EP0821345B1 - Method to determine the fundamental frequency of a speech signal - Google Patents

Description

The invention relates to a method for extracting fundamental frequency of a speech signal.

Current signal processing techniques digital speech have the essential object of extracting the basic parameters, in order to improve the quality, by improving the signal-to-noise ratio, and where appropriate, to determine the origin of the speaker, with a view to example of authentication of the latter.

Among the aforementioned fundamental parameters, the fundamental frequency is one of the parameters that characterize the better the voice of a given speaker and that allows therefore to contribute to the certain authentication of it.

Many processes for extracting the fundamental frequency from a speech signal have been proposed. For a general overview of the proposed techniques, one can usefully refer to the work published by W. HESS entitled "Pitch determination of speech signais: algorithms and methods", Springer-Verlag, New-York 1983.

• Les méthodes -temporelles telles que celles mettant en oeuvre un processus d'autocorrélation avec écrêtement central et comparaison des pics à une valeur de seuil ou celles désignées par AMDF, ces dernières ayant été décrites par R.BOITE et M.KUNT dans l'ouvrage intitulé "Traitement de la parole", pages 193-195, Presses polytechniques romandes, Lausanne 1987, sont relativement peu coûteuses en temps de calcul car elles ne nécessitent pas la mise en oeuvre d'opérations arithmétiques de multiplication. Toutefois, elles manquent de précision et il est nécessaire, en conséquence, de procéder à un suréchantillonnage du signal de parole, afin d'obtenir une précision convenable, ce qui, bien entendu, entraíne une augmentation notable du temps de calcul effectif.
  Parmi ces méthodes, le document KEIKICHI HIROSE ET AL : "A SCHEME FOR PITCH EXTRACTION OF SPEECH USING AUTOCORRELATION FUNCTION WITH FRAME LENGTH PROPORTIONAL TO THE TIME LAG", ICASSP-92, SPEECH PROCESSING 1, SAN FRANCISCO, MAR 23-26, 1992, Vol.1, IEEE pages 149-152, XP 000341105, décrit un processus d'extraction de la fréquence fondamentale d'un signal de parole par autocorrélation.
• Les méthodes fréquentielles sont, au contraire, basées sur l'analyse de la structure harmonique du spectre d'énergie en fonction de la fréquence du signal de parole. Parmi celles-ci, la méthode dite du peigne, décrite par P.MARTIN dans l'article intitulé "Extraction de la fréquence fondamentale par intercorrélation avec une fonction peigne", publiée aux Journées d'Etude Parole 12, pp. 221-232, 1981, consiste à calculer la fonction d'intercorrélation entre le spectre du signal numérique de parole et une fonction en peigne, pour différentes valeurs de la distance entre les dents du peigne. Le maximum de la fonction d'intercorrélation est obtenu pour une distance entre deux dents consécutives du peigne, égale à la fréquence fondamentale du signal à analyser. Cette méthode présente une bonne fiabilité mais elle est relativement complexe, dans la mesure où elle nécessite un prélèvement fréquentiel consistant à ne retenir que les maxima du spectre et les valeurs adjacentes. En outre, il est nécessaire d'effectuer une interpolation afin d'augmenter la précision du résultat.

The aforementioned techniques or methods can be classified into two main families.

• -Temporal methods such as those implementing an autocorrelation process with central clipping and comparison of peaks at a threshold value or those designated by AMDF, the latter having been described by R.BOITE and M.KUNT in the work entitled "Speech processing", pages 193-195, Presses polytechniques romandes, Lausanne 1987, are relatively inexpensive in computation time because they do not require the implementation of arithmetic multiplication operations. However, they lack precision and it is necessary, therefore, to proceed to an oversampling of the speech signal, in order to obtain a suitable precision, which, of course, results in a significant increase in the effective calculation time.
  Among these methods, the document KEIKICHI HIROSE ET AL: "A SCHEME FOR PITCH EXTRACTION OF SPEECH USING AUTOCORRELATION FUNCTION WITH FRAME LENGTH PROPORTIONAL TO THE TIME LAG", ICASSP-92, SPEECH PROCESSING 1, SAN FRANCISCO, MAR 23-26, 1992, Vol.1, IEEE pages 149-152, XP 000341105, describes a process of extracting the fundamental frequency from a speech signal by autocorrelation.
• Frequency methods are, on the contrary, based on the analysis of the harmonic structure of the energy spectrum as a function of the frequency of the speech signal. Among these, the so-called comb method, described by P.MARTIN in the article entitled "Extraction of the fundamental frequency by intercorrelation with a comb function", published in the Journées d'Etude Parole 12, pp. 221-232, 1981, consists in calculating the intercorrelation function between the spectrum of the digital speech signal and a comb function, for different values of the distance between the teeth of the comb. The maximum of the intercorrelation function is obtained for a distance between two consecutive teeth of the comb, equal to the fundamental frequency of the signal to be analyzed. This method has good reliability but it is relatively complex, insofar as it requires a frequency sampling consisting in retaining only the spectrum maxima and the adjacent values. In addition, it is necessary to perform an interpolation in order to increase the accuracy of the result.

où

L = M/k, M désignant le nombre de points du spectre

X(l) désigne le logarithme du spectre d'énergie.

