EP1944755B1 - Modification of a voice signal - Google Patents

Abstract

The method involves decomposing a voice signal in a parametric part and a non parametric residue according to an excitation-filter model, and estimating a temporal envelope of the residue. Acoustic characteristics of the parametric part and the residue are modified according to modification instructions. A new temporal envelope is determined for the modified residue according to the modification instructions. The modified voice signal from the modified residue and the modified parametric part is synthesized with the new temporal envelope. Independent claims are also included for the following: (1) a program comprising instructions to perform an acoustic characteristics modifying method of a voice signal (2) a device for modifying acoustic characteristics of a voice signal.

Description

The present invention relates to the modification of the speech and more particularly to the modification of the acoustic parameters of decomposed speech signals into a parametric part and a non-parametric part.

It is known to decompose the speech signals according to so-called "filter-excitation" models. In these models, speech is considered a glottic excitation transformed by a filter representing the vocal channel.

The excitation is obtained by inverse filtering of the speech signal. It sometimes includes a part that is also parametric and a residue. The residue corresponds to the difference between the excitation and the corresponding parametric modeling.

When modifying the speech signals, the frequency, rhythm or timbre information is changed through the parameters of the model.

However, these modifications cause audible distortions, in particular because of a lack of control of the temporal coherence, in particular during changes in fundamental frequency or timbre.

For example, the document " Applying the Harmonic Plus Noise Model in Concatenative Speech Synthesis ", IEEE Transactions on Speech and Audio Processing, vol 9 (1), pp. 21-29, January 2001 by Y. Stylianou , plans to use a harmonic model plus noise, or model NNM, with a temporal modulation of the noisy part so that it fits in a natural way to the deterministic part. However, this method does not preserve the temporal coherence of the deterministic part.

Another approach is to have a model of the glottal source compact enough that the pace of the glottal signal can be controlled during changes in the signal. Such an approach is described for example in the document " Toward a high-quality singing synthesizer with vocal texture control ", Stanford University, 2002 by HL Lu . Nevertheless, such a model does not capture all the information of the glottal signal. Residual information must be kept and its modification raises the problem of lack of temporal coherence mentioned above.

In the document " Time-scale modification of complex acoustic signals ", ICASSP 1993, vol.1, pp.213-216, 1993 by TF Quatieri, RB Dunn and TE Hanna , there is proposed a method of modifying speech signals to preserve both the spectral envelope and the temporal envelope. This method is applied only to the modification of the duration of acoustic signals and is not practical insofar as it is theoretically not possible to guarantee the existence of a solution satisfying both properties simultaneously. In addition, there is no convergence result of the proposed algorithm and therefore this method does not provide sufficient control over the characteristics of the resulting signal.

Thus, there is no technique for modifying the speech signals while ensuring good coherence at the time level.

One of the objectives of the present invention is to allow such a modification.

• une décomposition du signal en une partie paramétrique et un résidu non paramétrique ;
• une estimation de l'enveloppe temporelle du résidu ;
• une modification de caractéristiques acoustiques de la partie paramétrique et du résidu selon des consignes de modification ;
• une détermination d'une nouvelle enveloppe temporelle pour le résidu modifié par application desdites consignes de modification à l'enveloppe temporelle estimée; et
• une synthèse d'un signal de parole modifié à partir de la partie paramétrique modifiée et du résidu tel que modifié et avec la nouvelle enveloppe temporelle.

For this purpose, the present invention as defined by claim 1, relates to a method for modifying the acoustic characteristics of a speech signal, characterized in that it comprises:

• a decomposition of the signal into a parametric part and a nonparametric residue;
• an estimate of the temporal envelope of the residue;
• a modification of the acoustic characteristics of the parametric part and the residue according to modification instructions;
• determining a new time envelope for the modified residue by applying said modification instructions to the estimated time envelope; and
• a synthesis of a modified speech signal from the modified parametric part and the residue as modified and with the new time envelope.

Thanks to the specific processing carried out on the temporal characteristics of the residue, the temporal coherence of the modified signal is improved.

