JPH01201700A - Voice analyzing and synthesizing system - Google Patents

Abstract

PURPOSE:To transmit a predictive residual by a low bit rate, and also, to obtain a synthetic voice having a high quality by converting the predictive residual to a zero crossing wave and transmitting it to a decoder side together with sound source strength information. CONSTITUTION:A pulse train being a predictive residual signal which is generat ed and reconstituted by a residual signal reconstituting device 7 is supplied to an 8-order PARCOR synthesizing filter 9, and with respect to this filter 9, as well, a PARCOR type lattice filter is used, and an 8-order decoded PARCOR coefficient Km from a coefficient decoder 8 is supplied to the filter 9. An input signal of the filter 9 corresponds roughly to a predictive residual from a PARCOR analytical filter 3 of an encoder side, and from this PARCOR synthesizing filter 9, a digital reproducing synthetic voice is obtained. This digital reproducing synthetic voice is returned to an analog sound signal by a D/A converter 10, and in an output terminal 11, a sound signal having a high quality being similar to an input voice is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech analysis and synthesis method using linear predictive analysis.

[Summary of the invention]

This invention utilizes a linear predictive analysis filter such as Percoll (
A linear prediction analysis is performed using an analysis filter (PARCOR), and the prediction residual obtained from this analysis filter is converted into a zero-cross wave, that is, each sample is expressed as one + or - bit, and transmitted to a decoder at a low bit rate. On the decoder side, a pulse corresponding to the sound source intensity is generated at the position where the sign of the zero-crossing wave is reversed, and this pulse drives a linear predictive synthesis filter, making it possible to obtain high-quality synthesized speech. be.

[Conventional technology]

As a speech analysis and synthesis method, the analysis and synthesis method based on the linear predictive analysis method has traditionally been used because it has many advantages.
Many (used.

In this method, on the encoder side (analysis side), speech is subjected to linear predictive analysis using, for example, a Percoll-type linear predictive analysis filter, and a prediction residual is obtained as a filter output. The correlation of this prediction residual (modified correlation function) shows a large correlation at an integer multiple of the fundamental period (pitch period) in the case of speech,
In the case of no sound, there is no noticeable inaudience.

Therefore, the encoder side calculates the correlation function of the prediction residual, determines whether there is speech or not, and detects the pitch period, and these pieces of information are transmitted to the decoder side (speech synthesis side).

Furthermore, the encoder side detects the sound source strength (power value), and this information is also transmitted to the decoder side.

A pulse generator and a white noise generator 2g are provided on the decoder side, and upon receiving the above information from the encoder side, the pulse period of the pulse generator is made to match the transmitted pitch period. Then, for a voice with voice, a pulse from the pulse generator whose amplitude is controlled based on information on the sound source intensity is supplied to the synthesis filter. On the other hand, for non-speech, white noise from a white noise generator whose amplitude is controlled based on information on the sound source intensity is supplied to the synthesis filter. Then, a synthesized voice is obtained from the synthesis filter (see Digital Technology Series ``Digital Speech Processing'' by Sadahiro Furui, published by Tokai University Press).

The method described above in which a sound source is modeled by voice presence, no voice, and sound source intensity has a single pitch period pulse on the decoder side, resulting in poor sound quality, and the precision of pitch period extraction when voice is present. There is a problem with the amount of calculation.

In order to improve the sound quality, multiple pulse generators are prepared on the decoder side, and on the other hand, the encoder side generates information on the generation positions of multiple pulses and their amplitudes, and this is transmitted to the decoder side. A driving method has been proposed.

However, this multi-pulse driving method has problems with the method of searching for the pulse generation position, and cannot improve the conventional problems with pitch period extraction.

As a method to improve this problem, the prediction residual itself is sent from the encoder side to the decoder side, and with this prediction residual,
A possible method is to drive a synthesis filter on the decoder side.

