JPS61128299A - Voice analysis/analytic 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] [Industrial Field of Application] The present invention relates to a speech analysis/analysis synthesis method, particularly a so-called A-
b-S (Analysis-by-5ynthe
sis), the parameters of the vocal cord sound source waveform model are determined so that the mean squared error is minimized.

Speech that analyzes or analyzes and synthesizes speech by combining vocal cord source waveform models with linear predictive analysis methods.

This relates to analysis/analysis synthesis methods.

[Conventional technology and problems]

BACKGROUND ART Various methods have been devised in the past in order to compress the amount of voice-related information as much as possible during voice recognition, transmission, storage, etc., and to make it possible to reproduce high-quality voice from the voice-related information. One such method is a waveform encoding method that encodes the audio waveform as it is, such as ADPCM.

In addition, as a solution to this, there is a so-called analysis and synthesis method in a narrow sense by Pokogu.

In the case of the waveform encoding method, the audio signal is subjected to linear prediction analysis to obtain linear prediction coefficients and prediction errors, and the prediction errors are quantized. During reproduction, the quantized prediction error is driven by a filter using the linear prediction coefficient obtained by analysis. The distortion of the reproduced audio caused by this waveform encoding method is due to the quantization of prediction errors, and high-quality reproduced audio can be obtained. but.

The amount of information is 1, for example l 6 Kbps ~ 64 Kb
ps, and the amount of information regarding audio is considerably large.

In the case of the analysis-synthesis method, the speech generation mechanism is modeled, and attention is focused on the sound source signal and the acoustic filter characteristics of the articulator. For example, the source signal of a voiced sound is approximated as a periodic impulse train, and the source signal of an unvoiced sound is approximated as white noise.

According to this, for example, speech is represented by voiced/unvoiced sound discrimination information, pitch frequency regarding a one-period sound source, amplitude information, and i-type filter characteristics. In other words, it can be seen that the prediction error is modeled, and the audio information can be compressed to about 1.2 Kbps to 9.6 Kbps, for example. However, the quality of the synthesized speech is considerably lower than that of the waveform encoding method described above.

In analyzing or synthesizing speech, a method is desired that has a small amount of information regarding the speech and can obtain synthesized speech of high quality similar to the waveform encoding method described above.

[Means for solving problems]

The present invention aims to solve the above-mentioned problems, and when modeling the sound source waveform, instead of approximating the sound source with pulses and noise signals, a vocal fold sound source waveform model such as a Rosenberg waveform is used.

Then, the four parameters of pinch period, rise time, fall time, and amplitude for defining this vocal cord sound source waveform model are calculated using A-b-S (Analysis-b
y-Synthesis). Namely, the speech analysis/analysis and synthesis method of the present invention is a speech analysis/analysis and synthesis method that analyzes speech or analyzes and synthesizes speech based on information modeling a sound source waveform. It has a speech synthesis system that generates a speech signal by a linear prediction filter driven by a sound source signal defined by four types of parameters related to time and amplitude, and includes a selection means for sequentially selecting the four types of parameters, and a selection means for sequentially selecting the four types of parameters. means for determining an error between the synthesized speech signal obtained by the linear prediction filter and the input speech signal with respect to the four selected parameters; means for optimizing the four types of parameters and determining the four types of parameters, and performing voice analysis or voice analysis and synthesis based on the four types of parameters and the linear prediction coefficients. It is a feature.

[Effect]

The present invention provides a speech synthesis system that generates a speech signal by a linear prediction filter driven by a sound source signal of a vocal cord sound source waveform model defined by four parameters: pitch period, rise time, fall time, and amplitude, After selecting the above four types of parameters for the input audio signal, the above four types are calculated using the A-b-3 method, which repeats the procedure of performing linear predictive analysis to find the error between the 1 synthesized audio signal and the input audio signal. Determine optimal parameters. and.

These four types of parameters and linear prediction coefficients are taken as information regarding speech. If the four types of parameters and linear prediction coefficients described above are received or stored as necessary and synthesized using the model described above, high-quality speech can be analyzed and synthesized with a small amount of information. An embodiment will be described below with reference to one drawing.

