JPS60260098A - Drive signal generation for vioce synthesization - 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 Field of the Invention The present invention relates to a method for generating a driving signal for speech synthesis to obtain high quality synthesized speech in a method for analyzing and synthesizing speech.

Structure of the conventional example and its problems In the conventional speech analysis and synthesis method, as shown in Fig. 1a and b, a window function of a fixed length, for example, 30m5, is applied to the discrete speech signal.
Spectral parameters (linear prediction coefficients) that express the spectral emotion of speech are obtained from a finite number of data extracted by applying a long Hamming window, etc.

Percoll coefficients or line spectra), the sound source parameters (amplitude, pinch period, and voiced/unvoiced judgment) expressing sound source information are separated and extracted, and the extracted parameters are used to restore the original audio signal. ing.

The spectral parameters define the transfer characteristics of the vocal tract filter, and the source parameters define the driving signal of the vocal tract filter.

A speech signal has a periodic voiced part and a noisy unvoiced part, but the voiced/unvoiced determination parameter is used to switch the excitation function (drive waveform) of the vocal tract filter between voiced and unvoiced sounds. be.

Normally, when synthesizing voiced sounds, a pulse waveform or triangular waveform is used as the vibration function, and when synthesizing unvoiced sounds,
Random pulses are used.

The spectral parameters are determined so that the spectrum of the residual signal obtained by passing the audio signal through the vocal tract inverse filter becomes white. Further, as the sound source parameters, the amplitude is extracted from the residual signal by energy calculation, and the presence or absence of periodicity (voiced/unvoiced determination) and the pinch period are extracted by the autocorrelation method. Therefore, when synthesizing speech, it is sufficient to generate a drive signal corresponding to the residual signal obtained during analysis from the sound source parameters and input it to the vocal tract filter. In this case, a common method is to use a pulse waveform having a uniform vector distribution as a drive signal when synthesizing voiced sounds, and to control the repetition period and amplitude of the pulse waveform. This is because the spectrum of the residual signal is whitened when extracting the spectral parameters.
This is because it is ideal to drive with a signal having a white spectrum even during synthesis.

However, in actual speech analysis, the number of inverse filter stages is about 8 to 10 stages, and the inverse filter model does not necessarily match the speech signal generation model, so the spectrum of the residual signal is not necessarily ideally whitened. It is not something that will be done. Therefore, the residual signal contains spectral information that cannot be expressed by spectral parameters, and one cause of deterioration in the quality of synthesized speech is that this residual signal is replaced by repeated pulses or triangular waves. Conventionally, the spectral information contained in the residual signal is utilized by taking the long-term average power spectrum of the residual signal obtained by inverse filtering the audio signal, and performing an inverse Fourier transform with the phase term set to zero. A time waveform was used as a driving wave. However, in non-musical signals such as speech, it is known that phase information contributes to the timbre of the transitional part from the unvoiced part to the voiced part. This method had problems with the reproducibility of the timbre of one synthesized voice, such as the synthesized voice becoming nasal.

Purpose of the Invention The present invention eliminates the problems of the above-mentioned conventional example,
It is an object of the present invention to provide a method for generating a drive signal for speech synthesis that improves the reproducibility of the timbre of synthesized speech.

SUMMARY OF THE INVENTION In order to achieve the above object, the present invention derives a representative residual signal from the predicted residual signal of the voiced part of an audio signal, connects its autocorrelation function for each pitch period, and calculates the voiced section. This is used as a drive signal and has the effect of improving the reproducibility of the timbre of synthesized speech.

DESCRIPTION OF EMBODIMENTS A method of generating a drive signal according to the present invention will be described below. In the present invention, first, a representative residual signal is derived from the residual signal. The representative residual signal here refers to the length of one analysis section (10 to 20 m) in which the spectral information of the original residual signal is saved.
5) is the time waveform.

Literature (Kushiki, Arai et al., “PAIRCOR-type speech synthesis LS
"Consideration of drive waveform in I" Acoustical Society of Japan Speech Study Group 8
81-40, October 1981), using a 1-pitch-long residual waveform extracted from the frame with the maximum energy in the residual signal of the vowel part as the drive signal produces better quality than conventional pulse drive. It has been shown that speech can be synthesized.

In addition, another document (Okiya, Arai et al. ``L using general-purpose DIP
'SP Speech Synthesizer' Lecture Paper of the Acoustical Society of Japan.

1-7-7, October 1982), the symmetrical waveform obtained by inverse FFT (with the phase term set to O) of the long-term average power spectrum of the residual signal of the voiced part is used as the driving wave. It has been shown that

In the present invention, based on the above example, (1) the power spectrum of one frame in which the energy is maximum in the residual signal of a vowel (for example /a/), or (2) the long-term average of the residual signal of a voiced part The power spectrum is stored in the representative residual signal.

In order to obtain a time waveform (that is, a representative residual signal) that preserves the power spectrum of (1) and (2) above, it is sufficient to apply a phase term and perform inverse I''I''T.

