FR2741744A1 - Energy evaluation method for speech signal in low bit rate vocoder - Google Patents

Abstract

The method involves performing analysis of signals transmitted between transmission and reception vocoders to predict the linear response characteristic. A filter in the transmitter circuit (3) has an impulse response which is the inverse of the receiver synthesis filter response. The receiver synthesis filter is preferably a linear phase digital filter. The filter estimates (4) the residual frequency spectrum. The estimate allows the receiver vocoder impulse response filter to be set to the optimum frequency response curve to reduce the residual spectrum frequency.

Description

 The present invention relates to a method and a device for evaluating the energy of the speech signal by sub-band for low bit rate vocoder.

 It applies in particular to the production of linear prediction vocoders according to which the speech signal is divided into frames of fixed duration, and where a data packet representative of the speech is transmitted to a reception vocoder during each frame considered.

 In these vocoders, the synthesis of the speech signal is obtained on the receiver side by filtering, by a synthesis filter, an excitation signal obtained by the combination in adequate proportion of a white noise and a train of periodic pulses, the whole having a generally flat frequency spectrum.

 A disadvantage of these vocoders is that when the bit rate of the speech sample packets is reduced to 2400 bits / s and less, the speech signal reconstituted at synthesis has a "synthetic" sound. This stems from the fact that the analysis by linear prediction which is used for low flow rates is essentially designed to properly represent the resonances and not the anti-resonances produced in particular by the pronunciation of the so-called "nasal" vowels.

 The object of the invention is to overcome the aforementioned drawback by proposing a method for generating additional data to modulate the energy of the excitation signal of the synthesis filter in order to improve the quality of reproduction of the anti-resonances, in especially the frequency dips in the low frequencies where they are most noticeable.

 It also relates to a device for implementing the above method.

Other characteristics and advantages of the invention will become apparent from the following description given with reference to the appended drawings which represent:
Figures 1 and 2 an illustration of the method according to the invention in the form of flowcharts.

 Figure 3 a receiving device for the implementation of the method according to the invention.

 FIG. 4 is a graph illustrating a distribution of the central frequencies of the different sub-bands making up the filter of the excitation signal of the synthesis filter.

 FIG. 5 is a graph illustrating the impulse responses of elementary filters used in the filtering of the excitation signal of the synthesis filter.

 FIG. 6 of the compared frequency responses of the filter of the excitation signal and the residual signal.

 The method according to the invention consists in estimating the frequency spectrum of a prediction error signal from the transmission vocoder, so as to act upon reception on the coefficients of a filter, to filter the excitation signal from the synthesis filter so that the speech signal reconstituted by the synthesis filter has a shape which is as close as possible to the speech signal applied to the input of the transmission vocoder.

avec m=1...M où Fo est une fréquence de référence fixée par exemple à 1 000 Hz.To reproduce the frequency spectrum well, especially in the low frequencies, the method according to the invention aims to predetermine a set of M frequencies gm distributed according to a semi logarithmic law of the form

with m = 1 ... M where Fo is a reference frequency fixed for example at 1000 Hz.

 The M desired values of the frequency response corresponding to the set of frequencies gm are noted below Dm.

The excitation signal is filtered by a filter H (z) with linear phase, with 2p + 1 coefficients whose impulse response H (z) is of the form
H (Z) = ZP (hp ZP + ... + hlZ + ho + hlZ I .. + hpZ P) (2)
According to a preferred embodiment of the invention, this filter is constructed according to a linear combination of K elementary subband filters with linear phase and of the same propagation time of impulse response group Hk (Z) such that

pour k=1...K avec Fr 1/2=o et FK+1/2=9M. To obtain an approximately flat frequency response of the filter H (z), for example to within 0.5 dB, each kth filter is dimensioned in the manner of a band pass filter of central frequency Fk and of band pass Bk = Fk + 1t2-Dk-1/2, the frequencies Fk being themselves distributed on the same semi-logarithmic scale as the frequencies gm and such that

for k = 1 ... K with Fr 1/2 = o and FK + 1/2 = 9M.

pour i = O...p, relation qui donne généralement satisfaction.The coefficient hi k of rank i of each kth filter which represents the i th sample located from the center of the impulse response of the filter can be defined by the relation

for i = O ... p, relation which generally gives satisfaction.

