JPH0690638B2 - Speech analysis method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a voice analysis method for extracting characteristic parameters such as vocal tract parameters, pitch parameters, and amplifier parameters from a voice waveform.

BACKGROUND ART Generally, as shown in FIG. 4, a speech analysis and synthesis system extracts a characteristic parameter by analyzing a waveform of an input original voice Vi, and a speech reproduction code in which the characteristic parameter is encoded. A voice analyzer X that outputs an SC, a code storing and transmitting means Y that stores the voice reproduction code SC in a memory and appropriately reads and transmits it, and synthesizes voice based on the input voice reproduction code. , And a voice synthesizer Z for outputting synthetic voice from the speaker SP. As this speech analysis and synthesis method, there is a method using a linear separation equivalent circuit model that separates the generation of a sound source and the articulation due to the resonance of the vocal tract. Using the residual waveform obtained through the filter with the inverse characteristic of the equivalent filter as the source waveform,
The original voice is reproduced by controlling the vocal tract articulatory equivalent filter formed based on the vocal tract parameters extracted by the vocal tract articulatory equivalent filter. If the residual waveform obtained in the process of extracting the vocal tract parameters by the speech analyzer X is used as the source waveform as it is, the synthesized speech waveform obtained by passing this source waveform through the vocal tract articulatory equivalent filter is the original speech waveform. Is exactly the same as However, when the residual waveform is used as the sound source waveform in this manner, a large amount of voice reproduction information (particularly, a lot of information for reproducing the sound source waveform is required), which complicates the storage and transmission system of the voice reproduction information. In addition, there is a problem that the system becomes large and a practical system cannot be obtained. Therefore, conventionally, in order to compress the audio reproduction information, based on three characteristic parameters for controlling the sound source, a pitch parameter P, a voiced / unvoiced switching parameter U / V, and an amplifier parameter A extracted from the residual waveform. , An approximate waveform of the residual waveform is generated as a sound source waveform. In this case, as compared with the case where the residual waveform is used as the sound source waveform as it is, a large amount of information compression can be performed. Therefore, although the hardware can be easily implemented and the cost can be reduced, a characteristic parameter (pitch parameter P , The approximate waveform generated based on the amplifier parameter A) is used as the sound source waveform, so that the synthesized speech waveform is significantly deteriorated. FIG. 5 shows an example of the speech synthesizer Z. The sound source generation unit 10 is formed by an unvoiced sound generation unit 11, a voiced sound generation unit 12, a sound source switching unit 13, and an amplitude control unit 14 for controlling the sound source. A sound source signal having a waveform set by the characteristic parameters P, V / U, A is generated. The sound source signal output from the sound source generation unit 10 is input to a vocal tract articulatory equivalent filter 15 whose filter characteristic is set by a vocal tract parameter K (for example, partial correlation coefficients k _{1 to} Kn), and a macro frequency characteristic. Is added to the voice spectrum envelope to obtain a synthesized voice signal. Here, in the unvoiced sound generator 11 of the sound source generator 10, an unvoiced sound source signal composed of white noise is generated, and the voiced sound generator 12
In this case, the voiced sound source signal of a predetermined cycle is generated by repeating the waveform for one pitch stored in the memory at intervals set by the pitch parameter P. On the other hand, the sound source switching unit 13 is configured to switch whether to output unvoiced sound or voiced sound according to the voiced / unvoiced switching parameter V / U, and the amplitude control unit 14 sets the strength (that is, amplitude) of the unvoiced sound. The sound source signal is output by controlling based on the amplifier parameter A.

