JPS5852239B2 - Coding method for parameters of linear predictive speech analysis and synthesis system - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention generally relates to a narrowband transmission system for audio, and more particularly to a narrowband transmission system for transmitting audio using a linear predictive analysis and synthesis system.

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). The receiving side adds spectrum information to the sound source signal to synthesize and reproduce audio information.

As a prior art of this technology, the patent application filed in 1973 by the applicant
There is -117293.

A characteristic feature of the linear predictive speech analysis and synthesis method is that H(Z)=A is used to express spectral information (vocal tract information).
/, 5゜αiZl is used, a partial autocorrelation coefficient (partial autocorrelation coefficient) ki equivalent to the coefficient αi is transmitted, and a periodic pulse is used as the sound source. Alternatively, it is possible to use only one of the white noises or a mixture of both at an appropriate ratio.

Therefore, the information to be transmitted includes the coefficient αi of the all-pole function or an equivalent coefficient (partial autocorrelation coefficient ki), the average amplitude or average energy of the sound source, and the characteristics of the sound source, ie, whether it is pulsed or white noise. In the case of a pulsed sound source, the pulse generation cycle is required.

Spectrum information (vocal tract information) H(Z)=A/13.

What is expressed by an all-pole function of the form α, Zl is that in the audio time-series signal St, the signal St at a certain point is determined by the p signal values p St 4 (i=1 to p) just before that point. The form Σα1St-i, i=O, corresponds to optimal prediction in the sense of least squares error.

Furthermore, the fact that prediction is performed in this manner means that there is a strong correlation between adjacent signals in the time series signal St.

Note that αi is called a linear prediction coefficient. On the other hand, regarding the sound source information, first find the linear prediction coefficient of the time series signal St,
The difference between the predicted signal S't and the original signal St is considered as the sound source signal εt, and necessary information, that is, the size and characteristics of the sound source are determined from ε.

As an equivalent procedure, the sound source signal εt can be obtained by successively removing adjacent correlation components from the time-series signal Sj.

Each of the above information is encoded and transmitted.

If we consider the partial autocorrelation coefficient here, it expresses and transmits the shape of the vocal tract using parameters distributed between 1 and 10, and the vocal tract is expressed by 808 parameters ranging from 1 to 10. In this case, a low-order parameter such as 2 in k1~ (corresponding to the part of the vocal tract close to the mouth) is more distributed in the part whose absolute value is close to 1, and a higher-order parameter such as 8 in k3~ (corresponding to the part of the vocal tract close to the mouth) (corresponding to the back of the throat) is known to have a distribution close to normal distribution.

Since low-order parameters have a greater influence on speech intelligibility, in conventional technology, in order to accurately encode parts where the absolute value is close to 1, all parameters are
Encoding was performed after emphasizing the parts where the absolute value was close to 1 through non-linear processing using '%ness.

However, with this method of encoding, when the number of encoded bits is small, the result is often 0 when decoding a high-order partial autocorrelation coefficient, resulting in the apparent loss of the upper layer. This results in deterioration in the quality of the synthesized speech.

Note that in low-bit transmission of about 2400 BPS, the number of bits allocated to encoding high-order partial autocorrelation coefficients is at most 3 bits.

Therefore, the present invention aims to improve the above-mentioned drawbacks of the conventional technology, and its purpose is to improve the quality of synthesized speech in a linear predictive speech analysis and synthesis system.
For low-order partial autocorrelation coefficients, encoding is performed that emphasizes the portion whose absolute value is close to 1, and for high-order partial autocorrelation coefficients, encoding is performed that emphasizes the portion that is close to O.

Examples will be described below.

Figure 1 shows an example of a linear prediction type speech analysis and synthesis method. In this system, a linear prediction coefficient (partial autocorrelation coefficient ki, which is equivalent to ti and has characteristics suitable for transmission) is determined and transmitted. There is.

Further, as a procedure for obtaining the sound source signal εt from the time-series signal Sj, a method is used in which correlated components are removed one after another from adjacent signals.

Only one of pulses and white noise is used as the sound source.

In FIG. 1, in the analysis section, partial autocorrelators 3 to 10 correspond to the part for determining spectrum information and sound source signal εt,
The sound source information extraction circuit 11 corresponds to a part that analyzes sound source information.

On the other hand, in the synthesis section, the sound source generation circuit 51 corresponds to the part that generates the sound source signal, and the synthesis filters 52 to 59 correspond to the part that adds spectral information to the sound source.

The square circuit 12, the absolute value circuit 13, and the square root circuit 14 in the analysis section are necessary circuits for nonlinear encoding, and the nonlinear conversion result, code information of ki, and sound source information are input to the encoder 15. .

The square root circuit 32 and the square circuit 33 are circuits for nonlinearly converting the output of the decoder 31, and the multipliers 34 to 41 are k
This is a circuit for adding code information of i.

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(g passes through the low-pass filter 1, it is analyzed by the AD converter 2 as a digital sampling data series St.

First, the input data sequence Sj is input to the 8-stage partial autocorrelators 3 to 1.
By passing through 0, it is converted into a prediction error signal sequence εt from which low-order correlation components have been removed.

