresponds to the period PI of the next valid frame.

These models are then applied to the modeling described above. The resulting models are denoted Mod (PO) and Mod (Pl) respectively for the segments of the previous frame (t-1) and the following frame (t + 1).

In a preferred embodiment, the modeling applied corresponds to a discrete cosine spectral transform (DCT). 120 respective coefficients resulting from the spectral transformation of the Xo and Xi vectors are then calculated. These 120 coefficients are respectively stored as Mod (PO) and Mod (Pl) vectors.

In the example shown in FIG. 4B, the first estimate of the period P of the synthesis frame corresponds to the average of the PO periods of the previous valid frame and PI of the next valid frame. The optimization criterion of the value AN imposes the number of NBL blocks (six blocks in the example represented in FIG. 4B), as well as the number of samples per block TBL (five samples per block in the example represented on FIG. Figure 4B, which is the average of the number

<Desc / Clms Page number 29>

 samples in a PO period and the number of samples in a PI period).

Of course, FIGS. 4A to 4C illustrate the signals only as explanatory diagrams. The segments SEG (PO) and SEG (Pl) comprise seven and five samples. In practice, they should comprise at least twenty samples, which corresponds to the fundamental frequency 400 Hz, the highest in a speech signal.

A preferred embodiment is described below for forming the samples of the synthesis blocks.

Referring again to FIG. 6, after obtaining in step 66 the parameters of the Mod (PO) and Mod (P1) models from the SEG (PO) and SEG (P1) segments (steps 64 and 65) , in the NBL block sequence, an index n (0 <n NBL-1) is assigned to each block. In step 67, the parameters S (n) k of a block are synthesized as a function of the Mod (PO) and Mod (Pl) models, of the index n of this block and of the number of blocks to be synthesized NBL. In an advantageous embodiment, the function S (n) corresponds to an interpolation of the parameters of the models Mod (PO) and Mod (Pl), as described below.

For a block of index n, it is estimated, by preferentially linear interpolation, intermediate parameters Mod (S (n)) which should be attributed to this block, according to the models Mod (PO) and Mod (Pl).

<Desc / Clms Page number 30>

Si, pour estimer les modèles Mod (PO) et Mod (Pl), il a été utilisé préalablement une transformée spectrale, telle qu'une transformée en cosinus discrète, on procède, après l'estimation des paramètres intermédiaires Mod [S (n)] à une transformée spectrale inverse (telle qu'une transformée en cosinus discrète inverse IDCT) pour obtenir les valeurs binaires des échantillons S (n) k que doit comprendre chaque bloc de synthèse.

So, for all ke [0, TBL-1],

If, in order to estimate the Mod (PO) and Mod (Pl) models, a spectral transform, such as a discrete cosine transform, has been used beforehand, proceed after estimating the intermediate parameters Mod [S (n) ] to an inverse spectral transform (such as an IDCT inverse discrete cosine transform) to obtain the binary values of the samples S (n) k to be included in each synthesis block.

Thus, for all k <= [0, TBL-1], S (n) k = IDTC (Mod [S (n) k D.

When all the blocks are synthesized (NBL blocks in all), it is necessary to delete or add the number of samples defined by the optimization criterion AN. The position of the samples to be added or removed is here chosen randomly over the whole number of samples already synthesized. For the addition of a sample, the binary values of this sample correspond to a weighting (preferably to the average) of the binary values of the samples that surround it. For a deletion of a sample, the following sample in the frame to be synthesized is preferably assigned binary values which correspond to a weighting (preferably to the average) of the binary values of the deleted sample and of its own binary values before the deletion.

<Desc / Clms Page number 31>

In the following, we describe, in a preferred embodiment, a smoothing of the binary values of the beginning and end of block samples, applied to re-stick each of the blocks.

Pour tout ne [l, NBL-l],

So that no artefact comes to degrade the quality of the reconstituted signal, a smoothing is carried out between the binary values of the last samples of a block and the binary values of the first samples of a following block. Preferably, the smoothing performed is a smoothing type called "variable forgetting factor". With reference to FIG. 6, step 68 provides for the replacement of the binary values of the samples S (n) i by the smoothed values SI (n) i, in particular for the start and end of block samples. Preferably, this smoothing is carried out according to the following equations:

For all [l, NBL-l],

Dans ces équations, NLIS correspond au nombre d'échantillons sur lequel chaque lissage entre deux blocs est effectué. Il s'agit d'un entier inférieur à la taille de chaque bloc TBL. Le terme a (i) correspond au facteur d'oubli du lissage effectué. Avantageusement, de bons

In these equations, NLIS is the number of samples on which each smoothing between two blocks is performed. This is an integer less than the size of each TBL block. The term a (i) corresponds to the forgetting factor of the smoothing performed. Advantageously, good

<Desc / Clms Page number 32>

résultats sont obtenus pour un lissage faisant intervenir cinq échantillons (NLIS=5).

results are obtained for smoothing involving five samples (NLIS = 5).

It should be noted that the first synthesis block (n = 0) does not undergo smoothing at the beginning of the block. Indeed, a smoothing of the same type is performed between the frames preceding and following the erroneous frame, as shown in FIG.

Referring to FIG. 6, step 69 provides for the use of the last samples of the previous frame (t-1) in step 70, as well as the use of the first samples of the following frame (t + 1 ) in step 71, to smooth the transition with the values of the first samples and the last samples of the synthesis frame. In step 69, the forgetting factor a (i) and the smoothed values SR '(k) i which replace the old values of these samples SR (k) i are then calculated.

Like the smoothing between blocks, the smoothing between reconstructed frames and valid frames is carried out according to the following equations, given by way of example:

<Desc / Clms Page number 33>

 The number of samples to smooth NLIS is strictly less than the number of samples per NTR frame. In the preceding equations, we note that it is the first points of the next valid frame (t + 1) that are smoothed. Alternatively, it may, of course, be expected to smooth rather the last samples of the synthesis frame tr.

Advantageously, if the smoothing between frames is performed on the values of five samples (NLIS = 5), no artifact degrades the quality of the reconstituted signal SR, finally obtained in step 73 of FIG. variable forgetting applied to glue the synthesis frame tr frames that frame, as shown in Figure 7, thus provides a satisfactory result.

In the example shown in FIG. 4C, the tr synthesis frame was constructed from the TBL samples in the NBL blocks of FIG. 4B, in particular by selecting only a portion of these samples (by deleting two samples in FIG. example shown).

The values of the intermediate samples between the valid frames and the synthesis frame were smoothed, as previously described.

Reference is now made to FIG. 8A, which represents a reconstructed female speech signal, to be compared with the same speech signal, without reconstruction (FIG. 8B).

In the example shown, the reconstituted frame is

<Desc / Clms Page number 34>

 preceded and succeeded by valid voiced frames. Referring to Figure 8A, it was necessary to synthesize seven successive blocks to reconstruct the tr-frame. Comparing FIGS. 8A and 8B, there is a satisfactory regularity of the speech signal and no non-linearity. In particular, very few differences in the form of the signals are remarkable between the synthesis frame and the frame of the original signal without loss.

In the example shown in FIG. 9A, the reconstructed frame succeeds a valid voiced frame and precedes a valid unvoiced frame. It is therefore a loss of frame during a speech / noise transition (end of word).

Here, the speech signal is masculine. Referring to Figure 9A, the transition is smooth on the four synthesis blocks. The modification of the voiced signal to the unvoiced signal is performed without any non-linearity. Nevertheless, a difference in the extension of the voicing of the synthesis frame is to be noted in FIG. 9A. Indeed, while the original signal is voiced over a longer period, the voicing is gradually attenuated in the synthesis frame.

Above, an example of synthesis of a single tr substitution frame has been described in detail. If two frames received successively are erroneous, the number of samples lost in all corresponds to two frame sizes (2 * NTR) and the synthesis of the blocks is carried out as described above, using the first valid frame

<Desc / Clms Page number 35>

 preceding (t-1) and the next first valid frame (t + 2), as shown in step 34 of FIG.

FIGS. 10A and 10B show the case of a loss of two successive frames framed by an unvoiced frame and a voiced frame. It is therefore a noise / speech transition (beginning of the word). Here, the speech signal is masculine. Referring to FIG. 10A, it can be seen that the transition between the unvoiced signal and the voiced signal is regular over the duration of the two reconstructed frames. However, the voicing of the synthesis frame starts from the second block (sample 300) in the first synthesis frame, whereas, in the original signal, the voicing only begins with the sample 500 (corresponding to the fifth block of synthesis). However, the evolution of the signal seems satisfactory when listening.

Thus, more generally, the present invention offers an excellent quality of synthesis between two voiced frames. It also offers a satisfactory synthesis of a transition signal between a voiced area and an unvoiced area. According to another advantage provided by the present invention, it is possible to generate several synthetic frames (two successive frames as shown in FIG. 10A, or more).

Of course, the present invention is not limited to the embodiment described above by way of example it extends to other variants.

<Desc / Clms Page number 36>

 Thus, it will be understood that the synthesizer device within the meaning of the present invention may as well be separated from the decoder, as shown in FIG. 1, or may also be an integral part of a decoder, as a synthesizer module for missing or erroneous frames.