In one embodiment of the invention, said signal decomposition is a decomposition according to an excitation-filter type model. Such a decomposition makes it possible to obtain a residue corresponding to a glottic excitation.

Advantageously, the estimation of the temporal envelope of the residue comprises the estimation of a first envelope, then a temporal smoothing of this first envelope. This embodiment makes it possible to obtain a better estimate of the temporal envelope.

In a particular embodiment, the method further comprises temporal normalization of the residue based on the estimate of the time envelope. This makes it possible to obtain an expression of the residue that is substantially independent of its temporal characteristics.

In a particular embodiment, the temporal normalization of the residue comprises dividing the residue by estimating the time envelope.

In another embodiment, determining a new time envelope for the residue includes changing parameters of the time envelope of the residue according to said modification instructions and applying the modified time envelope to the normalized residual.

In one embodiment, the estimation of the temporal envelope and the determination of a new temporal envelope are combined.

Advantageously, the modification of acoustic characteristics comprises a modification of fundamental frequency information and duration of the parametric part and the residue.

In addition, the invention also relates to a program for implementing the method described above as defined in claim 9 and a corresponding device as defined in claim 10.

• la figure 1 représente de manière générale un organigramme du procédé de l'invention ; et
• les figures 2A à 2D représentent différents stades de traitement d'un signal de parole.

The invention will be better understood in the light of the description given by way of example and with reference to the figures in which:

• the figure 1 generally represents a flowchart of the process of the invention; and
• the Figures 2A to 2D represent different stages of processing a speech signal.

The process shown with reference to figure 1 starts with a step of analysis of the speech signal which comprises a decomposition 12 according to an excitation-filter model, that is to say a decomposition of the speech signal into a parametric part and a non-parametric part, called a residue and corresponding to a part of the glottal excitation.

A common practice for the implementation of step 12 is the use of linear prediction techniques such as those described in the J. Makhoul, "Linear Prediction: A Tutorial Review," Proceedings of the IEEE, vol. 63 (4), pp. 561-580, April 1975 .

In the embodiment described as an example, the decomposition 12 of the speech signal s (n) is performed using a self-regression, or AR model, of the following form: $$s\left(\mathrm{not}\right)={\displaystyle \underset{k=1}{\overset{p}{\Sigma }}}{\mathrm{at}}_{k}s\left(\mathrm{not}-k\right)+e\left(\mathrm{not}\right)\mathrm{.}$$

In this equation, the terms a _k denote the coefficients of an AR type filter modeling the vocal tract and the term e (n) is the residual signal relative to the excitation part, with n a signal frame index. Note that if the order of the model is large enough then e (n) is not correlated with s (n).

This is formally written E [e (n) s (n-m)] = 0 for any integer m, where E [.] Denotes the expected value.

In practice, typical orders of 10 and 16 are chosen for speech signals sampled respectively at 8 and 16 kHz.

où r est la fonction d'autocorrélation définie par : r(m) = E[s(n)s(n-m)].By multiplying the previous equation on the left and on the right by s (nm) and passing to the mathematical expectation, we arrive at the Yule-Walker equations defined by: $$r\left(m\right)=-{\displaystyle \underset{k=1}{\overset{p}{\Sigma }}}{\mathrm{at}}_{k}r\left(m-k\right)$$

where r is the autocorrelation function defined by: r (m) = E [s (n) s (nm)].

An estimator of r (m) is given by: $$r\left(m\right)=\frac{1}{\mathrm{NOT}-p}{\displaystyle \underset{k=1}{\overset{\mathrm{NOT}-p}{\Sigma }}}s\left(\mathrm{not}\right)s\left(\mathrm{not}-m\right)\mathrm{.}$$

