JPH0284699A - Voice analyzing and synthesizing device - Google Patents

Abstract

PURPOSE:To transmit the waveforms with a relatively small information quantity and to improve the sound quality which is like the buzzer sound specific to a VOCODER system and lacks personality and naturalness by using optimum sound source vectors to form the sound source waveforms within an analysis frame and forming synthetic waves. CONSTITUTION:A representative sound source forming circuit 8 segments the representative sound source waveforms 2b of a pitch period length from the predicted residual waveforms 2a within the analysis frame of the specified length of the vocal voices. Sound source selecting circuits 10, 11 select the optimum sound source vectors to obtain the synthesized waveform synthesized from this representative sound source by using the envelope parameter of the above-mentioned frame and the synthesized waveform of the smallest waveform distortion from the sound source vector groups in code note memories 9, 18. A sound source forming circuit 19 forms the sound source waveforms 2d by arraying the optimum sound source vectors at every pitch period within the above-mentioned analysis frame. Synthesizing filter circuits 10, 11, 22 and driven by the predicted residual signal 2a, by which the efficient transmission of the predicted residual signal is executed. The synthetic sounds of good quality are efficiently synthesized in this way with the relatively small transmission information quantity and the synthetic sounds which are natural and have the ample personality is obtd.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a speech analysis and synthesis device.

[Conventional technology]

As a method for compressing audio information by analyzing and synthesizing audio, the analysis side extracts the characteristic parameters (hereinafter referred to as envelope parameters) representing the frequency spectrum envelope shape of the input audio waveform and the pinch period, and the synthesis side extracts the pinch period and the frequency spectrum envelope shape of the input audio waveform. A method for generating synthetic speech waveforms using pitch periods (
(called a vocoder) is known. In this vocoder method, voiced/unvoiced discrimination is performed in the analysis section, and if it is determined that the sound is voiced, pinch extraction is further performed.

Then, in the synthesis section, if the voiced/unvoiced discrimination result is a voiced sound, an impulse train that repeats at a pitch period;
If the sound is unvoiced, synthesized speech is obtained by using white noise as a driving sound source and driving a synthesis filter using an envelope parameter. Although this method has the advantage of producing synthesized speech with sufficient clarity for practical use with a relatively small amount of transmitted information, the sound quality is poor and lacks individuality and naturalness compared to human-generated speech, and it is often used for long periods of time. It has the disadvantage of producing a buzzer sound that cannot be heard.

As a conventional method for improving sound quality, there is a method for improving a voiced sound source using an impulse train equivalent sound source as shown in the following document (1), for example.

[Impulse train equivalent sound source for high-quality speech synthesis] Transactions of the Institute of Electronics and Communication Engineers (1985-11) Vol. J68-^
N 0 , 11 --- (11) Although the frequency spectrum of the predicted residual signal of the voiced section is almost flat for the entire analysis frame length, Pi
The frequency spectrum of the FJtA time has a structure in which a high-frequency emphasis section and a low-frequency emphasis section repeat every pitch period. Material f
l) In the (71) method, a speech sound source (called an impulse train equivalent sound source) is generated in which the frequency spectrum of an extremely short period of time is given fluctuations that repeat every pitch period while maintaining the flatness of the frequency spectrum of the entire analysis frame. However, by using this, the quality of voiced sound sections, which greatly interferes with the sound quality of synthesized speech, could be improved.