 The coefficients) k which define H (Z) are calculated so that the frequency response of the filter H (Z) is as close as possible on the frequencies gm to the desired frequency responses Dm.

soit en application de la relation (3)

The filter H (Z) being linear phase its frequency response is of the form

either in application of relation (3)

(8) entre la réponse en fréquence souhaitée (Dm) et la réponse effective du filtre (H(gm).The optimal values of the coefficients Xk are those which minimize
Total square error

(8) between the desired frequency response (Dm) and the effective response of the filter (H (gm).

pour k=1...K
En appelant par

la réponse en fréquence du kème filtre à la fréquence gm le système d'équation devient

pourk=I...K soit encore

pour k=1...K.This is obtained for example by execution of the known least squares method consisting in canceling all the derivatives of E with respect to each of the K coefficients Xk and by solving a system of K equations defined by the relation:

for k = 1 ... K
By calling by

the frequency response of the kth filter at the frequency gm the equation system becomes

pourk = I ... K be again

for k = 1 ... K.

pour k=1...K par le fait que les coefficients ak,m sont constants et ne dépendent que de la forme exacte des filtres Hk(Z) et des fréquences de référence g,. In order to minimize the computational load, the solution of the system of equations, defined by the previous relation, takes place by transforming the previous relation into the relation

for k = 1 ... K by the fact that the coefficients ak, m are constant and depend only on the exact form of the filters Hk (Z) and the reference frequencies g ,.

 If the filters are orthogonal, that is to say have disjoint frequency responses, the coefficient Xk assigned to the filter of rank k is simply the average of the desired frequency responses to the frequencies "covered" by the filter.

pour m=1...M1 dans laquelle les wj sont les coefficients d'une fenétre de pondération destinée à lisser la réponse en fréquence, par exemple

i=O. . r et Ri est le résultat d'un calcul d'autocorrélation jusqu'à l'ordre r d'échantillons de signal résiduel tel que

i=O. . r. The calculation of the coefficients Xk takes place on the emission side according to steps 1 to 7 of the method represented on the flow diagram of FIG. 1. In steps 1 and 2 the coefficients of the synthesis filter are quantified by applying one of the analysis methods known from the speech signal by linear prediction and as described for example on pages 107 to 137 of the book
MM. René BOITE and Murat KUNT published by Presses Polytechniques
Romandes with the title "Speech processing". Step 3 consists in modeling a filter A (z) of the speech signal by the coefficients of the reception synthesis filter calculated in step 2, to reproduce in the vocoder d transmitting a filter whose impulse response is the opposite of that of the synthesis filter used on reception. The signal resulting from the filtering of stage 3 of the speech signal is a signal whose frequency spectrum is in principle very close that of the excitation signal of the synthesis filter of the reception vocoder whose spectrum is generally flat, since it results from an adequate combination of white noise and a periodic pulse train. This difference between the two spectra is used in steps 4 to 7 to calculate the coefficients Ak of the filter of the excitation signal of the synthesis filter in steps 5, 6 and 7. Step 4 consists in estimating a smoothed version of the frequency spectrum of the s residual ignal obtained following the execution of step 3, so that the desired frequency responses denoted Dm in FIG. 1 do not accidentally fall into narrow hollows or narrow bumps in the frequency spectrum of the residual signal. the desired frequency response takes place over a determined number M of frequencies by performing a discrete cosine transformation defined by the relation

for m = 1 ... M1 in which the wj are the coefficients of a weighting window intended to smooth the frequency response, for example

i = O. . r and Ri is the result of an autocorrelation calculation up to the order r of samples of residual signal such that

i = O. . r.

pour k=1...K où K désigne le nombre de filtres élémentaires utilisé pour la construction du filtre utilisé côté synthèse.The coefficients Xk of the filter H (z) are obtained by solving a system of equations defined by the relation

for k = 1 ... K where K denotes the number of elementary filters used for the construction of the filter used on the synthesis side.

 The coefficients Ak are then calculated in step 5 by applying the relation (13) and quantified in step 7, possibly after normalization in step 6, before being transmitted.

