EP1526508A1 - Method for the selection of synthesis units - Google Patents

Abstract

The method involves determining a value of a mean fundamental frequency for information segment e.g. speech segment. A sub-set of synthesis units whose mean fundamental frequency values are closer to the fundamental frequency value, is selected. Multiple proximity criteria are applied to the selected synthesis units for determining a synthesis unit representative of the information segment. The proximity criteria includes the fundamental frequency or pitch, spectral distortion, and/or energy profile. An independent claim is also included for the utilization of a synthesis unit selecting method for the selection and/or coding of the synthesis units for a very low bandwidth voice coder.

Description

The invention relates to a method for selecting units of synthesis.

It concerns for example a selection and coding method of synthesis units for a very low speed speech coder, for example less than 600 bits / sec.

The indexing techniques of natural speech units have recently allowed the development of synthesis systems from the particularly powerful text. These techniques are now studied in the context of very low speech rate coding, together with algorithms borrowed from the field of speech recognition, Ref [1-5]. The main idea is to identify in the speech signal to be coded a almost optimal segmentation in elementary units. These units can be units obtained from a phonetic transcription, which has the disadvantage of having to be corrected manually for an optimal result, or automatically according to spectral stability criteria. From this type of segmentation, and for each segment, we look for unity synthesis in a dictionary obtained during a phase prior learning, and containing reference synthesis units.

The coding scheme used is to model the space acoustic of the speaker (or speakers) by Markov models hidden (HMM or Hidden Markov Models). These dependent models or independent of the speaker are obtained during a learning phase prior to using algorithms identical to those used in the speech recognition systems. The essential difference lies in the fact that the models are learned on vectors grouped by classes automatically and not in a supervised way from a phonetic transcription. The learning procedure then consists of automatically obtain the segmentation of learning signals (for example using the so-called temporal decomposition method), group segments obtained in a finite number of classes corresponding to the number of HMM models that we wish to build. The number of models is directly related to the desired resolution for represent the acoustic space of the speaker (s). Once obtained, these models make it possible to segment the signal to be coded using an algorithm from Viterbi. Segmentation allows to associate to each segment, the index of class and its length. This information is not sufficient to model spectral information, for each of the classes a realization of spectral trajectory is selected from among several so-called synthesis. These units are extracted from the learning base at its segmentation using HMM models. It is possible to take into account context for example by using multiple subclasses allowing for take into account transitions from one class to another. A first index indicates the class to which the segment belongs, a second index specifies the subclass to which it belongs as the index of class of the previous segment. The subclass index is therefore not transmit, and the class index must be stored for the next segment. The subclasses thus defined make it possible to take into account the different transitions to the class associated with the segment in question. To the information spectral one adds the information of prosody, that is to say the value of pitch and energy parameters and their evolutions.

In order to achieve a very low bitrate encoder, it is necessary to optimize the bit allocation and thus the bit rate between parameters associated with the spectral envelope and the prosody information. A classic method is initially to select the unit closest to a spectral point of view then, once the unit selected, to code the prosody information, either independently of the selected unit.

The method according to the present invention proposes a novel method of selecting the closest synthesis unit together with modeling and quantification of additional information necessary at the level of the decoder for the reproduction of the speech signal.

pour un segment d'information considéré :

• déterminer la valeur F0 de la fréquence fondamentale moyenne pour le segment d'information considéré,
• sélectionner un sous-ensemble d'unités de synthèse défini comme étant celui dont les valeurs moyennes de pitch sont les plus proches de la valeur de pitch F0,
• appliquer un ou plusieurs critères de proximité aux unités de synthèse sélectionnées pour déterminer une unité de synthèse représentative du segment d'information.

The invention relates to a method for selecting synthesis units of information that can be decomposed into synthesis units. he
has at least the following steps:

for a segment of information considered:

• determine the value F0 of the average fundamental frequency for the segment of information considered,
• selecting a subset of synthesis units defined as being the one whose average pitch values are closest to the pitch value F0,
• applying one or more proximity criteria to the selected synthesis units to determine a summary unit representative of the information segment.