Figure 3 is a block diagram showing the conventional method shown in this document (11).First, the audio waveform is input to the audio waveform input terminal 3.
The signal is inputted to the envelope parameter extraction circuit 4, the voiced/unvoiced discrimination circuit 5, and the pinch extraction circuit 6 in the analysis section 1, respectively. The envelope parameter extraction circuit 4 calculates envelope parameters from the audio waveform, and transmits the parameters to the synthesis filter circuit 22 in the synthesis section 2 via the envelope parameter transmission path 15. The voiced/unvoiced discrimination circuit 5 discriminates whether or not the speech waveform is a voiced sound section, outputs the discrimination result to the pitch extraction circuit 6, and further outputs the discrimination result to the synthesis unit 2 via the voiced/unvoiced discrimination result transmission line 16. It is output to the sound source switching circuit 21 of. The pinch extraction circuit 6 performs pitch period analysis from the voice waveform when the voiced/unvoiced discrimination result is a voiced section, and transmits the extracted pinch data to the pitch data transmission line 17.
Impulse train equivalent sound source generation circuit 2 in synthesis unit 2 via
Output to 4. The impulse train equivalent sound source generating circuit 24 in the synthesizing section 2 repeatedly generates an impulse train equivalent sound source having a predetermined l pitch period length for each pitch period and outputs it to the sound source switching circuit 21 . The unvoiced sound source generation circuit 20 generates white noise and outputs it to the sound source switching circuit 21). The sound source switching circuit 21 leaves behind the impulse train equivalent sound source when the voiced/unvoiced discrimination result is a voiced sound, selects white noise when the voiced sound is not a voiced sound, and sends the selected sound source waveform to the synthesis filter circuit 22. Output. The synthesis filter circuit 22 generates a synthesized waveform using the envelope parameter inputted from the envelope parameter extraction circuit 4 in the analysis section 1 and the sound source waveform, and outputs it from the synthesized waveform output terminal 23.

(IKg to be solved by the invention) In the conventional speech analysis and synthesis device using the impulse train equivalent sound source described above, it is possible to obtain a slight reduction effect on the buzzer-like sound quality of the synthesized speech, but originally it depends on the speaker and phoneme. Since a changing sound source is represented by a fixed sound source, there is little recovery of individuality and naturalness, and there is a problem in that sufficient sound quality improvement cannot be obtained.

An object of the present invention is to transmit the sound source information included in the predicted residual waveform with a relatively small amount of information, and to improve the sound quality that is characteristic of the vocoder method, which has a buzzing sound and lacks individuality and naturalness.

[Means for Solving the Problems] The speech analysis and synthesis device according to the present invention extracts one pitch from the predicted residual waveform of the frame when it is determined that the frame is a voiced section for each analysis frame of a certain length. A representative sound source extraction circuit that extracts a representative sound source waveform with a period length, a codebook memory that stores a finite number of sound source vectors, and a sound source vector that is most equivalent to the representative sound source extracted by the representative sound source extraction circuit from among these sound source vectors. a sound source selection circuit that selects a sound source vector as an optimal sound source vector, a sound source generation circuit that generates a sound source waveform in the analysis frame using the optimal sound source vector in a synthesis section; It is configured to include a synthesis filter that generates a synthesized waveform.

[Effect]

The representative sound source generation circuit in this invention cuts out a representative sound source waveform with a pinch period length from the predicted residual waveform in an analysis frame of a certain length of voiced speech, and the sound source selection circuit synthesizes it from this representative sound source using the envelope parameter of the frame. selects an optimal sound source vector from the group of sound source vectors in the codebook memory to obtain a combined waveform with the smallest waveform distortion and a combined waveform with the smallest waveform distortion;
The sound source generation circuit generates a sound source waveform by arranging the optimal sound source vectors for each pitch period within the analysis frame.

Since the human voice signal is reproduced by driving the synthesis filter circuit with the prediction residual signal, it is clear that if this prediction residual signal can be efficiently transmitted, a synthesized sound that is closer to the input voice signal can be obtained. It is. Looking at the predicted residual waveform in the voiced interval, which greatly affects the sound quality of the vocoder, the approximate shape is good (the predicted waveform has a structure that repeats with a pitch period, and one pitch length residual within one analysis frame). If only the signal is cut out and transmitted, and the synthesizer generates the sound source waveform by arranging it for each pinch period, it is possible to significantly reduce the redundancy contained in this prediction residual signal and transmit it. Furthermore, by using a method that selects the optimal excitation vector for the prediction residual signal by cutting out the code length vector prepared in advance, efficiency can be improved by adding several bits of transmission information per analysis frame (this excitation information transmission can be realized.