〔Example〕

FIG. 1 is a block diagram of the configuration of an embodiment of the present invention, FIG. 2 is an explanatory diagram of a vocal cord sound source waveform model, FIG. 3 is a waveform diagram of an example in which the present invention is applied to synthesized speech, and FIG. FIG. 5 shows a waveform diagram of an example of the conventional method for comparison with the conventional method, and FIG. 5 shows a waveform diagram of an example of applying the present invention to natural speech.

In FIG. 1, reference numeral 1 indicates a pitch period estimator.

2 is an optimal parameter determination section, 3 is a parameter selection section, 4
5 represents a vocal cord sound source waveform generation section, and 5 represents a linear prediction analysis section.

The present invention does not use an impulse as a 5-period sound source in modeling the sound source for 9-analysis synthesis, but uses a vocal cord sound source waveform model. For example, the voices of the nine people vary in various ways, such as clear voices and suggestive voices. This may be due to the influence of differences in sound sources, and it is difficult to obtain reasonable results when approximating the -temporal impulse. By using the vocal cord sound source waveform model, approximation can be further improved. Note that it may be approximated by a triangular wave.

The vocal cord sound source waveform has a shape as shown in FIG. 2, for example. This waveform g(n) is expressed by a linear equation.

■ When tl<tst2, g(n) = 0■
When 12<1≦t3.

■ When t2<t5t3.

4-t32 This waveform can be expressed by four parameters: pitch period T, rise time ratio R2, fall time ratio F, and amplitude A, as shown below.

T= t4- tl R= (t3- t2) /T F= (t4- t3) /T A=α The pitch period estimator 1 shown in FIG.
is estimated from the input speech using various conventionally known means. The estimated pitch period is supplied to the optimal parameter determining section 2. Further, appropriate initial values are determined in advance for the parameters of the rise time ratio R, the fall time ratio F, and the amplitude A, and these values are provided to the optimum parameter determining section 2. The parameter selection unit 3 first selects these four types of parameters and outputs them to the vocal cord sound source waveform generation unit 4.

The vocal cord sound source waveform generation unit 4 calculates the pitch period T, the rise time ratio R5, the fall time ratio F, and the amplitude A.
From the two parameters, a vocal cord sound source waveform signal as shown in FIG. 2 is synthesized and output. This output signal is 2
Although not shown in the drawings, corrections are made in consideration of so-called radiation characteristics as necessary, and the data is supplied to the m-type predictive analysis unit 5.

The linear prediction analysis unit 5 calculates a prediction error from this synthesized speech signal and the input speech signal, and outputs its linear prediction coefficient.

The obtained prediction error is fed back to the optimal parameter determination unit 2.

In order to reduce this prediction error, the optimal parameter determining section 2 instructs the parameter selecting section 3 to gradually change the parameters that define the vocal cord sound source waveform. The parameter selection unit 3 selects a parameter that takes a value different from the previous parameter and outputs it to the vocal cord sound source waveform generation unit 4. This procedure is repeated to determine the four optimal parameters. That is, by using the so-called A-b-3 method, the pitch period T and the ratio R1 of the rise time are
Define four parameters: fall time ratio F and amplitude A. By extracting parameters using the A-b-3 method, more accurate analysis is possible than when determining parameters using, for example, an inverse filtering method.

According to linear predictive analysis that assumes an all-pole model for the audio signal, the audio signal s (n) is expressed by a linear equation.

Here, +ai is a prediction coefficient, and p is a prediction order.

a,+, is a gain. g (n) is assumed to be a white noise sequence.

However, according to the GLPC method, which combines a vocal cord sound source waveform model with a linear predictive analysis method, g(n) is a known waveform and does not have a flat spectrum.

That is, the audio signal s(n) is expressed by a linear equation.

Here e (n) is a white noise sequence, and the error Eg to be minimized is of order one.

By the optimum parameter determination unit 2 shown in FIG.

The parameter that minimizes this error Eg is determined. The linear prediction coefficient a, is the audio signal 5(n) and g(
The optimization D+ all+1 with respect to n) becomes the gain of g(n) that minimizes the error Eg.