For example, information is a segment of speech to code and one uses as proximity criteria the fundamental frequency or pitch, or the spectral distortion, and / or the energy profile and a step of merge of the criteria used to determine the synthesis unit representative.

The method comprises for example a coding step and / or a pitch correction step by modifying the synthesis profile.

The step of encoding and / or correcting the pitch may be a Linear transformation of the pitch profile.

The method is for example used for the selection and / or the coding of synthesis units for a very low bit rate speech coder.

• le procédé permet d'optimiser le débit alloué à l'information de prosodie dans le domaine de la parole.
• il permet de conserver, lors de la phase de codage, l'intégralité des unités de synthèse déterminées lors de la phase d'apprentissage avec cependant un nombre de bits constant pour coder l'unité de synthèse.
• Dans un schéma de codage indépendant du locuteur, ce procédé offre la possibilité de couvrir l'ensemble des valeurs de pitch possibles (ou fréquences fondamentales) et de sélectionner l'unité de synthèse en tenant compte en partie des caractéristiques du locuteur.
• La sélection peut s'appliquer à tout système basé sur une sélection d'unités et donc aussi à un système de synthèse à partir du texte.

The invention particularly has the following advantages:

• the method optimizes the bit rate allocated to prosody information in the speech domain.
• it makes it possible to retain, during the coding phase, the entirety of the synthesis units determined during the learning phase, with however a constant number of bits for coding the synthesis unit.
• In a speaker-independent coding scheme, this method offers the possibility of covering all of the possible pitch values (or fundamental frequencies) and of selecting the synthesis unit taking into account in part the characteristics of the speaker.
• The selection can be applied to any system based on a selection of units and therefore also to a system of synthesis from the text.

• La figure 1 un schéma de principe de sélection de l'unité de synthèse associée au segment d'information à coder,
• La figure 2 un schéma de principe d'estimation des critères de similarité pour le profil du pitch,
• La figure 3 un schéma de principe d'estimation des critères de similarité pour le profil énergétique,
• La figure 4 un schéma de principe d'estimation des critères de similarité pour l'enveloppe spectrale,
• La figure 5 un schéma de principe du codage du pitch par correction du profil de pitch de synthèse.

Other features and advantages of the invention will appear better on reading the description which follows of an exemplary non-limiting embodiment of the appended figures which represent:

• FIG. 1 is a block diagram for selecting the synthesis unit associated with the information segment to be coded,
• FIG. 2 is a schematic diagram of estimation of the similarity criteria for the pitch profile,
• Figure 3 a schematic diagram of estimation of the similarity criteria for the energy profile,
• FIG. 4 is a schematic diagram of estimation of the similarity criteria for the spectral envelope,
• FIG. 5 is a schematic diagram of pitch coding by correction of the synthesis pitch profile.

In order to better understand the idea implemented in the presents the invention, the following example is given for illustrative purposes and in no way for a method implemented in a vocoder, in particular the selection and coding of synthesis units for a very speech coder low flow.

As a reminder, at the level of a vocoder, the speech signal is analyzed frame to frame to extract characteristic parameters (spectral parameters, pitch, energy). This analysis is classically using a sliding window set on the horizon of the frame. This frame has a duration of the order of 20 ms, and the update is done with a shift of the analysis window of the order of 10 ms to 20 ms.

During a learning phase, a set of models of Hidden Markov (HMM, Hidden Markov Model) are learned. They allow model speech segments (set of successive frames) can be associated with phonemes if the learning phase is supervised (phonetic segmentation and transcription available) or to spectrally stable sounds in the case of a segmentation obtained from Automatic way. 64 HMM models are used here, which make it possible recognition phase to associate with each segment the model index HMM identified, and therefore the class to which it belongs. HMM models are also used with the help of a Viterbi type algorithm to be coding phase the segmentation and classification of each of the segments (belonging to a class). Each segment is therefore identified by an index between 1 and 64 which is transmitted to the decoder.