There are two methods of providing a phase term: a method of providing 0 phase and a method of providing a long-term average phase. In the former case, the time waveform is symmetrical and the energy is concentrated in the center, so
It is not suitable for the purpose of obtaining a representative residual signal. In the latter case, when calculating the phase angle for a short period of time (for each moxibustion analysis frame), there is a drawback that it is necessary to determine which quadrant the phase angle is in from the signs of the real and imaginary parts. Therefore, in the present invention, (1)' The phase angle of one frame at which the energy of the residual signal of the vowel part (for example, /a/) is maximum is used.

(2) ′ The residual signal of the voiced part is calculated by F for each analysis frame.
After the FT, the cosine and sine values obtained from the average of the real part over all frames and the average of the imaginary part over all frames are used.

In the case of (1)' above, each frequency component (cosine value C05 (θK) and sine value sin (θK) for each KJ is expressed as follows. However, +"KJbK is the real part and imaginary part of the component. In the case of (2)', the average value over all frames is used instead of the above aK+bK, so aK+')K in equation [1] is used to calculate the cosine value and the sine value. From equation [1], the real component a' and imaginary component b' to be subjected to inverse FFT are as follows.Here, PK is the power spectrum of one frame (maximum energy) of the vowel residual P a I K + or the average power spectrum of the voiced part residual Prr, , K.

In the present invention, since the representative residual signal is derived by performing inverse FFT on equation [3], there is no need to determine which quadrant the phase angle is in, and there is no need to calculate the phase angle itself. Therefore, there is an advantage that inverse trigonometric function operations and trigonometric function operations are not necessary.

Next, a method of deriving the drive signal from the representative residual signal will be explained.

Now, drive signal E(n), r+=o, 1.2. -,
is generated as the output of the linear filter of equation [4] which is excited by an impulse every pitch period.

E (n)−Σβi xn−i ・・−=−[4)1+
0 Here, ■ is the number of samples, and is set to a number that allows B (n) to converge by the next pitch period. For example, 10KHz
When sampling, l-31 for female voices and I-6 for male voices.
It is sufficient to set it to about 3. In the present invention, the autocorrelation coefficient of the representative residual signal derived previously is used as β1 in equation [4]. Therefore, the drive signal of formula [4] has the advantage that the spectral information of the original residual signal is preserved and the pitch period can be freely controlled.

In actual hardware, n = O~(■
This can be realized very easily by storing the waveforms up to -1) in a memory and reading them out repeatedly at every pinch cycle.

FIGS. 2(a), 2(b), and 2(c) show the results of auditory evaluation of the drive signal according to the present invention. As described above, in the present invention, four types of representative residuals can be derived depending on the power spectrum and how the phase is given. In Figure 2, (A1) is the power spectrum of (11) and the phase of (1)', (A2) is the power spectrum of (1) and the phase of (2)', and (vl) is (2)
The power spectrum of and the phase of (1)′, (V2) is (2
) and the phase of (2)′ are shown. 1 weather forecast message (male voice) of about 6 seconds
Sampled at 0KHz, window length 30m5. Speech analyzed and synthesized by LSP using F/L/-mushi7 and 10m5 was evaluated using the 7-step rating scale method. In Figure 2, (a) is an evaluation of richness, (b) is an evaluation of ease of listening, and (C) is an evaluation of whether the synthesized speech sounds human-like and resembles the original speech. , all of them have a one-step higher evaluation than the conventional pulse drive, indicating that the effects of the present invention are significant.

Effects of the Invention The present invention has the above-mentioned configuration and uses cosine and sine values obtained from the average of the real part and the average of the imaginary part shown below, so there is no need to calculate the phase angle itself. This has the effect of simplifying calculations.

(b) Since the drive signal is derived from the representative residual signal that preserves the power spectrum and phase information of the original residual signal, the amount of information is greater than when using a symmetrical waveform with the phase term set to 0 ( 2x) This has the effect of significantly improving the quality of synthesized speech.

FIG. 2 is a diagram showing audible evaluation results of speech synthesized using drive signals according to a drive signal generation method for speech synthesis according to an embodiment of the present invention.

Name of agent: Patent attorney Toshio Nakao and 1 other person No. 1
Figure Hinf 17”!, i' Blind?

Claims (1)

[Claims] The power spectrum of the frame where the power of the vowel residual signal is maximum or the average power spectrum over all frames of the voiced part residual signal, and the phase spectrum (each frequency component) of the frame where the power of the vowel residual signal is maximum (cosine and sine values for each frequency component) or cosine and sine values for each frequency component derived from the average values over all frames of the real and imaginary parts obtained by short-time Fourier transform of the voiced part residual signal. For speech synthesis, the driving signal is generated as an impulse response of an all-zero filter whose coefficient is the autocorrelation coefficient of the obtained time waveform (representative residual signal) by performing inverse Fourier transform using Drive signal generation method.