L'inconvénient de cette méthode réside dans le fait que l'amplitude des pics harmoniques décroít en fonction de la fréquence, avec une pente de l'ordre de -12 dB/octave. Bien qu'un processus de pré-accentuation permette de relever le niveau des harmoniques de fréquence élevée, certains pics harmoniques présentent un niveau d'énergie plus faible que d'autres en raison de la contribution des formants, ce qui provoque des erreurs fréquentes dans l'estimation de la valeur de la fréquence fondamentale.Another method, designated by spectral compression method, has been published by NOLL (1970), see the work by W. HESS previously cited on pages 414-417. This method, based on an analysis of the harmonic structure of the energy spectrum as a function of the frequency of the speech signal, consists in compressing the energy spectrum of the speech signal along the frequency axis, by factors successive integers, then adding the compressed spectra obtained to the initial spectrum. These operations allow, in principle, to obtain a significant maximum, which results from the coherent contribution of the harmonics of the fundamental frequency after compression. The extraction of the fundamental frequency then consists in seeking the maximum of the logarithm of the harmonic product defined by:

or

L = M / k, M denoting the number of points of the spectrum

X (l) denotes the logarithm of the energy spectrum.

The disadvantage of this method lies in the fact that the amplitude of the harmonic peaks decreases as a function of the frequency, with a slope of the order of -12 dB / octave. Although a pre-emphasis process raises the level of high frequency harmonics, some harmonic peaks have a lower energy level than others due to the contribution of formants, which causes frequent errors in estimating the value of the fundamental frequency.

The subject of the present invention is the implementation of a process for extracting the fundamental frequency of a speech signal in which frequency extraction fundamental is obtained with increased reliability.

Another object of the present invention is the implementation using a method for extracting the fundamental frequency of a speech signal in which the process actual extraction of the fundamental frequency may be conditional upon detection of voicing or the absence of voicing of the sounds constituting the signal of speech.

Another object of the present invention is finally the implementation of a frequency extraction process fundamental of a speech signal in which the value of extracted fundamental frequency is further subjected to a post-processing process, of the learning type, in order to eliminate any improbable or outliers.

avec m = [1,2,...,p], soumettre le signal de parole préaccentué à un filtrage de type passe-bas et à un sous-échantillonnage, pour engendrer un signal de parole filtré sous-échantillonné, calculer, par compression spectrale, à partir du signal de parole filtré sous-échantillonné et à partir des coefficients cepstraux pour chaque trame courante d'une succession de trames de même recouvrement de durée, la fréquence fondamentale, maximum de rang k, d'une fonction P(k) représentative de la différence entre un deuxième ensemble des valeurs X₂(k) du logarithme du spectre d'énergie et l'ensemble des valeurs H(k) du spectre de fréquences lissé, ladite fonction vérifiant la relation :

avec L = M₂/k, k variant entre une première et une deuxième valeur représentatives d'une bande de fréquences basses comprises entre 70 et 450 Hz, ladite fonction P(k) présentant un maximum pour k=F0, valeur extraite de la fréquence fondamentale du signal de parole.The method of extracting the fundamental frequency from a speech signal, a succession of digital samples, which is the subject of the present invention, is remarkable in that it comprises at least the steps consisting in subjecting this speech signal to a process. pre-emphasis, to generate a pre-emphasized speech signal, calculating, from the pre-emphasized speech signal, for each current frame of a succession of frames each corresponding in duration to a determined number N of samples, two consecutive frames each having a duration recovery in number of consecutive consecutive samples at most equal to 50/100 of the number N of samples, a first set of values X ₁ (k) of the logarithm of the energy spectrum by Fourier transform on a number M ₁ of points, calculate, from this first set of values, a determined number p of first cepstral coefficients C (m), by applying a transform in disc cosine repeat at said values X ₁ (k) over a number of these values at least equal to half the number N of samples constituting the current frame, this transform verifying the relation:

with m = [1,2, ..., p], subject the pre-emphasized speech signal to low-pass type filtering and to sub-sampling, to generate a filtered sub-sampled speech signal, calculate, by spectral compression, from the filtered sub-sampled speech signal and from cepstral coefficients for each current frame of a succession of frames of the same duration overlap, the fundamental frequency, maximum of rank k, of a function P ( k) representative of the difference between a second set of values X ₂ (k) of the logarithm of the energy spectrum and the set of values H (k) of the smoothed frequency spectrum, said function verifying the relationship:

with L = M ₂ / k, k varying between a first and a second value representative of a low frequency band between 70 and 450 Hz, said function P (k) having a maximum for k = F0, value extracted from the fundamental frequency of the speech signal.

The process which is the subject of the present invention finds in particular application to speech recognition and identification of speakers from signatures sound.

