JPH01304499A - System and device for speech synthesis - Google Patents

Abstract

PURPOSE:To obtain excellent speech quality by storing a sound source signal and a spectrum parameter as to a unit speech, controlling the rhythm of the sound source signal and synthesizing a speech by using the spectrum parameter, and correcting the spectrum of the synthesized speech by a filter. CONSTITUTION:The signal is analyzed in speech units, the sound source signal is stored in a storage part 110, and the spectrum parameter is stored in a storage part 350. The storage part 110 selects a required unit speed according to input control information from a terminal 100 and generates a corresponding predicted residue signal. A control part 150 uses information for pitch variation in control information, varies the pitch of the residue signal in each pitch section in a vowel section according to a pitch start position, and adjusts the section length in pitch units according to the specification of information. A storage part 350 outputs a corresponding LPC parameter ai and a synthesizing film 200 generates a synthesized speech (x) by using the predicted residue signal and parameter (a). A calculation part 300 calculates a parameter bi for distortion correction from the signal (a) and synthesized speech (x) and a filter 250 makes corrections with the signals (x) and bi to output a speech x' of good quality from a terminal 360.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention stores sound source signals and spectral parameters,
The present invention relates to a speech synthesis method and apparatus for controlling the prosody (pitch, amplitude, two-time length, etc.) of a sound source signal and driving a synthesis filter using the sound source signal to synthesize speech.

[Conventional technology]

As a method of arbitrary word speech synthesis, a part of the prediction residual signal obtained using linear prediction analysis etc. is used as a sound source signal, and this sound source signal drives a synthesis filter composed of linear prediction coefficients to generate speech. A method of synthesizing is known. This method is described, for example, in a paper (Reference 1) by Mr. Sato entitled "(SYMPLE) Speech synthesis based on cvc and sound source elements" (Acoustical Society of Japan Speech Study Group Materials 583-69, 1984). In the method of Reference 1, a prediction residual signal obtained by linear predictive analysis of the original speech is used as a sound source signal in an unvoiced section, and a typical one of the vowel sections is used in a voiced section. The prediction residual signal cut out from the pitch period section is used as the sound source for the voiced sound section to drive the synthesis filter to synthesize speech.In this method, the impulse train is used as the sound source for the voiced sound section.
It is said that the sound quality in unvoiced sections is improved compared to methods that use noise signals.

[Problem to be solved by the invention]

Speech synthesis, especially arbitrary word synthesis, connects unit sounds to synthesize speech, but in order to create natural intonation similar to what humans do when speaking, prosodic information or prosodic rules are used to synthesize speech. It is necessary to change the pitch period of the audio signal or sound source signal. However, in the method of Document 1, when the pitch period of the residual signal, which is the sound source of the voiced section, is changed, the pitch period of the original voice whose synthesis filter coefficients are analyzed differs from the pitch period of the voice to be synthesized. Therefore, mismatching occurs between the changed pitch of the residual signal and the spectral envelope of the synthesis filter, and the spectrum of the synthesized speech is greatly distorted, resulting in major problems such as the synthesized speech being greatly distorted and the clarity significantly reduced. Ta.

Further, this problem was sometimes more severe when the pitch period was greatly changed for a female speaker with a short pitch period.

Regarding this problem, there is a known method to improve it to some extent by shifting the peak position of the formant in the low range of the spectrum envelope so that it matches the position of the pitch frequency when synthesizing. For example, the spectral envelope synthesis method considering the pitch structure by Mr. Osaka et al.
(Ref. 2). However, the method of Document 2 shifts the formant peak position to the position of the changed pitch frequency, so it is not a serious improvement method, but has the problem that clarity and sound quality deteriorate due to the shift of the formant position. A new occurrence had occurred.

Furthermore, in the method of Document 1, in the vowel section, the predicted residual signal of a typical one pitch section of the same vowel section is basically repeatedly used, so the spectrum of the residual signal in the vowel section is It was not possible to sufficiently express temporal changes in phase, and the sound quality deteriorated in vowel intervals.