 The execution of the method in the reception part of the vocoder takes place according to steps 8 to 10 of the flow diagram of FIG. 2 consisting in carrying out a dequantification, according to step 8, of the coefficients Xk. a calculation according to step 9 of the coefficients of the filter H (z) defined by relations (3) and (5) and a filtering according to step 10 of the excitation signal of the synthesis filter by the filter H (z) thus defined.

 A corresponding reception device for implementing the method according to the invention is shown in FIG. 3. This device comprises, in a known manner, a synthesis filter 11 excited, through a gain control device 12, by an excitation signal supplied alternatively through a voiced / unvoiced sound switch 13 by a noise source 14 and a pulse source 15. For the reproduction of the voice signal the response of the synthesis filter 11 is controlled by the coefficients ak obtained in step 2 of quantification shown in fig 1. A complementary filter H (z) 16 having the characteristics defined above is interposed according to the invention between the switch 13 and the gain control device 12.

 By way of example, for a choice of a sampling frequency of 8000Hz and a frequency of 9m = 4000 Hz, the filter H (z) can be formed as shown in the diagram in FIG. 4 by a set K of six elementary filters, centered respectively on the frequencies 500, 1,000, 1,500, 2,500 and 3,500 Hz. Using an H (z) filter with 2p + 1 coefficients as defined above, the impulse response of each of the 6 filters is for p = 16 that of FIG. 5. The results obtained by the different operations are shown in FIG. 6. The dotted curve A shows the frequency spectrum of the residual signal. In this example, the spectrum is obtained from a calculation of a fast Fourier transform on N = 256 points.

 The curve B, formed of small circles represents the values Dm of the desired frequency response calculated on M = 30 points shows a greater density in the low frequencies. This frequency response corresponds to a smoothed version of the spectrum of the residual signal.

 The points 6 in number represented by asterisks correspond to the normalized values of the coefficients Xk. They represent with something near, an exaggerated version of the desired frequency response comprising only 6 points instead of 30.

 Finally, the curve in solid lines C is the final frequency response of the filter H (z) obtained. Its impulse response appears very close to the desired frequency response.

 For the implementation of the method according to the invention, it is possible to use signal processing processors suitably programmed according to the steps of the method previously described.

Claims (9)

 1. Method for evaluating the energy of a speech signal by subband between a transmission vocoder and a reception predictor with linear prediction, characterized in that it consists in performing on the transmission side a filtering (3) of the signal of speech applied to the input of the vocoder by a filter having an impulse response opposite to the impulse response of the reception synthesis filter, estimating (4) the frequency spectrum of the residual signal obtained and determining (9) in the vocoder receiving the impulse response H (z) of a filter to filter the excitation signal of the synthesis filter so that its frequency response curve is as close as possible to the frequency spectrum of the residual signal.  2. Method according to claim 1, characterized in that the filter H (z) of the excitation signal of the synthesis filter is a digital linear phase filter.  3. Method according to claim 2, characterized in that the filter H (z) of the excitation signal is constructed according to a linear combination of a determined number of K elementary filters with linear phase and of the same group propagation time to share the bandwidth of the filter H (z) of the signal of excitation in sub-bands.  4. Method according to claim 3, characterized in that the central frequencies FK of the sub-bands are distributed according to the same semi loaarithmiaue scale by checking a relation of the form

 with F1.112 = O and FK + 1/2 = 9M  5. Method according to claim 4, characterized in that each hitk coefficient of rank i of each kth filter is determined by a relation of the form

 6. Method according to claims 3 to 5, characterized in that the coefficients XK representing the gain of the elementary filters are determined by minimizing the total quadratic error between the desired frequency response defined by the frequency spectrum of the residual signal (Dm ) and the effective filter response (Hgm) of the excitation signal.  7. Method according to claim 6, characterized in that the desired frequency response Dm takes place over a determined number M of frequencies by execution of a discrete cosine transformation defined by the relation

 and Ri is the result of an autocorrelation calculation up to the order r of residual signal samples.

 8. Method according to claims 3 and 7 characterized in that the coefficients Xk of the filter of the excitation signal are obtained by solving the system of equations defined by the relation

 9. Device for implementing the method according to any one of claims 1 to 8, characterized in that it comprises one or more suitably programmed signal processing microprocessors.