In practice, only the first p + 1 values of the autocorrelation function are necessary for the estimation of the filter coefficients a _k . The expression of this last equation in matrix form leads to the resolution of the following linear system: $$\left[\begin{array}{cccc}r\left(0\right)& r\left(1\right)& \cdots & r\left(p-1\right)\\ r\left(1\right)& r\left(0\right)& \cdots & r\left(p-2\right)\\ ⋮& ⋮& \ddots & ⋮\\ r\left(p-1\right)& r\left(p-2\right)& \cdots & r\left(0\right)\end{array}\right]\left[\begin{array}{c}{\mathrm{at}}_1\\ \mathrm{<figure-callout\; id="2"\; label="r"\; filenames="imgf0002.png"\; state="\{\{state\}\}">r</figure-callout>}\left(2\right)\\ ⋮\\ {\mathrm{at}}_p\end{array}\right]=\left[\begin{array}{c}r\left(1\right)\\ \mathrm{<figure-callout\; id="2"\; label="r"\; filenames="imgf0002.png"\; state="\{\{state\}\}">r</figure-callout>}\left(2\right)\\ ⋮\\ r\left(p\right)\end{array}\right]\mathrm{.}$$

Thus, the estimation of the coefficients amounts to the inversion of a Toeplitz matrix, which can be achieved using conventional procedures and in particular using the algorithm described by J. Durbin, "The fitting of time-series models", Rev. Inst. Int. Statistics .

As a variant, the decomposition 12 makes it possible to obtain, for excitation, a parametric model in addition to the residue.

For example, the excitation-filter decomposition is carried out using a priori information on the excitation. Thus, the excitation can be modeled by integrating information related to the speech production process, in particular via a parametric model of the derivative of the glottal flow wave (DODG), such as for example the LF model proposed by Liljencrants and Fant in "A four-parameter model of glottal flow", STL-QPSR, vol. 4, pp. 1-13, 1985 . This model is entirely defined by the data of the fundamental period T0, of three shape parameters which are an open quotient of periods, an asymmetry coefficient and a return phase coefficient, of a position parameter corresponding to the instant of closure of the glottis and a term b ₀ characterizing the amplitude of the DODG.

où u(n) désigne le signal correspondant au modèle LF de la DODG.In this context, the speech signal can be represented by the following ARX-LF exogenous self-regression model: $$s\left(\mathrm{not}\right)={\displaystyle \underset{k=1}{\overset{p}{\Sigma }}}{\mathrm{at}}_{k}s\left(\mathrm{not}-k\right)+b_{0}u\left(\mathrm{not}\right)+e\left(\mathrm{not}\right)$$

where u (n) is the signal corresponding to the LF model of the DODG.

The simultaneous estimation of the parameters of the DODG and the parameters related to the filter is delicate, in particular because the optimization according to the shape and position parameters is a non-linear problem. However, when T0 and u are fixed, the optimization according to the parameters a _k and b ₀ is a classical linear problem, for which a least squares estimator can be obtained analytically. On the basis of observation, an effective method has been proposed by D. Vincent, O. Rosec, and T. Chonavel, in the publication "Estimation of LF glottal source parameters based on ARX model", Interspeech'05, pp. 333-336, Lisbon, Portugal, 2005 .

• des paramètres caractérisant complètement la DODG selon le modèle LF;
• des paramètres de filtre a_k ;
• le résidu e(n) correspondant à l'erreur de modélisation liée au modèle ARX-LF.

In this embodiment, at the end of the estimation procedure, the method delivers:

• parameters characterizing DODG completely according to the LF model;
• filter parameters a _k ;
• the residue e (n) corresponding to the modeling error related to the ARX-LF model.

In general, at the end of step 12, the method delivers a modeling of the speech signal s (n) in the form of a parametric part and a residue that is non-parametric.

The analysis step 10 then comprises an estimate 14 of the temporal envelope of the residue.

où H désigne l'opération de transformation de Hilbert.In the embodiment described, the temporal envelope is defined as the module of the analytical signal and is obtained by a so-called Hilbert transformation. Thus, the temporal envelope d (t) of the residue e (t) is written: $$d\left(t\right)=\left|x_e\left(t\right)\right|{\mathrm{with}x}_e\left(t\right)=e\left(t\right)+\mathit{iH}\left(e\left(t\right)\right),$$

where H denotes the Hilbert transformation operation.

