JPS5855519B2 - speech synthesizer - Google Patents

Description

[Detailed Description of the Invention] The present invention relates to a speech synthesis device that synthesizes a speech signal using at least two characteristics of the speech signal, namely, the spectral envelope characteristic and the speech fundamental frequency, and particularly relates to a speech synthesis device that synthesizes a speech signal using at least two characteristics of the speech signal, namely, the spectral envelope characteristic and the speech fundamental frequency. The aim is to provide high-quality synthesized speech.

A conventional speech synthesis device of this type will be explained using FIG.

Sound source information such as the amplitude of the sound, the fundamental period, so-called pitch T, and the voiced/unvoiced signal that distinguishes between voiced and unvoiced sounds is stored in the storage device 11 and corresponds to the sound signal to be synthesized by the sound source signal control device 12. A set of sound source signals are read out from the storage device 11 at regular intervals.

The read fundamental period T is sent to the pulse generator 13 as voiced/
The unvoiced signal is supplied to a switching device 14, and the voice amplitude is supplied to a variable gain amplifier 15.

The pulse generator 13 generates a pulse for each input basic period T, and the switching device 15 selects the output of the pulse generator 13 when the input voiced/unvoiced signal indicates a voiced sound, and selects the output of the pulse generator 13 when the input voiced/unvoiced signal indicates a voiced sound. Select the output of the white line sound generator 16 and apply it to the amplifier 1.
Supply to 5.

Amplifier 15 amplifies the output of switching device 14 according to its amplitude.

The output is the digital filter's speech synthesis filter 17.
is input.

On the other hand, parameters representing the spectral envelope of the audio signal are stored in the storage device 18.

As this parameter, for example, PARCOR coefficient (k
i)"j=1 (References, Japanese Acoustics Complete Edition, Acoustic Engineering Course 7 "Speech" by Kazuo Nakata, pp. 93-97) can be used.

This parameter is sent to the parameter control device 1 at regular intervals.
9 and sent to the filter 17, and the characteristics of the speech synthesis filter 17 are controlled according to this parameter.

The output of the filter 17 is a digital to analog converter 21.
The signal is converted into an analog signal and sent out from the output terminal 22 as a synthesized audio signal.

In such a conventional speech synthesis device, the sound source signal control device 1
2 and the parameter control device 19 operated independently.

This causes a mismatch between the sound source signal and the spectral envelope,
In some cases, the intended spectrum envelope could not be achieved.

That is, FIG. 2A shows the spectral envelope of speech described by, for example, 10 PARCOR coefficients (kl) 14 =1 and the corresponding Fourier spectrum.

This spectral envelope is determined to optimally approximate the Fourier spectrum represented by the frequency structure of audio (References, Journal of the Institute of Electronics and Communication Engineers Vol. 53-A,
A;, 1.1970, Saito, Saito "Estimation of speech spectral density and formant frequency using statistical methods").

Therefore, the digital filter 1 for reproducing the spectral envelope characteristic is used to reproduce the sound source signal having a fundamental frequency that matches this frequency structure.
7, the sound source signal and the spectral envelope match, and the spectral envelope of the resulting synthesized speech also becomes close to the spectral envelope of the original speech.

These problems do not arise when the sound source information and spectral envelope information are extracted by analyzing the sound, and these are recombined to reproduce the original sound.

However, when changing only the sound source information independently, for example, the first
As shown in the figure, the sound source information of the storage device 11 is stored in the storage device 1.
If the spectral envelope information is generated independently from the spectral envelope information of No. 8, it will be a big problem.

The spectral envelope parameters corresponding to the spectral envelope shown in FIG. 2A are given to the digital filter 17,
This filter 17 has a harmonic structure shown in FIG. 2A, that is, a fundamental frequency F.

= 350 Hz and a different harmonic structure, e.g. fundamental frequency F.

When a sound source signal of =230 Hz is added, the spectral envelope and Fourier spectrum of the synthesized speech are as shown in FIG. 2B.

As is clear from comparing these Figures 2 A and B,
In FIG. 2B, the clear peak 23 in the low frequency region that existed in FIG. 2A is lost, and the intended spectral envelope is not obtained.

Therefore, synthesized speech having a spectral envelope as shown in FIG. 2B has a large spectral distortion and is of degraded quality.

Therefore, if you try to synthesize a voice signal that is different from the fundamental period (pitch) of the voice signal, for example, higher or lower, using parameters that indicate the spectral envelope characteristics obtained from an actual voice signal, the quality of the synthesized voice will be quite low. It has become degraded.

The present invention provides a speech synthesis device that eliminates the quality deterioration of synthesized speech caused by mismatch between the sound source signal and the spectral envelope, and makes it possible to realize a spectral envelope that is audibly close to the intended spectral envelope. be.

According to the present invention, the parameters representing the corresponding spectral envelope characteristics are converted so that the given audio fundamental frequency and the peak frequency of the spectral envelope of the audio signal to be obtained, for example, the 17th ormant frequency, have a correlation.

FIG. 3 shows an example of a speech synthesis device according to the present invention, and parts corresponding to those in FIG. 1 are given the same reference numerals.

In this invention, a matching device 24 is provided, and the fundamental frequency T is transmitted to the matching device 24 from the control device 12.
Parameters are supplied from 9 respectively.

This matching device 24 adjusts the spectral envelope described by the parameters to the fundamental frequency F.

The parameters are converted to match the frequency structure of (=1/T).