[Embodiments of the invention]

Hereinafter, one embodiment of the present invention will be explained with reference to FIGS. 1 and 2. In FIG. 0, 7 is an inverse filter circuit, 8 is a representative sound source extraction circuit, 9 and 18 are codebook memories, and 10.11
and 22 is a synthesis filter circuit, 12 is a waveform distortion calculation circuit,
13 is a comparison circuit, 14 is a sound source vector number transmission line, 19
is a one frame long sound source generation circuit. 1~6.15~1
7. 20 to 23 are the same as the conventional example, so their explanation will be omitted. 2a is a predicted residual waveform, 2b is a representative sound source waveform, 2C is a selected sound source vector, and 2d is a 1-frame long sound source waveform.

The inverse filter circuit 7 in the analysis unit 1 performs inverse filter processing on the speech waveform input via the input terminal 3 using the envelope parameters extracted by the envelope parameter extraction circuit 4, and calculates the obtained prediction residual. The waveform 2a is output to the representative sound source extraction circuit 8. When the voiced/unvoiced discrimination circuit 5 determines that the voiced/unvoiced segment is a voiced sound section, the representative sound source extraction circuit 8 extracts a residual waveform of one pitch period length from the predicted residual waveform 2a as a representative sound source waveform 2b. The representative sound source waveform 2b is output to the synthesis filter circuit IO. An example of a method for cutting out processing is to cut out a waveform with a length of one pinch cycle from the zero intersection immediately before the part with the maximum amplitude in the residual waveform, or to cut out a waveform with a length of one pinch cycle so that the part with the largest amplitude in the predicted residual waveform is near the beginning. The code length memory 9 in the analysis unit 1 can reduce the number of sound source vector groups by cutting out the sound source vectors so that only the sound source vectors that have large amplitude parts near the beginning are stored. The stored finite number of sound source vectors are sequentially output to the synthesis filter circuit 11.

The synthesis filter circuits 10 and 11 each generate a synthesized waveform using the input sound source vector and the envelope parameter, and output the synthesized waveform to the waveform distortion calculation circuit 12. The waveform distortion calculation circuit 12 calculates the waveform distortion between the composite waveform obtained using the representative sound source waveform 2a and the composite waveform obtained using each sound source vector stored in the codebook memory 9, and The result is output to the comparison circuit 13. Comparison circuit 1
3 compares the waveform distortions, and outputs the number of the excitation vector giving the minimum waveform distortion to the codebook memo IJ 18 in the synthesis section 2 via the excitation vector number transmission line 14. The codebook memory 18 converts the selected excitation vector 2c specified by the input excitation vector number into a one frame length excitation generator circuit 19.
Output to. The one-frame long sound source generation circuit 19 generates a one-frame long sound source waveform 2d by arranging the selected sound source vectors 2c at pitch period intervals inputted via the pitch data transmission line 17, and sends this to the sound source switching circuit 21. Output. The sound source switching circuit 21 selects the one frame long sound source waveform 2d when the voiced/unvoiced discrimination result inputted via the voiced/unvoiced discrimination result transmission line 16 is a voiced sound,
If the sound is not a voiced sound, white noise is selected and the selected sound source waveform is output to the synthesis filter circuit 22. Then, the synthesis filter circuit 22 generates a synthesized waveform using the envelope parameter input from the envelope parameter extraction circuit 4 in the analysis section 1 and the sound source waveform, and outputs it from the synthesized waveform output terminal 23.

A desired number of finite sound source vectors in the codebook memory 9 are selected from a set of representative sound source waveforms extracted from many input audio signals using the representative sound source extraction circuit 8 in the analysis section 2. Make it easy in advance. The selection method is, for example, to configure a synthesis filter circuit with an average spectral envelope shape of a voiced sound section, and to minimize the distortion in the synthesized waveform obtained by driving this synthesis filter circuit with the set of representative sound source waveforms. A clustering method based on the following criteria can be used.

