JPS5941597B2 - Speech analysis and synthesis device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech analysis and synthesis device using linear prediction coefficients, which reduces deterioration of synthesized speech due to quantization of transmission parameters, and which is suitable for use in transmission paths with high transmission error rates. This invention relates to a speech analysis and synthesis device for obtaining synthesized speech.

In general, a so-called linear predictive speech analysis and synthesis device (for example, J(5HNR) ,
HASKEW, J.M.

KELLY, ROBERTM, KELLY, JR.
ANDTHOMASH, MCKINNEY'''Resu
ltsofa Study of the Linear Pre
dictionV0c0derpartEEETRANSA
CTIONONCOMMUNICATIONS, VOL
, COM-21.

A6.9 SEPTEMBER, 1973 and Fumitada Itakura
In Shu Saito's "Speech Information Compression Using Maximum Likelihood Spectrum Estimation" Journal of the Acoustical Society of Japan, Vol. 27, No. 9 (Eng. 971)), in order to transmit information on the analyzed waveform from the analysis side to the synthesis side, For example, five types of transmission parameters: linear prediction coefficient, pitch frequency, short-time average power, normalized prediction residual power which is prediction residual power when short-time average power is normalized to 1, and voiced/unvoiced determination signal, or A decoding parameter of the five types of parameters, for example, a composite parameter of pitch frequency and voiced/unvoiced determination signal, a composite parameter of short-time average power and normalized predicted residual power, or a part of the speech waveform or predicted residual waveform, etc. is used. Generally, in this type of speech analysis and synthesis apparatus, in order to improve the economical efficiency of the transmission path, it is necessary to transmit the transmission parameters using as few bits as possible within the allowable range in relation to the quality of the synthesized speech. Furthermore, it is necessary to make the analysis period of the waveform to be analyzed as large as possible. Among the transmission parameters used in a speech analysis and synthesis device using linear prediction coefficients, the normalized predicted residual power or the composite parameter of short-time average power and normalized predicted residual power are temporally sensitive compared to other transmission parameters. It is experimentally known that the rate of change is generally extremely large.

Conventionally, this type of speech analysis/synthesis device requires the accuracy of quantization of linear prediction coefficients, which are spectrum parameters, and the quantization of a composite parameter of normalized predicted residual power or short-time average power and normalized predicted residual power. The amplitude reproducibility of synthesized speech often deteriorates even when using linear prediction coefficients quantized with sufficient accuracy to represent the spectrum. There was the first drawback.

In addition, the temporal change rate of the normalized predicted residual power or the mixed parameter of short-time average power and normalized predicted residual power is significantly large compared to other transmission parameters, and the analysis cycle of the analyzed waveform is linear. It needs to be shorter than the analysis period required to analyze other transmission parameters such as prediction coefficients.
The second drawback is that it causes an increase in transmission capacity. Furthermore, due to the influence of transmission errors, conventional devices have been configured such that one or both of the linear prediction coefficient and the normalized prediction residual power, or a composite parameter of the linear prediction coefficient, the short-time average power, and the normalized prediction residual power. If one or both of these are deformed, the amplitude reproducibility of the synthesized speech deteriorates, and especially when the amplitude of the synthesized speech increases, the third drawback is that it gives a sense of strangeness to the listener of the synthesized speech. I had it. The first drawback is unique to conventional speech analysis and synthesis devices of this type, and although it is possible to select the optimal one in terms of the relationship between the quality of the synthesized speech and the transmission capacity, it cannot essentially be alleviated. . In order to alleviate the second drawback, the analysis frame period for determining transmission parameters other than the normalized predicted residual power or the composite parameter of short-time average power and normalized predicted residual power is lengthened, and the normalized predicted residual power is increased. Alternatively, by shortening the analysis frame period for obtaining a composite parameter of short-time average power and normalized predicted residual power, and interpolating the transmission parameter with a long analysis frame period in accordance with the analysis interval of the transmission parameter with a short analysis frame period. , the relationship between the linear prediction coefficient and the composite parameter of the normalized predicted residual power or the short-term average power and the normalized predicted residual power collapses, causing deterioration of the sound quality of the synthesized sound.
The third drawback, like the first drawback, is unique to this type of conventional speech analysis and synthesis apparatus, and cannot be essentially improved. The purpose of the present invention is to reduce the deterioration in the quality of synthesized speech due to quantization of transmission parameters related to the transmission interval and quantization accuracy of transmission parameters, and further reduce the occurrence of abnormal amplitude of synthesized speech due to the influence of transmission path errors. An object of the present invention is to provide a speech analysis and synthesis device using linear prediction coefficients that can be used.

