JPS597996A - Phoneme piece editing pattern voice synthesizer - 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 The present invention relates to a speech analysis and synthesis method, particularly to a phoneme segment editing type speech analysis and synthesis method.

The phoneme segment editing speech analysis and synthesis method extracts representative phoneme data from the original speech signal in units of pinch cycles based on the strong similarity between adjacent waveforms of speech, especially voiced sounds, and extracts the extracted phoneme segment data from the original speech signal. This is a method of synthesizing a desired audio signal by sequentially connecting multiple times while repeating the audio synthesis control information.

FIG. 1 shows a part of the speech signal waveform synthesized by the phoneme segment editing speech analysis and synthesis method. Figure 1 shows the phoneme P.
Repeat Hm three times, then connect the phoneme piece PHB, and
It shows an audio signal obtained by repeating HB twice.

The phoneme segment editing type speech analysis and synthesis method synthesizes a speech signal by sequentially connecting phoneme segment data according to speech synthesis control information, so it is a PARCOR method.

Compared to parameter analysis and synthesis methods such as the LSP method and the formant synthesis method, this method has the feature that the synthesis procedure allows voice synthesis to be easily realized using a general-purpose microprocessor or the like.

However, in this method, as shown in Figure 1, the phoneme waveform and the pinch period change rapidly at the connection points of different phoneme pieces, so synchronous noise sounds due to the repetition of phoneme pieces occur, and smooth There was a problem that it was difficult to obtain an audio signal.

In order to improve such problems, it has been proposed in the past to insert an interpolated phoneme segment obtained by interpolation calculation between two phoneme segments.

That is, the i-th data value of the preceding phoneme piece PHA of the phoneme piece data details obtained by sampling the audio signal at a constant sampling period is PHA(i).
(i=1.2,..., NA, where NA is PHA
), and the first data value of the subsequent phoneme piece PHB is PHB(i) (1==-+.

2,..., NB, where NB is the number of PHB data)
, the preceding phoneme PHA and the following phoneme PH
The first data value PHI (
i) is obtained from equation (1).

PHI (i) = f (PH input (i), P HB (
i) l ・・・・・・・・・(1) However, f
(A, B) indicates the interpolation function of A and B.

Here, the interpolation of two phoneme pieces is determined by linear interpolation, and if the number of interpolation phoneme pieces to be inserted between the two phoneme pieces is M, then the number of interpolation phoneme pieces to be inserted between the two phoneme pieces is i-th data value PHI(i.

h) is obtained from equation (2).

The data value P HB (i) of the following phoneme is (2)
By setting h=M-)-1 in the equation, it can be more easily determined, so PHB will be referred to as an interpolated phoneme in a broad sense. Furthermore, M' defined by equation (3) will be referred to as interpolation repetition. Using M', equation (2) can be expressed as equation (4).

M'=M+1 ・・・・・・・・・
-(3) However, h: 1, 2...1M'.

Next, in equation (4), phoneme pieces PHA(i) and P H
The equation expressed by Fourier transforming B(i) is defined as equation (6).

However, Ckk and OBk are phoneme pieces PHA and PHB
This is the frequency spectrum of the harmonics.

In the equal sample number phoneme segment editing type and speech synthesis method, the number of phoneme segment data of phoneme segments to be interpolated is first set equal to a predetermined data number N. In principle, in order to equalize the number of data for phoneme segments with different pinch periods, it is possible to sample phoneme segments while varying the clock cycle when sampling the phoneme segments so that the number of data for the phoneme segment remains constant. .

However, in practice, it is extremely difficult to change the sampling clock period of a phoneme segment according to the pitch period, so after sampling a phoneme piece with a constant sampling clock period τ, for example, PROCEEDIN
GS OF THE IEEE Magazine Volume 69 No. 3
'; 3 (March 1981), pages 300 to 331
,E,CROCHIERE and L,R,RA B I
“INTERPOLATIO” written by NER
NANDDECIMATION OF DIGITA
L 5IeNALS-ATUTORIAL REVI
The number of data is increased or decreased by interpolating or thinning out the takes using the method described in detail in the paper titled EWl'ttl' to reach a predetermined number of data.

At this time, in equation (5), the number of data N of phoneme piece PHA
Since the data number NB of the phoneme segment PHB that follows the sound element PHB is equal to the number of samples N, equation (6) can be rearranged to become the following equation.

However, ΔC*ak=Cak−Cmu.

Therefore, by setting the number of phoneme piece data equal to the predetermined number N of data, in equation (4), (
6) The interpolated phoneme piece PHI(i,
It can be seen that the harmonic spectrum of h) is linearly interpolated.

However, the sangrijig time τJ at this time is obtained by the following equation.

Here, INT(x) indicates an integer conversion of X.

By interpolating the interpolated data PHI (co, h) at the sampling time τj determined from equation (7), linear interpolation i is also performed on the frequency axis.

The present invention provides a speech synthesis method which performs memory compression without impairing synthesized sound quality in the above-mentioned speech synthesis method using phoneme segment editing with an equal number of samples.

The present invention will be explained in detail below.

In natural speech, the fundamental pitch frequency varies over a range of about twice as much, as shown in FIG. At this time, if the equal number of samples is made to match the number of data possessed by the lowest fundamental pitch frequency, the data in the shaded area in FIG. 3 will be wasted. In fact, if the equal number of samples is selected near the average value of the number of data as shown in FIG. 4, the data in the shaded area will be wasted and the data in the lattice area will be partially deleted.

In this way, determining the equal number of samples has a large effect on the synthesized sound quality.

Therefore, in the present invention, the number of equal samples is determined for each small section according to the change in the number of basic pitch period data, and this is shown in FIG. In this way, data compression can be performed without changing the synthesized sound quality by converting the number of samples into conformance.

The present invention has the above configuration, and according to the present invention, data compression can be performed without changing the synthesized sound quality.

[Brief explanation of the drawing]

Figure 1 is a diagram of the synthesized speech waveform synthesized by the phoneme segment editing type speech synthesis method, Figure 2 is a diagram showing the relationship between the phoneme segment number and the number of phoneme segment data of natural speech, and Figures 3 and 4 are respectively FIG. 6 is a diagram showing data that is wasted when the number of samples is equalized. FIG. 6 is a diagram showing data that is wasted when the number of samples is equalized in the phoneme segment editing type speech synthesis method according to an embodiment of the present invention.

Claims (1)

[Claims] The phoneme piece data are connected sequentially according to the speech synthesis control information, and an interpolated phoneme piece obtained by interpolation calculation is inserted between two phoneme pieces, and the phoneme piece data is connected sequentially according to the basic pitch frequency of the voice or the frequency characteristics of the phoneme piece data. A phoneme segment editing type speech synthesis method characterized by determining a constant number of samples corresponding to each section.