The decoder uses this index to find the synthesis unit in the dictionary built during the learning phase. Units of synthesis that make up the dictionary are simply the sequences of parameters associated with the segments obtained on the learning corpus.

A dictionary class contains all units associated with the same model HMM. Each synthesis unit is therefore characterized by a sequence of spectral parameters, a sequence of pitch value (pitch profile), a sequence of gains (energy profile).

In order to improve the quality of synthesis, each class (from 1 to 64) of the dictionary is subdivided into 64 subclasses, where each subclass contains synthesis units that are temporally preceded by a segment belonging to the same class. This approach allows take into account the past context, and thus improve the recovery of the areas transients from one unit to another.

The present invention relates in particular to a method of selection of a multicriteria synthesis unit. The process allows example to take into account simultaneously the pitch, the spectral distortion, and pitch and energy evolution profiles.

1) Extraire le pitch moyen F₀ (fréquence fondamentale moyenne) sur le segment à coder composé de plusieurs trames. Le pitch est par exemple calculé pour chaque trame T, les erreurs de pitch sont corrigées en tenant compte de l'ensemble du segment afin d'éliminer les erreurs de détection voisé/non voisé, et le pitch moyen est calculé sur l'ensemble des trames voisées du segment.
Il est possible de représenter le pitch sur 5 bits, en utilisant par exemple un quantificateur non uniforme (compression logarithmique) appliqué à la période de pitch.
La valeur du pitch de référence est par exemple obtenue à partir d'un générateur de prosodie dans le cas d'une application en synthèse.

2) la valeur de pitch moyen F₀ étant ainsi quantifiée, sélectionner un sous-ensemble d'unités de synthèse SE dans la sous-classe considérée. Le sous-ensemble est défini comme étant celui dont les valeurs moyennes de pitch sont les plus proches de la valeur de pitch F₀.
Dans la configuration précédente cela conduit à retenir de façon systématique les 32 unités les plus proches selon le critère du pitch moyen. Il est donc possible de retrouver ces unités au niveau du décodeur à partir du pitch moyen transmis.

3) Parmi les unités de synthèse ainsi sélectionnées, appliquer un ou plusieurs critères de proximité ou de similarité, par exemple le critère de distorsion spectrale, et/ou le critère de profil d'énergie et/ou le critère de pitch pour déterminer l'unité de synthèse. Lorsque l'on utilise plusieurs critères, une étape de fusion 3b) est réalisée pour prendre la décision. L'étape de fusion des différents critères est réalisée par combinaison linéaire ou non-linéaire. Les paramètres utilisés pour réaliser cette combinaison peuvent être obtenus par exemple sur un corpus d'apprentissage en minimisant un critère de distorsion spectrale sur le signal re-synthétisé. Ce critère de distorsion peut avantageusement inclure une pondération perceptuelle soit au niveau des paramètres spectraux utilisés, soit au niveau de la mesure de distorsion. Dans le cas d'une loi de pondération non linéaire il est possible d'utiliser un réseau connexionniste (MLP, Multi Layer Perceptron par exemple), de la logique floue, ou une autre technique.

4) Etape de codage du pitch

The selection method for a speech segment to be encoded comprises, for example, the selection steps schematized in FIG.

1) Extract the average pitch F ₀ (average fundamental frequency) on the segment to be encoded composed of several frames. The pitch is for example calculated for each frame T, the pitch errors are corrected taking into account the entire segment in order to eliminate the voiced / unvoiced detection errors, and the average pitch is calculated on all voiced frames of the segment.
It is possible to represent the pitch on 5 bits, using for example a non-uniform quantizer (logarithmic compression) applied to the pitch period.
The value of the reference pitch is for example obtained from a prosody generator in the case of an application in synthesis.

