JPH03216699A - Sound source data generating method of sound synthesizer - Google Patents

Abstract

PURPOSE:To reduce rough noises by reanalyzing residue information by setting an analytic window centering on a mean peak position in order, and segmenting the residue information by the reanalysis according to a peak position and generating sound source data. CONSTITUTION:The analytic window centering on the averaged peak position sequence is set in a residue waveform in order to perform resegmentation and reanalysis, and the residue obtained as a result is segmented uniformly according to the previous peak position sequence to generate a sound source file. Consequently, a sound source wave shifts in peak position, pitch by pitch, and the residue waveform has no phase variance in peak position. The residual waveform in each pitch section is averaged axially in an in-phase state and the added mean and weighted mean of the residue waveform B to be processed and other residue waveforms A and C are calculated. Consequently, this method is suitable for the synthesis of a female voice, etc., rough noises in the synthesized voice are reduced, and the synthesized voice which has good quality on the whole is obtained.

Description

[Detailed Description of the Invention] A. BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesis device using a rule synthesis method, and more particularly to a method for generating sound source data.

B. Summary of the Invention The present invention provides a speech synthesis device that uses residual information obtained from a plurality of audio waveforms as sound source information. This process reduces the rough noise in the synthesized speech by emphasizing the fluctuations in the impulsiveness and the impulsiveness of weakly impulsive sections.

C. Conventional speech synthesis devices using technical rule synthesis method divide input character strings into words and phrases by syntactic analysis, and each one has intonation. The accent is determined, words and phrases are broken down into syllables and even phonemes, sound source waves and articulatory filter parameters are determined for each syllable or phoneme, and synthesized speech is obtained as the response output of the articulatory filter to the sound source waves. .

In such a speech synthesis device, there are a method that uses impulses and noise as sound source information, and a method that uses residual information. Among these methods, the method that uses the residual as sound source information calculates articulatory parameters by linear predictive analysis of the speech waveform.
A speech waveform is input to an articulation filter based on these parameters, a residual waveform is obtained from its output, and this residual waveform is used as sound source information by sampling and encoding. Furthermore, when cutting out the audio waveform, the original waveform is multiplied by a window function (Hamming window, Hanning window, etc.) to prevent sudden changes from occurring at both ends of the cutting section.

D. Problems to be Solved by the Invention In a method that uses residual as a sound source, when information compression is applied to sampling and encoding of the residual waveform in order to compress sound source information, the synthesized speech using this sound source information is human speech. There was a problem in which the sound was significantly out of character.

On the other hand, as the original speech waveform for obtaining sound source information, in addition to consonant vowels (CV waveform), vowel ten consonants (VC waveform) and common vowels (V waveform) are mixed in the transition interval, and from this mixed waveform. When obtaining residual information and using it as a sound source, there may be large variations in the amplitude and shape of the sound source for each pitch interval (especially when analyzing female voices), and as a result, the synthesized speech contains rough noise. There was a problem that overall high-quality synthesized speech could not be obtained.

For example, in the residual waveform diagram of a female voice, as shown in FIG. 4, there are variations in the time T0 to T3 from the reference point L0 to t3 for each pitch to the peak (impulse) position, and there is no peak value. -However, there is a large variation from pitch to pitch, and there are also a section 2 where impulses appear strongly and a section 3 which is almost noise (weak impulsiveness). Such variations in peak positions and peak values, as well as the presence of weakly impulsive sections, are considered to be the main causes of roughness and noise in synthesized speech.

An object of the present invention is to provide a method for generating sound source data in which rough noise is reduced in a speech synthesis device that obtains residual information from a plurality of speech waveforms and uses it as a sound source.

E. Means for Solving the Problems In order to achieve the above object, the present invention calculates the cross-correlation between the complete impulse train and the residual information for each pitch section of the residual information obtained by analyzing the speech waveform, and calculates the cross-correlation between the complete impulse train and the residual information. Calculate the peak position for each pitch section from the position where is the maximum, and re-analyze the residual information by sequentially taking an analysis window around the averaged peak position obtained by averaging the peak position sequence, The present invention is characterized in that the residual information resulting from this reanalysis is extracted based on the peak position to generate sound source data.