[Explanation of other embodiments, explanation of examples of diversion to other uses] In the above embodiments, an example was described in which each arithmetic process is realized within a circuit, but this can be implemented using software using a general-purpose arithmetic device such as a signal processing processor. It may be realized by processing, and
As a sound source selection method, we have described a method of selecting a sound source that minimizes the distortion in the synthesis filter output waveform, but this can be done by representing the envelope parameter and the sound source waveform as a frequency spectrum such as DFT, and minimizing the distortion on the frequency axis. It is also possible to select a method.

〔Effect of the invention〕

As described above, according to the present invention, it is possible to transmit voiced sound source information that can efficiently synthesize a high-quality synthesized sound with a relatively small amount of transmitted information, and it is possible to transmit voiced sound source information that can efficiently synthesize high-quality synthesized sounds with a relatively small amount of transmitted information, and it is possible to transmit voiced sound source information that can efficiently synthesize high-quality synthesized sounds with a relatively small amount of transmitted information. The synthesized sound of the vocoder, which was a problem with devices using the vocoder, has been improved and a more natural and individualized synthesized sound can be obtained.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a speech analysis and synthesis device according to an embodiment of the present invention, FIG. 2 is a schematic diagram showing representative sound source waveform extraction and one-frame long sound source waveform generation in the embodiment, and FIG. 1 is a block diagram showing a conventional speech analysis and synthesis device. In the figure, 1 is an analysis section, 2 is a synthesis section, 3 is a speech waveform input terminal, 4 is an envelope parameter extraction circuit, 51 is a voiced/unvoiced discrimination circuit, and 6 is a pitch extraction circuit, 7 is an inverse filter circuit,
8 is a representative sound source extraction circuit, 9 and 18 are codebook memories, 1
0.11 and 22 are synthesis filter circuits, 12 is a waveform distortion calculation circuit, 13 is a comparison circuit, 14 is a sound source vector number transmission line, 15 is an envelope parameter transmission line, 16 is a voiced/unvoiced discrimination result transmission line, and 17 is a pitch a data transmission line; 19 is a 1 frame length sound source generation circuit; 20 is an unvoiced sound source generation circuit; 21
2 is a sound source switching circuit, 23 is a composite waveform output terminal, and 24 is an impulse train equivalent sound 'a generating circuit. 2a is the predicted residual waveform, 2b is the representative sound source waveform, 2c is the selected sound source vector, 2
d is a one frame long sound source waveform. Agent Masuo Oiwa Figure 2

Claims (1)

[Claims] In a speech analysis and synthesis device that compresses the amount of information in a speech waveform, the input speech waveform is inverse filtered in units of analysis frames of a certain length using the spectral envelope parameters of this input speech waveform, and the predicted residual waveform of the analysis frame is obtained. an inverse filter circuit, a representative sound source extraction circuit that extracts one pitch cycle length of the predicted residual waveform obtained by the inverse filter circuit and uses it as a representative sound source waveform of the analysis frame; a codebook memory that stores sound source vectors; a synthesis filter circuit configured using the frequency feature parameters; and a synthesis waveform obtained by driving this synthesis filter circuit with the representative sound source waveform extracted by the representative sound source extraction circuit. The analysis unit includes an excitation selection circuit that selects an optimal excitation vector from within the code width memory so as to minimize distortion between the synthesis filter and a synthesized waveform obtained by driving the synthesis filter with an excitation vector within the codebook memory; a 1-frame long sound source generation circuit that generates a sound source waveform for the analysis frame by repeatedly arranging the optimal sound source vectors determined by the analysis unit for each pitch cycle within the analysis frame;
A speech analysis and synthesis device characterized in that the synthesis section includes a synthesis filter circuit that uses the sound source waveform generated by the one-frame long sound source generation circuit as a driving source and obtains a synthesized waveform using the spectral envelope parameter.