• la figure la représente un organigramme illustratif de l'ensemble des étapes permettant la mise en oeuvre du procédé objet de la présente invention ;
• la figure 1b représente un organigramme illustratif d'une variante de mise en oeuvre avantageuse du procédé objet de la présente invention, dans laquelle certaines étapes sont conduites en parallèle ou, le cas échéant, sous système d'exploitation multitâche afin de permettre un mode opératoire en temps réel, sans toutefois nécessiter une puissance de calcul très importante ;
• la figure 2a représente un détail de réalisation d'une succession d'étapes élémentaires permettant une mise en oeuvre optimale de l'étape terminale de calcul par compression spectrale de la fréquence fondamentale du signal de parole du procédé, objet de la présente invention, illustré conformément à la figure 1a ou 1b ;
• la figure 2b représente une série de signaux obtenus dans le domaine fréquentiel suite à la mise en oeuvre des étapes élémentaires illustrées en figure 2a ;
• les figures 3a, 3b, 3c et 3d représentent un mode opératoire de formatage de trames d'échantillons, constitutifs du signal de parole, un processus de discrimination des trames courantes en fonction d'un critère relatif au caractère voisé ou non voisé de chaque trame courante, un mode d'établissement de ce critère et un abaque d'attribution d'un indice de voisement de segments temporels constitutifs de chaque trame respectivement ;
• la figure 4 représente un schéma synoptique de l'architecture d'un dispositif permettant la mise en oeuvre du procédé, objet de la présente invention, à partir d'un micro-ordinateur hôte et d'un processeur de signal numérique spécialisé ou dédié connectés par une liaison de type BUS.

It will be better understood on reading the description and on observing the drawings below in which:

• FIG. 1a represents an illustrative flowchart of all the steps allowing the implementation of the method which is the subject of the present invention;
• FIG. 1b represents an illustrative flowchart of an advantageous variant implementation of the method which is the subject of the present invention, in which certain steps are carried out in parallel or, where appropriate, under a multitasking operating system in order to allow an operating mode in real time, without however requiring very large computing power;
• FIG. 2a represents a detail of the realization of a succession of elementary steps allowing an optimal implementation of the terminal calculation step by spectral compression of the fundamental frequency of the speech signal of the process, object of the present invention, illustrated according to Figure 1a or 1b;
• FIG. 2b represents a series of signals obtained in the frequency domain following the implementation of the elementary steps illustrated in FIG. 2a;
• FIGS. 3a, 3b, 3c and 3d represent an operating mode for formatting sample frames constituting the speech signal, a process for discriminating current frames according to a criterion relating to the voiced or unvoiced character of each frame current, a mode of establishment of this criterion and an abacus of allocation of a voicing index of time segments constituting each frame respectively;
• FIG. 4 represents a synoptic diagram of the architecture of a device allowing the implementation of the method, object of the present invention, from a host microcomputer and a specialized or dedicated digital signal processor connected by a BUS type connection.

A more detailed description of the extraction process the fundamental frequency of a speech signal, object of the present invention, will now be given in connection with Figures 1a and 1b.

As will be seen in Figure la, the signal of speech on which we want to extract the fundamental frequency, in accordance with the process object of the present invention is for example an analog signal representative of distinct words and syllables, this signal analog being transformed into a succession of samples digital, the speech signal, in its digital form, being designated by sp in Figure 1a.

As it also appears in the above figure, the speech signal sp is then subjected to a pre-emphasis process making it possible to generate a pre-accented speech signal, noted spp. The pre-emphasis process is a conventional type process, which, as such, will not be described in detail. This process consists of an overall pre-emphasis, which in fact consists in applying an increasing gain value with frequency to compensate for the attenuation of the harmonics of higher rank. By way of nonlimiting example, it is indicated that the global pre-emphasis process can consist in applying to the speech signal sp a transfer function of the type: G (z) = 1 - z -1 .

In the above-mentioned relation, it is indicated that z = e ^jω where ω = 2πf, f denoting the instantaneous frequency of the speech signal.

The process which is the subject of the present invention, as well as represented in FIG. 1a, then consists, in a step b), formatting the pre-emphasized speech signal spp. This formatting operation actually consists of constitute the pre-accentuated speech signal spp in frames successive each with N samples and corresponding at a duration of these N samples, two consecutive frames each having a number of duration overlaps consecutive consecutive samples at most equal to 50/100 of number N of constituent samples of each frame.

The aforementioned step b) also consists in calculating, on each current frame designated by T _q , a first set of values, denoted X ₁ (k) of the logarithm of the energy spectrum for the frame considered by application of a transform of Fourier on a number M ₁ of points.

In a practical way, it is indicated that the number M ₁ of points on which the Fourier transform is applied is chosen so that the Shannon theorem is satisfied. By way of nonlimiting example, it is indicated that for frames constituted by 256 successive samples and for a duration of each current frame equal to 32 ms, the number M ₁ of points can be taken equal to 128.

The aforementioned step b), represented in FIG. 1a, then makes it possible to have the first set of values, noted {X ₁ (k)}.

As shown in the above-mentioned figure, the method which is the subject of the present invention then consists in carrying out in a step c) the calculation, from the first set of values {X ₁ (k)}, a determined number p of first coefficients cepstrals denoted C (m) of the logarithm of the energy spectrum defined by the first set of values {X ₁ (k)}.

Dans cette relation, on indique que m est un entier prenant les valeurs = [1,2,...,p], p désignant le nombre de premiers coefficients cepstraux calculé et retenu pour la mise en oeuvre du procédé objet de la présente invention. A titre d'exemple non limitatif, on indique que p peut être limité à 16.The above cepstral coefficients verify the relation:

In this relation, it is indicated that m is an integer taking the values = [1,2, ..., p], p designating the number of first cepstral coefficients calculated and retained for the implementation of the method object of the present invention . By way of nonlimiting example, it is indicated that p can be limited to 16.

At the end of step c), we thus have the above-mentioned cepstral coefficients, which will allow the implementation of the following stages of the process which is the subject of the invention, as shown in Figure 1a.