2) the average pitch value F ₀ is thus quantized, select a subset of synthesis units SE in the subclass considered. The subset is defined as the one whose average pitch values are closest to the pitch value F ₀ .
In the previous configuration this leads to systematically retain the 32 closest units according to the average pitch criterion. It is therefore possible to find these units at the decoder from the average pitch transmitted.

3) Among the synthesis units thus selected, apply one or more proximity or similarity criteria, for example the spectral distortion criterion, and / or the energy profile criterion and / or the pitch criterion to determine the synthesis unit. When using several criteria, a merger step 3b) is performed to make the decision. The melting step of the different criteria is performed by linear or non-linear combination. The parameters used to achieve this combination can be obtained for example on a training corpus by minimizing a spectral distortion criterion on the re-synthesized signal. This distortion criterion may advantageously include a perceptual weighting either at the level of the spectral parameters used or at the level of the distortion measurement. In the case of a nonlinear weighting law it is possible to use a connectionist network (MLP, Multi Layer Perceptron for example), fuzzy logic, or another technique.

4) Step coding the pitch

The method may comprise in a variant embodiment a pitch coding step by correction of the synthesis pitch profile detailed below.

The criterion relating to the evolution profile of the pitch makes it possible in part to consider the voicing information. It is however possible to disable when the segment is totally unvoiced, or the subclass selected is also unvoiced. Indeed, we can notice mainly three types of subclasses: the subclasses containing majority of voiced units, those containing predominantly unvoiced units, and subclasses containing predominantly units mixed.

The method according to the invention is not limited to optimizing the flow rate allocated to prosody information but also allows to keep for the coding phase the entirety of the synthesis units obtained during the learning phase with a constant number of bits to encode the unit of synthesis. Indeed the synthesis unit is characterized by both the value pitch and its index. This approach allows in a scheme of speaker-independent coding to cover all pitch values possible and to select the synthesis unit taking into account in part characteristics of the speaker, there is indeed for the same speaker a correlation between the pitch variation range and the characteristics of the vocal tract (especially the length).

It can be noticed that the principle of selection of units described can apply to any system whose operation is based on a selection of units and therefore also to a system of synthesis from the text.

A1) sélectionner dans la sous-classe identifiée du dictionnaire des unités de synthèse et à partir de la valeur moyenne du pitch, les N unités les plus proches au sens du critère du pitch moyen. La suite du traitement se fait alors sur les profils de pitch associés à ces N unités. Le pitch est extrait lors de la phase d'apprentissage sur les unités de synthèse, et lors de la phase de codage sur le signal à coder. Les méthodes possibles pour l'extraction du pitch sont nombreuses, cependant les méthodes hybrides, combinant un critère temporel (AMDF, Average Magnitude Difference Function, ou autocorrélation normalisée) et un critère fréquentiel (HPS, Harmonic Power Sum, structure en peigne, ...) sont potentiellement plus robustes.

A2) aligner temporellement les N profils avce celui du segment à coder, par exemple par interpolation linéaire des N profils. Il est possible d'utiliser une technique d'alignement plus optimale basée sur un algorithme de programmation dynamique (DTW ou Dynamic Time Warping). L'algorithme s'applique sur les paramètres spectraux, les autres paramètres pitch, énergie, etc sont alignés de manière synchrone aux paramètres spectraux. Dans ce cas il faut transmettre les informations relatives au chemin d'alignement.

A3) calculer N mesures de similarités, entre les N profils de pitch alignés et le profil de pitch du segment de parole à coder pour obtenir les N coefficients de similarité {rp(1), rp(2), ....rp(N)}. Cette étape peut être réalisée au moyen d'une intercorrélation normalisée. L'alignement temporel peut être un alignement par ajustement simple des longueurs (interpolation linéaire des paramètres). L'utilisation d'une simple correction des longueurs des unités de synthèse permet notamment de ne pas transmettre d'information relative au chemin d'alignement, le chemin d'alignement étant partiellement pris en compte par les corrélations des profils de pitch et d'énergie.Dans le cas de segments mixtes (co-existence au sein d'un même segment de trames voisées et non voisées), l'utilisation des trames non voisées pour lesquelles le pitch est arbitrairement positionné à zéro permet de tenir compte dans une certaine mesure de l'évolution du voisement.La figure 3 schématise le principe d'estimation des critères de similarité pour le profil énergétique.
Le procédé comporte par exemple les étapes suivantes :