In the above, the synthesizer device operates on the basis of decoded frames. As a variant, especially when the CELP coding is used, the synthesizer device can operate on partially coded data, in particular between two intermediate decoding steps. For example, the synthesizer device can obtain the fundamental periods in the previous and next valid frames (pitchs PO and PI), as well as parameters LPC (Mod (PO) and Mod (Pl)) associated with these periods or segments of multiple periods of these periods, this information is generally available in the bit stream encoded by a type CELP encoding (eg type G 723.1). The synthesis can therefore be carried out by estimating appropriate block sizes as a function of the pitch periods PO and PI which are given in the coded bitstream, the synthesis of the blocks then taking place by an interpolation from Mod (PO) to Mod. (Pl).

The present invention is advantageously applicable to any other type of coding. It finds for example an interesting application, but not limited to, the synthesis of frames encoded initially by time coders whose structure lends itself less well a priori to the concealment of error packets.

<Desc / Clms Page number 37>

 A particular embodiment has been described above in which the number of samples assigned to each synthesis block is substantially constant from one synthesis block to another. In a variant, the first synthesis block and the last synthesis block comprise a number of samples respectively corresponding to a first period PO (if the previous valid frame is voiced) and to a second period PI (if the next valid frame is voiced), while the number of samples per block varies, preferably continuously, in the succession of blocks. This embodiment finds an advantageous application to IMBE coding.

In the example described above, the choice of samples to be added or deleted to obtain a number of samples in the synthesis frame corresponding to the number of samples lost from the erroneous frame, is made randomly. In a more sophisticated variant, the detection of valid voiced or unvoiced frames is used as follows: if the two preceding and following valid frames are voiced, the deletion or the addition of points is done randomly; if only one of the frames is voiced (start of word or end of word), the deletion or addition of samples is preferably carried out in a part of the synthesis frame which is adjacent to the valid unvoiced frame . Nevertheless, an advantageously random character is retained for adding or deleting samples. Preferably, the addition or deletion of the samples does not occur at intervals

<Desc / Clms Page number 38>

 signal, which could lead to audible overdetermination.

In the example described above, a linear interpolation between Mod (PO) and Mod (Pl), to synthesize the blocks, gives satisfactory results in listening when the valid frames are voiced on either side of the wrong frame. Nevertheless, especially if one of the frames is unvoiced, the synthesis of the blocks can be based on a different model of a linear interpolation. Indeed, in case of end of word (or beginning of word), it is not necessary to continue too long (or to start too early) the voicing during the synthesis of the substitution frame. Thus, in a more sophisticated variant, in the allocation of the parameters Mod [S (n)] to synthesize the blocks, the degree of voicing of the preceding and following valid frames is taken into account.

Typically, if one of the frames is unvoiced, the coefficients of its model will be preponderant for the estimation of the parameters associated with the blocks.

The present invention finds an interesting application, but not limited to the processing of speech signals transmitted by telephony.

Claims (29)

1. A method of synthesizing a speech signal represented by successive frames of samples, wherein: a) identifying at least one erroneous frame (Te) and at least one first valid frame (t-1); ) preceding the erroneous frame and a second valid frame succeeding the erroneous frame (t + 1), b) forming a succession of synthesized blocks (NBL) as a function of the valid frames preceding (t-1) and succeeding (t + 1) the erroneous frame, each block comprising a number of samples (TBL) calculated according to a fundamental period of the speech signal in at least one of the valid frames (PO, PI), and c) a frame of synthesis (tr) using at least a portion of the samples of the blocks. The method of claim 1, wherein the blocks formed are speech signal synthesis blocks. 3. Method according to one of claims 1 and 2, wherein: - one obtains values of first (ModPO) and second parameters (ModPl) of a model applied to first and second portions (SEG (PO), SEG (P1)) respectively of the first frame and the second frame, and - the samples of a block (S (n)) are formed by applying to this block said model with intermediate parameters (Mod [S (n)] ) between the first and second parameters. <Desc / Clms Page number 40> 4. Method according to claim 3, in which each block of the succession is identifiable by a block index (n), and in which the intermediate parameters of a block are estimated as a function of the index of the block (n), the total number of blocks in the succession (NBL) and said first and second parameters (ModPO, ModPl). 5. Method according to claim 4, wherein the intermediate parameters of a block are estimated by involving an interpolation between said first and second parameters. 6. Method according to one of claims 3 to 5, wherein said model comprises the allocation of spectral parameters (DCT) associated with a portion of speech signal. 7. Method according to one of claims 3 to 6, wherein one obtains binary values of the samples of a block by involving an inverse spectral transform (IDCT) applied to the intermediate parameters of this block. 8. Method according to one of the preceding claims, wherein there is obtained respective degree of voicing information (VO / NV) of the speech signal in the first frame and in the second frame. The method of claim 8, wherein, if the speech signals are substantially unvoiced in the first and second frames, the frame of synthesis <Desc / Clms Page number 41>  substantially corresponds to one of the first or second valid frames (56). 10. Method according to one of the preceding claims, wherein the number of samples per block (TBL) is substantially constant in said succession of blocks, and in which is calculated an estimate (P) of a fundamental period of the signal of speech in the synthesis frame according to a fundamental period of the speech signal in at least one of the valid frames (PO, PI).  The method of claim 10, taken in combination with one of claims 8 and 9, wherein, if the speech signal is substantially voiced in only one of the first and second valid frames, the estimate of the period (P) in the synthesis frame substantially corresponds to a period value of the speech signal in the valid frame that is voiced (54; 57). The method according to one of claims 8 to 11, taken in combination with one of claims 3 to 7, wherein, if the speech signal is substantially voiced in only one of the first and second valid frames, the first and second portions are of durations corresponding to said period of the speech signal in the voiced valid frame (54; 57). The method of claim 10, taken in combination with one of claims 8, 9, 11 and 12, wherein, if the speech signal is substantially voiced in the first and second frames, with a first <Desc / Clms Page number 42>  fundamental period (PO) in the first frame (t-1) and a second fundamental period (PI) in the second frame (t + 1), the estimate of the period (P) in the synthesis frame corresponds substantially to a a value selected from a set including the average of the first and second periods, the maximum between the first and second periods and the minimum between the first and second periods (53).