FIG. 3 shows the results of applying the present invention to one synthesized speech and analyzing it using the above-mentioned GLPC in order to evaluate the method according to the present invention.

The waveform in FIG. 3(a) is the vocal cord sound source waveform used for 1 synthesis, and FIG. 3(b) is the voice signal synthesized thereby. FIG. 3(c) shows the vocal cord sound source waveform estimated by GLPC for the speech signal shown in FIG. 3(b), and the speech signal resynthesized by this is
) is shown.

Figure 3 (8) shows the error signal between the audio signal in Figure 3 (b), which was the subject of analysis, and the resynthesized audio signal in Figure 3 (d), and the resulting SN ratio is 12.7dB
It becomes.

For comparison with FIG. 3, FIG. 4 shows an example of approximating a sound source model using a conventional periodic impulse train for the same synthesized speech.

Figure 4 (a L (b) is the same as Figure 3 (a), (b)
and the vocal cord sound source waveforms used for synthesis, respectively.

The resultant synthesized audio signal to be analyzed is shown. Focusing only on the pitch period and amplitude, the linear prediction filter is driven by an impulse train as shown in Fig. 4(C), and the resynthesized speech signal obtained thereby is as shown in Fig. 4(d).
) is the signal shown.

The error between this signal and the original audio signal of FIG. 4(b) is shown in FIG. 4(e). The SN ratio resulting from this is 3.
It is 2dB. In comparison, it can be seen that in the case of the method according to the present invention, the SN ratio is significantly improved.

Of course, according to the present invention, similarly good results can be obtained with respect to natural speech. Figure 5 shows the vowel /a
An example in which the present invention is applicable to a natural speech signal of / is shown.

FIG. 5(a) shows the speech waveform of the vowel /a/ which was the subject of analysis, and FIG. 5(b) shows the speech signal analyzed and resynthesized using the present invention. The error is shown in Figure 5 (C
) As shown in the diagram, it is extremely small. Figure 5(d) shows
At that time, the vocal cord sound source waveform estimated by GLPC is shown. Incidentally.

The waveform shown in FIG. 5(e) is a vocal cord sound source waveform obtained by inverse filtering.

The present invention is particularly effective for voiced sounds, and when analyzing and synthesizing an unvoiced sound part9, for example, only the conventional waveform encoding method or analysis synthesis method is used for that part. The present invention can be implemented in combination with conventionally used methods.

〔Effect of the invention〕

As explained above, according to the present invention, it is possible to efficiently compress the amount of information related to audio and to significantly reduce the amount of information compared to the waveform encoding method. It becomes possible to obtain synthesized speech of higher quality than analytically synthesized speech.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the configuration of an embodiment of the present invention, FIG. 2 is an explanatory diagram of a vocal cord sound source waveform model, FIG. 3 is a waveform diagram of an example in which the present invention is applied to synthesized speech, and FIG. FIG. 5 shows a waveform diagram of an example of the conventional method for comparison with the conventional method, and FIG. 5 shows a waveform diagram of an example of applying the present invention to natural speech. In the figure, 1 is a pitch period estimation section, 2 is an optimal parameter determination section, 3 is a parameter selection section, 4 is a vocal cord sound source waveform generation section,
5 represents a linear prediction analysis section.

Claims (1)

[Claims] In a speech analysis/analysis synthesis method that analyzes speech or analyzes and synthesizes speech based on information modeling a sound source waveform, a sound source signal defined by at least four parameters related to pitch period, rise time, fall time, and amplitude is used. a speech synthesis system that generates a speech signal by a linear prediction filter driven by a linear prediction filter, a selection means for sequentially selecting the four types of parameters, and a selection means for sequentially selecting the four types of parameters selected by the selection means, the linear prediction filter means for determining the error between the synthesized speech signal obtained by 1. A speech analysis/analysis synthesis method, characterized in that the speech analysis/analysis synthesis method is characterized in that the speech analysis or speech analysis/synthesis is performed based on the four types of parameters and the linear prediction coefficients.