Here, when synthesizing a voiced sound, the waveform for one pitch stored in the voiced sound generation unit 11 is read at a predetermined cycle by the pitch parameter P to generate a voiced sound source signal. , In order to reduce the above-mentioned deterioration as much as possible, each characteristic parameter P, A should be set so that the synthesized speech becomes as close as possible to the original speech in consideration of the sound source waveform generated by the waveform. It may be necessary to determine

By the way, in the conventional speech analysis method, as a method for obtaining the pitch parameter P, a waveform processing method (waveform envelope method, zero crossing method), a correlation processing method (autocorrelation method, modified correlation method, AMDF method), a spectrum processing method. (Cepstrum method, period histogram method) was adopted, but with those methods, it is only trying to find the pitch in the original speech, and the approximate waveform of the residual waveform is used as the sound source. The situation on the side of the speech synthesizer Z used as a waveform is not considered at all, and moreover, the pitch extraction accuracy cannot be said to be sufficiently high. Therefore, when the speech synthesizer Z performs speech synthesis using the characteristic parameters extracted by the conventional method, there is a problem that it is impossible to obtain a synthesized speech faithful to the original speech.

[Object of the Invention] The present invention has been made in view of the above points, and an object thereof is to provide a speech analysis method capable of extracting a characteristic parameter for obtaining a synthesized speech faithful to an original speech. Especially.

DISCLOSURE OF THE INVENTION (Embodiment) FIG. 1 shows a voice analyzer X using the method of the present invention, which is a vocal tract parameter extracting means 1 for extracting a vocal tract simulation filter characteristic and a vocal tract parameter extracting means 1. By using the vocal tract parameters extracted from the original voice by the original voice to the residual waveform extracting means 2 and the residual waveform extracting means 2 each having a filter whose filter characteristic is opposite to the filter characteristic of the vocal tract articulatory equivalent filter. The cross-correlation waveform between the residual waveform extracted by passing the waveform and the waveform for one pitch used when synthesizing the voiced sound is obtained, and the pitch parameter P with the distance between two peaks in the cross-correlation waveform as the pitch period is obtained. An amplifier parameter extracting means for extracting the amplifier parameter A based on the energy comparison between the pitch parameter extracting means 3 and the waveform of the residual waveform and the one-pitch waveform used in speech synthesis. And 4,
And the parameter coding means 5 for coding the characteristic parameters K, P, A extracted by the respective extraction means 1, 3, 4 and outputting as the audio reproduction code SC. Here, the feature of the speech analysis method according to the present invention is that a linearly separated equivalent circuit model that separates generation of a sound source and articulation due to resonance of the vocal tract is used, and is stored in advance when synthesizing a voiced sound.
A waveform of a pitch is used to generate a sound source waveform using a pitch parameter P corresponding to the repetition interval and an amplifier parameter A corresponding to the intensity, and the sound source waveform is set with a filter characteristic by a vocal tract parameter K. Characteristic parameters K to be input to a speech synthesizer Z, which is adapted to obtain a synthesized speech signal by passing through a vocal tract articulation equivalent filter
In a voice analysis method similar to the conventional example in which P and A are extracted from the original voice, the vocal tract parameter K extracted from the original voice is used to obtain the original voice through a filter set to the inverse filter characteristic of the vocal tract articulation equivalent filter. When extracting the pitch parameter P from the residual waveform obtained, a cross-correlation waveform between the residual waveform and a waveform for one pitch used when synthesizing voiced sound is obtained, and the distance between two peaks in the cross-correlation waveform is calculated. The pitch period is used.

The operation of the embodiment will be described below. FIG. 2 is a flow chart showing the operation.
Is input to the vocal tract parameter extracting means 1 for extracting the vocal tract simulation filter characteristic, and the vocal tract parameter K is extracted.
Further, this original voice Vi is applied to the residual waveform extracting means 2 including a filter whose filter characteristic is opposite to that of the vocal tract articulatory equivalent filter by using the vocal tract parameter K extracted by the vocal tract parameter extracting means 1. The residual waveform as input is extracted by inverse filtering as shown in FIG. This residual waveform is input to the pitch parameter extraction means 3 to obtain a cross-correlation waveform between the residual waveform and the one-pitch waveform used when synthesizing the voiced sound as shown in FIG. 3 (b). That is, the cross-correlation waveform is a cross-correlation coefficient string in which time-series cross-correlation coefficients of a one-pitch waveform used for synthesis of voiced sound and a residual waveform obtained from original speech are arranged. A pitch parameter whose pitch period is the interval between two peaks in the correlation waveform is extracted. The pitch parameter P thus extracted is used as the voice synthesizer Z.
The waveform of the voiced sound source generated based on the 1-pitch waveform and the pitch parameter P is a desirable sound source waveform because the 1-pitch waveform used when synthesizing the voiced sound is extracted. The approximate waveform is very close to the waveform, and it is possible to obtain a synthesized voice that is faithful to the original voice.