Further, the present invention is characterized in that the extracted residual information is averaged in the time axis direction by aligning the phases of the residual waveforms for each pitch section.

F. The variation in the peak position of the action residual waveform is determined by calculating the peak position for each pitch interval by cross-correlation with the complete impulse for each pitch interval of the residual waveform, and averaging this peak position by averaging processing. By cutting out and reanalyzing using an analysis window centered on the peak position, a residual waveform with low variation in peak position is obtained.

In addition, by performing averaging processing on the residual waveform with reduced peak position variations by aligning the phase for each pitch section, fluctuations in peak values for each pitch and impulsiveness of sections with weak impulsiveness are emphasized.

G. Embodiment FIG. 1 is a processing procedure diagram showing an embodiment of the method of the present invention. In step Sl, similar to conventional residual information generation, voice feature parameters are obtained from a mixed waveform of a plurality of voice waveforms, and residuals are extracted as sound source information. Step 82
~s7 is the peak (impulse-like) for the residual information
Waveform processing is performed so that the waveform does not vary significantly from frame to frame. For this reason, first, for the residual waveform (& in FIG. 2) obtained in frame i, the reference point t for each pitch section is
A complete impulse train (b in FIG. 2) shifted by a fixed time ΔL from 0 to t4 is prepared, and the cross-correlation between this complete impulse train and the residual waveform is calculated (step S2). From this calculation, the shift width that maximizes the correlation coefficient is set to
+ binding, and this Xt is determined for each pitch section (step S3). This calculated value is (Xt+Δt) for frame i
Let it be the peak position from the reference point t, at .

Peak position sequence (xt+Δ
t) performs averaging processing in the frame direction so that the peak position changes smoothly (step S4). This averaging process aims to reduce variations due to errors in extracting peak positions, for example, in the case of female voice waveforms, in sections where the impulsiveness of the residual is weak or when the variations are very large.

Note that the averaged peak position sequence may not be the actual peak point of the residual cross-section shape, but the residual peak should originally appear at that position (the peak should appear at the pitch interval), and this The variation is reduced by setting the position to the peak position.

Next, the residual waveform is re-cut (step SS) and re-analyzed by sequentially taking an analysis window around the averaged peak position sequence (step S6). The resulting residuals are uniformly cut out based on the previous peak position sequence to create a sound source file (step S7). As a result, the peak position of each pitch of the sound source wave changes smoothly, and there is no phase variation in the peak position in the residual waveform.

Next, the residual waveforms for each pitch section are averaged in the time axis direction with their respective phases aligned (step S8).
. In this averaging process, as shown in FIG. Impulsivity is improved even in pitch sections that are generally weak in impulsiveness, and transitions in impulse amplitude are also smoothed.

H. Effects of the Invention As described above, according to the present invention, it is possible to reduce the dispersion of the peak position in each pitch section of the residual information, and further to smooth out fluctuations in the peak value and emphasize the impulsiveness of the section with weak impulsiveness. Since this method generates sound source data by using the method, it is possible to apply it to synthesis of female voices, etc., to reduce rough noise in the synthesized speech, and to obtain synthesized speech of good quality overall.

[Brief explanation of drawings]

FIG. 1 is a processing procedure diagram showing one embodiment of the method of the present invention, and FIG.
The figure shows a residual waveform and a complete impulse waveform, FIG. 3 shows an averaged waveform of the residual waveform, and FIG. 4 shows a conventional residual waveform. Figure 1 Gali Ryo 1l old edition of Jitsui Seisho 1j

Claims (2)

[Claims] (1) Calculate the cross-correlation between the complete impulse train and the residual information for each pitch section of the residual information obtained by analyzing the audio waveform, and calculate the peak position for each pitch section from the position where the correlation is maximum. Then, the residual information is re-analyzed by sequentially taking an analysis window centered around the averaged peak position obtained by averaging the peak position sequence, and the residual information resulting from this re-analysis is calculated based on the peak position. A method for generating sound source data for a speech synthesizer, characterized in that the sound source data is generated by cutting out the sound source data. (2) A sound source data generation method for a speech synthesizer, characterized in that the extracted residual information is averaged in the time axis direction by aligning the phases of the residual waveforms for each pitch section.