Following step c) above, the process which is the subject of the The present invention consists, in a step d), in submitting the speech signal pre-emphasized spp to type filtering low pass and downsampling to generate a sub-sampled filtered speech signal, noted spf.

In Figure 1a, there is shown a connection in mixed line between step c) and step d), this link in mixed line indicating an operation performed on the signal spp pre-accentuated speech available after step a) of global pre-emphasis. We understand in particular that the speech signal in digital form sp, consisting in fact of a burst of successive words by example, the pre-emphasized speech signal spp can be memorized after the pre-emphasis stage performed in step a), and that, of course, step d) can be realized from the pre-emphasized speech signal spp previously cited.

In general, it is indicated that the filtering low pass type can be achieved with a low pass filter with a cut-off frequency equal to 2 kHz by means of a finite impulse response filter, called RIF filter, at 47 coefficients. The filtered signal from the above filtering can then be subjected to subsampling, the subsampling can be achieved by decimation, for delivering the noted down sampled filtered speech signal spf.

The above-mentioned step d) is then followed, as shown in FIG. 1a, by a step e) consisting in calculating by spectral compression the maximum fundamental frequency of rank k of a function P (k) representative of the difference between a second set of values X ₂ (k) of the logarithm of the energy spectrum of the subsampled filtered speech signal spf, and of the set of values H (k) of the smoothed frequency spectrum obtained from the cepstral coefficients available at the end of step c) previously mentioned in the description.
The function P (k) checks the relation:

In general, step e) shown in Figure la also comprises a frame in formatting step of N ₂ samples, with N ₂ = N / 2, two consecutive frames being in overlying N _2/2 samples of the filtered sub-sampled speech signal spf, the formatting being of course similar to the formatting applied at the start of step b) to the pre-emphasized speech signal spp.

The formatting step carried out in step e) is then followed by an effective step of calculating the second set of values {X ₂ (k)} of the logarithm of the energy spectrum, this calculation being carried out by application of a Fourier transform on a number M ₂ of points for each current frame obtained at the end of the formatting carried out. The second set of values {X ₂ (k)} is advantageously calculated via a fast Fourier transform FFT applied to M ₂ = 2048 points using the method of filling with zeros.

The step of calculating the second set of values {X ₂ (k)} is then followed by a step of calculating the smoothed frequency spectrum H (k) from the cepstral coefficients C (m) available from the end of l step c), the connection between step c) and step e) in FIG. 1a being shown in phantom for this reason. The smoothed spectrum H (k) is calculated by applying a cosine transform to the p cepstral coefficients available.

The step of calculating the smoothed frequency spectrum is then followed by a step of calculating the function P (k) verifying the relation previously cited in the description. In this relation, we indicate that L is equal to M ₂ / k for k varying between a first and a second value representative of a low frequency band between 70 and 450 Hz. The function P (k) then has a maximum for p = F ₀ , value extracted from the fundamental frequency of the speech signal.

The fundamental frequency extraction process of a speech signal, object of the present invention, allows, by spectral compression, by the calculation of the product harmonic of the difference between the energy spectrum of the speech signal and the spectrum of the signal smoothed, to eliminate the contribution of the formants and extracting the structure harmonic of the fundamental frequency of the speech signal.

In the embodiment of FIG. 1a, we have shown, by way of nonlimiting example, an embodiment sequential type, steps a) to e) can be executed successively. We understand in particular that, from a on the other hand, the pre-emphasized speech signal spp, and that, on the other share, cepstral coefficients, especially p cepstral coefficients used, can be stored at the outcome of step c) respectively after step a) to allow the sequential implementation of the steps b) to e) previously mentioned.

However, and in order not to overload unnecessarily the calculation processor used for the implementation of steps a) to e) above, but in order to facilitate the execution of the above steps in real time, the process object of the present invention can be implemented in an alternative embodiment as shown in FIG. 1b, in parallel, steps b), c) being carried out sequentially, in parallel with steps d) and e) from of the pre-emphasized speech signal spp. This embodiment as shown in Figure 1b, is made possible by due to the fact that steps b) and c) are qualitatively independent of steps d) and e) and can be carried out in parallel on the pre-emphasized speech signal spp.

Regarding the formatting substeps performed in steps b) and e) on the speech signal pre-emphasized spp, respectively on the speech signal filtered subsampled spf, we indicate that these stages of formatting can be performed by appropriate addressing on the pre-emphasized speech signal spp, respectively the spf subsampled filtered speech signal. Well understood, the realization of the substep of calculating the smoothed frequency spectrum H (k) of step e) is conditional the availability of p cepstral coefficients C (m) at the end of step c).

The implementation of the alternative embodiment of the process object of the present invention as shown in Figure 1b does not prejudge the mono structure or multiprocessor of the device allowing the implementation of the process which is the subject of the present invention, a structure single processor with multitasking operating system which can of course be envisaged, as will be described later in the description.

Furthermore, it is indicated that, in another alternative embodiment, the method which is the subject of the present invention may consist in calculating a single set of values, denoted X (k), of the energy spectrum of the speech signal at 1 'step c) on a number M of points equal for example to 2048, that is to say the value M = M ₂ the largest previously described in the description, and to store this set of values. The number M ₁ = 128 of values used for the calculation of the cepstral coefficients in step c) can then be obtained by decimation from the set of values X (k). However, it is indicated that this other embodiment, although equivalent to the embodiment described with calculation of the first set of values X ₁ (k) then of the second set of values X ₂ (k), has the drawback of requiring keeping in memory all the values X (k) during the entire execution time of the calculation process for each of the current frames, which causes a memory congestion harmful to the management of all the resources of calculation.

