JPS6170599A - Voice reproduction 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 [Technical Field 1] The present invention relates to an audio reproduction method using a linear prediction method such as a PARCOR method or an LPG method.

1 Volume technology 1 Currently, audio synthesis rLIcht using linear prediction methods such as PARCOR method? ! R types are commercially available. Speech synthesis using this type of linear prediction method is as follows:
Prediction coefficient Ak that minimizes the two-east average value of residual iZn
is calculated, and if the predicted order p is sufficiently large (p≧8),
Taking advantage of the fact that the residual Z++ is an impulse train corresponding to the sound source and the notch in the voiced section, and close to white noise in the unvoiced section, the residual waveform is converted into an impulse train with a Pino thousand period and an average residual intensity in the voiced section. (In some cases, triangular skin, 1
Pinch residual l-rind etc. are also used), and information is compressed while maintaining good sound quality by replacing the average residual intensity with bright white noise in unvoiced sections.

The second section is a simulated representation of the residual waveform.In the unvoiced section of (ai), the voice parameters are determined so as to minimize the root mean square value of the white noise-like residual; b)
In the voiced sound section, an impulse-like residual is added to ν, so the speech parameters in this case 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, it is thought that it occupies the slope part of the root mean square value.
) Since the voice parameters are determined so that they are raised by the impulse of the root mean square value and are at the solid line position () than in the case of the figure, the 2P and mean value of only white noise-like residuals will no longer be the minimum.

However, in the process of extracting speech parameters, correlation remains in the white noise-like part that occupies most of one cycle.
By minimizing this white noise-like residual, the features of the vocal tract are extracted from this part, so it is unreasonable for the white noise-like residual to become no longer minimum due to the impulse.I) It is thought that the restoration of the original sound deteriorates accordingly. When synthesizing Mako's speech, only the impulse train is used as the sound source waveform in the voiced section, so the speech parameters are determined so that the white noise residual becomes zero at the analysis stage. It is desirable to leave it there.

[Object of the Invention 1 The present invention has been made in view of the above problems,
In the predictive analysis of a speech reproduction method that approximates the residual waveform obtained by the linear prediction method using an impulse sequence and white random noise, the speech parameters are set so as to minimize the root mean square value of only the white noise-like residual in the voiced speech section. The purpose of this is to improve audio restoration performance. = [Disclosure of the Invention 1 The present invention provides a speech reproduction method in which the residual waveform resulting from linear prediction is approximated by impulse sequence sealing and white noise. A residual waveform is generated using the waveform information, this residual waveform is subtracted from the original audio waveform, linear predictive analysis is performed again on this waveform to obtain each audio parameter, and these audio parameters and the above residual waveform are The speech waveform is synthesized using information, and prior to prediction, the root mean square value of only the impulses, that is, D in (1+)l in FIG. 2 is canceled out.

FIG. 1 is a block diagram showing one embodiment of the present invention. In the figure, the original audio passes through a low-pass filter 1 and then passes through the A/D-
After conversion 2, it is stored in memory 3. Next, this original audio waveform S1. 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, and the audio parameter Ak and the residual waveform Z at this time. is discarded, and only the residual information P is used. These P upstreams U are added to the sound source generating circuit 6 to generate a residual waveform Zl consisting only of a pure impulse train and white particles. This residual waveform is stored in memory 3.
The second prediction 7 is performed on the thus obtained waveform S1 by subtracting it from the original speech waveform S. The audio parameters Ak, P, and U parameters obtained in the second prediction 7 (same as P, U, and U obtained previously) 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 Z2 that is exactly the same as Zl, and the second is sent to the digital filter 1.
0 and synthesizes the speech waveform S2 using the speech parameter Ak, and passes it through the D/N converter 11 and the low-pass filter 12 to produce synthesized speech.

In this way, the original speech waveform S. When the second prediction is made by subtracting the residual waveform ZI from the waveform 3 and l, the root mean square value of the impulse, that is, D in Figure 2 (b>) cancels out for the following reason. According to the current sound source (
If prediction is performed using only the waveform Z1 as input, then Z, is only an impulse train in the voiced section, and f1.
Since the time occupied in one impulse cycle is small, almost no sampling is required, and all the audio parameters become substantially zero, and the residual becomes the same impulse train as the sound source. Therefore, “source waveform Z1
When the waveform S, (=S, -Z,) obtained by superimposing the inverted waveform <-2,> on the original speech waveform S, is manually predicted, the softened waveform of the impulse train is superimposed on the residual. Therefore, the root mean square value of the original impulse is canceled out, and only the white noise-like residual, which is a pure error component, remains in this part by minimizing the root mean square value. This almost completely eliminates the correlation between the two.

[Effect of the invention 1 As mentioned above, the present invention first generates a residual waveform using the residual waveform information obtained by predicting the original speech waveform, and then subtracts this residual waveform from the original speech waveform. The waveform is predicted again to obtain each voice parameter, and the voice waveform is synthesized using these voice parameters and the residual salt waveform information. Speech parameters can be extracted so as to minimize the root mean square value of the residuals of This has the advantage that it can be improved by a new configuration.

[Brief explanation of the drawing]

FIG. 1 is a block circuit diagram showing one embodiment of the present invention, and FIG.
Figure (II) (lJ) is a waveform diagram showing the same operation as above. 1 is a low-pass filter, 2 is an A/D converter, 3 is memory, 4
1 is a microcombiner, 5 is a first prediction circuit, 6 is a sound source generation circuit, 7 is a second prediction circuit, 8 is a storage or transmission circuit, 9 is a sound source generation circuit, 10 is a dinotal filter, 11 is a D/At converter, 12 is an I[L* filter, S,
is the original audio waveform, Sl is the superimposed waveform, S2 is the reproduced audio waveform, and Z. is the residual waveform, Z, and Z2 are the approximated sound 2! (Residual) $L type. Agent Patent Attorney Ishi 1) Written amendment to the 7th procedure (voluntary) December 29, 1980

Claims (1)

[Claims] (1) In a speech reproduction method that approximates a residual waveform by linear prediction with an impulse train and white noise, first performs a linear prediction analysis on the original speech waveform, generates a residual waveform using the obtained residual waveform information, A voice characterized in that each voice parameter is obtained by performing linear predictive analysis on the waveform obtained by subtracting this residual waveform from the original voice waveform, and a voice waveform is synthesized using these voice parameters and the residual waveform information. Playback method.