[Problem to be solved by the invention]

However, in the case of this method of transmitting prediction residuals, in order to obtain high-quality playback audio, it is necessary to set a sufficient number of quantization bits for the prediction residuals, which has the disadvantage of increasing the bit rate. be. For example, if the number of quantization bits is 4 bits, the bit rate is 32 pits/s.
It becomes ec.

In view of this point, it is an object of the present invention to make it possible to transmit prediction residuals at a low bit rate and to obtain high-quality synthesized speech.

[Means to solve the problem]

In this invention, the prediction residual is converted into a zero-crossing wave and transmitted to the decoder side together with the sound source intensity information.

On the decoder side, a pulse having a peak value corresponding to the sound source intensity information is generated at a position where the positive and negative signs of the zero-crossing wave are reversed. The pulse train is then supplied to a linear prediction synthesis filter.

[Effect]

In the zero-crossing waveform, one sample is expressed as one pit, such that when it is greater than zero, that is, positive, it is marked as "1," and when it is smaller, that is, negative, it is marked as "0," and so on. Therefore, the prediction residual information is transmitted in one pit.

On the decoder side, the synthesis filter is driven by a pulse corresponding to the sound source intensity, which is generated at a position where the sign of the zero-crossing wave is inverted, rather than by a zero-crossing wave generated from this prediction residual. In other words, although there is a possibility that m-sonification distortion may occur when the zero-crossing wave is converted, since the residual signal is reconstructed from the zero-crossing wave, high-quality reproduced synthesized speech can be obtained.

〔Example〕

FIG. 1 is a block diagram of two examples of a speech analysis and synthesis system according to the present invention, and FIG. 2 is a waveform diagram for explaining the same.

The input audio through the input terminal (1) (see the waveform diagram in Figure 2) is supplied to the A/D converter (2), for example 3K.
llz, and the sampled value is converted into a digital code value. This digital code value is fed to a Parcoll analysis filter (3), in this example of order 8, and to a Nosy Coal analysis autocorrelator (4). A Percoll type city filter is used as the Percoll analysis filter (3).

In the Percoll analysis autocorrelator (4), the analysis order is 8th order,
With an analysis frame length of 29 m sec, analysis calculations are performed by so-called sequential calculations, and the Percoll count and sound source intensity (power) are determined. The Percoll coefficient km (m is the order) is then supplied to the analysis filter (3) via the parameter quantizer (5). Further, the Percoll count km and the sound source intensity are quantized in a parameter quantizer (5). In this case, the number of quantized bits Km of the Percoll coefficient km is non-uniform, since the larger the order m is, the smaller the increase in distortion due to the decrease in the number of bits Krn is, so a large amount of information is allocated to the Percoll coefficient km with a lower order. Perform bit allocation. This makes it possible to reduce distortion even if the total number of bits remains the same. By the way, in this example, (Kl, K2. Ks, Ks, Ks, K
s, Ky, Km) = (16, 12, 12, 8, 4
, 4, 4, 4) bits.

A prediction residual (see waveform diagram in FIG. 2B) is obtained from the Percoll analysis filter (3), which is supplied to a zero-cross reduction circuit (6). This zero-cross reduction circuit (6) converts the residual signal into a zero-cross wave (see FIG. 2C).

That is, a value greater than zero, that is, a positive value, is encoded as "l", a value less than zero, that is, a negative value, is encoded as "0", and the encoded data (see Figure 2 C) is obtained from this. .

The above is the configuration of the encoder (speech analysis side).

The encoded data from the zero-cross reduction circuit (6) is supplied to the residual signal reconstructor (7) of the decoder (speech synthesis side). The transmission bit rate of this encoded data is 8000
It becomes bit/sec.

The quantized 8th-order Percoll coefficient km and sound source strength information from the parameter quantizer (5) are supplied to the coefficient decoder (8) of the decoder. The transmission rate at this time is 36
00 bits/sec.

Next, to explain the decoder side, the residual signal reconstructor (7) consists of a positive/negative pulse generator (71) and a gain adjustment circuit (72) for determining amplitude strength.