A more detailed description of the betting process of step e) of the process, object of the present invention, as shown in Figures 1a and 1b, will now given in connection with Figure 2a.

According to the above-mentioned figure, step e) of calculation by spectral compression consists, as mentioned previously in the description, in performing a step e ₁ ) comprising the formatting in frames of N ₂ samples from the speech signal filtered under- sampled spf and calculation of the second set of values X ₂ (k) of the logarithm of the energy spectrum by application of a Fourier transform on a number M ₂ of points on a frequency band between 0 and 2 KHz.

Dans cette relation, k prend les valeurs [0,1,2,...M₂] et M₂ est égal à Q/4 avec Q = 8192.Sub-step e ₁ ) and followed by sub-step e ₂ ) consisting in calculating the spectral envelope H (k) or smoothed frequency spectrum of the current frame on the frequency band between 0 and 2 kHz over the same number M ₂ of points, by applying to the first p-1 cepstral coefficients of a cosine transform verifying the relation:

In this relation, k takes the values [0,1,2, ... M ₂ ] and M ₂ is equal to Q / 4 with Q = 8192.

Sub-step e ₂ ) is followed by sub-step e ₃ ) consisting in calculating the difference, denoted D (k) = X ₂ (k) - H (k).

Sub-step e ₃ ) is itself followed by a sub-step e ₄ ) consisting in calculating the function P (k) by spectral compression of the difference D (k) on the low frequency band between 70 and 450 Hz. The function P (k) is none other than the harmonic product of the difference D (k). This calculation is carried out for L = M ₂ / k, k varying for representative values of 70 to 450 Hz, that is to say in the low frequency band mentioned above.

Finally, sub-step e ₄ ) is itself followed by a sub-step e ₅ ) carrying out the extraction of the maximum of the function P (k) for the value of k representative of the value F ₀ , fundamental frequency of the speech signal.

Sub-step e ₅ ) can be carried out from a program for sorting the successive values of the function P (k) in the aforementioned low frequency band. The sorting program is a conventional type program for searching for a maximum value among several values.

In FIG. 2b, diagrams have been shown successively in a W-frequency energy space successively relating to the short-term spectrum between 0 and 2 KHz of a frame of a speech signal, the frame having a duration of 32 ms over 2048 points, this diagram possibly corresponding to a frame obtained following the formatting sub-step carried out in sub-step e ₁ ) of FIG. 2a, the spectral envelope obtained by cosine transform applied to the first 16 cepstral coefficients, this envelope representing only the contribution of the formants, that is to say the smooth spectrum H (k) obtained at the end of sub-step e ₂ ) of FIG. 2a for example, the difference D (k) between the two preceding spectra, difference in which only the fundamental frequency structure of the speech signal remains, the contribution of the formants being eliminated, this diagram corresponding to the values D (k) of the difference obtained at the end of the sub- and ape e ₃ ) of FIG. 2a, then, finally, the curve obtained by spectral compression of the fundamental frequency structure of the speech signal between 70 and 450 Hz, this function having a significant maximum or peak value for the frequency F ₀ , this latter diagram corresponding to the implementation of sub-steps e ₄ ) and e ₅ ) of FIG. 2a.

The process which is the subject of the present invention can normally be implemented on a continuous stream or pseudo-continuous of words or syllables constituting a signal of speech.

However, extensive investigations have shown the value of implementing a discrimination process between voiced and unvoiced frames because sampling of unvoiced frames is likely cause errors in frequency evaluation fundamental of the speech signal due to the fact that for unvoiced frames, the sounds do not result from a periodic vibration of the vocal cords, these frames not voiced not being significant of the fundamental frequency of this speech signal.

For this purpose, and following the sub-step consisting of submit pre-emphasized speech signal spp, respectively the spf subsampled filtered speech signal at the frame formatting sub-step, the method the present invention may advantageously consist, moreover, to discriminate, among all the successive frames, the voiced frames and unvoiced frames then to be eliminated each unvoiced frame. In fact, unvoiced frames do not are not physically removed from the frame sequence common. These unvoiced frames are discriminated by assigning a fundamental frequency value thereto arbitrary, zero value, as will be described later in the description.

Thus, as represented in FIG. 3a, the constitution of these signals in successive frames of N respectively N ₂ samples can be carried out in a conventional manner by reception and storage of these samples at specific addresses of a random access memory for example, then sequential reading , as shown in FIG. 3a, successive frames, with reading for example of the frame of rank q-1 by simultaneous reading of the N corresponding samples, then reading at the end of the frame duration, ie 32 ms, of the frame of subsequent rank q corresponding to N samples overlapping N / 2 samples with respect to the previous frame of rank q-1, and so on for the frame of rank q + 1 and the following frames. This reading process can advantageously be carried out by simple addressing in reading of the memory containing the samples of the speech signal. As shown in FIG. 3b, the formatting in frames having been carried out on one or the other signal as described in relation to FIG. 3a, the process of discrimination between voiced frames and unvoiced frames can consist, starting from the current frame T _q , in a step 100, to apply a criterion 101 of discrimination between current frames, voiced or unvoiced. On a negative response to the aforementioned criterion 101, the current frame T _q is assigned an arbitrary value of fundamental frequency, zero value for example, in a step 102, whereas on the contrary, on a positive response to criterion 101, the current frame is kept in step 103 for processing according to the calculation process to extract the fundamental frequency from the speech signal. The succession of current frames kept in step 103 is then subject, as a function of the signal considered spp, respectively spf, to the calculation of the first set of values X ₁ (k) or X ₂ (k) respectively, within the framework of the implementation of step b) or step e), or sub-step e ₁ ), of FIGS. 1a, 1b or 2a.