Advantageously, the estimate 14 comprises a smoothing of the temporal envelope of the residue. This provides a better estimate especially for voiced sounds for which the envelope is periodic period T ₀ , with T ₀ denoting the inverse of the fundamental frequency f ₀ . For example, a k-mode cepstral modeling of the envelope can be used. It is written in the form: $$\ln \left(d\left(\mathrm{not}\right)\right)=\frac{1}{2}{\displaystyle \underset{k=-K}{\overset{K}{\Sigma }}}{\mathrm{vs}}_{\mathrm{<figure-callout\; id="2"\; label="k\; exp"\; filenames="imgf0002.png"\; state="\{\{state\}\}">k\; </figure-callout>}}\exp \left(2{\mathit{i\pi knf}}_0/f_s\right)+\epsilon \left(\mathrm{not}\right)$$

avec $$\mathrm{d}=\frac{1}{2}[\ln \left(\mathit{d}\left(-\mathit{N}\right)\right),\cdot \cdot \cdot ,\ln \left(\mathit{d}\left(\mathit{N}\right)\right){]}^{\mathit{T}}$$

M _{n+(N+1),k+(K+1)} = exp(2iπknf ₀/f_s ), n ∈ {-N,···,N}, k ∈ {-K,···,K}
et $$c=[c_{-K},\dots ,c_K{]}^T\mathrm{.}$$

The cepstral coefficients c _k are then estimated by minimizing ε (n) in the least squares sense. More precisely, the preceding equation is written in the following matrix form: $$\mathrm{d}\mathrm{=}\mathrm{Mc}\mathrm{+}\mathrm{\epsilon }\mathrm{,}$$

with $$\mathrm{d}=\frac{1}{2}[\ln \left(\mathit{d}\left(-\mathit{NOT}\right)\right),\cdot \cdot \cdot ,\ln \left(\mathit{d}\left(\mathit{NOT}\right)\right){]}^{\mathit{T}}$$

M _{n + ( N +1) , k + ( K +1)} = exp (2 i π knf ₀ / f _s ), n ∈ {- N , ··· , N }, k ∈ {- K , ·· ·, K }
and $$\mathrm{vs}=[{\mathrm{vs}}_{-K},...,{\mathrm{vs}}_K{]}^T\mathrm{.}$$

In these equations, the exponent T represents the transposition operator.

où H désigne l'opérateur de transposition hermitienne. L'enveloppe correspondante s'écrit de la façon suivante : $$\hat{d}\left(n\right)=\frac{1}{2}{\left({\displaystyle \sum _{k=-K}^K}{\hat{c}}_k\exp \left(2{\mathit{i\pi knf}}_0/f_s\right)\right)}_{\mathrm{.}}$$

The optimal solution in the least squares sense is then $$\mathbf{vs}=({\mathbf{M}}^H\mathbf{M}{)}^{-1}{\mathbf{M}}^H\mathbf{d}$$

where H denotes the Hermitian transposition operator. The corresponding envelope is written as follows: $$\hat{d}\left(\mathrm{not}\right)=\frac{1}{2}{\left({\displaystyle \underset{k=-K}{\overset{K}{\Sigma }}}{\widehat{\mathrm{vs}}}_{\mathrm{<figure-callout\; id="2"\; label="k\; exp"\; filenames="imgf0002.png"\; state="\{\{state\}\}">k\; </figure-callout>}}\exp \left(2{\mathit{i\pi knf}}_0/f_s\right)\right)}_{\mathrm{.}}$$

Once the temporal envelope of the estimated residue, the method comprises a step 16 of temporal normalization of the residue. In this document, temporal normalization means obtaining a substantially invariant residue at the time level, more precisely, obtaining a residue whose temporal envelope is constant.

In the embodiment described, step 16 is implemented by dividing the residue by the expression of the temporal envelope according to the following equation: $$\tilde{e}\left(\mathrm{not}\right)=\frac{e\left(\mathrm{not}\right)}{\hat{d}\left(\mathrm{not}\right)}\mathrm{.}$$

In parallel with the analysis 10, the method comprises a step 18 for determining instructions for modifying the speech signal. These instructions can be of two types.

