JPH0576639B2 - - Google Patents

Description

[Detailed Description of the Invention] [Technical Field] The present invention relates to an audio reproduction method using a linear prediction method such as the PARCOR method or the LPC method.

[Background Art] Currently, various speech synthesis ICs using linear prediction methods such as the PARCOR method are commercially available. Speech synthesis using this type of linear prediction method assumes that the synthesized sound S(n) is S(n) = _p 〓 ^k=1i Ak・S(n−k)+Zn, and that the root mean square value of the residual Zn is the minimum. Find the prediction coefficient Ak such that if the prediction order p is sufficiently large (p≧8), the residual Zn will be an impulse train corresponding to the sound source pit in the voiced section, and close to white noise in the unvoiced section. The residual waveform is an impulse train with a pitch period and an average residual intensity (in some cases, a triangular wave, a 1-pitch residual waveform, etc. is also used) in a voiced interval, and an average residual intensity in an unvoiced interval. By replacing each with white noise, information is compressed while maintaining good sound quality.

Figure 2 schematically shows the residual waveform, a
In the unvoiced sound section of b, the speech parameters are determined to minimize the root mean square value of the white noise-like residual, but in the voiced sound section of b, an impulse-like residual is added, so in this case The speech parameters of are determined to minimize the root mean square value of the entire residual including impulses and white noise. However, since the impulse part has a large amplitude, 2
Therefore, in figure b, the voice parameters are set so that the zero level is raised by the impulse D of the root mean value than in figure a, and comes to the position of the solid line. is determined, so the root mean square value of only the white noise-like residual will no longer be the minimum.

However, in the process of extracting audio parameters, 1
Correlation remains in the white noise-like part that occupies most of the period, and vocal tract features are extracted from this part by minimizing this white noise-like residual, so the impulse It would be unreasonable for the white noise-like residual to be no longer at its minimum because of this, and it is thought that the restoration of the original sound would deteriorate accordingly.
Furthermore, when synthesizing speech, only the impulse train is used as the sound source waveform in the voiced section, so from this point of view it is important to determine the speech parameters so that the white noise residual becomes zero at the analysis stage. is desirable.

[Object of the Invention] The present invention has been made in view of the above-mentioned problems, and is a method for predicting a voice reproduction method in which a residual waveform obtained by a linear prediction method is approximated by an impulse sequence and white random noise. The objective is to determine speech parameters so as to minimize the root mean square value of only the white noise-like residual of , thereby improving speech restoration performance.

[Disclosure of the Invention] Accordingly, the present invention provides a speech reproduction method in which a residual waveform obtained by linear prediction is approximated by an impulse sequence and white noise. A residual waveform consisting only of an impulse train and white noise is generated, and this residual waveform is subtracted from the original speech waveform to perform linear predictive analysis again to obtain each speech parameter. The speech waveform is synthesized using the difference waveform information, and the root mean square value of only the impulses, ie, D in FIG. 2b, is canceled out prior to prediction.

FIG. 1 is a block diagram showing one embodiment of the present invention. In the figure, the original sound is passed through a low-pass filter 1, A/D converted 2, and then stored in a memory 3. Next, this original speech waveform S ₀ is read out from the memory 3 and subjected to linear predictive analysis by the microcomputer 4. In the first prediction 5, the audio parameter Ak is extracted and the pitch parameter P and amplitude parameter U are obtained as residual information, but the audio parameter Ak and the residual waveform Z ₀ at this time are discarded, and the residual Information P and U
only is used. These P and U are added to the sound source generation circuit 6 to generate a residual waveform Z ₁ consisting only of a pure impulse train and white noise. This residual waveform is subtracted from the original speech waveform S ₀ read out from the memory 3, and the second prediction 7 is performed using the waveform S ₁ thus obtained. The audio parameters Ak, P, and U parameters obtained in the second prediction 7 (the same applies to P and U obtained in the previous time) are stored or transmitted 8 as audio information. On the playback side, P and U are added to the sound source generation circuit 9 to generate a sound source waveform Z ₂ that is exactly the same as Z ₁ , and this is added to the digital filter 10 to synthesize the sound waveform S ₂ using the sound parameter Ak. It passes through a converter 11 and a low-pass filter 12 to produce synthesized speech.

When the second prediction is made using the waveform S ₁ obtained by subtracting the residual waveform Z ₁ from the original speech waveform S ₀ , the root mean square value of the impulse, that is, Fig. 2b
The reason why D in is canceled out is as follows. Assuming that prediction is performed using only the sound source (residual) waveform Z ₁ as input, Z ₁ consists only of impulse trains in the voiced sound section, and since the time occupied by one impulse period is small, it is hardly used for sampling. Therefore, all sound parameters become substantially zero, and the residual becomes the same impulse train as the sound source. _Therefore , _the waveform S ₁ ( ₌
When prediction is performed using S ₀ −Z ₁ ) as input, the inverted waveform of the impulse train is superimposed on the residual, so
The root mean square value of the original impulse is canceled out, and by minimizing the root mean square value of only the white noise-like residual, which is a pure error component, the correlation remaining in this part is almost completely eliminated. It can be removed.

[Effects of the Invention] As described above, the present invention first generates a residual waveform consisting of only an impulse train and white noise using residual waveform information obtained by predicting an original speech waveform, and then generates a residual waveform consisting of only an impulse train and white noise. is subtracted from the original speech waveform, and the prediction is performed again to obtain each speech parameter.
Since the speech waveform is synthesized using these speech parameters and the residual waveform information, the root mean square value of the white noise-like residual is minimized regardless of whether it is a voiced sound section or an unvoiced sound section. The advantage of this method is that it is possible to extract speech parameters in a very simple manner, and that it is possible to improve the original sound restoration performance when reproducing speech by approximating the sound source waveform with an impulse train and white noise. be.

[Brief explanation of the drawing]

FIG. 1 is a block circuit diagram showing one embodiment of the present invention, and FIGS. 2a and 2b are waveform diagrams showing the same operation. 1 is a low-pass filter, 2 is an A/D converter, 3 is a memory, 4 is a microcomputer, 5 is a first prediction circuit, 6 is a sound source generation circuit, 7 is a second prediction circuit, and 8 is storage or transmission. circuit, 9 is a sound source generation circuit, 10 is a digital filter, 11 is a D/A converter, 12 is a low-pass filter, S ₀ is an original audio waveform, S ₁
is the superimposed waveform, _S2 is the reproduced audio waveform, _Z0 is the residual waveform, and _Z1 and _Z2 are the approximated sound source (residual) waveforms.

Claims (1)

[Claims] 1. In a speech reproduction method that approximates a residual waveform resulting from linear prediction using an impulse train and white noise, the original speech waveform is first subjected to linear predictive analysis, and the obtained residual waveform information is used to generate a residual waveform consisting only of an impulse train and white noise. Generate a difference waveform, subtract this residual waveform from the original audio waveform, perform linear predictive analysis on the waveform to obtain each audio parameter, and synthesize the audio waveform using these audio parameters and the residual waveform information. An audio reproduction method characterized by: