EP0594480A1 - Speech detection method - Google Patents

Abstract

The method of the invention consists in: - detecting a voiced frame; - searching for the noise frames preceding this voiced frame; - constructing an auto-regressive model of the noise and a mean noise spectrum; - whitening the frames preceding the voicing; - searching for the actual start of speech in the whitened frames; - ridding the voiced frames of noise and parametrising them; and - searching for the actual end of speech.

Description

The present invention relates to a speech detection method.

• (1) On peut travailler sur l'amplitude instantanée par référence à un seuil déterminé expérimentalement et confirmer la détection de parole par une détection de voisement (voir article "La discrimination parole- bruit et ses applications" de V. PETIT/F. DUMONT, paru dans la Revue Technique THOMSON-CSF - Vol. 12 - N°4, déc. 1980).
• (2) On peut aussi travailler sur l'énergie du signal total sur une tranche temporelle de durée T, en seuillant, toujours expérimentalement, cette énergie à l'aide d'histogrammes locaux, par exemple, et confirmer ensuite à l'aide d'une détection de voisement, ou du calcul de l'énergie minimale d'une voyelle. L'utilisation de l'énergie minimale d'une voyelle est une technique décrite dans le rapport "AMADEUS Version 1.0" de J.L. GAUVAIN du laboratoire LIMSI du CNRS.
• (3) Les systèmes précédents permettent la détection de voisement, mais non pas du début et de la fin effectifs de la parole, c'est-à-dire la détection des sons fricatifs non voisés (/F/, /S/, /CH/) et des sons plosifs non voisés (P/, /T/, /Q/). Il faut donc les compléter par un algorithme de détection de ces fricatives. Une première technique peut consister en l'utilisation d'histogrammes locaux, comme le préconise l'article "Problème de détection des frontières de mots en présence de bruit additifs" de P. WACRENIER, paru dans le mémoire de D.E.A. de l'université de PARIS-SUD, Centre d'Orsay.

When trying to determine the effective start and end of speech, various solutions are possible:

• (1) We can work on the instantaneous amplitude by reference to a threshold determined experimentally and confirm the speech detection by a voice detection (see article "Speech-noise discrimination and its applications" by V. PETIT / F. DUMONT , published in the THOMSON-CSF Technical Review - Vol. 12 - N ° 4, Dec. 1980).
• (2) We can also work on the energy of the total signal over a time slice of duration T, by thresholding, still experimentally, this energy using local histograms, for example, and then confirming using d '' voicing detection, or calculation of the minimum energy of a vowel. The use of the minimum energy of a vowel is a technique described in the report "AMADEUS Version 1.0" by JL GAUVAIN of the LIMSI laboratory of the CNRS.
• (3) The preceding systems allow the detection of voicing, but not of the effective start and end of speech, that is to say the detection of unvoiced fricative sounds (/ F /, / S /, / CH /) and unvoiced plosive sounds (P /, / T /, / Q /). They must therefore be supplemented by an algorithm for detecting these fricatives. A first technique may consist in the use of local histograms, as recommended in the article "Problem of detection of word boundaries in the presence of additive noise" by P. WACRENIER, published in the DEA thesis of the University of PARIS-SUD, Center d'Orsay.

Other techniques similar to the previous ones and relatively close to that exposed here, were presented in the article "A Study of Endpoint Detection Algorithms in Adverse Conditions: Incidence on a DTW and HMM Recognizer" by J.C. JUNQUA / B. REAVES / B. MAK, at the 1991 EUROSPEECH Congress.

In all of these approaches, much is made of heuristics, and few powerful theoretical tools are used.

Much more work on noise reduction similar to that presented here, and in particular the book "Speech Enhancement" by JS LIM published by Prentice-Hall Signal Processing Series "Suppression of Acoustic Noise in Speech Using Spectral Substraction "by SF BOLL, published in the journal IEEE Transactions on Acoustics, speech, and signal processing, vol. ASSP-27, N ° 2, April 1989, and "Noise Reduction For Speech Enhancement In Cars: Non-Linear Spectral Subtraction / Kalman Filtering" by P. LOCKWOOD, C. BAILLARGEAT, JM GILLOT, J. BOUDY, G. FAUCON published in the review EUROSPEECH 91. We will only mention denoising techniques in the spectral domain, and it will be referred to below as "spectral" denoising by abuse of language.

In all of these works, the close relationship between detection and denoising is never really highlighted, except in the article "Suppression of Acoustic Noise in Speech Using Spectral Subtraction", which offers an empirical solution to this problem.

However, it is obvious that a denoising of the speech, when one does not have two channels of recordings, requires the use of frames of “pure” noise, not polluted by the speech, which requires to define a detection tool capable of distinguishing between noise and noise + speech.

The subject of the present invention is a method for detecting and denoising speech which makes it possible to detect as safely as possible the beginnings and end times of speech signals whatever the types of speech sounds, and which makes it possible to denois the most Signals thus detected can be effectively detected, even when the statistical characteristics of the noise affecting these signals vary widely.

The method of the invention consists in carrying out in a low-noise environment a detection of voiced frames, and in detecting a vocal nucleus to which a confidence interval is attached.

In a noisy environment, after having carried out the detection of at least one voiced frame, we search for noise frames preceding this voiced frame, we build an autoregressive model of noise and an average spectrum of noise, we whiten by rejector filter and we denoised by spectral denoiser the frames preceding the voicing, one searches for the effective start of the speech in these bleached frames, one extracts from the noisy frames included between the effective start of the speech and the first voiced frame the acoustic vectors used by the voice recognition system , as long as voiced frames are detected, these are denoised and then parameterized with a view to their recognition (that is to say that the acoustic vectors adapted to the recognition of these frames are extracted), when no detects more voiced frames, we search for the effective end of the speech, we denois then we configure the frames between the last tr voiced soul and the effective end of speech.

In what follows, when it is a question of parameterizing the frames, it will be understood that the acoustic vector (or equivalent, the acoustic parameters) used by the recognition algorithm is extracted from the frame.

An example of such acoustic parameters are the cepstral coefficients well known to specialists in speech processing.

In the following, we will understand by bleaching, the application of the rejector filtering calculated from the autoregressive model of noise, and by denoising, the application of the spectral denoiser.

Spectral whitening and denoising do not apply sequentially, but in parallel, whitening allowing the detection of unvoiced sounds, denoising an improvement in the quality of the voice signal to be recognized.

• ^* Filtrage de réjection, principalement utilisé pour la détection de fricatives, en vertu de ses propriétés de blanchiment.
• ^* Filtrage de Wiener en particulier, utilisé pour débruiter le signal de parole en vue de sa reconnaissance. On peut aussi utiliser une soustraction spectrale.

Thus, the method of the invention is characterized by the use of theoretical tools allowing a rigorous approach to the detection problems (voicing and fricatives), by its great adaptability, because this method is above all a process local to the word. The statistical characteristics of the noise can change over time, the process will remain capable of adapting to it, by construction. It is also characterized by the development of detection expertise from the results of signal processing algorithms (this minimizes the number of false alarms due to detection, taking into account the specific nature of the speech signal ), by denoising processes coupled with speech detection, by a "real time" approach, and this, at all levels of analysis, by its synergy with other techniques for processing the voice signal, by the use of two different denoisers:

• ^* Rejection filtering, mainly used for the detection of fricatives, by virtue of its bleaching properties.
• ^* Wiener filtering in particular, used to denoise the speech signal for recognition. We can also use spectral subtraction.

• - Le niveau "élémentaire" qui met en oeuvre des algorithmes de traitement du signal qui sont en fait les éléments de base de tous les traitements de niveau supérieur.

In the process of the invention, three levels of processing must therefore be distinguished:

• - The "elementary" level which implements signal processing algorithms which are in fact the basic elements of all higher level processing.

Thus, the "elementary" level of voicing detection is an algorithm for calculating and thresholding the correlation function. The result is assessed by the higher level.

• - Le niveau intermédiaire d'expertise élabore des détections de voisements et de début de parole "intelligentes", compte tenu de la détection "brute" fournie par le niveau élémentaire. L'expertise peut faire appel à un langage informatique approprié, type Prolog.
• - Le niveau "supérieur" ou utilisateur gère en temps réel les différents algorithmes de détection, de débruitage et d'analyse du signal vocal. Le langage C, par exemple, est approprié à l'implémentation de cette gestion.

These treatments are implemented on signal processing processors, for example of the DSP 96000 type.

• - The intermediate level of expertise develops "intelligent" voice detection and speech detection, taking into account the "raw" detection provided by the elementary level. Expertise can use an appropriate computer language, such as Prolog.
• - The "higher" or user level manages the different algorithms for detecting, denoising and analyzing the voice signal in real time. The C language, for example, is suitable for the implementation of this management.

The invention is described in detail below according to the following plan. We first describe the algorithm which makes it possible to appropriately chain the various signal processing techniques and necessary expertise.

It will be assumed at this highest level of processing in the design hierarchy, that reliable detection and denoising methods are available, including all the signal processing algorithms, all the expertise, necessary and sufficient. This description is therefore very general. It is even independent of the expertise and signal processing algorithms described below. It can therefore be applied to other techniques than those described here.

We then describe the expertise of voice detection, start and end of speech, using elementary level algorithms, some examples of which are cited.

Finally, the methods used for speech detection and denoising are described.

It is the results of these techniques (voiced, unvoiced speech, etc.) that are used by higher levels of processing.

Conventions and vocabulary used.

We will call frame, the elementary time unit of processing. The duration of a frame is conventionally 12.8 ms, but can, of course, have different values (achievements in mathematical language). The treatments use discrete Fourier transforms of the processed signals. These Fourier transforms are applied to all of the samples obtained on two consecutive frames, which corresponds to performing a Fourier transform over 25.8 ms.

When two Fourier transforms are consecutive in time, these transforms are calculated, not on four consecutive frames, but on three consecutive frames with an overlap of a frame. This is illustrated by the following diagram:

We first describe here the functioning of the algorithm at the design level closest to the user.

The preferred embodiment of the invention is described below with reference to the analysis of signals from very noisy avionic environments, which makes it possible to have starting information which is the micro alternation that use the pilots. This information indicates a time zone close to the signal to be processed.

However, this alternation can be more or less close to the actual start of speech, and therefore we can only give it a small amount of credit for any precise detection. It will therefore be necessary to specify the actual start of speech from this first information.

First, we look for the first voiced frame located around this half-duplex. This first voiced frame is sought first of all from the N1 frames which precede the half-cycle (N1 = about 30 frames). If this voiced frame is not found among these N1 frames, then the voicing is sought on the frames which follow the alternation, as they arise.

• ^* constituée de bruit seul
• ^* constituée de bruit + respiration
• ^* constituée de bruit + fricative ou occlusive non voisée.

As soon as the first voiced frame is found by this method, we will initialize the denoisers. For this, it is necessary to highlight frames made up solely of noise. These noise frames are sought from among the N2 frames which precede the first voiced frame (N2 = approximately 40 frames). Indeed, each of these N2 frames is either:

• ^* consisting of noise alone
• ^* consisting of noise + breathing
• ^* consisting of noise + fricative or occlusive not voiced.

The assumption made is that the noise energy is on average lower than that of noise + respiration, itself lower than that of noise + fricative.

So, if we consider among the N2 frames, the one with the lowest energy, it is very likely that this frame consists only of noise.

From the knowledge of this frame, we look for all those which are compatible with it, and compatible 2 to 2, in the sense given below, in the paragraph "Compatibilities between energies".

• ^* Modèle autorégressif du bruit permettant de construire le filtrage réjecteur qui blanchit le bruit.
• ^* Spectre moyen de bruit pour débruitage spectral.

. When the noise frames have been detected, we build the two noise models which will be used later:

• ^* Autoregressive noise model allowing to build the rejection filter which whitens the noise.
• ^* Average noise spectrum for spectral denoising.

These models are described below.

The noise models having been constructed, the N3 frames which precede the voicing are whitened (by rejection filter and denoised (by spectral denoiser) and among which the effective start of speech is sought (N3 = around 30). that N3 is less than N2. This detection is done by fricative detection and is described below.

When the start of speech is known, all the frames between the start of speech and the first voiced frame are denoised, then these frames are parameterized for recognition. As these frames are denoised and configured, they are sent to the recognition system.

Since the effective start of speech is known, we can continue to process the frames following the first voiced frame.

Each acquired frame is no longer bleached but only denoised, then configured for recognition. A voicing test is carried out on each frame.

If this frame is voiced, the acoustic vector is effectively sent to the recognition algorithm.

If it is not voiced, we seek if it is in fact the last frame of the vocal nucleus in progress.

If it is not the last frame of the vocal nucleus, we acquire a new frame and we repeat the process, until we find the last voiced frame.

When the last voiced frame is detected, the N4 frames which follow this last voiced frame are bleached (N4 = about 30 frames), then the effective end of speech is sought among these N4 bleached frames. The process associated with this detection is described below.

When the effective end of speech is detected, the frames between the end of voicing and this end of speech are denoised, then parameterized and sent to the pure voice recognition system for processing.

When the last speech frame has been denoised, parameterized and sent to the recognition system, all the processing parameters are reset, for the processing of the next speech.

As we can see, this process is local to the spoken speech (that is to say, it treats each sentence or each set of words without "gap" between words), and therefore makes it possible to be very adaptive to any change in noise statistics, especially since we use adaptive algorithms for auto-regressive noise modeling, and relatively sophisticated theoretical models for the detection of noise frames and the detection of fricatives.

In the absence of an alternation, the process is launched as soon as a voicing is detected.

A significant simplification of the method described above is possible when the processed signals are not very noisy. The use of denoising and whitening algorithms can then prove useless, even harmful, when the noise level is negligible (laboratory atmosphere). This phenomenon is known, especially in the case of denoising, where denoising a very noisy signal can induce a distortion of speech detrimental to good recognition.

• - dans la suppression du débruitage spectral pour reconnaissance en vue d'éviter toute déformation de la parole, ne compensant pas le gain en rapport signal sur bruit que l'on pourrait obtenir par débruitage, et préjudiciable alors à une bonne reconnaissance.
• - dans l'éventuelle suppression du filtre de blanchiment (et donc du calcul du modèle autorégressif du bruit, ce qui implique aussi la suppression du module de confirmation de bruit). Cette suppression n'est pas forcément nécessaire en milieu peu bruité. Des essais préalables sont préférables pour en décider.

The simplifications are:

• - In suppressing spectral denoising for recognition in order to avoid any distortion of speech, not compensating for the gain in signal to noise ratio that could be obtained by denoising, and thus detrimental to good recognition.
• - in the eventual removal of the whitening filter (and therefore of the calculation of the autoregressive noise model, which also implies the removal of the noise confirmation module). This removal is not necessarily necessary in low noise environments. Pre-testing is preferable to decide.

We will now describe in detail the expertise procedures for voicemail detection and fricative detection

These expert procedures use well-known signal processing and detection tools, which are all basic automata, the ability of which is to decide roughly whether the processed frame is voiced or not, is a unvoiced fricative or unvoiced plosive frame ...

Expertise consists in combining the different results obtained using these tools, so as to highlight coherent sets, forming the vocal nucleus for example, or blocks of fricative sounds (plosives), not seen.

By nature, the implementation language for such procedures is preferably PROLOG.

Unlike the process described above, this expertise is the same whether the environment is noisy or not.

For the voicing detection expertise, use is made of a known voicing detection process, which, for a given frame, decides whether this frame is voiced or not, by returning the value of the "pitch" associated with this frame. The "pitch" is the frequency of repetition of the voicing pattern. This pitch value is zero, if there is no voicing, and not zero otherwise.

This elementary voicing detection is done without using results relating to the preceding frames, and without predicting results relating to future frames.

As a vowel nucleus can consist of several voiced segments, separated by unvoiced holes, an expertise is necessary, in order to validate or not a voicing.

• Règle 1 : Entre deux trames voisées consécutives ou séparées d'un nombre relativement faible de trames (de l'ordre de trois ou quatre trames), les valeurs de pitch obtenues ne peuvent différer de plus d'un certain delta (environ ± 20 Hz en fonction du locuteur). Par contre, lorsque l'écart entre deux trames voisées excède un certain nombre de trames, la valeur de pitch peut évoluer très vite.
• Règle 2 : Un noyau vocalique est constitué de trames voisées entrecoupées de trous. Ces trous doivent vérifier la condition suivante: la taille d'un trou ne doit pas excéder une taille maximale, qui peut être fonction du locuteur et surtout du vocabulaire (environ 40 trames). La taille du noyau est la somme du nombre de trames voisées et de la taille des trous de ce noyau.
• Règle 3 : le début effectif du noyau vocalique est donné dès que la taille du noyau est suffisamment grande (environ 4 trames).
• Règle 4: la fin du noyau vocalique est déterminée par la dernière trame voisée suivie d'un trou excédant la taille maximale permise pour un trou dans le noyau vocalique.

We will now set out the general rules of expertise.

• Rule 1: Between two voiced frames consecutive or separated by a relatively small number of frames (of the order of three or four frames), the pitch values obtained cannot differ by more than a certain delta (approximately ± 20 Hz depending on the speaker). On the other hand, when the difference between two voiced frames exceeds a certain number of frames, the pitch value can change very quickly.
• Rule 2: A vocal core consists of voiced frames interspersed with holes. These holes must verify the following condition: the size of a hole must not exceed a maximum size, which can be a function of the speaker and especially of the vocabulary (about 40 frames). The size of the core is the sum of the number of voiced frames and the size of the holes in this core.
• Rule 3: the effective start of the vocal nucleus is given as soon as the size of the nucleus is sufficiently large (approximately 4 frames).
• Rule 4: the end of the vocal nucleus is determined by the last voiced frame followed by a hole exceeding the maximum size allowed for a hole in the vocal nucleus.

Expertise process

The above rules are used as described below, and when a pitch value has been calculated.

• On valide ou non la valeur du pitch calculée, en fonction de la valeur du pitch de la trame précédente et de la dernière valeur non nulle du pitch, et ce en fonction du nombre de trames séparant la trame actuellement traitée et celle du dernier pitch non nul. Ceci correspond à l'application de la règle 1.

First part of the expertise:

• The calculated pitch value is validated or not, as a function of the pitch value of the preceding frame and of the last non-zero value of the pitch, and this as a function of the number of frames separating the frame currently processed and that of the last pitch not no. This corresponds to the application of rule 1.

Second part of the expertise:

This second part of the expertise is broken down according to different cases.

• - On incrémente la taille possible du noyau, qui vaut donc 1
• - Le début possible du noyau vocalique est donc la trame actuelle
• - La fin possible du noyau vocalique est donc la trame actuelle

Case 1: First voiced frame:

• - We increment the possible size of the kernel, which is therefore worth 1
• - The possible beginning of the vocal nucleus is therefore the current frame
• - The possible end of the vocal nucleus is therefore the current frame

Case 2: The current frame is seen as well as the previous one.

• - On incrémente le nombre possible de trames voisées du noyau
• - On incrémente la taille possible du noyau
• - La fin possible du noyau peut être la trame actuelle qui est aussi la fin possible du segment. Si la taille du noyau est suffisamment grande (environ quatre trames, comme précisé ci-dessus). Et Si le début effectif du noyau vocalique n'est pas connu. Alors :
• - le début du noyau est la première trame détectée comme voisée.

We therefore treat a voiced segment.

• - We increment the possible number of voiced frames of the kernel
• - We increase the possible size of the nucleus
• - The possible end of the kernel can be the current frame which is also the possible end of the segment. If the kernel size is large enough (about four frames, as specified above). And if the actual start of the vocal nucleus is not known. So :
• - the beginning of the kernel is the first frame detected as voiced.

This corresponds to the implementation of rule 3.

Case 3: the current frame is not seen, while the previous frame is.

• - On incrémente la taille du trou, qui passe à 1

We are processing the first frame of a hole.

• - We increase the size of the hole, which goes to 1

Case 4: The current frame is not seen and neither is the previous frame.

We are treating a hole.

• - We increase the size of the hole.

If the hole size exceeds the maximum size allowed for a vowel core hole,

So :

If the effective start of voicing is known,

So :

the end of the vocal nucleus is the last voiced frame determined before this hole. We stop the expertise and reset all the data for the processing of the next speech. (see rule 4)

If the actual start of speech is still unknown,

So :

The expertise is continued on the following frames after reinitialization of all the parameters used, because those which have been updated previously are not valid.

Otherwise, this hole may be part of the vowel nucleus and we cannot yet make a final decision.

Case 5: The current frame is seen and the previous one is not.

• - On incrémente le nombre de trames voisées du noyau.
• - On incrémente la taille du noyau.

Si le trou que l'on vient de finir peut faire partie du noyau vocalique, (c'est-à-dire si sa taille est inférieure à la taille maximale autorisée pour un trou du noyau selon la règle 2).We just finished a hole, and we're starting a new voiced segment.

• - We increment the number of voiced frames of the kernel.
• - We increase the size of the nucleus.

If the hole that we have just finished can be part of the vowel nucleus, (that is to say if its size is less than the maximum size authorized for a nucleus hole according to rule 2).

So :

- The size of this hole is added to the current size of the core.

- The hole size is reset, for processing of the next unvoiced frames.

• - le début du voisement est le début du segment voisé précédant le trou que l'on vient de terminer. Sinon, ce trou ne peut pas faire partie du noyau vocalique :

If the effective start of voicing is not yet known, And If the size of the kernel is now sufficient (Rule 3), Then:

• - the start of voicing is the beginning of the voiced segment preceding the hole that we have just finished. Otherwise, this hole cannot be part of the vowel core:

If the effective start of the voicing is known,

So :

• - the end of the vocal nucleus is the last voiced frame determined before this hole. We stop the expertise and reset all the data for the processing of the next speech. (see rule 4).

If the effective start of voicing is still not known,

So :

• - The expertise is continued on the following frames after reinitialization of all the parameters used, because those which have been updated previously are not valid.

This procedure is used for each frame, and after calculating the pitch associated with this frame.

Unvoiced speech detection expertise.

Here we use a process known per se for detection of unvoiced speech.

This elementary voicing detection is done without using results relating to the preceding frames, and without predicting results relating to future frames.

• - d'un seul segment fricatif comme dans "chaff'
• - d'un segment fricatif suivi d'un segment occlusif comme dans "stop"
• - d'un seul segment occlusif comme dans "parole"

Unvoiced speech signals placed at the start or end of speech can be made up:

• - of a single fricative segment as in "chaff '
• - a fricative segment followed by an occlusive segment as in "stop"
• - of a single occlusive segment as in "speech"

There is therefore the possibility of holes in the set of unvoiced frames.

In addition, such fricative blocks should not be too large. Also, an expertise intervening after the detection of these sounds is necessary.

In the following, for abuse of language, the term fricative will refer just as well to unvoiced fricatives as to unvoiced plosives.

General rules of expertise.

• Règle 1 : la distance entre le noyau vocalique et la première trame fricative détectée ne doit pas être trop grande (environ 15 trames maximum) .
• Règle 2 : la taille d'un bloc fricatif ne doit pas être trop grande. Ceci signifie de manière équivalente, que la distance entre le noyau vocalique et la dernière trame détectée comme fricative ne doit pas être trop grande (environ 10 trames maximum).
• Règle 3 la taille d'un trou dans un bloc fricatif ne doit pas excéder une taille maximale (environ 15 trames maximum). La taille totale du noyau est la somme du nombre de trames voisées et de la taille des trous dans ce noyau.
• Règle 4: le début effectif du bloc fricatif est déterminé dès que la taille d'un segment est devenue suffisante, et que la distance entre le noyau vocalique et la première trame de ce segment fricatif traité n'est pas trop grande, conformément à la règle 1. Le début effectif du bloc fricatif correspond à la première trame de ce segment.
• Règle 5 : la fin du bloc fricatif est déterminée par la dernière trame du bloc fricatif suivie d'un trou excédant la taille maximale autorisée pour un trou dans le noyau vocalique, et lorsque la taille du bloc fricatif ainsi déterminé n'est pas trop grande conformément à la règle 2.

The expertise presented here is similar to that described above in the case of voicing. The differences are mainly due to the taking into account of the new parameters which are the distance between the vowel nucleus and the fricative block, and the size of the fricative block.

• Rule 1: the distance between the vocal nucleus and the first detected fricative frame must not be too large (about 15 frames maximum).
• Rule 2: the size of a fricative block should not be too large. This equivalently means that the distance between the vocal nucleus and the last frame detected as fricative must not be too large (about 10 frames maximum).
• Rule 3 the size of a hole in a fricative block must not exceed a maximum size (about 15 frames maximum). The total size of the core is the sum of the number of voiced frames and the size of the holes in this core.
• Rule 4: the effective start of the fricative block is determined as soon as the size of a segment has become sufficient, and the distance between the vowel nucleus and the first frame of this treated fricative segment is not too great, in accordance with the rule 1. The actual start of the fricative block corresponds to the first frame of this segment.
• Rule 5: the end of the fricative block is determined by the last frame of the fricative block followed by a hole exceeding the maximum size authorized for a hole in the vocal nucleus, and when the size of the fricative block thus determined is not too large in accordance with rule 2.

Conduct of the expertise.

This expertise is used to detect the fricative blocks preceding or following the vowel nucleus. The benchmark chosen in this expertise is therefore the vocal core.

In the case of the detection of a fricative block preceding the vowel nucleus, the processing is done starting from the first voicing frame, therefore by "going back" in time. Also, when we say that a frame i follows a frame j (previously treated), we have to understand by this: vis-à-vis this first frame of the vowel nucleus. In reality, the frame j is chronologically posterior to the frame i. What is called the start of the fricative block in the appraisal described below, is in fact, chronologically, the end of this block, and what is called the end of the fricative block, is in fact the chronological start of this block. The distance between the vocal nucleus and the frame detected as fricative is the distance between the first frame of the voiced block and this fricative frame.

In the case of the detection of a fricative block situated after the vocal nucleus, the processing is done after the last voiced frame, and therefore follows the natural chronological order, and the terms of the expertise are perfectly adequate.

• - On incrémente la distance entre le segment voisé et le bloc fricatif. Cette distance ainsi calculée est un minorant de la distance entre le bloc fricatif et le noyau vocalique. Cette distance sera figée dès que la première trame de fricative sera détectée.

Case 1: As long as there is no fricative detection, we are in a hole which follows the vowel nucleus and precedes the fricative block.

• - The distance between the voiced segment and the fricative block is increased. This distance thus calculated is a minus of the distance between the fricative block and the vowel nucleus. This distance will be fixed as soon as the first frame of fricative is detected.

• - On initialise la taille du bloc fricatif à 1.
• - On fige la distance entre le bloc voisé et le bloc fricatif.

Case 2: First detection of fricative, We begin to treat a fricative segment.

• - The size of the fricative block is initialized to 1.
• - The distance between the voiced block and the fricative block is frozen.

If the distance between the vocal nucleus and the fricative block is not too great (in accordance with rule 2).

• - Le début possible du bloc fricatif peut être la trame actuelle.
• - La fin possible du bloc fricatif peut être la trame actuelle. Si la taille du bloc fricatif est suffisamment grande Et Si le début effectif du bloc fricatif n'est pas encore connu, alors :
• - le début du noyau peut être confirmé.

So :

• - The possible start of the fricative block can be the current frame.
• - The possible end of the fricative block can be the current frame. If the size of the fricative block is sufficiently large And If the effective start of the fricative block is not yet known, then:
• - the beginning of the nucleus can be confirmed.

It will be noted that this If (in "If the size of the fricative block is sufficiently large") is useless if the minimum size for a fricative block is greater than a frame, but when one seeks to detect occlusive in noisy medium, these these can appear only for the duration of a single frame. We must therefore take the minimum size of a fricative block equal to 1, and keep this condition.

If the distance between the vowel nucleus and the fricative block is too great (see rule 2).

• - On réinitialise pour le traitement de la prochaine élocution.
• - On sort du traitement.

There is no acceptable fricative block.

• - We reset for the processing of the next speech.
• - We are leaving treatment.

As the test on the distance between the vocal nucleus and the fricative block is carried out as of the first detection of fricative, it will not be repeated in the following cases, especially since if this distance is too great here, the procedure is stopped for this speech.

Case 3: The current frame and the previous frame are both fricative frames.

• - La fin possible du bloc fricatif est la trame actuelle.
• - On incrémente la taille du bloc fricatif.

We are in the process of processing a frame which is right in the middle of an acceptable fricative segment (located at a correct distance from the vowel core in accordance with rule 1).

• - The possible end of the fricative block is the current frame.
• - We increase the size of the fricative block.

If the size of the fricative block is large enough (see rule 4).

And if the size of this block is not too large (see rule 2).

• - le début du noyau peut être confirmé comme étant le début de ce segment fricatif. Cas 4: La trame actuelle n'est pas une fricative contrairement à la trame précédente. On est en train de traiter la première trame d'un trou situé à l'intérieur du bloc fricatif.
• - On incrémente la taille totale du trou (qui devient égale à 1). Cas 5 : Ni la trame actuelle ni la précédente ne sont des trames de fricatives. On est en train de traiter une trame située en plein dans un trou du bloc fricatif.
• - On incrémente la taille totale du trou.

And if the actual start of the fricative block is not yet known, then:

• - the beginning of the nucleus can be confirmed as the beginning of this fricative segment. Case 4: The current frame is not a fricative unlike the previous frame. We are currently processing the first frame of a hole located inside the fricative block.
• - We increment the total size of the hole (which becomes equal to 1). Case 5: Neither the current nor the previous frame are fricative frames. We are currently processing a frame located right in a hole in the fricative block.
• - We increase the total size of the hole.

If the current size of the fricative block increased by the size of the hole is greater than the maximum size authorized for a fricative block (rule 2).

Or If the size of the hole is too large.

If the start of the fricative block is known,

• - La fin du bloc fricatif est la dernière trame détectée comme fricative.
• - On réinitialise toutes les données de manière à traiter la prochaine élocution.

so :

• - The end of the fricative block is the last frame detected as fricative.
• - We reset all the data in order to process the next speech.

• - on réinitialise toutes les données, même celles qui ont été précédemment actualisées, car elles ne sont plus valides. On traite alors la prochaine trame.

If not :

• - we reset all data, even those that were previously updated, because they are no longer valid. We then process the next frame.

Otherwise, this hole may be part of the fricative block and we cannot yet make a final decision. Case 6: The current frame is a fricative frame unlike the previous frame. The first frame of a fricative segment located after a hole is processed.

- We increase the size of the fricative block.

If the current size of the fricative block increased by the size of the previously detected hole is greater than the maximum size authorized for a fricative block, Or If the size of the hole is too large,

• - La fin du bloc fricatif est alors la dernière trame détectée comme fricative.
• - On réinitialise toutes les données de manière à traiter la prochaine élocution. Sinon,
• - On réinitialise toutes les données, même celles qui ont été précédemment actualisées, car elles ne sont pas valides. On traite alors la prochaine trame.

then: If the start of the fricative block is known, then:

• - The end of the fricative block is then the last frame detected as fricative.
• - We reset all the data in order to process the next speech. If not,
• - We reset all data, even those that were previously updated, because they are not valid. We then process the next frame.

• - La taille du bloc fricatif est augmentée de la taille du trou
• - La taille du trou est réinitialisée à O Si la taille du bloc fricatif est suffisamment grande Et Si cette taille n'est pas trop grande Et Si le début effectif du bloc fricatif n'est pas connu

Otherwise, (the hole is part of the fricative segment).

• - The size of the fricative block is increased by the size of the hole
• - The size of the hole is reset to O If the size of the fricative block is sufficiently large And If this size is not too large And If the effective start of the fricative block is not known

• - Le début du noyau peut être confirmé.

So :

• - The beginning of the nucleus can be confirmed.

Simplification in the case of a slightly noisy environment.

In the event that the user considers that the environment is insufficiently noisy to require the preceding sophisticated treatments, it is possible not only to simplify the expertise presented above, but even to eliminate it. In this case, speech detection will be reduced to a simple detection of the vocal nucleus to which a confidence interval expressed in number of frames is attached, which proves to be sufficient to improve the performance of a speech recognition algorithm. It is thus possible to start recognition ten, or even fifteen frames before the start of the vocal nucleus, and to complete it, ten, or even fifteen frames after the vocal nucleus.

Signal Processing Algorithms.

The calculation procedures and methods described below are the constituents used by the expertise and management algorithms. Such functions are advantageously implemented on a signal processor and the language used is preferably the assembler.

For the detection of voices in low noise environments, an interesting solution is the thresholding of the A.M.D.F. (Average Magnitude Difference Function), the description of which can be found, for example, in the book "Speech processing" by R. BOITE / M. KUNT published by Presses Polytechniques Romandes.

• D(k) ≦ 2(Γ_x(0) - Γ_x(k))^½. Cette fonction présente donc des "pics" vers le bas, et doit donc être seuillée comme la fonction de corrélation.

The AMDF is the function D (k) = In 1 x (n + k) - x (n) 1. This function is bounded by the correlation function, according to:

• D (k) ≦ 2 (Γ _x (0) - Γ _x (k)) ^½ . This function therefore has downward "peaks", and must therefore be thresholded like the correlation function.

Other methods based on the calculation of the signal spectrum can be envisaged, for equally acceptable results (article "speech processing" mentioned above). However, it is interesting to use the AMDF function, for simple questions of calculation costs.

In noisy environments, the AMDF function is a distance between the signal and its delayed form. However, this distance is a distance which does not admit an associated scalar product, and which therefore does not allow the concept of orthogonal projection to be introduced. However, in a noisy environment, the orthogonal projection of the noise can be zero, if the axis of projection is well chosen. AMDF is therefore not an adequate solution in noisy environments.

• r_x(k) = E[x(n)x(n-k)], donc r_x(k) = E[s(n)s(n-k)] + E[b(n)b(n-k)] = r_s(k) + r_b(k) Comme le bruit est blanc : r_x(O) ^{= r}s(^O) ⁺ r_b(O) et r_x(k) = r_s(k) pour k ≠0 .

The method of the invention is then based on correlation, because the correlation is a scalar product and performs an orthogonal projection of the signal on its delayed form. This method is therefore more robust to noise than other techniques, such as AMDF. Indeed, suppose that the observed signal is x (n) = s (n) + b (n) where b (n) is white noise independent of the useful signal s (n). The correlation function is by definition:

• r _x (k) = E [x (n) x (nk)], so r _x (k) = E [s (n) s (nk)] + E [b (n) b (nk)] = r _s (k) + r _b (k) As the noise is white: r _x (O) ^{= r} s ( ^O ) ⁺ r _b (O) and r _x (k) = r _s (k) for k ≠ 0.

Noise whiteness in practice is not a valid assumption. However, the result remains good approximation as soon as the noise correlation function decreases rapidly, and for k sufficiently large, as in the case of pink noise (white noise filtered by a bandpass), where the correlation function is a cardinal sine, therefore practically zero as soon as k is sufficiently large.

We will now describe a pitch calculation and pitch detection procedure applicable to noisy environments as well as to noisy environments.

Let x (n) be the processed signal where n e {0, ..., N-1}.

In the case of AMDF, r (k) = D (k) = En | x (n + k) - x (n).

In the case of correlation, the mathematical expectation allowing access to the correlation function can only be estimated, so that the function r (k) is: r (k) = K Σ _{0 ≦ n ≦ N -1} x (n) x (nk) where K is a calibration constant.

In both cases, we theoretically obtain the value of the pitch by proceeding as follows: r (k) is maximum in k = 0. If the second maximum of r (k) is obtained in k = k _o , then the value of the voicing is F _o = F _e / k ₀ where F _e is the sampling frequency.

However, this theoretical description must be revised in practice.

• k₁ = F_e/F₀ = 10000/100 = 100 et celui correspondant à 333 Hz vaut k₂ = F_e/F₀ = 10000/333 = 30.

Indeed, if the signal is known only on samples 0 to N-1, then x (nk) is taken zero as long as n is not greater than k. There will therefore not be the same number of calculation points from one value k to the other. For example, if the pitch range is taken equal to [100 Hz, 333 Hz], and this, for a sampling frequency of 10 KHz, the index k ₁ corresponding to 100 Hz is equal to:

• k ₁ = F _e / F ₀ = 10000/100 = 100 and the one corresponding to 333 Hz is k ₂ = F _e / F ₀ = 10000/333 = 30.

The pitch calculation for this range will therefore be from k = 30 to k = 100.

If for example there are 256 samples (2 frames of 12.8 ms sampled at 10 KHz), the calculation of r (30) is done from n = 30 to n = 128, i.e. on 99 points and that of r (100 ) from n = 100 to 128, i.e. on 29 points.

The calculations are therefore not homogeneous between them and do not have the same validity.

For the calculation to be correct, the observation window must always be the same regardless of k. So that if n-k is less than 0, it is necessary to have stored in memory the past values of the signal x (n), so as to calculate the function r (k) on as many points, whatever k. The value of the constant K no longer matters.

This is detrimental to the calculation of the pitch only on the first frame actually voiced, since, in this case, the samples used for the calculation come from an unvoiced frame, and are therefore not representative of the signal to be processed. However, from the third consecutive voiced frame, when working, for example, in frames of 128 points sampled at 10 KHz, the pitch calculation will be valid. This supposes, in general, that a voicing lasts at least 3x12.8 ms, which is a realistic assumption. This hypothesis must be taken into account during the appraisal, and the minimum duration to validate a voiced segment will be 3x12.8 ms in this same appraisal.

This function r (k) being calculated, it is then a question of thresholding it. The threshold is chosen experimentally, according to the dynamics of the signals processed. Thus, in an example of application, where the quantization is done on 16 bits, where the dynamics of the samples does not exceed ± 10000, and where the calculations are done for N = 128 (Sampling frequency of 10 KHz), we have chosen Threshold = 750000. But remember that these values are only given as an example for particular applications, and must be modified for other applications. In any case, this does not change the methodology described above.

We will now describe the method of detecting noise frames.

• 1) bruit seul
• 2) bruit + fricative non voisée
• 3) bruit + respiration.

Apart from the vocal nucleus, the signal frames that can be encountered are of three types:

• 1) noise alone
• 2) noise + voiceless fricative
• 3) noise + breathing.

The detection algorithm aims to detect the start and end of speech from a whitened version of the signal, while the denoising algorithm requires knowledge of the average noise spectrum. To build the noise models which will make it possible to whiten the speech signal for the detection of unvoiced sounds as described below, and to denoise the speech signal, it is obvious that it is necessary to detect the noise frames , and confirm them as such. This search for noise frames is done among a number of Ni frames defined by the user once and for all for its application (for example for Ni = 40), these Ni frames being located before the vocal nucleus.

Recall that this algorithm allows the implementation of noise models, and is therefore not used when the user judges the noise level insufficient.

• Une variable aléatoire X sera dite positive lorsque Pr{ X < 0 } «1 .
• Soit X_o la variable centrée normalisée associée à X. On a :
• Pr { X < 0 } = Pr { X_o <-m/σ} où m = E[X] et σ² = E[(X-m)²].

We will first define the "positive" Gaussian random variables:

• A random variable X will be said to be positive when Pr {X <0} "1.
• Let X _{o be} the normalized centered variable associated with X. We have:
• Pr {X <0} = Pr {X _o <-m / σ} where m = E [X] and σ ² = E [(Xm) ² ].

As soon as m / σ is large enough, X can be considered positive.

When X is Gaussian, we denote by F (x) the distribution function of the normal distribution, and we have: Pr {X <0} = F (-m / σ) for X ∈ N (m, σ ² )

Résultat fondamental :

• Si X = X₁/X₂ où X₁ et X₂ sont toutes deux des variables aléatoires gaussiennes, indépendantes, telles que X₁ ∈ N(m₁ ; σ₁ ²) et X₂ ∈ N( m₂ ; σ₂ ²), on pose m = m₁/m₂, α₁ ⁼ m₁/σ₁, a₂ ⁼ m₂/σ₂.

