JPS5855992A - Voice analysis/synthesization system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates generally to narrowband speech transmission systems, and more particularly to linear predictive analysis and synthesis systems.

The linear predictive speech analysis and synthesis method was developed for the purpose of highly efficient digital transmission of speech signals, and its basic idea can be expressed generally by dividing speech information into sound source information and spectral information (voice information). On the receiving side, spectrum information is added to the sound source signal to synthesize and reproduce audio information. The characteristic feature of the linear predictive speech analysis and synthesis method is that H(Z): and
The method is to use either periodic pulses or white noise alone, or a mixture of both at an appropriate ratio7, as sound source information. Therefore, the information to be transmitted includes the coefficient αi of the all-pole function or an equivalent coefficient, the average amplitude or average energy of the sound source, information on the characteristics of the sound source, that is, whether it is pulsed or white-tone, and the pulseness. A period for generating pulses in the case of a sound source is required.

What is expressed by an all-pole function of the form is that in the audio time series signal St, the signal St at a certain point is p
α1S (-
i, which corresponds to being optimally predicted in the least squares error sense. Furthermore, making predictions in this manner means that there is a strong correlation between adjacent signals in the time series signal SL. Note that αl is called a linear prediction coefficient. On the other hand, regarding the sound source information, first obtain the linear prediction coefficient from the time series signal St, and then combine the signal Slt predicted from this K with the original signal St.
The difference between the two is considered to be the sound source signal ε, and the necessary information, that is, the size and characteristics of the sound source, is determined from ε. As a procedure equivalent to this, the sound source signal εt can be obtained by successively removing adjacent correlation components from the time-series signal St.

Now, when speech is analyzed and synthesized using the conventional linear prediction type analysis and synthesis method, synthesized speech with sufficient intelligibility can be obtained, but when naturalness and individuality are taken into consideration, the sound quality is not necessarily satisfactory. The main reason for this is that a pulse train or white noise is used as the sound source characteristic, but it is unreasonable to treat the sound source signal εt of the actual speech waveform as a pulse train, especially in the case of voicing, and furthermore, the phonological Although it is possible to completely restore the original voice by directly encoding and transmitting the signal εtF<< to the spectrum characteristic, it is impossible to realize narrowband transmission because the amount of information to be transmitted is too large.

The present invention therefore aims to eliminate the above-mentioned drawbacks of the prior art and is characterized in that in the case of voicing, the source signal ε
A part of the waveform extracted from t or its impulse response is encoded and transmitted, and when synthesizing a voice waveform, the transmitted waveform is used as a pulse symbol to synthesize voice. This will be explained in detail below with reference to the drawings.

Figure 1 shows an example of a linear predictive speech analysis and synthesis method. In this system, a partial autocorrelation coefficient that is equivalent to the linear predictive coefficient αi and has characteristics suitable for transmission is determined and transmitted. are doing. Further, as a procedure for obtaining the sound source signal εt from the time-series signal St, a method is used in which correlated components are removed one after another from adjacent signals. As the sound source, one of the sound source waveforms extracted from the prediction error signal sequence εt in the white noise analysis section is used.

In FIG. 1, in the analysis section, partial autocorrelators 3 to 2 correspond to a part for obtaining spectral information and sound source signal εt, a register 13, an autocorrelator 14, and a maximum value detection circuit 15.
, the divider 16, and the determination circuit 17 correspond to the part for obtaining sound source information. On the other hand, in the synthesis section, the sound source waveform register n, multiplier U, white noise generator 5, switch 26, and square root circuit correspond to the part where the multiplier generates the sound source signal. Corresponds to the part added to the sound source.

Note that the maximum value detection circuit 101, the normalization circuit 102, and the sound source waveform register path are important parts of the present invention that are not found in conventional systems.

Next, the operation of the system shown in FIG. 1 will be explained.

This system is divided into an analysis section and a synthesis section.

First, in the analysis section, after the input audio signal v (t) passes through the low-frequency P-wave device 1, it is analyzed by the AD converter 2 as a digital sampling data series St.