In a first case, a target has been defined for each of the parameters to be modified. This is particularly the case in speech synthesis where many algorithms for predicting the duration, the fundamental frequency or even the energy of the signals exist. For example, fundamental frequency and energy values can be estimated for the beginning and the end of each syllable or each phoneme of the utterance. Similarly, the duration of each syllable or phoneme can be predicted. Given these digital targets and the speech signal, modifying coefficients can be obtained by relating the measurement made on the signal to the value of the corresponding predicted target.

In a second case, such targets are not available, but it is possible to define a set of modification coefficients for modifying the desired parameters. For example, a fundamental frequency modification coefficient of 0.5 makes it possible to divide by 2 the perceived voice height. Note that these modification coefficients can be defined globally over the entire statement or more locally, for example at the scale of a syllable or a word.

The method then comprises a step 20 of modifying the speech signal s (n) according to the previously determined instructions.

The modifications made concern the fundamental frequency, the duration and the energy of the speech signals. In addition, when an analysis using a DODG is implemented because a source-filter decomposition is available, changes in the voice quality parameters can be made by altering the open quotient, the asymmetry coefficient , or the return phase coefficient.

The modification step 20 firstly comprises a modification 22 of the parametric portion of the pattern corresponding to the speech signal and the normalized residue.

In the embodiment described, this modification relates to the fundamental frequency as well as the duration and is implemented in a conventional manner with a technique known as TD-PSOLA (in English). Time Domain Pitch Synchronous Overlap and Add) as described in the publication "Non-parametric techniques for pitch-scale and time-scale modification of speech", Speech Communication, Vol. 16, pp. 175-205, 1995 by E. Moulines and J. Laroche .

This technique makes it possible to jointly modify the duration and the fundamental frequency with the respective coefficients α (t) and β (t).

With reference to Figures 2A to 2D , the main steps in the operation of the TD-PSOLA technique are illustrated.

The Figure 2A represents the speech signal to be modified s (n). During a step 24, this signal is segmented into so-called pitch-synchronous frames, that is to say that each segment has a duration corresponding to the inverse of the fundamental frequency of the signal.

Indeed, the glottal closure instants, also called analysis instants, are located in the vicinity of the energy maxima of the speech signal and the TD-PSOLA treatment allows a good preservation of the characteristics of the speech signal in the vicinity of the extremities. segments obtained by pitch-synchronous analysis. Thus, when these moments are spotted with a satisfactory accuracy, the performances of TD-PSOLA are optimized. Such pitch-synchronous segmentation is obtained, for example, by time delay techniques or from the method proposed by D. Vincent, O. Rosec, and T. Chonavel, in the publication "Glottal closure instant estimation using an appropriateness measure of the source and continuity constraints", IEEE ICASSP'06, vol. 1, pp. 381-384, Toulouse, France, May 2006 .

Advantageously, this pitch-synchronous marking step is performed offline, that is to say not in real time, which reduces the calculation load for implementation in real time.

• pour un allongement de durée, certains segments sont dupliqués afin d'augmenter artificiellement le nombre d'impulsions glottiques ;
• pour une réduction de la durée, certains segments sont supprimés;
• pour une augmentation de la fréquence fondamentale, c'est-à-dire un rendu plus aigu, les instants d'analyse sont rapprochés, ce qui nécessite éventuellement la duplication de segments pour conserver la durée totale ; et
• pour une diminution de la fréquence fondamentale, c'est-à-dire un rendu plus grave, les instants d'analyse sont écartés, ce qui nécessite éventuellement la suppression de segments pour conserver la durée totale.

Depending on the desired modification factors for the fundamental frequency and the duration, the instants separating the segments are modified according to the following rules:

• for an extension of duration, certain segments are duplicated in order to artificially increase the number of glottal pulses;
• for a reduction of the duration, certain segments are deleted;
• for an increase of the fundamental frequency, that is to say a more acute rendering, the analysis instants are brought closer, which possibly requires the duplication of segments to preserve the total duration; and
• for a decrease in the fundamental frequency, that is to say a more serious rendering, the analysis instants are discarded, which possibly requires the removal of segments to maintain the total duration.