An immediate essential property is that the sum X of N independent positive Gaussian variables X _i ∈ N (m _i ; σ _i ² ) remains a positive Gaussian variable:

Fundamental result:

• If X = X ₁ / X ₂ where X ₁ and X ₂ are both independent Gaussian random variables, such as X ₁ ∈ N (m ₁ ; σ ₁ ² ) and X ₂ ∈ N (m ₂ ; σ ₂ ² ), we set m = m ₁ / m ₂ , α ₁ ⁼ m ₁ / σ ₁ , a ₂ ⁼ m ₂ / σ ₂ .

où U(x) est la fonction indicatrice de R⁺ :

• U(x) = 1 si x ≦ 0 et U(x)=0 si x < 0
• Dans toute la suite, on posera :
  
  
  de sorte que : f_x(x) = f(x,m |α₁,α₂).U(x) Soit
  
  On pose P(x,y |α,β) = F[h(x,y |α,β)]. On a alors :

When α ₁ and α ₂ are large enough to assume X ₁ and X ₂ positive, the probability density f _x (x) of X = X ₁ / X ₂ can then be approximated by:

where U (x) is the indicator function of R ⁺ :

• U (x) = 1 if x ≦ 0 and U (x) = 0 if x <0
• In the following, we will ask:
  
  
  so that: f _x (x) = f (x, m | α ₁ , α ₂ ) .U (x) Let
  
  We set P (x, y | α, β) = F [h (x, y | α, β)]. We then have:

Pr {X <x} = P (x, m | α ₁ , α ₂ ) f (x, y | α, β) = ∂P (x, y | α, β) / ∂x and f (x, y | α ₁ , α ₂ ) = ∂P (x, m | α ₁ , α ₂ ) / ∂x Special case: a = β. We will pose: f _α (x, y) = f (x, y | α, β), h _α (x, y) = h (x, y | α, β) and P _α (x, y) = P (x, y | α, β)

• (1) Signal à énergie déterministe : Soient les échantillons x(0),...x(N-1) d'un signal quelconque, dont l'énergie est déterministe et constante, ou approximée par une énergie déterministe ou constante. On a donc U = Σ_0≦n≦N-1x(n)²∈N(Nµ, 0) où µ = (1/N)Σ_0≦n≦N-1x(n)²

We will describe below some basic models of "positive" Gaussian variables usable in the following.

• (1) Signal with deterministic energy: Let the samples x (0), ... x (N-1) of any signal, whose energy is deterministic and constant, or approximated by a deterministic or constant energy. We therefore have U = Σ _{0 ≦ n ≦ N-1} x (n) ² ∈N (Nµ, 0) where µ = (1 / N) Σ _{0 ≦ n ≦ N-1} x (n) ²

• (1/N) E_0≦n≦N-1x(n)²# E[x(n)²] =A²/2. Pour N assez grand, U peut être assimilé à NA²/2 et donc à une énergie constante.
• (2) Processus Blanc Gaussien : Soit un processus blanc et gaussien x(n) tel que σ_x ² = E[x(n)²] . Pour N suffisamment grand, U ⁼ Σ_0≦n≦N-1x(n)² ∈ N(Nσ_x ² ; 2Nσ_x ⁴). Le paramètre a est a = (N/2)^1/2
• (3) Processus Gaussien Bande Etroite : Le bruit x(n) est issu de l'échantillonnage du processus x(t), lui-même issu du filtrage d'un bruit blanc gaussien b(t) par un filtre passe bande h(t) : x(t) = (h*b)(t), en supposant que la fonction de transfert du filtre h(t) est :

Let us take as an example the signal x (n) = Acos (n + θ) where θ is equidistributed between [0.2π]. For N sufficiently large, we have:

• (1 / N) E _{0 ≦ n ≦ N-1} x (n) # ² E [x (n) ^2] = A ^2/2. For large enough N, U can be likened to NA ^2/2 and thus a constant energy.
• (2) Gaussian White Process: Let be a white and Gaussian process x (n) such that σ _x ² = E [x (n) ² ]. For N sufficiently large, U ⁼ Σ _{0 ≦ n ≦ N-1} x (n) ² ∈ N (Nσ _x ² ; 2Nσ _x ⁴ ). Parameter a is a = (N / 2) ^1/2
• (3) Narrow Band Gaussian Process: The noise x (n) comes from the sampling of the process x (t), itself resulting from the filtering of a white Gaussian noise b (t) by a band pass filter h ( t): x (t) = (h * b) (t), assuming that the transfer function of the filter h (t) is:

H (f) = U _{[- (f0-B / 2, -f0 + B / 2]} (f) + U _{[f0-B / 2, f0 + B / 2]} (f), where U denotes the characteristic function of the interval in index and f _o the central frequency of the filter.

We therefore have U ∈ N (Nσ _x ² , 2σ _x ⁴ Σ _{0 ≦ i ≦ N-1.0 ≦ j ≦ N-1} g _{f0, B, Te} (ij) ² ) with g _{f0, B, Te} (k ) ⁼ cos (2πtkf ₀ T _e ) sin _c (πkBT _e ) The parameter a and a = N / [2Σ _{0 ≦ i ≦ N-1.0 ≦ j ≦ N-1} g _{f0, B, Te} (ij) ² ] ^½ - Subsampling of a Gaussian process: This model is more practical than theoretical. If the correlation function is unknown, we know however that: lim k → + ∞ Γ _x (k) = 0. Therefore, for k large enough such that k> k _o , the correlation function tends to 0. Also, at instead of processing the sequence of samples x (0) ... x (N-1), can we process the subsequence x (0), x (k _o ), x (2k _o ), ... , and the energy associated with this sequence remains a positive Gaussian random variable, provided that there remain in this subsequence enough points to be able to apply the approximations due to the central limit theorem. Compatibility between energies.

Let Ci = N (m ₁ , σ ₂ ) and C ₂ = N (m ₂ , σ ₂ ² ) We set: _m = m ₁ / m ₂ , α ₁ = m ₁ / σ ₁ and α ₂ ⁼ m ₂ / σ ₂ . α ₁ and a ₂ are large enough that the random variables of Ci and C ₂ can be considered as positive random variables.

Let (U, V) where (U, V) belong to (C ₁ CU ₂ ) X (C ₁ CU ₂ ).

As before, U and V are assumed to be independent.

We set U ≡ V ⇔ (U, V) ∈ (C ₁ XC ₁ ) U (C ₂ UC ₂ ).

Let (u, v) be a value of the couple (U, V). If x = u / v, x is a value of the random variable X = U / V.

Let s> 1. 1 / s <x <s ⇔ We decide that U ≡ V is true, which will be the decision D = D ₁ x <1 / s or x> s ⇔ We decide that U ≡ V is false, which will be the decision D = D ₂ This decision rule is therefore associated with 2 hypotheses:

H ₁ ⇔U≡V is true, H ₂ ⇔U≡V is false. We will set 1 = [1 / s, s].

The detection rule is also expressed according to: x ∈ I⇔D = D ₁ , x ∈ RI ⇔ D = D ₂ We will say that u and v are compatible when the decision D = D ₁ is taken.

This decision rule admits a correct decision probability, the expression of which will in fact depend on the value of the probabilities Pr {H ₁ } and Pr {H ₂ }. However, these probabilities are generally not known in practice.

We therefore prefer a Neyman-Pearson type approach, since the decision rule is reduced to two hypotheses, seeking to ensure a certain value fixed a priori for the probability of false alarm which is:

• Soit { u₁,..., u_n} un ensemble de valeurs de variables aléatoires gaussiennes positives. On dira que ces valeurs sont compatibles entre elles, si et seulement si les u_i sont compatibles 2 à 2.

The choice of signal and noise models determines α ₁ and a ₂ . We will see that then m appears to be homogeneous at a signal-to-noise ratio which will be fixed heuristically. The threshold is then fixed so as to ensure a certain value of P _fa . Special case: α ₁ = α ₂ = α. It then comes: P _fa = P _α (s, m) - P _α (1 / s, m) Compatibility of a set of values:

• Let {u ₁ , ..., u _n } be a set of values of positive Gaussian random variables. We will say that these values are compatible with each other, if and only if the u _i are compatible 2 to 2.

Signal and noise models used by the method of the invention.

Hypothèse 2 : Le signal utile est perturbé par un bruit additif noté x(n), que l'on suppose gaussien et en bande étroite. On suppose que le processus x(n) traité est obtenu parfiltrage bande étroite d'un bruit blanc gaussien. La fonction de corrélation d'un tel processus est alors :

In order to apply the procedures corresponding to the previous theoretical reminders, it is necessary to fix a noise and signal model. We will use the following example. This model is governed by the following hypotheses: Hypothesis 1: We suppose not to know the useful signal in its form, but we will make the following hypothesis: V the value s (0), ..., s (N-1) of s (n), the energy S = (1 / N) Σ _{0 ≦ n ≦ N} - ₁ s (n) 2 is bounded by µ _s ² , as soon as N is sufficiently large, so that:

Hypothesis 2: The useful signal is disturbed by an additive noise denoted x (n), which is assumed to be Gaussian and in narrow band. We suppose that the process x (n) processed is obtained by narrow band filtering of a white Gaussian noise. The correlation function of such a process is then:

2σ_x ⁴Σ_{0≦i≦N1,0jN-1}g_f0,B,Te(i-j)²) Le paramètre a de cette variable est a = N/[2Σ_{0≦i≦N-1,0≦j≦N-1}g_f0,B,Te(i-j)²]^½ Hypothèse 3 : Les signaux s(n) et x(n) sont alors supposés indépendants. On suppose que l'indépendance entre s(n) et x(n) implique la décorrélation au sens temporel du terme, c'est-à-dire que l'on peut écrire :If we consider N samples x (n) of this noise, and we set: g _fO , _B , _Te (k) = cos (2πkf ₀ T _e ) sin _c (πkBT _e ), we have:

2σ _x ⁴ Σ _{0 ≦ i ≦ N1,0jN-1} g _{f0, B, Te} (ij) ² ) The parameter a of this variable is a = N / [2Σ _{0 ≦ i ≦ N-1.0 ≦ j ≦ N -1} g _{f0, B, Te} (ij) ² ] ^½ Hypothesis 3: The signals s (n) and x (n) are then assumed to be independent. We suppose that the independence between s (n) and x (n) implies decorrelation in the temporal sense of the term, that is to say that we can write:

• E[s(n)x(n)]/(E[s(n)²]E[x(n)²])^1/2 lorsque les processus sont ergodiques.
• Soit u(n) = s(n) + x(n) le signal total, et U = Σ_0≦n≦N-1-u(n)2.
• On peut alors approximer U par : U = Σ_0≦n≦N-1s(n)² + Σ_0≦n≦N-1x(n)²
• Comme on a : Σ_0≦n≦N-1s(n)² ≧ µ_s ²,
• on aura : U ≧ Nµ_s ² + Σ_0≦n≦N-1x(n)².
• Hypothèse 4 : Comme nous supposons que le signal présente une énergie moyenne bornée, nous supposerons qu'un algorithme capable de détecter une énergie µ_s ², sera capable de détecter tout signal d'énergie supérieure. Compte tenu des hypothèses précédentes, on définit la classe Ci comme étant la classe des énergies lorsque le signal utile est présent. Selon l'hypothèse 3, U ≧ Nµ_s ² + Σ_0≦n≦N-1x(n)², et selon l'hypothèse 4, si on détecte l'énergie Nµ_s ² + Σ_0≦n≦N-1x(n)², on saura détecter aussi l'énergie totale U.

This correlation coefficient is only the expression in the time domain of the spatial correlation coefficient defined by:

• E [s (n) x (n)] / (E [s (n) ² ] E [x (n) ² ]) ^1/2 when the processes are ergodic.
• Let u (n) = s (n) + x (n) be the total signal, and U = Σ _{0 ≦ n ≦ N-1-} u (n) 2.
• We can then approximate U by: U = Σ _{0 ≦ n ≦ N-1} s (n) ² + Σ _{0 ≦ n ≦ N-1} x (n) ²
• As we have: Σ _{0 ≦ n ≦ N-1} s (n) ² ≧ µ _s ² ,
• we will have: U ≧ Nµ _s ² + Σ _{0 ≦ n ≦ N-1} x (n) ² .
• Hypothesis 4: As we assume that the signal has a bounded average energy, we will assume that an algorithm capable of detecting an energy µ _s ² , will be able to detect any signal of higher energy. Taking into account the preceding hypotheses, the class Ci is defined as being the class of energies when the useful signal is present. According to hypothesis 3, U ≧ Nµ _s ² + Σ _{0 ≦ n ≦ N-1} x (n) ² , and according to hypothesis 4, if we detect the energy Nµ _s ² + Σ _{0 ≦ n ≦ N- 1} x (n) ² , we will also be able to detect the total energy U.

According to hypothesis 2,

Donc C₁ ⁼ N(Nµ_s ²+Nσ_x ², 2σ_x ⁴ Σ_{0≦i≦N-1,0≦j≦N-1}g_f0,B,Te(i-j)²) et le paramètre a de cette variable vaut α₁ = N(1+r)/[2Σ_{0≦i≦N-1,0≦j≦N-1}g_f0,B,Te(i,j)²]^½, où r = µ_s ²/σ_x ² représente le rapport signal à bruit. C₂ est la classe des énergies correspondant au bruit seul. D'après l'hypothèse 2, si les échantillons de bruit sont x(0),...,x(M-1), il vient V = (1/M)Σ_0≦n≦M-1x(n)² ∈ N (Mσ_x ², 2σ_x ⁴ Σ_{0≦i≦M-1,0≦j≦M-1}g_f0,B,Te(i-j)²). Le paramètre a de cette variable est :

On a donc : C₁ = N(m₁,σ₁ ²) et C₂ = N(m₂,σ₂ ²), avec: m₁ ⁼ Nµ_s ² + Nσ_x ², m₂ = Ma_x ², σ₁ = σ_x ²[2Σ_{0≦i≦N-1,0≦j≦N-1}g_f0,B,Te(i-j)^2½ et σ₂ = σ_x ²[2Σ_{0≦i≦M-1,0≦j≦M-1}g_f0,B,Te(i-j)]²]^½. D'où m = m₁/m₂ = (N/M)(1+r), α₁ - m₁/σ₁ =N(1-r)/2Σ_{0≦i≦N-1,0≦j≦N-1}g_f0,B,Te(i-j)²]^½ et σ₂ = m₂/σ₂ = M/[2Σ_{0≦i≦M-1,0≦j≦M}- l9fp g,Te(I-J)2J1/2. On remarquera que : - si le bruit d'origine est blanc et gaussien, les hypothèses précédentes restent encore valables. Il suffit de remarquer qu'alors g_f0,B,Te(k) = o_o(k). Les formules précédentes s'en trouvent simplifiées :

avec :

D'où

Il est possible de tendre vers un tel modèle en sous-échantillonnant le bruit, et ne prenant du bruit qu'un échantillon sur k₀ échantillons où k₀ est tel que : V k > k_o, r_x(k) → 0. - la notion de compatibilité entre énergies ne se met en place que conditionnellement à la connaissance a priori du paramètre m, donc du rapport signal à bruit r. Celui-ci peut être fixé de manière heuristique à partir de mesures préliminaires des rapports signaux à bruit que présentent les signaux que l'on ne veut pas détecter par l'algorithme de confirmation de bruit, ou fixé de manière péremptoire. La seconde solution est utilisée de préférence. En effet, l'objet de ce traitement vise à mettre en évidence, non pas toutes les trames de bruit, mais seulement quelques unes présentant une forte probabilité de n'être constituées que de bruit. On a donc tout intérêt à ce que l'algorithme soit très sélectif. Cette sélectivité s'obtient en jouant sur la valeur de la probabilité de fausse alarme que l'on décide d'assurer et qui sera donc choisie très faible (la sélectivité maximale étant établie pour P_FA = 0, ce qui conduit à un seuil nul et à aucune détection de bruit, ce qui est le cas extrême et absurde). Mais cette sélectivité s'obtient aussi par le choix de r : choisi trop grand, on risque de considérer des énergies comme représentatives du bruit, alors que ce sont des énergies de respiration, par exemple, présentant un rapport signal à bruit inférieur à r. A contrario, choisir un r trop petit peut limiter la P_FA accessible, qui serait alors trop forte pour être acceptable.

So C ₁ ⁼ N (Nµ _s ² + Nσ _x ² , 2σ _x ⁴ Σ _{0 ≦ i ≦ N-1.0 ≦ j ≦ N-1} g _{f0, B, Te} (ij) ² ) and the parameter a of this variable is worth α ₁ = N (1 + r) / [2Σ _{0 ≦ i ≦ N-1.0 ≦ j ≦ N-1} g _{f0, B, Te} (i, j) ² ] ^½ , where r = µ _s ² / σ _x ² represents the signal to noise ratio. C ₂ is the class of energies corresponding to noise alone. According to hypothesis 2, if the noise samples are x (0), ..., x (M-1), it comes V = (1 / M) Σ _{0 ≦ n ≦ M-1} x (n ) ² ∈ N (Mσ _x ² , 2σ _x ⁴ Σ _{0 ≦ i ≦ M-1.0 ≦ j ≦ M-1} g _{f0, B, Te} (ij) ² ). The parameter a of this variable is:

We therefore have: C ₁ = N (m ₁ , σ ₁ ² ) and C ₂ = N (m ₂ , σ ₂ ² ), with: m ₁ ⁼ Nµ _s ² + Nσ _x ² , m ₂ = Ma _x ² , σ ₁ = σ _x ² [2Σ _{0 ≦ i ≦ N-1.0 ≦ j ≦ N-1} g _{f0, B, Te} (ij) ^2½ and σ ₂ = σ _x ² [2Σ _{0 ≦ i ≦ M-1, 0 ≦ j ≦ M-1} g _{f0, B, Te} (ij)] ² ] ^½ . Hence m = m ₁ / m ₂ = (N / M) (1 + r), α ₁ - m ₁ / σ ₁ = N (1-r) / 2Σ _{0 ≦ i ≦ N-1.0 ≦ j ≦ N-1} g _{f0, B, Te} (ij) ² ] ^½ and σ ₂ = m ₂ / σ ₂ = M / [2Σ _{0 ≦ i ≦ M-1.0 ≦ j ≦ M} - l9fp g, Te (IJ ) 2J1 / 2. Note that: - if the original noise is white and Gaussian, the previous assumptions are still valid. It suffices to note that then g _{f0, B, Te} (k) = o _o (k). The previous formulas are simplified:

with:

From where

It is possible to tend towards such a model by subsampling the noise, and only taking noise from one sample out of k ₀ samples where k ₀ is such that: V k> k _o , r _x (k) → 0. - the concept of compatibility between energies is set up only conditionally on a priori knowledge of the parameter m, therefore of the signal to noise ratio r. This can be fixed heuristically from preliminary measurements of the signal-to-noise ratios presented by the signals that one does not want to detect by the noise confirmation algorithm, or fixed peremptorily. The second solution is preferably used. Indeed, the object of this processing aims to highlight, not all the noise frames, but only a few having a high probability of being made up only of noise. We therefore have every interest in making the algorithm very selective. This selectivity is obtained by playing on the value of the probability of false alarm that we decide to ensure and which will therefore be chosen very low (the maximum selectivity being established for P _FA = 0, which leads to a zero threshold and no noise detection, which is the extreme and absurd case). But this selectivity is also obtained by the choice of r: chosen too large, there is a risk of considering energies as representative of the noise, whereas these are breathing energies, for example, having a signal to noise ratio less than r. Conversely, choosing an r that is too small can limit the accessible P _FA , which would then be too high to be acceptable.

Given the previous models, and the calculation of the threshold having been made, the following algorithm for detecting and confirming noise is then applied, essentially based on the concept of compatibility, as described above.