14. The method according to one of claims 8 to 13, taken in combination with one of claims 3 to 7, wherein the first and second portions are respective durations corresponding to the first and second periods (53). 15. Method according to one of claims 10 to 14, wherein determining a number of blocks (NBL) to form corresponding to the integer closest to a fraction of the number of samples to be provided in the synthesis frame (NTR) on the number of samples in the estimated period (P) of the synthesis frame (tr). 16. The method of claim 15, wherein the number of samples provided per block (TBL) corresponds to the rounding to the integer closest to the fraction of the number of samples to predict in the synthesis frame (NTR). ) on the number of blocks to be formed (NBL). 17. Method according to one of the preceding claims, wherein one adds or deletes samples in the blocks so as to obtain a number of samples. <Desc / Clms Page number 43>  total in the blocks corresponding substantially to the number of samples (NTR) to be provided in the synthesis frame. 18. The method of claim 17, taken in combination with one of claims 15 and 16, wherein it is estimated a difference (AN) between the number of samples to predict in the synthesis frame (NTR) and the number of blocks to be formed (NBL) multiplied by the number of samples in the estimated period (P) for the synthesis frame, and in which: - only a part of the samples are kept in the blocks to construct the synthesis frame if said difference (AN) is negative, or - we keep all the samples of the blocks, to which we add additional samples, to build the synthesis frame, if the difference (AN) is positive. 19. Method according to one of claims 17 and 18, wherein the samples to be added or removed are selected substantially randomly. 20. The method according to one of claims 17 to 19, wherein each additional sample is assigned binary values corresponding to a weighting of respective binary values of samples preceding and succeeding the additional sample in the synthesis frame. 21. Method according to one of claims 17 to 20, wherein neighboring samples, preceding and / or succeeding <Desc / Clms Page number 44>  deleted samples are assigned binary values corresponding to a weighting of the binary values of the neighboring samples before deletion and the binary values of the deleted sample. 22. Method according to one of the preceding claims, wherein at least one smoothing is performed between the respective binary values of the last sample of a block and the first sample of the next block. 23. Method according to one of the preceding claims, wherein smoothing is carried out between the respective binary values of the last sample at least one valid frame (tl) preceding the erroneous frame and the first sample at least of the frame of synthesis (tr), and / or between the respective binary values of the last sample at least of the synthesis frame (tr) and the first sample of at least one valid frame succeeding the erroneous frame (t + 1). 24. Method according to one of claims 22 and 23, wherein performs a smoothing with variable forgetting factor. 25. The method according to one of claims 22 to 24, wherein the smoothed binary values of a sample are calculated according to a weighting (a (i)) with the binary values of at least one preceding sample. 26. Synthesizer device of a speech signal represented by successive frames of samples, <Desc / Clms Page number 45>  characterized in that it comprises means for implementing the method according to one of claims 1 to 25. 27. Device according to claim 26, characterized in that it further comprises connection means (SD) intended to be connected to a decoder (2) of an encoded data stream of speech signals (FC '), and in that said connecting means is adapted to receive a succession of decoded frames. 28. Device according to claim 26, characterized in that it further comprises means for receiving a stream of coded speech signal data (FC '), and at least partial decoding means (COD) of this flow. 29. Device according to one of claims 26 to 28, characterized in that it comprises a memory capable (10) of storing at least three frames, which allows to synthesize at least two successive erroneous frames.