A detailed description of these rules can be found in the publication " Non-parametric techniques for pitch-scale and time-scale modification of speech ", Speech Communication, vol 16, pp. 175-205, 1995 by E. Moulines and J. Laroche .

At the end of this step, the signal comprises an integer number of segments or frames, each of a duration corresponding to a period which is the inverse of the modified fundamental frequency, as shown in FIG. Figure 2B .

The modification processing then comprises a windowing 26 of the signal around the analysis instants, that is to say the moments separating the segments. During this windowing, a portion of the windowed signal around this instant is selected for each analysis instant. This portion of the signal is called a "short-term signal" and extends, in the example, over a period corresponding to twice the fundamental period modified as shown with reference to FIG. Figure 2C .

The modifying processing finally comprises a summation 28 of the short-term signals which are recentered on the synthesis instants and added as shown with reference to FIG. 2D figure .

In a variant, step 22 is performed using a HNM ( Harmonic plus Noise Model ) type technique , or a phase vocoder type technique. The fundamental frequency and duration changes can also be made by different techniques.

In the following, the modified normalized residue, that is to say the normalized residue whose fundamental frequency and / or time information has been modified, is denoted by ẽ ^modif ( n ).

The method then comprises a step 30 of modifying the temporal envelope of the residue. More precisely, this step makes it possible to substitute the original temporal characteristics of the residue with temporal characteristics in accordance with the desired modifications.

Step 30 begins with a determination 32 of new temporal characteristics of the residue. In the example, it is the modification of the temporal envelope of the residue, as obtained at the end of step 14.

• une modification de la fréquence fondamentale ; et
• une modification des paramètres liés à la qualité vocale.

As indicated above, considering a pitch-synchronous frame of the signal, two types of modifications can be operated together or not:

• a change in the fundamental frequency; and
• a change in the parameters related to voice quality.

The modification of the fundamental frequency consists of a modification of the temporal envelope to make it coherent with the normalized residual whose fundamental frequency has been modified beforehand.

One embodiment of such a modification consists of a dilation / contraction of the original time envelope d (n) in order to preserve the general shape.

l'enveloppe temporelle modifiée d^modif s'écrit alors de la manière suivante : $$d^{\mathit{modif}}\left(n\right)=\exp \left(\frac{1}{2}{\displaystyle \sum _{k=-K}^K}{\hat{c}}_k\exp \left(2{\mathit{i\pi knf}}_0^{\mathit{modif}}/f_s\right)\right)\mathrm{.}$$

Given the modified fundamental frequency value $f_0^{\mathit{Modified}},$

the modified temporal envelope d ^modif is then written as follows: $$d^{\mathit{Modified}}\left(\mathrm{not}\right)=\exp \left(\frac{1}{2}{\displaystyle \underset{k=-K}{\overset{K}{\Sigma }}}{\widehat{\mathrm{vs}}}_{\mathrm{<figure-callout\; id="2"\; label="k\; exp"\; filenames="imgf0002.png"\; state="\{\{state\}\}">k\; </figure-callout>}}\exp \left(2{\mathit{i\pi knf}}_0^{\mathit{Modified}}/f_s\right)\right)\mathrm{.}$$

When changes to voice quality related parameters are made, the shape of the time envelope must be changed. For example, when changes in the open coefficient are made, different expansion / contraction factors should be applied to the open and closed portions of the glottal cycle, respectively.

avec $T_e^{\mathrm{mod}\mathit{if}}<{\mathit{T}}_{\mathit{0}},$

avec T₀ qui est la longueur d'un cycle glottique dont l'instant de fermeture coïncide avec l'origine des temps et une phase ouverte originale de durée T_e. Dans ce cas, pour conserver la même période fondamentale, il convient de dilater le signal selon les coefficients suivants : $${\mathit{\alpha }}_1=\frac{{\mathit{T}}_0-{\mathit{T}}_{\mathit{e}}^{\mathit{modif}}}{{\mathit{T}}_0-{\mathit{T}}_{\mathit{e}}}\mathrm{pour\; la\; phase\; fermée};$$

$${\mathit{\alpha }}_2=\frac{{\mathit{T}}_{\mathit{e}}^{\mathit{modif}}}{{\mathit{T}}_{\mathit{e}}}\mathrm{pour\; la\; phase\; ouverte}\mathrm{.}$$