The search and confirmation of the noise frames is done among a number of frames Ni defined by the user once and for all for its application (for example Ni = 40), these frames being located before the vocal core. The following hypothesis is made: the energy of the noise frames alone is on average lower than that of the noise + respiration and signal noise frames. The frame having the minimum energy among the Ni frames is therefore assumed to consist only of noise. We then look for all the frames compatible with this frame in the sense recalled above, using the aforementioned models.

The noise detection algorithm will search, among a set of frames T ₁ , ..., T _n , for those which can be considered as noise.

Let E (T ₁ ), ..., E (T _n ), the energies of these frames, calculated in the form: E (T _i ) = Σ _{0 ≦ n ≦ N-1} u (n) 2 where u ( n) are the N samples constituting the frame T.

We make the following assumption: the frame with the lowest energy is a noise frame. Let To be this frame.

• L'ensemble des trames de bruit est initialisé: Bruit={To}

The algorithm works as follows:

• All the noise frames are initialized: Noise = {To}

For i describing {E (T ₁ ), ..., E (T _n )} - {E (T _i0 )} Do If E (T) is compatible with each Noise element:

Noise = Noise U {E (T _i )} End for Autoregressive Model of noise.

Since the noise confirmation algorithm provides a certain number of frames which can be considered as noise with a very high probability, we seek to construct, from the data of the temporal samples, an autoregressive model of the noise.

If x (n) designates the noise samples, we model x (n) in the form: x (n) = Σ _{1 ≦ i} p _P a _i x (ni) + b (n), where p is the order of the model, the a _i , the coefficients of the model to be determined and b (n) the modeling noise, assumed to be white and Gaussian if we follow a maximum likelihood approach.

This type of modeling is widely described in the literature, notably in "Spectrum Analysis A Modern Perspective", by S.M. KAY and S.L. MARPLE Jr, published in Proceedings of the IEEE, Vol. 69, No. 11, November 1981.

As for the algorithms for calculating the model, many methods are available (Burg, Levinson-Durbin, Kalman, Fast Kalman ...).

The methods of the Kalman and Fast Kalman type are preferably used, see articles "Transverse Adaptive Filtering" by O. MACCHI / M.BELLANGER published in the journal Signal Processing, Vol. 5, N ° 3, 1988 and "Signal analysis and adaptive digital filtering" by M.BELLANGER published in the CNET-ENST Collection, MASSON, which present very good real-time performance. But this choice is not the only possible one. The order of the filter is for example chosen equal to 12, without this value being limiting.

Rejector filtering

Let u (n) = s (n) + x (n) be the total signal, composed of the speech signal s (n) and the noise x (n).

Let the filter H (z) = 1 - Σ _{1 ≦ i ≦ p} a _i z ^-i .

Applied to the signal U (z), we obtain H (z) U (z) = H (z) S (z) + H (z) X (z).

Or: H (z) X (z) = B (z) => H (z) U (z) = H (z) S (z) + B (z)

The rejector filter H (z) whitens the signal, so that the signal at the output of this filter is a speech signal (filtered therefore distorted), added to a generally white and Gaussian noise.

The signal obtained is in fact unsuitable for recognition, because the rejector filter distorts the original speech signal.

However, the signal obtained being disturbed by a practically white and Gaussian noise, it follows that this signal is very interesting for carrying out a detection of the signal s (n) according to the theory set out below, according to which the signal is kept wide band obtained, or it is filtered beforehand in the band of fricatives, as described below (see "detection of fricatives").

It is for this reason that we use this rejection filtering after auto-regressive modeling of the noise. Average noise spectrum.

As we have a certain number of frames confirmed as being noise frames, we can then calculate an average spectrum of this noise, so as to implement a spectral filtering, of the spectral subtraction or WIENER filtering type.

We choose for example WIENER filtering. Also, we need to calculate C _xx (f) = E [| X (f) | 2] which represents the average noise spectrum. As the calculations are digital, we only have access to FFTs of digital signals weighted by a weighting window. In addition, the spatial average can only be approximated.

Let X ₁ (n) ..., X _M (n) be the M + 1 FFT of the M noise frames confirmed as such, these FFT being obtained by weighting the initial time signal by an adequate apodization window.

C _xx (f) = E [| X (f) 2] is approximated by:

The performance of this estimator is given for example in the book "Digital signal processing" by L.RABINER / C.M.RADER published by IEEE Press.

With regard to the Wiener filter, a few classic results are recalled below, explained in particular in the work "Speech Enhancement" by J.S. LIM published by Prentice-Hall Signal Processing Series.

Let u (t) = s (t) + x (t) be the total signal observed, where s (t) denotes the useful signal (speech) and x (t) the noise. In the frequency domain, we obtain: U (f) = S (f) + X (f), with obvious notations.

We then look for the filter H (f), so that the signal ^ S (f) = H (f) U (f) is the closest to S (f) within the meaning of the standard L ₂ . We therefore seek H (f) minimizing: E [| S (f) - ^ S (f) | ² ].

We then demonstrate that: H (f) = 1 - (C _XX (f) / C _UU (f)) where C _xx (f) = E [| X (f) 2] and C _uu (f) = E [ | U (f) ² |].

This type of filter, because its expression is directly frequency, is particularly interesting to apply as soon as the parameterization is based on the calculation of the spectrum. Implementation by smoothed correlogram.

In practice, C _xx and C _uu are not accessible. We can only estimate them. A procedure for estimating C _xx (f) has been described above.

C _uu is the average spectrum of the total signal u (n) available only on a single frame. In addition, this frame must be configured so that it can intervene in the recognition process. There is therefore no question of carrying out any average of the signal u (n) all the more since the speech signal is a particularly non-stationary signal.

It is therefore necessary to construct, from the data of u (n), an estimate of C _uu (n). We then use the smoothed correlogram.

We then estimate C _uu (n) by: ^ C _UU (k) = Σ _{0 ≦ n ≦ N - 1} F (kn) | X (n) | 2

where F is a smoothing window constructed as follows, and N the number of points allowing the calculation of FFT: N = 256 points for example.

• f(n) = a_o + a₁cos(2πn/N) + a₂cos(4πn/N). Ces fenêtres sont largement décrites dans l'article précité: "On the Use of Windows for Hamming Analysis with the Discrete Fourier Transform de F.J.HARRIS paru dans Proceedings of the IEEE, Vol.66, N° 1, January 1978.

We choose a smoothing window in the time domain:

• f (n) = a _o + a ₁ cos (2πn / N) + a ₂ cos (4πn / N). These windows are widely described in the aforementioned article: "On the Use of Windows for Hamming Analysis with the Discrete Fourier Transform by FJHARRIS published in Proceedings of the IEEE, Vol.66, N ° 1, January 1978.

The function F (k) is then simply the Discrete Fourier Transform of f (n).

^ C _UU (k) = Σ _{0 ≦ n ≦ N} - ₁ F (kn) | X (n) | ² appears as a discrete convolution between F (k) and V (k) = | X (k) ² |, so that ^ C _UU = F _* V

Let ^ c _{UU be} the FFT- ¹ of ^ C _UU . ^ c _UU (k) = f (k) v (k) where v (k) is the FFT- ¹ of V (k).

• (1) Calcul de v(k) par FFT inverse de V(n) = |X(n) 2
• (2) Calcul du produit f.v
• (3) FFT directe du produit f.v qui aboutit à ^C_UU

We therefore calculate ^ C _UU (k) according to the so-called smoothed correlogram algorithm:

• (1) Calculation of v (k) by inverse FFT of V (n) = | X (n) 2
• (2) Calculation of the product fv
• (3) Direct FFT of the product fv which leads to ^ C _UU

Rather than applying the same estimator for the noise and the total signal, the method of the invention applies the algorithm of the previous smoothed correlogram to the average noise spectrum M _xx (n).

^ C _XX (k) is therefore obtained by:

Le filtre de Wiener H(f) est donc estimé par la suite des valeurs : ^H(n) = 1 - (^C_XX(n)/^C_UU(n))

The Wiener filter H (f) is therefore estimated by the following values: ^ H (n) = 1 - (^ C _XX (n) / ^ C _UU (n))

The denoised signal has the spectrum: ^ S (n) = A _H ( _n ) _U ( _n )

An FFT- ¹ can possibly make it possible to recover the denoised time signal.

The denoised spectrum ^ S (n) obtained is the spectrum used for parameterization with a view to recognizing the frame.

To detect unvoiced signals, the procedures described above are also used, since there are energies representative of the noise (see above the noise detection algorithm). Activity detection. Let Ci = N (m ₁ , σ ₁ ² ) and C ₂ = N (m ₂ , σ ₂ ² ).

Since we have an algorithm capable of highlighting values of random variables belonging to the same class, of class C ₂ (for example), and this, with a very low probability of error, it then becomes much easier to decide, by observation of the couple U / V, if U belongs to the class Ci or to the class C ₂ . So there are two distinct possible hypotheses,

correspondant à deux décisions possibles distinctes :

Décision optimale.

corresponding to two distinct possible decisions:

Optimal decision.

We set: m = m ₁ / m ₂ , α ₁ ⁼ m ₁ / σ ₁ and a ₂ ⁼ m ₂ / a ₂ .

Let be a couple (U, V) of random variables, where we suppose that V ∈ C ₂ and U ∈ C ₁ UC ₂ . U and V are assumed to be independent. By observing the variable X = U / V, we seek to make a decision between the following two possible: "C ₁ XC ₂ ", C ₂ XC2 ".

We therefore have two hypotheses:

Let p = Pr {U∈C ₁ }.

The decision rule is expressed in the following form:

où

Le seuil optimal est celui pour lequel P_c(s,m |α₁,α₂) est maximal. On résout donc l'équation :The probability of correct decision P _c (s, m | α ₁ , α ₂ ) is then:

or

The optimal threshold is that for which P _c (s, m | α ₁ , α ₂ ) is maximum. We therefore solve the equation:

∂P _c (s, m | α ₁ , α ₂ ) / ∂s = 0 ⇔ pf (s, m | α ₁ , α ₂ ) - (1-p) f (s, 1 α ₂ , α ₂ ) = 0

Neyman-Pearson approach

In the previous approach, we assumed to know the probability p. When this probability is unknown, a Neyman-Pearson type approach can be used.