The sound source information extraction circuit 11 receives εt as input and obtains sound source information (average amplitude of the sound source, characteristics of the sound source, pulse period, etc.).

In the above configuration, the analyzed information is, for example, 20m5
encoded and transmitted over a 2400 BPS line,
Combined on the receiving side.

The information transmitted in one encoding is, for example, 8 items of partial autocorrelation coefficient 1 to 8, and 3 items of sound source information, totaling 1.
11 items, organized into 48-bit words.

The number of bits assigned to the partial autocorrelation coefficient for encoding 1 to 8 is, for example, kl (7 bits), k2 (6 bits), k3 (5 bits), k4 (4 bits), k, (4 bits).
bit), k6 (4 bits), k7 (3 bits), k8
(3 bits).

Among the partial autocorrelation coefficients, low-order 1 and 2 are squared by the square circuit 12.
After being subjected to nonlinear processing, the signal is sent to the encoder 15.

Further, the higher-order signals 3 to 8 pass through an absolute value circuit 13, undergo nonlinear processing by a square root circuit 14, and then are sent to an encoder 15.

The input/output characteristics of the square circuit 13 and the square root circuit 14 are as shown in Figures 2A and B, with the square root circuit emphasizing the part where the absolute value of the input is close to 1, and the square root circuit emphasizing the part of the input close to O. It is clear that

As for the configuration of the square circuit and the square root circuit, calculation by a microcomputer or table conversion by a read-only memory is possible.

Note that the partial autocorrelation coefficient has a sign (positive or negative) of 1 to 8.
is transmitted separately, and only the absolute value is subjected to square or square root processing.

The encoded signal is sent to the combining section of the receiving section via a modulator (MODEM) 21, a communication line 22, and a demodulator (MODEM) 23.

In the synthesis section, the decoder 31 decodes the code, and the square root circuit 32 and the square circuit 33 convert the code into the square circuit 1 of the analysis section.
2 and the results of the nonlinear processing by the square root circuit 14 are restored to perform inverse nonlinear processing.

Note that since the code information is transmitted separately, the multipliers 34 to
41, code information is added.

In this way, partial autocorrelation coefficients J (i=1 to 8) and sound source information are obtained.

In the synthesis section, the sound source generation circuit 5 is
1 operates to generate a prediction error signal sequence εt.

Next, adjacent correlation components are added to the signal sequence εt by synthesis filters 52 to 59 having the partial autocorrelation coefficient ki as a coefficient in a process opposite to that of the analysis section, to generate a digital data sequence St.

Thereafter, St passes through a DA converter 60 and a low-pass filter 61 and is output as synthesized speech v(t).

FIG. 3 shows the configuration of one stage of partial autocorrelators 3 to 10.

The partial autocorrelator is a delay circuit 71 within the unit sample time, and the part for calculating the partial autocorrelation coefficient is an adder 72.7.
3. Consists of square circuits 74, 75, adders 76, 77, averaging filters 78, 79, and divider 80,
When the signal series Xt, yt are input, the output from the adder 76 is 4Xtyt, and the output from the adder 77 is 2Xtyt.
t2+2yt2, which means that the correlation component and energy magnitude of the signals Xt and yt are determined.

As explained above, in the present invention, low-order partial autocorrelation coefficients are nonlinearly transformed and encoded using a square characteristic, and high-order partial autocorrelation coefficients are encoded using a square root characteristic.

The low-order partial autocorrelation coefficients are statistically more distributed in the region where the absolute value is close to 1, and the higher-order partial autocorrelation coefficients are statistically more distributed in the region where the absolute value is close to O.

Therefore, by using the nonlinear encoding method described above, it is possible to suppress the quantization error from the original coefficients during encoding and decoding, and it is possible to suppress the quantization error from the original coefficients when encoding and decoding. tanl about the partial autocorrelation coefficient
Higher quality speech can be obtained than when encoding with the l-1 characteristic.

Furthermore, the square characteristic and the square root characteristic can be easily implemented into devices, and this point is also superior to conventional nonlinear encoding systems.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of the speech analysis/synthesis system according to the present invention, Figs. 2A and 2B are diagrams showing the input/output characteristics of the square circuit and the square root circuit, and Fig. 3 is the partial autocorrelator 1. This is an example of the structure of stages. 3 to 10; partial autocorrelator; 11; sound source information extraction circuit;
12, 33; Square circuit; 14, 32; Square root circuit.

Claims (1)

[Claims] 1. In a method in which a time-series audio signal is decomposed into partial autocorrelation coefficients and sound source information, each of which is encoded and transmitted, low-order partial autocorrelation coefficients have a square characteristic. , a parameter encoding method for a linear predictive speech analysis and synthesis system, which is characterized in that parameters of higher order are nonlinearly encoded using square root characteristics. 2. The partial autocorrelation coefficient has eight transfer parameters of 1 to 8, and kl and 2 are nonlinearly encoded with a square characteristic, and k3 to 8 are nonlinearly encoded with a square root characteristic. Invention of Section.