For example, we make a modification of the open quotient so that the duration of the open phase becomes $T_e^{\mathrm{mod}\mathit{yew}}$

with $T_e^{\mathrm{mod}\mathit{yew}}<{\mathit{T}}_{\mathit{0}},$

with T ₀ which is the length of a glottal cycle whose closing moment coincides with the origin of the times and an original open phase of duration T _e . In this case, to maintain the same fundamental period, the signal must be dilated according to the following coefficients: $${\mathit{\alpha }}_1=\frac{{\mathit{T}}_0-{\mathit{T}}_{\mathit{e}}^{\mathit{Modified}}}{{\mathit{T}}_0-{\mathit{T}}_{\mathit{e}}}\mathrm{for\; the\; closed\; phase};$$

$${\mathit{<figure-callout\; id="2"\; label="\alpha "\; filenames="imgf0002.png"\; state="\{\{state\}\}">\alpha </figure-callout>}}_2=\frac{{\mathit{T}}_{\mathit{e}}^{\mathit{Modified}}}{{\mathit{T}}_{\mathit{e}}}\mathrm{for\; the\; open\; phase}\mathrm{.}$$

où la fonction g est définie par : $$g\left(t\right)=\{\begin{array}{ll}\frac{T_0-T_e^{\mathit{modif}}}{T_0-T_e}t& \mathit{pour\; t}\in \left[0,T_0-T_e\right]\\ T_0-T_e^{\mathit{modif}}+\frac{T_e^{\mathit{modif}}}{T_e}\left(t\left(T_0-T_e\right)\right)& \mathit{pour\; t}\in \left[T_0-T_e,T_0\right]\end{array}\mathrm{.}$$

Mathematically, this amounts to determining a temporal envelope of the following form: $$d^{\mathit{Modified}}\left(t\right)=\exp \left(\frac{1}{2}{\displaystyle \underset{k=-K}{\overset{K}{\Sigma }}}{\widehat{\mathrm{vs}}}_{\mathrm{<figure-callout\; id="2"\; label="k\; exp"\; filenames="imgf0002.png"\; state="\{\{state\}\}">k\; </figure-callout>}}\exp \left(2\mathit{i\pi kg}\left(t/T_0^{\mathit{Modified}}\right)\right)\right),$$

where the function g is defined by: $$\mathrm{boy\; Wut}\left(t\right)=\{\begin{array}{ll}\frac{T_0-T_e^{\mathit{Modified}}}{T_0-T_e}t& \mathit{for\; t}\in \left[0,T_0-T_e\right]\\ T_0-T_e^{\mathit{Modified}}+\frac{T_e^{\mathit{Modified}}}{T_e}\left(t\left(T_0-T_e\right)\right)& \mathit{for\; t}\in \left[T_0-T_e,T_0\right]\end{array}\mathrm{.}$$

Of course, other types of modification of voice quality parameters are possible according to similar principles.

Step 30 then comprises a determination of the new residue. In the example, this new residue is obtained by multiplying the residue ẽ ^modif (n) by the modified envelope d ^modif .

The original residue was therefore normalized, modified and then combined with the new temporal envelope. This makes it possible to ensure the coherence of its temporal envelope with the modifications of fundamental frequency and / or of vocal quality.

In the embodiment described, the excitation is merged with the residue, which corresponds to the case where the residue is obtained by simple inverse linear filtering and where the excitation comprises no parametric part.

In the case where the excitation is composed of a glottic source that can be modeled by a parametric model and of a residual, it is advisable to make the same type of modification on the glottic source thus parameterized by adjusting the parameters of fundamental frequency and of voice quality.

The method finally comprises a step 40 of synthesis of the modified signal. This synthesis consists of a filtering of the signal obtained at the end of step 20 by the filter of the vocal tract as defined in step 12. Step 40 also comprises an addition - overlap of the frames thus filtered. This synthesis step is conventional and will not be described in more detail here.