We define the probabilities of non-detection and false alarm:

We have:

We then set P _fa or P _nd , to determine the value of the threshold.

In order to apply activity detection as described above to the case of speech, it is necessary to establish an energy model of the unvoiced signals compatible with the assumptions which govern the proper functioning of the methods described above. We are therefore looking for a model of the energies of unvoiced fricatives / F /, / S /, / CH) and of unvoiced plosives / P /, / T /, / Q /, which make it possible to obtain energies whose statistical law is approximately a Gaussian.

Model 1.

The sounds / F /, / S /, / CH / are located spectrally in a frequency band which extends from approximately 4 KHz to more than 5KHz. The sounds / P /, / T /, / Q / as short phenomena in time, are spread over a wider band. In the chosen band, it is assumed that the spectrum of these fricative sounds is relatively flat, so that the fricative signal in this band can be modeled by a narrow band signal. This can be realistic in certain practical cases without resorting to the bleaching described above. In most cases, however, it is a good idea to work on a whitened signal to ensure a suitable narrowband noise model.

By accepting such a narrow band noise model, we therefore have to process the ratio of two energies which can be processed by the methods described above.

Let s (n) be the speech signal in the band studied and x (n) the noise in this same band. The signals s (n) and x (n) are assumed to be independent.

The class Ci corresponds to the energy of the total signal u (n) = s (n) + x (n) observed on N points, the class C ₂ corresponds to the energy V of the noise alone observed on M points.

De même :The signals being Gaussian and independent, u (n) is a signal itself Gaussian, so that:

Similarly:

V = Σ _{0 ≦ n ≦ M-1} y (n) ² ∈ N (Mσ _x ² , 2σ _x ² Σ _{0 ≦ i ≦ M-1.0 ≦ j ≦ M-1} g _{f0, B} (ij) ² ) , where y (n) denotes, we recall, another value noise x (n) over a time slice different from that where u (n) is observed. We can therefore apply the above theoretical results with:

Note that m = (N / M) (1 + r) where r = σ _x ² / σ _x ² ultimately denotes the signal to noise ratio.

To completely solve this problem, it is necessary to be able to know the signal-to-noise ratio r as well as the probability of the presence p of the useful signal. What appears to be a limitation here is common to the other two models discussed below.

Model 2.

As in the case of model 1, we seek to detect only unvoiced fricatives, therefore to detect a signal in a particular band.

Here, the model of the fricative signal is not the same as previously. We assume that the fricatives have the minimum energy µ _s ² = E _{0 ≦ n ≦ N-1} s (n) 2 known, thanks for example to learning, or estimated.

The voiced sound is independent of the noise x (n) which is here Gaussian narrow band.

• V = Σ_0≦n≦N-1 y(n)² ∈ N( Mσ_x ²,2Tr(C_x,M ²)) où C_x,_M désigne la matrice de corrélation du M-uplet : ^t (y(0),...,y(M-1))

If y (n), for n between 0 and M-1, denotes another noise value x (n) on a time slot distinct from that where the total signal is observed u (n) = s (n) + x (n), we will have:

• V = Σ _{0 ≦ n ≦ N-1} y (n) ² ∈ N (Mσ _x ² , 2Tr (C _{x, M} ² )) where C _x , _M denotes the correlation matrix of the M-tuplet: ^t (y ( 0), ..., y (M-1))

Regarding the energy U = E o _{≦ n ≦ N-1} u (n) 2 of the total signal, this can be expressed according to:

This result is obtained by supposing that the independence between s (n) and x (n) is expressed by decorrelation in the temporal sense of the term, that is to say that we can write:

As V '= Σ _{0 ≦ n ≦ N-1} x (n) ² ∈ N (Nσ _x ² , 2Tr (C _x , N ² )) where C _{x, N} denotes the correlation matrix of the N-tuplet: ^t ( x (0), ..., x (N-1)), we then have:

We can therefore apply the above theoretical results with:

Note that m = (N / M) (1 + r) where r = µ _s ² / σ _x ² ultimately designates the signal to noise ratio. The same remark as that of Model 1, concerning the signal to noise ratio r and the probability p of presence of the useful signal, is valid here.

Model 3.

In this model, we seek to perform a detection of all the unvoiced signals, with a Gaussian white noise hypothesis.

The narrowband signal model used previously is therefore no longer valid. We can therefore only assume that we are dealing with a broadband signal for which we know the minimum energy µ _S ^2.

So it comes:

To use this model, the noise must be Gaussian white. If the original noise is not white, we can approach this model by actually sub-sampling the observed signal, that is to say by considering only one sample out of 2, 3, or even more, depending on the noise autocorrelation function, and assuming that the speech signal thus subsampled still has detectable energy. But it is also possible, and this is preferable, to use this algorithm on a signal whitened by a rejector filter, since then the residual noise is approximately white and Gaussian.

The preceding remarks concerning the a priori value of the signal to noise ratio and the probability of the presence of the useful signal, still remain valid.

Unvoiced sound detection algorithms.

Using the previous models, two algorithms for detecting unvoiced sounds are exposed below.

• Disposant d'énergies représentatives de bruit, on peut moyenner ces énergies de sorte que l'on obtient une énergie de "référence" de bruit. Soir E_o cette énergie. Pour N₃ trames T₁, ..., T_n qui précèdent la première trame voisée, on procède comme suit :
• Soient E(T₁), ..., E(T_n), les énergies de ces trames, calculées sous la forme E(T_i) = Σ_0≦n≦N-1u(n)² où u(n) sont les N échantillons constituant la trame T.

Algorithm 1:

• Having representative noise energies, these energies can be averaged so that a noise "reference" energy is obtained. Evening E _o this energy. For N ₃ frames T ₁ , ..., T _n which precede the first voiced frame, the procedure is as follows:
• Let E (T ₁ ), ..., E (T _n ), the energies of these frames, calculated in the form E (T _i ) = Σ _{0 ≦ n ≦ N-1} u (n) ² where u (n ) are the N samples constituting the frame T.

For E (T _i ) describing {E (T ₁ ), ..., E (T _n )}

To do

If E (T _i ) is compatible with E _o (Decision on the value of E (T _i ) / E ₀ ).

Detection on the T frame.

End for.

• Cet algorithme est une variante du précédent. On utilise pour E_o, soit l'énergie moyenne des trames détectées comme du bruit, soit la valeur de l'énergie la plus faible de toutes les trames détectées comme du bruit.

Algorithm 2:

• This algorithm is a variant of the previous one. E _o uses either the average energy of the frames detected as noise, or the value of the lowest energy of all the frames detected as noise.

• Pour E(T_i) décrivant {E(T₁), ..., E(T_n)}.

Then we proceed as follows:

• For E (T _i ) describing {E (T ₁ ), ..., E (T _n )}.

To do

If E (T _i ) is compatible with E _o (Decision on the value of E (T _i ) / E ₀ ).

Detection on the T frame.

Otherwise E _o = E (T).

End for

The signal-to-noise ratio r can be estimated or fixed in a heuristic manner, provided that some preliminary experimental measurements, characteristic of the field of application, are carried out, so as to fix an order of magnitude of the signal-to-noise ratio presented by the fricatives in the chosen band.

The probability p of the presence of unvoiced speech is also a heuristic datum which modulates the selectivity of the algorithm, in the same way as the signal-to-noise ratio. This data can be estimated according to the vocabulary used and the number of frames on which the search for unvoiced sounds is made.

Simplification in the case of a slightly noisy environment.

In the case of a weakly noisy environment, for which no noise model has been determined, by virtue of the simplifications proposed above, the theory recalled above justifies the use of a threshold, which is not linked bijectively to the signal to noise ratio, but which will be fixed in a completely empirical way.

An interesting alternative for environments where the noise is negligible, is to be satisfied with the detection of voicing, to eliminate the detection of unvoiced sounds, and to fix the start of speech at a few frames before the vocal nucleus (about 15 frames ) and the end of speech a few frames after the end of the vocal nucleus (about 15 frames).

Claims (12)

1. A method of detecting speech in noisy signals, characterized in that after having carried out in these signals the detection of at least one voiced frame, we search for noise frames preceding this voiced frame, we build a model autoregressive noise and an average noise spectrum, we blank by rejector filter and we denoised by spectral denoiser the frames preceding voicing, we search for the effective start of speech in these bleached frames, we extract noisy frames between the effective start of speech and the first voiced frame the acoustic vectors used by the voice recognition system, as long as voiced frames are detected, these are denoised and then parameterized with a view to their recognition, when voiced frames are no longer detected, we are looking for the effective end of the speech, we denois then we configure the frames between the last voiced frame and the effective end of the speech. 2. Method according to claim 1, characterized in that the bleaching is carried out by filtering rejector calculated from the autoregressive model of noise. 3. Method according to claim 2, characterized in that when the last speech frame has been parameterized, all the processing parameters are reset. 4. Method according to one of the preceding claims, characterized in that the signal frames to be processed are processed by Fourier transforms, and that when two transforms are consecutive in time, they are calculated on three consecutive frames with overlap d 'a frame. 5. Method according to one of the preceding claims, characterized in that the voicing detection is done, for each frame, using the value of the "pitch" associated with this frame. 6. Method according to claim 5, characterized in that the calculation of the pitch is validated after having recognized at the moint three voiced frames, ie 3 x 12.8 ms. 7. Method according to claim 5 or 6, characterized in that the calculation of the pitch is made from the correlation of the signal with its delayed form. 8. Method according to one of claims 5 to 7, characterized in that the detection of unvoiced sounds is done by thresholding. 9. Method according to one of the preceding claims, characterized in that to detect unvoiced speech, the distance between the vowel core and the fricative block is examined, and the size of this fricative block. 10. Method according to one of the preceding claims, characterized in that the average noise spectrum is obtained by Wiener filtering. 11. Method according to claim 10, characterized in that the smoothed correlogram algorithm is applied to the average noise spectrum. 12. Method according to claim 1, characterized in that in a low-noise environment, only a detection of voiced frames is carried out, and a detection of a vocal nucleus to which a confidence interval is attached.