A4) extraire les profils d'évolution de l'énergie pour les N unités sélectionnées comme indiqué précédemment, c'est-à-dire selon un critère de proximité du pitch moyen. Selon la technique de synthèse utilisée, le paramètre d'énergie utilisé peut soit correspondre à un gain (associé à un filtre de type LPC par exemple) ou une énergie (l'énergie calculée sur la structure harmonique dans le cas d'une modélisation harmonique/stochastique du signal). Enfin, l'estimation de l'énergie peut avantageusement se faire de manière synchrone du pitch (1 valeur d'énergie par période de pitch). Les profils énergétiques sont pré-calculés pour les unités de synthèse lors de la phase d'apprentissage.

A5) aligner temporellement les N profils avec celui des segments à coder, par exemple par interpolation linéaire, ou par programmation dynamique (alignement non-linéaire) de façon similaire à la méthode mise en oeuvre pour corriger le pitch.

A6) calculer N mesures de similarités, entre les profils des N valeurs d'énergie alignées et le profil d'énergie du segment de parole à coder pour obtenir les N coefficients de similarité {re(1), re(2), ...., re(N)}. Cette étape peut aussi être réalisée au moyen d'une intercorrélation normalisée. La figure 4 schématise le principe d'estimation des critères de similarité pour l'enveloppe spectrale.
Le procédé comporte les étapes suivantes :

A7) aligner temporellement les N profils,

A8) déterminer les profils d'évolution des paramètres spectraux pour les N unités sélectionnées comme indiqué précédemment, c'est-à-dire selon un critère de proximité du pitch moyen. Il s'agit ici tout simplement de calculer le pitch moyen du segment à coder, et de considérer les unités de synthèse de la sous-classe associée (indice HMM courant pour définir la classe, indice HMM précédent pour définir la sous-classe) qui ont un pitch moyen proche.

A9) calculer N mesures de similarités, entre la séquence spectrale du segment à coder et les N séquences spectrales extraites des unités de synthèse sélectionnées pour obtenir les N coefficients de similarité {rs(1), rs(2), ...., rs(N)}. Cette étape peut être réalisée au moyen d'une intercorrélation normalisée.

Figure 2 schematizes a principle of estimation of similarity criteria for the pitch profile.
The method comprises for example the following steps:

A1) select in the identified subclass of the dictionary of the synthesis units and from the average value of the pitch, the N units closest to the meaning of the average pitch criterion. Further processing is then done on the pitch profiles associated with these N units. The pitch is extracted during the learning phase on the synthesis units, and during the coding phase on the signal to be coded. The possible methods for the extraction of the pitch are numerous, however the hybrid methods, combining a temporal criterion (AMDF, Average Magnitude Difference Function, or standard autocorrelation) and a frequency criterion (HPS, Harmonic Power Sum, comb structure, .. .) are potentially more robust.

A2) align the N profiles temporally with that of the segment to be coded, for example by linear interpolation of the N profiles. It is possible to use a more optimal alignment technique based on a dynamic programming algorithm (DTW or Dynamic Time Warping). The algorithm applies to the spectral parameters, the other parameters pitch, energy, etc. are aligned synchronously with the spectral parameters. In this case it is necessary to transmit the information relating to the alignment path.