The present invention relates to a linear prediction type speech analysis/synthesis device, which includes means for obtaining normalized prediction residual power from linear prediction coefficients transmitted on the synthesis side.

A feature of the present invention is that speech is synthesized using normalized prediction residual power obtained from linear prediction coefficients affected by quantization errors and transmission path errors.

Therefore, it is possible to compensate for amplitude fluctuations in synthesized speech caused by deformation due to quantization of linear prediction coefficients and transmission path errors. Next, the present invention will be explained in detail with reference to the drawings.

The figure is a block diagram showing one embodiment of the present invention. In the figure, 101 indicates an analysis side, 102 a linear prediction coefficient transmission path, 103 a sound source information transmission path, and 104 a synthesis side. Speech waveform data is input to a linear prediction analyzer, 106 and a sound source parameter analyzer, 107 via a waveform input terminal, 105. A linear prediction analyzer 106 measures linear prediction coefficients from the audio waveform zeta and supplies the linear prediction coefficients to a linear prediction coefficient encoder 108 . The linear prediction coefficient encoder 108 quantizes the linear prediction coefficient and outputs it to the linear prediction coefficient transmission path 102. The sound source parameter analyzer 107 measures sound wave information consisting of pitch frequency, voiced/unvoiced discrimination signal, and short-time average power from the audio waveform data, and supplies it to the sound source information encoder 109. The sound source information encoder 109 quantizes the sound source information and outputs it to the sound source information transmission path 103. The linear prediction coefficients quantized by the linear prediction coefficient encoder 108 are supplied to the linear prediction coefficient decoder 110 via the linear prediction coefficient transmission path 102. Linear prediction coefficient decoder 110
decodes the linear prediction coefficient to which the transmission error has been added in the quantized linear prediction coefficient transmission line 102, and outputs the decoded linear prediction coefficient to the decoded linear prediction coefficient transmission line 111 and the decoded linear prediction coefficient transmission line 112. do. The normalized prediction residual power calculator 113 calculates the normalized prediction residual power from the decoded linear prediction coefficients supplied via the decoded linear prediction coefficient transmission path 111, for example, in the form of a partial autocorrelation coefficient. If so, set the order of the partial autocorrelation coefficient to PP (-T(1)
1 (i-th order partial autocorrelation coefficient) 2) and outputs it to the normalized predicted residual power transmission path 114.

The prediction residual power is the power of the prediction residual signal, that is, the power of the residual signal (prediction residual signal) after removing the signal consisting of the prediction coefficients. The normalized prediction residual power is obtained by dividing this prediction residual signal by the human input signal. It is well known that the normalized prediction residual power can be calculated from the linear prediction coefficient (partial autocorrelation coefficient) as described above without necessarily calculating the power of the prediction residual signal directly. For example, in the Proceedings of the 8th Symposium sponsored by the Institute of Telecommunications, Tohoku University, published in February 1971, Fumitada Itakura's paper ``Speech feature extraction using statistical methods''
It is shown by the formula (to) on pages -5-1 to -5 to 12.

The value obtained by recursively calculating Un in this equation up to the order P of the partial autocorrelation coefficient is the normalized predicted residual power. That is,
U1-UO (1-Kr), U2-U1 (1-KS),・
・・・・・・・・・・・・・・・, Up-Up-1 (1-
Up calculated as K2p) is the normalized predicted residual power. U here. represents the order of the partial correlation coefficient “0”, in other words, the normalized prediction residual power when no prediction is performed, and is 1.0 for the following reason. In other words, since no prediction is performed, the prediction residual signal matches the input audio signal. Therefore, the predicted residual power and the short-term average power, which is the power of the input audio signal, match, and the ratio of the two powers is 1.0. Note that Un is the partial self It is clear from the following that Un-1 of the order n of the correlation coefficient is the prediction residual power of the order n-1.

According to the formula (20) in the above document, Un.-1 is ε
This is the expected value of the square value of Ft, that is, the power of the signal ε(n-1). ε(n-1) is called a prediction error (or prediction residual, prediction error signal, prediction residual signal) and is commonly used. Therefore, Un-, and Un are the orders of the partial autocorrelation coefficients. 1
, n is the predicted residual power.

And the initial value U. It is clear that Un when =1.0 indicates the normalized prediction residual power. The normalized predicted residual power is input to the excitation signal generator 115 via the normalized predicted residual power transmission line 114.