Thus, the specific treatment of the temporal envelope of the residue makes it possible to obtain a modification ensuring a good temporal coherence.

Of course, other embodiments can be envisaged.

First, the residue can be broken down into subbands. In this case, steps 14, 16 and 20 are performed on all or part of the sub-bands considered separately. The final residue obtained is then the sum of the modified residues resulting from the different subbands.

In addition, the residue can be decomposed into a deterministic part and a stochastic part. In this case, steps 14, 16 and 20 are performed for each of the parts considered. Here again, the final residue obtained is then the sum of the modified deterministic and stochastic components.

In addition, these two variants can be combined, so that a separate processing on each sub-band and for each of the deterministic and stochastic components can be performed.

In another embodiment, the different steps of the invention can be performed in a different order. For example, the time envelope is changed before changes are made to the signal. Thus, the modifications are made on the residue with its new time envelope and not on the normalized residue as in the example described above.

According to another embodiment, the steps of normalizing the residue and determining new temporal characteristics are combined. In such an embodiment, the residue is directly modified by a time factor determined from its time envelope and modification instructions. This temporal factor makes it possible both to suppress the dependence of the residue with its original temporal characteristics and to apply new temporal characteristics.

Furthermore, the invention may be implemented by a program containing specific instructions which, when executed by a computer, cause the steps described above to be carried out.

The invention can also be implemented by a device comprising appropriate means, such as microprocessors, microcomputers and associated memories, or programmed electronic components.

Such a device can be adapted to implement any embodiment of the method described above.

Claims (11)

1. Method of modifying the acoustic characteristics of a speech signal (s(n)) characterized in that it comprises:

   - a decomposition (12) of the signal into a parametric part and a non-parametric residual (e(n));

   - an estimation (14) of the temporal envelope of the residual;

   - a modification (22) of acoustic characteristics of the parametric part and of the residual according to modification guidelines;

   - a determination (30) of a new temporal envelope for the modified residual by applying the said modification guidelines to the estimated temporal envelope; and

   - a synthesis (40) of a speech signal modified on the basis of the modified parametric part and of the residual as modified and with the new temporal envelope.
2. Method according to Claim 1, characterized in that the said decomposition of the signal is a decomposition according to a model of excitation-filter type.
3. Method according to either of Claims 1 and 2, characterized in that the estimation of the temporal envelope of the residual comprises the estimation of a first envelope and then a temporal smoothing of this first envelope.
4. Method according to any one of Claims 1 to 3, characterized in that it furthermore comprises a temporal normalization (16) of the residual as a function of the estimation of the temporal envelope.
5. Method according to Claim 4, characterized in that the temporal normalization of the residual comprises the division of the residual by the estimation of the temporal envelope.
6. Method according to Claim 4 or 5, characterized in that the determination of a new temporal envelope for the residual comprises a modification (32) of parameters of the temporal envelope of the residual according to the said modification guidelines and an application (34) of the modified temporal envelope to the normalized residual.
7. Method according to either of Claims 1 and 5, characterized in that the estimation of the temporal envelope and the determination of a new temporal envelope are merged.
8. Method according to any one of Claims 1 to 7, characterized in that the modification of acoustic characteristics comprises a modification of information regarding fundamental frequency and duration of the parametric part and of the residual.
9. Program for a device for modifying a speech signal (s(n)), characterized in that it comprises instructions which, when they are executed on a computer of this device, bring about the implementation of a method according to any one of Claims 1 to 8.
10. Device for modifying a speech signal, characterized in that it comprises:

    - means for decomposing the signal into a parametric part and a non-parametric residual (e(n));

    - means for estimating the temporal envelope of the residual;

    - means for modifying acoustic characteristics of the parametric part and of the residual according to modification guidelines;

    - means for determining a new temporal envelope for the modified residual by applying the said modification guidelines to the estimated temporal envelope; and

    - means for synthesizing a speech signal modified on the basis of the modified parametric part and of the residual as modified and with the new temporal envelope.
11. Device according to Claim 10, characterized in that it comprises means suitable for the implementation of a method according to any one of Claims 2 to 8.

Images