First, the input data series St is a 10-stage partial autocorrelator 3~
12, it is converted into a prediction error signal sequence C4 from which low-order correlation components have been removed. FIG. 2 shows an example of the prediction error signal sequence εt for the input signal v (tl. Note that FIG. 2(a) is for a voiced sound, and FIG. 2(13) is for an unvoiced sound.General Specifically, the prediction error signal sequence εt is periodic for voiced sounds such as vowels, and εt has a waveform close to white noise for unvoiced sounds.

On the other hand, the synthesis section assumes that εt is either white noise or the sound source waveform obtained by the analysis section, and the analysis section needs to determine the characteristics of εt.

Therefore, the prediction error signal series εt is sequentially stored in the register 13, the autocorrelator 14 calculates the autocorrelation coefficient once every 10 milliseconds, and the maximum value detection circuit 15 calculates the autocorrelation coefficient
The largest autocorrelation coefficient VMA other than the next autocorrelation coefficient v0
The order of the autocorrelation coefficient that gives the maximum value of The periodicity measure σ is obtained by dividing VMAXY by VMAXY, and in the judgment circuit 17, if σ is smaller than 05, it is considered to be voiced and is set to 1, and if σ is less than 05, it is not considered to be voiced, and O is the value of V/UV. Output as .

In the above process, partial autocorrelations determined by partial autocorrelators 3 to 12 at °C, coefficients kH (i=1 to 10),
Information on the energy vo of the prediction error signal, the fundamental period L, the shape of the sound source waveform extracted from the prediction error signal sequence εt, and the voiced/unvoiced determination V/UV is sent to the encoder 18 every 10 milliseconds.
is sent to a modem and modulator (MODEM') after being encoded.
The data is transferred from the communication line adder to the synthesis unit via 19.

Next, in the combining section, the signal from the communication line is sent to the modulator/demodulator (M
ODEM) 21, the decoder 22 calculates the partial autocorrelation coefficient Qi (i=l~10), the energy v0 of the prediction error signal, and the voiced/unvoiced judgment V/ of the fundamental period L1.
It is converted into UV and sound source waveform information.

The synthesis section includes a sound source waveform register that outputs the sound source waveform every fundamental period sent from the analysis section, and an amplifier 24 that amplifies the size of the sound source waveform by L times the length of the fundamental period.
A part for voiced sound consisting of a white noise generator 5 and a part for unvoiced sound using a white noise generator 5 are prepared to provide a sound source signal et for synthesizing speech, and the switch song is switched by the V/UV signal. to specify which one to use. Next, it is converted into an amplitude value, and the source signal e is multiplied by B0,000 in the amplifier to generate a negative prediction error signal sequence on the synthesis side. Next is the signal series? The synthesis filter 6, which has a partial autocorrelation coefficient for t as a coefficient, generates a digital data sequence t by adding adjacent correlation components through a process opposite to that of the analysis section.

Thereafter, St passes through a DA converter 39 and a low-frequency r wave generator 40 and is output as synthesized speech v (t).

FIG. 3 shows the configuration of one stage of partial autocorrelators 3 to 12. The partial autocorrelator is a delay circuit 41 within the unit sample time.
, partial autocorrelation, the part for calculating the number is adder 4
2, 43, square circuit 44, 45, adder 46° 47
, an averaging filter 48 #494, and a divider 50, and when the signal sequences xt and yt are input, the output from the adder 46 is 4xtyt, and the output from the adder 47 is 2xt2-1. −2 y p, and the magnitude of the correlation component and energy of the signals xi and yl are determined.

On the other hand, the inside of the averaging filter 49 is a digital low-frequency f-wave filter as shown in FIG.
, 52 , 53, delay circuit 511 within unit sample time,
55, 56, and multipliers 57, 58, 59, and has the function of time-averaging the input.

Therefore, the output from the averaging filter 48 is E(4xty
t), and the output E (xt2+
yl2), which is normalized by the correlation component l energy of the signals xt and yt as a result of the division by the divider 50, and the partial autocorrelation coefficient ki is obtained by the above process. Next, using this partial autocorrelation coefficient, a lattice-type circuit performs an operation to remove correlation components from the signal sequences xt and yt. 63. Furthermore, a terminal 64 is prepared to output the partial autocorrelation coefficient ktY.