With regard to the actual discrimination of voiced and unvoiced frames, it is indicated that this can consist, as shown in connection with FIG. 3c, of subdividing each current frame T _q into a number ST of contiguous frame segments successive, then establishing, for each of the frame segments, a voicing discrimination criterion. In FIG. 3c, four contiguous frame segments, denoted S ₁ to S ₄ , are shown, each frame segment therefore comprising 64 samples and occupying a duration of 8 ms.

According to a particularly advantageous, non-limiting embodiment, it is indicated that the voicing discrimination criterion may consist in assigning to each frame segment considered a voicing index whose value is between 0 and 1. Each voicing index is denoted Vs (1) to Vs (4) and is representative of the low frequency energy level of the frame segment S ₁ to S ₄ considered, according to a substantially linear law. Finally, each current frame T _q is classified as an unvoiced frame by comparison of a linear combination of the voicing indices of each segment with a determined threshold value. By way of nonlimiting example, it is indicated that the aforementioned linear combination of the voicing indices can consist in calculating the arithmetic mean of these indices and in comparing this arithmetic mean with the aforementioned threshold value ε, the criterion for comparing the combination linear writing:

Finally, as shown in Figure 3d, the value of each voicing index can be assigned based low frequency energy of each segment according to the abacus shown in the above figure. In the mode of realization studied for the implementation of the object process of the present invention, it is indicated that the index value of affected voicing is linear between the values 0 and 1 for low frequency energy values of each segment between -35 and -15 dB. These values may well heard to be changed.

Finally, errors can occur in the estimation the value of the fundamental frequency of the signal speech, these errors may be due to the presence in a same frame of voiced segments and unvoiced segments or of silences. These types of errors are referred to as errors of transition. Such errors can also occur in voiced or mixed low energy frames. In certain conditions, it is then possible to correct these errors when, when correction is not possible, the value of the fundamental frequency of the speech signal is taken arbitrarily equal to a fictitious value, the zero value, by convention for example, similarly the value assigned to unvoiced frames or frames of silence.

The process which is the subject of the present invention can then consist, in addition, in carrying out a post-processing of the value extracted from fundamental frequency of the signal speech.

This post-processing step can consist of example to establish a histogram of fundamental frequencies, to determine the range of frequency values most likely as well as lower bounds and higher of these values. Following establishment of the fundamental frequency histogram, the process post-processing can be to submit each value extracted from fundamental frequency to a sorting criterion by relation to the lower and upper value limits, for get sorted values representative of the evolution values extracted from fundamental frequency.

These sorted values can then be submitted to non-linear filtering to remove outliers.

si F₀(i) > B.Sup ou F₀(i) < B.Inf F0(i) = 0

sinon si |F₀(i) - F₀(i-1)| > γ F0(i) = 0

sinon si F₀(i) < B.Inf F0(i) = F0(i)*2

si F₀(i) > B.Sup ou F₀(i) < B.Inf F0(i) = 0

sinon si |F₀(i) - F₀(i-1)| > γ F0(i) = 0.

Dans le processus de calcul précité, l'indice i affecté aux valeurs de fréquence fondamentale désigne l'ordre successif des valeurs extraites, γ représente une valeur de seuil arbitraire à laquelle est comparée la différence entre deux valeurs de fréquence fondamentale successives de rang i et i-1.Thus, for a most probable frequency band between values denoted B.Sup respectively B.Inf, upper and lower value of the frequency band, and for successive fundamental frequency values denoted F ₀ (i), the correction process can be carried out according to the calculation steps below:
if F ₀ (i)> B.Sup F 0 (i) = F 0 (i) / 2

if F ₀ (i)> B. Super or F ₀ (i) <B. Inf F 0 (i) = 0

otherwise if | F ₀ (i) - F ₀ (i-1) | > γ F 0 (i) = 0

otherwise if F ₀ (i) <B.Inf F 0 (i) = F 0 (i) * 2

if F ₀ (i)> B. Super or F ₀ (i) <B. Inf F 0 (i) = 0

otherwise if | F ₀ (i) - F ₀ (i-1) | > γ F 0 (i) = 0.

In the abovementioned calculation process, the index i assigned to the fundamental frequency values designates the successive order of the extracted values, γ represents an arbitrary threshold value to which is compared the difference between two successive fundamental frequency values of rank i and i-1.

Following the non-linear filtering, the isolated zero values are then recalculated by linear interpolation, while the non-zero values isolated in the middle of a sequence of zeros are assigned to the value O by convention. Finally, statistical parameters such as the maximum and minimum F ₀ values as well as the average value can be calculated.

A description of a device allowing the setting implementation of the process which is the subject of the present invention will now given in connection with Figure 4.

The device shown in the above figure allows the implementation of the process, object of the present invention, previously described in the description. This device presents an architecture adapted to the implementation work of this process.