The sound source information quantized by the sound source information encoder 109 is supplied to the sound source information decoder 116 via the sound source information transmission line 103. The sound source information decoder 116 decodes the quantized sound source information and supplies the decoded sound source information to the excitation signal generator 115 via the decoded sound source information transmission line 117. The excitation signal generator 115 generates a filter excitation signal from the normalized prediction residual power and the decoded sound source information, and supplies it to the speech synthesis filter 119 via the filter excitation signal transmission path 118. The speech synthesis filter 119 synthesizes speech from the decoded linear prediction coefficients supplied via the decoded linear prediction coefficient transmission path 112 and the filter excitation signal, and outputs the synthesized speech to the waveform output terminal 120. Note that if the format of the decoded linear prediction coefficient is other than a partial autocorrelation coefficient, the normalized prediction residual calculator 113 converts it into a partial autocorrelation coefficient by linear operation, or by an equivalent means. It is clear that the normalized prediction residual power can be determined. In addition, transmission parameters such as short-time average power supplied via the sound source information transmission path 103 are also affected by the transmission path, but temporal changes in short-time average power, etc. are similar to temporal changes in normalized predicted residual power. It is relatively gentle, and even if it is smoothed on the receiving side, the effect on the quality of the synthesized sound is relatively small, and it is possible to easily reduce errors in the transmission path, and the purpose of the present invention is not hindered. However, the present invention is applicable to a voice excitation type linear predictive speech analysis and synthesis device (for example, B.S.APAL.M.R.SCHROEGER.V.
STOVRBELLTELEPHONELABORAT
ORIESMURRAYHILL,. N. JO7974
゛TVOlCeExcitedPredictiveC
OrdingSystemfOrLOwBitRate
TransmissionONOfSpeech'5IEE
ECATL0GNUMBER75CH09712CSC
It is clear that the present invention is applicable to BICC75/JUNEL6-18) and is not directly related to the method of transmitting sound source information.

Also, a speech analysis and synthesis device using a predictive residual waveform excitation method (for example, CHONGKWANUN.ANDD.THOMASMA)
GILL゛TheResidualExcitedLi
nearPredictiOnVOcOderwith
TransmissionOnRatEBelOw9.6K
bits/s″IEEETransactiOnsOn
Community OnslvOl. COM-23
, A6. l2, December 975),
The analysis side divides the prediction residual waveform by the normalized prediction residual power found from the linear prediction coefficients on the analysis side, compresses the amplitude fluctuation range of the prediction residual waveform, and then transmits the prediction residual waveform to the synthesis side. However, by multiplying the predicted residual waveform by the normalized prediction residual power obtained from the linear prediction coefficients on the synthesis side, the influence of transmission path errors on the linear prediction coefficients can be reduced.
It is clear that deterioration of the amplitude reproducibility of synthesized speech can be prevented.

[Brief explanation of drawings]

The figure is a block diagram for explaining an embodiment of the present invention. 101... Analysis side configuration, 102... Linear prediction coefficient transmission line, 103... Sound source information transmission line, 104... Synthesis side configuration, 105...・・・・・・
Waveform input terminal, 106...Linear prediction analyzer, 1
07...Sound source parameter analyzer, 108...
. . . Linear prediction coefficient encoder, 109 . . . Sound source information encoder, 110. ．． ．． ...Linear prediction coefficient decoder, 111...Decoding linear prediction coefficient transmission path, 11
2...Decoded linear prediction coefficient transmission path, 113...
...Normalized prediction residual power calculator, 114...
Normalized predicted residual power transmission line, 115... Excitation signal generator, 116... Sound source information decoder, 11
7...Decoded sound source information transmission line, 118...
- Filter excitation signal transmission line, 119...Speech synthesis filter, 120...Waveform output terminal.

Claims (1)

[Claims] 1. Measure parameters indicating the frequency spectrum of the input audio signal, such as linear prediction coefficients, at predetermined time intervals on the analysis side, and synthesize speech by determining the coefficients of the synthesis filter provided on the synthesis side based on the parameters. In the speech analysis and synthesis device, the synthesis side includes means for calculating normalized predicted residual power defined by a signal obtained by dividing the predicted residual power by the input signal power from the parameters, and a means for calculating the normalized predicted residual power defined by the signal obtained by dividing the predicted residual power by the input signal power. A speech analysis and synthesis device characterized in that excitation sound source information of a synthesis filter is used.