FIG. 5 shows synthesis filters 4...37, 38.
The configuration of one stage is shown. The synthesis filter is composed of a delay line 71 within the unit sample time, adders 72, 73, 74, and a multiplier 75 for multiplying the partial autocorrelation coefficient, and has the function of adding a correlation component to the signal. There is.

By the way, the important feature of the present invention is, as already mentioned,
The purpose of this invention is to use a part of the prediction error signal sequence εt or a sound source waveform obtained by encoding and transmitting its impulse response in place of the conventional pulse train in the voiced sound part when synthesizing the voice waveform. In this embodiment, a method using a part of the prediction error signal sequence εt as it is is used.The above procedure is performed by the maximum value detection circuit 101 shown in FIG.
This is realized by the 1 normalization circuit 102. That is, the maximum value detection circuit 101 discovers from the prediction error signal sequence εt stored in the register 13 an interval in which the energy of consecutive N points is maximum, and the waveforms of these consecutive N points are sent to the normalization circuit 102. After being energetically normalized, it is sent to the encoder ]8. According to the experimental results so far, by using a value of N of about 15,
Approximately 7 of the energy of the prediction error signal sequence εt within the fundamental period
0% is included in the above range, and if N≧15, it is practically a juice content.

FIG. 6 shows a maximum value detection circuit 101 and a normalization circuit 10.
A flowchart of a series of processing procedures according to 2 is shown.

First, the maximum value detection circuit 101 sequentially calculates the total energy of N consecutive points of the prediction error signal sequence εt stored in the register 13, and determines the position and corresponding position where the total energy is maximum for εt in the register 13. A signal sequence is obtained. Next, the prediction error signal sequence εt of consecutive N points having the maximum total energy is sent to the normalization circuit 102, normalized by energy, and encoded and transmitted as sound source information.

The result of this structure is that the lack of naturalness and individuality that was a problem with conventional analysis and synthesis methods is removed, and the quality of the synthesized speech is much better. . In addition, the system is extremely simple, requiring only a maximum value detection function, a normalization function, and a storage device to be slightly added to the conventional system. In addition, a general-purpose computer programmed with the main part of the operation shown in Figure 1 explained above is also available.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a linear predictive speech analysis and synthesis device according to the present invention, and FIG. 2 is an example of a prediction error signal sequence.
Figure 3 is the block tire diagram of the partial autocorrelator, and the fourth
5 is a block diagram of the averaging filter, FIG. 5 is a block diagram of the synthesis filter, and FIG. 6 is a flow chart of the processing of the maximum value detection circuit 101 and the normalization circuit 102. l...Low pass filter, 2...Analog-digital converter, 3-12
... Partial autocorrelator, 13 ... Register, 14 ... Song autocorrelator, 15 ... Maximum value detection circuit, 16 ... Division by four circuit, 17 ... ...determination circuit, I8/music/encoder 19, 21...modulator/demodulator, addition...transmission line, η...decoder, ru... sound source waveform register,
5...In color noise generator, song...switch, υ...curved square root circuit, 昂...multiplier, 29-38... ...Synthesis filter, 39...
・・・・・・Digital-to-analog converter, 40...
......Low pass filter, 101...
......Maximum value detection circuit, 102...
...Normalization circuit patent applicant International Telegraph and Telephone Co., Ltd. Patent application agent Megumi Yamamoto - Tortoise 3 figure a Tail 4 figure Tail 5 figure Associate An

Claims (1)

[Claims] Speech analysis and synthesis uses audio as an input signal and decomposes the time-series signal into linear prediction coefficients, sound source information into hail information, and spectral information and transmits them, and synthesizes these information on the receiving side to reproduce the original audio. In this method, the transmitting side transmits a part of the waveform of the prediction error signal obtained by linearly predicting the input signal or the corresponding impulse response, and the receiving side determines whether either this information or white noise is voiced or not. Therefore, the speech analysis and synthesis method is characterized in that it is used as a driving sound source during waveform synthesis, adds spectral information to the sound source, and reproduces the original speech.