A3) calculating N similarity measures between the N aligned pitch profiles and the pitch profile of the speech segment to be encoded to obtain the N similarity coefficients {rp (1), rp (2), .... rp ( NOT)}. This step can be performed by means of standardized intercorrelation. The time alignment can be a simple adjustment of the lengths (linear interpolation of the parameters). The use of a simple correction of the lengths of the synthesis units notably makes it possible not to transmit information relating to the alignment path, the alignment path being partially taken into account by the correlations of the pitch profiles and In the case of mixed segments (coexistence within the same segment of voiced and unvoiced frames), the use of unvoiced frames for which the pitch is arbitrarily set to zero makes it possible to take into account in a certain measurement of the evolution of voicing. Figure 3 schematizes the principle of estimating the similarity criteria for the energy profile.
The method comprises for example the following steps:

A4) extract the evolution profiles of the energy for the N units selected as indicated above, that is to say according to a criterion of proximity of the average pitch. According to the synthesis technique used, the energy parameter used can either correspond to a gain (associated with an LPC type filter for example) or an energy (the energy calculated on the harmonic structure in the case of harmonic modeling / stochastic signal). Finally, the energy estimate can advantageously be synchronously pitch (1 energy value per pitch period). The energy profiles are pre-calculated for the synthesis units during the learning phase.

A5) temporally align the N profiles with that of the segments to be encoded, for example by linear interpolation, or by dynamic programming (nonlinear alignment) similarly to the method implemented to correct the pitch.

A6) calculating N similarity measures between the profiles of the N aligned energy values and the energy profile of the speech segment to be coded to obtain the N similarity coefficients {re (1), re (2), .. .., re (N)}. This step can also be performed by means of standardized cross correlation. Figure 4 schematizes the principle of estimating similarity criteria for the spectral envelope.
The method comprises the following steps:

A7) temporally align the N profiles,

A8) determine the evolution profiles of the spectral parameters for the N units selected as indicated above, that is to say according to a criterion of proximity of the average pitch. It is simply a matter of calculating the average pitch of the segment to be coded, and of considering the synthesis units of the associated subclass (current HMM index to define the class, previous HMM index to define the subclass) which have a near average pitch.

A9) calculating N similarity measures, between the spectral sequence of the segment to be coded and the N spectral sequences extracted from the synthesis units selected to obtain the N coefficients of similarity {rs (1), rs (2), ...., rs (N)}. This step can be performed by means of standardized intercorrelation.

The similarity measure may be a spectral distance.

Step A9) comprises for example a step where the average the set of spectra of the same segment and the measure of similarity is an intercorrelation measure.

The spectral distortion criterion is for example calculated on harmonic structures resampled to constant pitch or resampled the pitch of the segment to be coded, after interpolation of initial harmonic structures.

The similarity criterion will depend on the spectral parameters used (for example the type of parameters used for the representation of the envelope). Several types of spectral parameters can be used, insofar as they make it possible to define a measure of distortion spectral. In the field of speech coding, it is common to use LSP or LSF parameters (LSP, Line Spectral Pair, LSF, Line Spectral Frequencies) derived from a linear prediction analysis. In the field speech recognition, cepstral parameters are usually used, and they can either be derived from a linear prediction analysis (LPCC, Linear Prediction Cepstrum Coefficients) or estimated from a filter bank often on a Mel or Bark perceptual scale (MFCC, Mel Frequency Cepstrum Coefficients). It is also possible in the as we use a sinusoidal modeling of the component harmonic of the speech signal, to directly use the amplitudes of the harmonic frequencies. These last parameters being estimated according to pitch can not be used directly to calculate a distance. The number of coefficients obtained is indeed variable depending on the pitch, unlike LPCC, MFCC or LSF parameters. Pre-treatment then consists in estimating a spectral envelope from the amplitudes harmonics (linear or polynomial spline interpolation) and to resample the envelope thus obtained, either by using the frequency of the segment to be coded, either by using a frequency fundamental constant (100 Hz for example). A fundamental frequency constant allows to pre-calculate all the harmonic structures of synthesis units during the learning phase. Re-sampling is then only on the segment to be coded. On the other hand, if we limit ourselves to linear alignment by linear interpolation, it is possible to average harmonic structures on all segments considered. The similarity measure can then be estimated simply from the average harmonic structure of the segment to be coded, and that of the synthesis unit considered. This measure of similarity can also be a standardized cross-correlation measurement. It can also be noted that resampling procedure can be done on a scale perceptual frequencies (Mel or Bark).