The positive and negative pulse generator (71) generates positive and negative pulses as shown in the second figure at positions where the positive and negative signs of the transmitted zero-cross wave (FIG. 2C) are reversed. That is, a positive pulse is generated at the rising position of the zero-crossing wave, and a negative pulse is generated at the falling position.

The output pulses from the positive and negative pulse generator (71) are supplied to a gain adjustment circuit (72), and the peak value of each pulse is adjusted. The peak value of this pulse is determined by the number of zero-crossings per unit time of the transmitted zero-crossing waves and the sound source intensity. That is, when the number of zero crosses is less than a predetermined value, the amplitude intensity of the zero cross wave is determined in proportion to the sound source intensity. However, when the number of zero crosses increases and exceeds a predetermined value, such as when there is no voice, for example, when the amplitude intensity is made proportional to the sound source intensity, the amplitude intensity becomes too large. Therefore, if the number of zero crossings exceeds a predetermined value, the sound source strength is multiplied by a coefficient a (a<l) to reduce the sound source strength, and the peak value of each pulse is determined accordingly.

By the way, in this case, the sound source strength is the analysis frame length 2Qm.
Since it changes every second, this 20ffl
During SeC, the amplitude intensity remains constant as shown by the solid line in FIG. In order to improve this and obtain higher quality reproduced synthesized speech, linear interpolation is performed from the sound source intensities before and after the section of analysis frame 2 length, as shown by the broken line in Figure 3, to find the sound source strength. The peak value of the pulse is determined accordingly.

The second diagram shows a pulse train in which the gain has been adjusted in this way and the peak value of each pulse has been determined.

The pulse train as the reconstructed predicted residual signal generated by the residual signal reconstructor (12) in this manner is supplied to the 8th-order Percoll synthesis filter (9). This synthesis filter (9) also uses a Percoll type lattice filter. This synthesis filter (9) includes a coefficient decoder (8).
) is supplied with the decoded 8th-order Barcall coefficient km.

Since the input signal of the synthesis filter (9) approximately corresponds to the prediction residual from the Percoll analysis filter (3) on the encoder side, digitally reproduced synthesized speech is obtained from this Percoll synthesis filter (9). This digitally reproduced synthesized voice is then returned to an analog voice signal by the D/A converter (10), and a high quality voice signal close to the input voice is obtained at the output terminal (11).

Note that the present invention is applicable not only to the Percoll type analysis and synthesis method but also to other linear predictive analysis and synthesis methods such as the maximum likelihood spectrum estimation method.

〔Effect of the invention〕

According to this invention, since the prediction residual is transmitted after being reduced to zero crosses, the prediction residual can be encoded with one bit, and low bit rate speech encoding can be realized. for example,
11.6 Kbit/s in the embodiment of FIGS. 1 and 3
This results in a low bit rate of ec.

Then, in the decoder, a prediction residual signal is reconstructed from the zero-crossing wave, and a synthesis filter is driven by the reconstructed signal, thereby obtaining high-quality synthesized speech.

[Brief explanation of the drawing]

Figure 1 is a system diagram of one embodiment of this invention, Figures 2 and 3
The figure is a diagram for explaining the same. (3) is a Percoll analysis filter, (5) is a parameter quantizer, (6) is a zero-cross wave converting circuit, (7) is a residual signal reconstructor, and (9) is a Percoll synthesis filter.

Claims (1)

[Claims] On the encoder side, input audio data is subjected to linear prediction analysis using a linear prediction analysis filter, a prediction residual is obtained from the linear prediction analysis filter, and this prediction residual is converted into a zero cross wave to obtain sound source intensity information. On the decoder side, a pulse train with a peak value corresponding to the above sound source strength information is generated at the position where the sign of the received zero-crossing wave is reversed, and this pulse train is supplied to a linear prediction synthesis filter, which synthesizes the signal. A speech analysis and synthesis method that obtains speech data.