As shown in the above figure, it includes a circuit 1 for sampling and conversion analog-to-digital analog speech signal entry into a series of digital samples. In addition, a host computer 2 is provided to allow driving of the succession of steps a) to e) of the process which is the subject of the present invention, as well as management and control peripheral organs such as in particular the circuit sampling 1 and analog-digital conversion, as will be described later in the description.

The device represented in FIG. 4 further comprises a dedicated digital signal processor 3 interconnected, on the one hand, by a link by BUS to the host microcomputer 2, and, on the other hand, by a link specific to the conversion circuit analog-digital 1, this digital signal processor 3 making it possible to carry out the operations of calculating the first set of values X ₁ (k) of the logarithm of the energy spectrum of the speech signal by Fourier transform on a number M ₁ of points, the calculation of the first cepstral coefficients, the low-pass filtering and the sub-sampling of the speech signal sp as well as the calculation of the second set of values X ₂ (k) of the logarithm of the energy spectrum, the calculation of l set of values H (k) of the smoothed frequency spectrum, the calculation of the function P (k) and the operation of extraction of the maximum of the function P (k) for k = F ₀ , value extracted from the frequency of the speech signal. The acquisition of the constituent samples of the speech signal sp is carried out by the host computer 2, via the signal processor 3.

In a nonlimiting embodiment, we indicates that the dedicated digital signal processor 3 can be made up of a MOTOROLA signal processor, referenced DSP56001, clocked at the clock frequency of 33 MHz. The host microcomputer 2 can advantageously be consisting of a PC-PENTIUM microcomputer, clocked at a clock frequency of 90 MHz and equipped with a operating system such as an operating system MS-WINDOWS multitasking. The digital signal processor dedicated 3 is a 24 bit fixed point processor, this type processor to perform calculations previously cited, for the implementation of steps a) to e) of process object of the present invention optimally. This signal processor 3 is in fact constituted by a central processing unit 30, denoted DSP-CPU, to which associated with a program memory space denoted P, referenced 31, and two data memory spaces, noted X and Y, with a capacity of 512 words each and referenced 32. The memory spaces P, X and Y are each accessible by three independent 24-bit BUSes, the addressing being done by three 16-bit BUSs for addressing separately each memory space which can therefore be extended at 64 k words.

For reasons of speed, the programs and calculation subroutines are executed within 512 words of the internal memory P, these programs or subroutines being previously loaded in the 8 k-words from memory P external. On instruction from the host microcomputer 2, a program or a subroutine can be transferred from the external memory to the internal memory to be executed. The data to be processed, data relating to the signal speech, as well as the calculation tables necessary for the calculation cepstral coefficients for example and the results intermediaries are stored in spaces X and Y 32 extended to 2 x 64 k-words.

The host microcomputer 2 has programs and subprograms to ensure dialogue with the dedicated digital signal processor 3 for performing loading code and data, reading data, code transfer, execution of one or more programs as well as the initialization of the analog-digital conversion module 1 to ensure acquisition and reproduction of the speech signal.

The assembly constituted by the conversion circuit analog-digital 1 and digital signal processor dedicated 3 is installed on an additional card, such than a card marketed by the DIGIMETRIE Company, under the PC-DSP56k / AD / MEM reference. This card, in addition to the digital signal processor DSP56001, has a analog-to-digital / digital-to-analog converter marketed by TEXAS INSTRUMENTS, under the reference TCL32040CN to ensure the acquisition of speech signals, this converter bearing the reference 10 in figure 4.

Taking into account such an architecture, it is indicated that the calculation time of the fundamental frequency, for 100 speech frames of duration 32 ms, is approximately 2.7 seconds, or 27 ms per frame of 32 ms. The calculation of the logarithm of the energy spectrum, ie the second set of values {X ₂ (k)} on M ₂ = 2048 points requires a calculation time of 14 ms. Given the complexity of the calculations made, the calculation times appear remarkably short. We also indicate that it is possible to perform these calculations in real time, since the effective calculation time of 27 ms per frame is less than the duration of each frame.

In order to improve system performance and with a view to ensuring parallel processing of the steps b), c) and d), e) of the process which is the subject of the present invention, as shown for example in Figure 1b, it is indicated that the host microcomputer can be configured from the MS-Windows operating system to operate in multitasking mode, which allows driving parallel operations in the aforementioned multitasking mode. A such a procedure is not essential but it allows optimize the use of computing resources.

It is finally understood that as regards the post-processing operations, these can be carried out at the level of the host microcomputer 2 insofar as the post-processing process, as described previously in the description according to the 'algorithm defined above, can be realized thanks to a program written by means of a language such as the language C for example, allowing a sufficient speed of treatment to ensure the correction of the successive fundamental values and frequencies extracted F ₀ (i).

Given the above architecture, it is indicated in particular that the method and the device, objects of the present invention, can advantageously be used so as to produce a speaker authentication system with a high probability of success. Indeed, it is understood in particular that the construction of the frequency histogram can be carried out, either generally for a determined number of speakers, or, on the contrary, for a particular speaker for which the frequency histogram is effectively representative of this speaker. It is of course the same as regards the value of the lower and upper limits, as well as, where appropriate, statistical parameters such as the values F _0max and F _0min and mean value of the fundamental frequency of the speech signal of this speaker. Of course, the aforementioned frequency histogram, for a determined speaker, can then be updated over time as a function of the evolution of the speaker's voice.