For the time alignment procedure it is possible to use either a Dynamic Time Warping (DTW) algorithm or to perform a simple linear interpolation (linear length adjustment). In the event that it is not desired to transmit additional information relating to the alignment path, it is preferable to use a simple linear interpolation of the parameters. Taking into account the best alignment is then partly achieved by the selection procedure.
Pitch coding by modifying the synthesis profile

According to one embodiment, the method comprises a step of pitch coding by modifying the synthesis profile. This consists of resynthesizing a pitch profile from that of the synthesis unit selected and a linearly variable gain over the duration of the segment code. It is then sufficient to transmit an additional value for characterize the correction gain over the entire segment.

ce qui conduit aux relations suivantes : a = (S4.S 2 - S5.S 1)(S2.S 2 - S3.S 1) et b = (S5.S 2 - S4.S 3) (S2.S 2 - S3.S 1) où

Le coefficient a, ainsi que la valeur moyenne du pitch modélisé sont quantifiés et transmis : aq = Q[a]

La valeur du coefficient b est obtenu au niveau du décodeur à partir de la relation suivante : bq = f 0q - Σa q .n.f 0S (n) N 〈f0S 〉 où 〈f_0S〉 est le pitch moyen de l'unité de synthèse.
Remarque : cette méthode de correction peut bien entendu s'appliquer au profil énergétique. The pitch reconstructed at the level of the decoder is given by the following equation: f 0 ( not ) = boy Wut ( not ). f 0S ( not ) = ( year + b ). f 0S ( not ) where f _0S ( n ) is the pitch at the frame of index n of the synthesis unit.
This corresponds to a linear transformation of the pitch profile.
The optimal values of a and b are estimated at the encoder level by minimizing the mean squared error:

which leads to the following relationships: at = (S 4 . S 2 - S 5 . S 1 ) (S 2 . S 2 - S 3 . S 1 ) and b = (S 5 . S 2 - S 4 . S 3 ) (S 2 . S 2 - S 3 . S 1 ) or

The coefficient a, as well as the average value of the modeled pitch are quantized and transmitted: at q = Q [ at ]

The value of the coefficient b is obtained at the level of the decoder from the following relation: b q = f 0q - Σ at q . not . f 0S (not) NOT < f 0S > where <f _0S > is the average pitch of the synthesis unit.
Note: This correction method can of course be applied to the energy profile.

Flow example associated with the encoding scheme

• Indice de classe sur 6 bits (64 classes)
• Indice de l'unité sélectionnée sur 5 bits (32 unités par sous-classe)
• Longueur du segment sur 4 bits (de 3 à 18 trames)

The information relating to the bit rate associated with the coding scheme previously described is as follows:

• 6-bit class index (64 classes)
• Index of the unit selected on 5 bits (32 units per subclass)
• Segment length on 4 bits (from 3 to 18 frames)

• FO moyen sur 5 bits
• Coefficient correcteur du profil de pitch sur 5bits
• Gain correcteur sur 5 bits

The average number of segments per second is between 15 and 20; which leads to a base rate of between 225 and 300 bits / sec for the previous configuration. At this base rate is added the flow required to represent the information of pitch and energy.

• Mean FO on 5 bits
• Corrective coefficient of the 5bits pitch profile
• 5-bit correction gain

The flow associated with the prosody is then between 225 and 300 bits / sec, which leads to an overall bit rate between 450 and 600 bits / sec.

References

[1] G. Baudoin, F. El Chami, "Corpus based very low bit rate speech coder", Proc. Conf. IEEE ICASSP 2003, Hong Kong, 2003.

[2] G. Baudoin, J. Cernocky, P. Gournay, G. Chollet, "Speech Coding at Low and Very Low Flow ", Annals of Telecommunications, Vol 55, N 9-10 Pages 421-456, Nov. 2000.

[3] G. Baudoin, F. Capman, J. Cernocky, F. El-chami, M. Charbit, G. Chollet, D. Petrovska-Delacretaz. "Advances in Very Low Bit Speech Coding Using Recognition and Synthetic Techniques", TSD '2002, pp. 269-276, Brno, Czech Republic, Sept 2002.

[4] K.Lee, R.Cox, 'A segmental coder based on a concatenative TTS', in Speech Communications, Vol. 38, pp 89-100, 2002.

[5] K.Lee, R.Cox, "A very low bit rate speech coder based on recognition / synthesis paradigm ", in IEEE on ASSP, Vol; 9, pp 482-491, July 2001.

Claims (14)

A method of selecting synthesis units of information that can be decomposed into synthesis units, characterized in that it comprises
less the following steps:
for a segment of information considered: determine the value F0 of the average fundamental frequency for the segment of information considered, selecting a subset of synthesis units defined as being the one whose average pitch values are closest to the pitch value F0, applying one or more proximity criteria to the selected synthesis units to determine a summary unit representative of the information segment. Method for selecting synthesis units according to claim 1, characterized in that the information is a speech segment to be encoded and in that the fundamental frequency or pitch, the spectral distortion, and / or or the energy profile and a step of merging the criteria is performed to determine the representative synthesis unit. A method of selecting units according to claim 1 characterized in that for a speech segment to code the reference pitch is obtained from a prosody generator. Method according to Claim 2, characterized in that the estimate of the similarity criterion for the pitch profile comprises at least the following steps: A1) selecting in the identified subclass of the dictionary of synthesis units and from the average value of the pitch, the N nearest units in the sense of the average pitch criterion, A2) align the N profiles temporally with that of the segment to be coded, A3) calculating N similarity measures between the N aligned pitch profiles and the pitch profile of the speech segment to be encoded to obtain the N similarity coefficients {rp (1), rp (2), .... rp ( NOT)}. Method according to Claim 2, characterized in that the similarity estimate for the energy profile comprises at least the following steps: A4) determine the evolution profiles of the energy for the N units selected according to a criterion of proximity of the average pitch. A5) temporally aligning the N profiles with that of the segment to be coded, A6) calculating N similarity measures between the N aligned energy profiles and the energy profile of the speech segment to be coded to obtain the N similarity coefficients {re (1), re (2), ... ., re (N)}. Method according to Claim 2, characterized in that the estimation of the similarity criteria for the spectral envelope comprises at least the following steps: A7) temporally aligning the N profiles with that of the segment to be coded, A8) determining the evolution profiles of the spectral parameters for the N units selected according to a criterion of proximity of the average pitch, A9) calculating N measures of the similarities between the spectral sequence of the segment to be coded and the N corresponding extracted spectral sequences of the speech segment to be coded to obtain the N coefficients of similarity {rs (1), rs (2), ... ., rs (N)}, Method according to one of Claims 4, 5 and 6, characterized in that the time alignment is a time alignment obtained by dynamic programming (DTW) or an alignment by linear adjustment of the lengths. Method according to one of Claims 4, 5 and 6, characterized in that the similarity measurement is a standardized cross-correlation measurement. Method according to claim 6, characterized in that the similarity measure is a spectral distance measurement. A method according to claim 6 characterized in that step A9) comprises a step where all the spectra of the same segment are averaged and in that the similarity measure is a cross correlation measure. Method according to claim 6, characterized in that the spectral distortion criterion is calculated on harmonic structures resampled to constant pitch or resampled to the pitch of the segment to be coded, after interpolation of the initial harmonic structures. Method according to one of Claims 1 to 11, characterized in that it comprises a coding step and / or a pitch correction step by modifying the synthesis profile. Method according to claim 12, characterized in that the step of encoding and / or correcting the pitch is a linear transformation of the pitch profile of origin. Use of the method according to one of claims 1 to 12 for the selection and / or coding of synthesis units for a very speech coder low flow.

Images