[Effects of the Invention] As described above, the present invention uses the linear separation equivalent circuit model for separating the generation of the sound source and the articulation due to the resonance of the vocal tract, and the waveform for one pitch stored in advance at the time of synthesizing the voiced sound. And the pitch parameter corresponding to the repetition interval,
A voice in which a synthesized voice signal is obtained by generating a sound source waveform using an amplifier parameter corresponding to intensity and passing the sound source waveform through a vocal tract articulatory equivalent filter whose filter characteristic is set by the vocal tract parameter. In the speech analysis method that extracts each characteristic parameter input to the synthesizer from the original speech, the vocal tract parameters extracted from the original speech are used to obtain the original speech through the filter set to the inverse filter characteristic of the vocal tract articulatory equivalent filter. When a pitch parameter is extracted from the residual waveform, a cross-correlation waveform between the residual waveform and a waveform for one pitch used when synthesizing a voiced sound is obtained, and a distance between two peaks in the cross-correlation waveform is determined by a pitch period. The pitch parameter extracted by the method of the present invention is one pitch used when synthesizing a voiced sound in the speech synthesizer. Since it is extracted in consideration of the waveform, the waveform of the voiced sound source signal generated based on this pitch parameter is an approximate waveform that is very close to the residual waveform which is the desired sound source waveform, and the synthesized speech faithful to the original speech. There is an effect that can be obtained. That is, the amplitude of the sound source input to the vocal tract articulatory equivalent filter at the time of synthesizing the voiced sound is determined by the magnification set by the amplifier parameter based on the waveform for one pitch stored in advance,
In the present invention, the repetition period is determined by the pitch period set by the pitch parameter, and in the present invention, the residual waveform obtained from the waveform of the original voice is associated with the waveform for one pitch that is the source of the synthesized sound source of voiced sound. Since the pitch parameter is determined by the pitch parameter, it is possible to obtain the pitch parameter that matches the characteristics of the sound source at the time of synthesis. In short, if the pitch parameter is determined only by the original voice, an accurate pitch parameter for the original voice can be obtained, but such a pitch parameter does not consider the sound source at the time of synthesis, and this pitch parameter is used. Although synthesizing a voice does not necessarily improve the quality of the synthesized voice, in the present invention, the pitch parameter is determined in association with the waveform for one pitch that is the source of the sound source at the time of synthesis. The effect of improving quality can be brought about.

[Brief description of drawings]

FIG. 1 is a block diagram of a speech analyzer using the method of the present invention,
2 and 3 are diagrams for explaining the operation of the same as above, FIG. 4 is a block diagram of a voice analysis / synthesis system, and FIG. 5 is a block diagram of a voice synthesizer. 1 is vocal tract parameter extraction means, 2 is residual waveform extraction means,
3 is a pitch parameter extracting means, 4 is an amplifier parameter extracting means, and 5 is a parameter coding means.

Claims (1)

[Claims] 1. A linear separation equivalent circuit model for separating generation of a sound source and articulation due to vocal tract resonance is used, and when synthesizing a voiced sound, a waveform for one pitch stored in advance is used to correspond to the repetition interval. A sound source waveform is generated using the pitch parameter and the amplifier parameter corresponding to the intensity, and the synthesized sound signal is obtained by passing the sound source waveform through a vocal tract articulatory equivalent filter whose filter characteristic is set by the vocal tract parameter. In the voice analysis method for extracting each of the above parameters to be input to the speech synthesizer from the original voice, in the filter set to the inverse filter characteristic of the vocal tract articulatory equivalent filter using the vocal tract parameters extracted from the original voice. When extracting pitch parameters from the residual waveform obtained through the original speech, a wave of one pitch used when synthesizing the residual waveform and voiced sound Cross correlation waveform calculated, speech analysis method, characterized in that the distance of the two peaks in the cross correlation waveform and pitch period of the.