Claims (9)

1. A method of extracting the pitch of a speech signal, being a series of digital samples, this method comprising at least the following steps:

   a) subjecting said speech signal to a pre-accentuation process in order to generate a pre-accentuated speech signal,

   b) from the pre-accentuated speech signal and for each current frame of a series of frames, each corresponding in duration to a given number N of sample, two consecutive frames each having a time overlap in terms of number of common consecutive samples at most equal to 50/100 of the number N of samples, computing a first set of values X₁(k) for the log-power spectrum by Fourier transform over a number M₁ of points;

   c) from said set of values, computing a predetermined number p of first cepstral coefficients C(m) by applying a discrete cosine transform to said values X₁(k) over a number of these values at least equal to half the number N of consecutive samples of said current frame, said transform verifying the equation:

   

   where m = [1, 2, ..., p] and C(m) denotes said cepstral coefficients;

   d) subjecting said pre-accentuated speech signal to lowpass filtering and sub-sampling in order to generate a sub-sampled, filtered speech signal;

   e) from said sub-sampled, filtered speech signal and from said cepstral coefficients, computing, by spectral compression, for each current frame in a series of frames having a same time overlap, the maximum pitch of rank k of a function P(k) representative of the difference between a second set of values X₂(k) for the log-power spectrum and the set of values H(k) of the smoothed frequency spectrum, said function verifying the equation:

   

   where L = M₂/k, k varying between a first and a second value representative of a band of low frequencies ranging between 70 and 450 Hz, said function P(k) having a maximum for k=F₀, a pitch value extracted from the speech signal.
2. Method as claimed in claim 1, characterised in that said step of computing by spectral compression consists in successively:

   computing from said sub-sampled, filtered speech signal, for each current frame, said second set of values X₂(k) of the log-power spectrum by Fourier transform over a number M₂ of points on a band of frequencies ranging between 0 and 2 kHz;

   computing the spectral envelope H(k), a smoothed spectrum of frequencies of said current frame on said band of frequencies ranging between 0 and 2 kHz over a same number M₂ of points, by applying to said p-1 first cepstral coefficients a cosine transform verifying the equation:

   

   where k = [0, 1, 2, ... , M₂] and M₂ = Q/4;

   computing the difference (D(k) = X₂(k) - H(k);

   computing the harmonic product representative of the function P(k) by spectrally compressing said difference D(k) on said band of low frequencies ranging between 70 and 450 Hz;

   determining, by a sorting process, the maximum of the function P(k) and the corresponding rank k=F₀, an extracted pitch value.
3. Method as claimed in claim 1 or 2, characterised in that, after the step which consists in framing the respectively filtered and sub-sampled pre-accentuated speech signal, it additionally consists in distinguishing between the voiced frames and the unvoiced frames in all the frames, the pitch extraction process being applied to the voiced frames.
4. Method as claimed in claim 3, characterised in that the step which consists in differentiating between the voiced frames and the unvoiced frames consists in:

   sub-dividing each frame into a number ST of successive contiguous frame segments;

   for each of said frame segments, establishing a voice discrimination criterion on the basis of a voice index ranging between 0 and 1 representing the low frequency power level of the frame segment considered in accordance with a substantially linear law;

   categorising each frame as an unvoiced frame by comparing a linear combination of the voiced indices of each segment with a given threshold value.
5. Method as claimed in one of claims 1 to 4, characterised in that, following the step of determining the maximum of rank k of said function P(k), where k=F⁰ represents the pitch value of the speech signal, and with a view to eliminating any aberrant pitch value and eliminating the risk of error due to the presence of transition errors generated by the existence of voiced segments, unvoiced segments or silences in a same frame and by the existence of low-power voiced or mixed frames, said process additionally consists in applying post-processing to said pitch value extracted from said speech signal, this post-processing step consisting in:

   establishing a histogram of pitches in order to determine the range of the most probable frequency values and the upper and lower boundary values of these values;

   applying a sorting criterion to each extracted pitch value relative to said lower and upper boundary values in order to obtain sorted values representative of the change in the extracted pitch values;

   subjecting these sorted values to non-linear filtering in order to suppress the aberrant values.
6. Method as claimed in claim 1, characterised in that steps a) to e) are performed in sequence.
7. Method as claimed in claim 6, characterised in that steps b) and c), respectively d) and e) are performed under a multi-tasking operating system, which enables pitch extraction to be performed in real time.
8. Device for extracting the pitch of a speech signal using the method as claimed in one of claims 1 to 7, this device comprising:

   means for sampling and converting, analogue to digital, a speech signal into a series of digital samples;

   a host micro-computer enabling steps a) to e) of the method to be performed and to manage and control peripheral units, in particular said sampling and analogue-digital conversion means;

   a digital signal processor inter-connected by a BUS link with said host micro-computer and enabling the operations to be run in order to compute the first set of values X₁(k) of the log-power spectrum by Fourier transform over a number M₁ of points, the first p cepstral coefficients, lowpass filtering and sub-sampling, the second set of values X₂(k) of the log-power spectrum, the set of values H(k) of the smoothed spectrum of frequencies, the function

   

   extraction of the maximum P(k) for k=F₀, being a pitch value extracted from the speech signal.
9. Use of the method and the device for extracting the pitch of a speech signal as claimed in one of claims 1 to 8 as a means of authenticating one or more speakers.

Images