That is, the peak frequency of the spectral envelope of the audio signal to be obtained and the fundamental frequency F from the control device 12.

The parameters from the control device 19 are converted so that they have a correlation with -1/T.

For example, the parameters stored in the storage device 18 are PA
When described by the RCOR coefficient (ki)i=, this P
From the ARCOR coefficients, a low frequency of the spectral envelope of the original speech, generally a peak frequency of 500 Hz or less, usually the frequency F1 of the first formant, and its bandwidth B1 are extracted.

This extraction can be performed by a known method.

Fl−nF'o (n is a positive integer) is calculated, and if the difference output is less than or equal to B1/2, the storage device 18
Use the parameters read from .

However, if FlnFo is B1/2 or more, the first
Considering the frequency F1' which is the formant frequency F plus a slight frequency ΔF, this l Fl' nFo 1 is B1
Set the one in which ΔF1 is the smallest among F1′ that is smaller than or equal to /2.

Let this F□′ be the first formant frequency, and let the other formant frequencies be S as they are, and from these, PARCO
Calculate R coefficient (ki'), =1.

This calculation can be performed using a known method.

This calculated PARCOR coefficient is applied to the speech synthesis filter 1.
7 as a parameter to control the characteristics.

Such parameter conversion can be performed not only on PARCOR coefficients but also on linear prediction coefficients derived by linear prediction analysis of speech, logarithmic cross-sectional area ratios, and even line spectrum parameters.

Since the tabular line spectrum parameter is itself a pseudo-formant, its first-order parameter and the given fundamental frequency F.

What is necessary is to find a value corresponding to the above-mentioned F1' from , and use this as the first-order parameter of the line spectrum parameter.

For example, if the spectral envelope and Fourier spectrum of the original speech signal are shown in FIG. 2A, and its first formant frequency is F1 = 350 Hz and its bandwidth B1 is 20 Hz, then the fundamental frequency F of the synthesized speech.

When 230 Hz is given, the spectral envelope of the synthesized speech obtained with the conventional device shown in Fig. 1 is shown in Fig. 2B.
As shown in the figure, it contains large spectral distortion.

However, in this invention, the matching device 24 allows F1'=
2 F.

-(Bl/2)=460H2 is obtained, and the parameters are converted so that F1/ of is the first formant frequency. The filter 17 is controlled by this parameter, and the spectral envelope of the synthesized speech obtained at that time is the fourth formant frequency. The result will be as shown in the figure.

In FIG. 4, a spectral peak 25, ie, the first formant, occurs at a low frequency, and the envelope is close to that shown in FIG. 2A, but the frequency of the first formant is different.

Therefore, it is possible to obtain audio of good quality and at the desired pitch (height).

As explained above, when the speech synthesis device according to the present invention is used to synthesize speech, the given fundamental frequency F.

In order to transform the spectral envelope so that it depends on the first
It is possible to obtain high quality synthesized speech having clear spectral peaks without losing the spectral peaks typified by formant F.

Due to this modification of the spectral envelope, the first formant frequency F1' of the synthesized speech becomes a frequency different from the first formant frequency F1 of the original speech of the given parameters, but the formant frequency changes depending on the fundamental frequency. This is clear from the observation results of natural speech shown in FIGS. 5 and 6, and in this respect, the synthesized speech obtained by the speech synthesizer of the present invention has a relationship between formant frequency and fundamental frequency that is close to that of natural speech. can be kept.

Figure 5 shows the formant distribution of Japanese vowels, and also shows the formant distribution of Japanese vowels.
The figure shows the formant distribution of English vowels, and in both cases, when the fundamental frequency changes, the four-formant frequency also changes.

As described above, the synthesized speech obtained by the apparatus of the present invention is superior to conventional synthesized speech in both its naturalness and clarity, and the quality improvement effect of the synthesized speech is also excellent.

[Brief explanation of the drawing]

Figure 1 is a block diagram showing a conventional speech synthesis device, Figure 2A is a curve diagram showing the spectral envelope and Fourier spectrum obtained by analyzing natural speech, and Figure 2B is a curve diagram showing the spectral envelope and Fourier spectrum obtained by analyzing natural speech. Fig. 2 is a curve diagram showing the spectral envelope and Fourier spectrum of synthesized speech synthesized using the same spectral envelope as in Fig. 2A; Fig. 3 is a block diagram showing an embodiment of the speech synthesis device according to the present invention; Fig. 4; is a curve diagram showing the spectral envelope and Fourier spectrum of synthesized speech obtained by the apparatus of the present invention, and FIGS. 5 and 6 are diagrams showing the relationship between fundamental frequency and formant in natural speech, respectively. 11: Storage device storing sound source information, 12: Sound source information control device, 13: Pulse generator, 14: Switching device, 15:
variable gain amplifier, 16: white line sound generator, 17: speech synthesis filter, 18 storage device storing parameters representing varnish vector envelope, 19: parameter control device, 21:
DA converter, 22: output terminal, 24: matching device.

Claims (1)

[Claims] 1. Setting the filter characteristics of the voice synthesis filter using a parameter representing the spectral envelope characteristic of the voice signal, driving the voice synthesis filter with a pulse signal of a frequency equal to the given voice fundamental frequency, and obtaining the voice signal from the filter output. In the speech synthesis device, means is provided for converting a parameter representing the corresponding spectral envelope characteristic so that the given speech fundamental frequency and the peak frequency of the spectral envelope of the speech signal to be obtained have a correlation. Characteristic speech synthesizer.