It is an object of the present invention to not only improve the conventional problems when synthesizing speech by changing the pitch period of a sound source signal and driving a synthesis filter, but also to provide speech synthesis that can obtain good sound quality even in vowel intervals. The objective is to provide a method and its equipment.

[Means to solve the problem]

The speech synthesis method of the present invention stores a sound source signal and spectral parameters for each unit speech, synthesizes speech using the spectral parameters while controlling the prosody of the sound source signal, and corrects the spectrum of the synthesized speech using a filter. It is characterized by

Further, the speech synthesis device of the present invention includes a sound source signal storage circuit that stores a sound source signal for each unit sound, a spectral parameter storage circuit that stores a spectral parameter representing a spectral characteristic for each unit sound, and a prosody of the sound source signal. a prosody control circuit for controlling the prosody control circuit; a synthesis circuit for synthesizing speech using the sound source signal whose prosody has been controlled and the spectral parameters; The present invention is characterized in that it includes a filter circuit that corrects the spectrum of synthesized speech.

[For cattle]

The present invention has a sound source signal for the entire unit voice section, regardless of voice presence or no voice, and also for correcting spectral distortion when voice is synthesized by changing the pitch of the sound source signal. It is characterized by using a correction filter.

FIG. 2 is a block diagram showing the operation of the present invention, in which the sound source signal storage unit 110 analyzes the sound signal for each unit sound (for example, CV, VC, etc.) to obtain a sound source signal, and stores this sound source signal for each unit sound. Remember it. In addition, the spectral parameters (order M+) obtained through analysis are stored in the spectral parameter storage section 350. Here, the explanation will proceed assuming that a well-known linear predictive analysis is used as the analysis method, and a prediction residual signal obtained by the linear predictive analysis is used as the sound source signal. However, as the spectrum parameters and sound source signals, other well-known and good ones can also be used. Also,
In the vowel section of the prediction residual signal, the start position for each pitch is also stored. Various types of linear prediction parameters can be considered as spectral parameters, but here LPC parameters will be used. Besides this, L
SP.

Other well-known parameters can be used, such as P A RCOR, formant, etc. The analysis may be performed using a predetermined fixed length frame (5 or 10 m5), or a pitch synchronized analysis synchronized with the pinch period may be used in the vowel section.

Furthermore, based on the control information input from the terminal 100, the sound source signal storage section 11.0 selects a necessary unit sound and outputs a prediction residual signal corresponding thereto.

The pitch control unit 150 uses information for changing the pitch among the control information to expand or contract the pitch of the residual signal for each pitch section based on the pitch start position in the vowel section. For the specific method, see the above document 1.
As described in , to lengthen the pitch period, fill in zeros at the end of the pitch section, and to shorten the pitch period, truncate the samples from the end of the pitch section. Further, the time length of the vowel section is adjusted on a pitch-by-pitch basis using the time length specified by the control information.

The spectral parameter storage unit 350 stores LPG parameters obtained in advance by linear predictive analysis for each unit voice. Then, according to the control information, select the unit voice and set the corresponding LPC parameter al.
(order Ml) is output.

The synthesis filter 200 has the following transfer characteristic and 5(z)=
−−・・・・・・(1〉1−ΣalZ
-' Output synthesized speech x (n) synthesized using the pitch-changed prediction residual signal and LPG parameters.

The correction spectral parameter storage unit 300 is l-5PC.
Calculate the correction spectral parameter b1 to correct the spectral distortion that occurs in the synthesized speech when the pitch is changed using the parameter al and the synthesized speech Do it like this.

First, the following power spectrum H''(z) is calculated using the LPG parameter cl + 3.

H2(z) = −−− ・・・・・・(2) Next, the LPG is
Analysis is performed and the spectrum parameter a1' (order: N
12) and use this to calculate the following power spectrum F
2 Calculate (z).

F2 (z)=--21111□ ・・・・・・(3) Next, the ratio between equations (1) and (2) is determined as follows.

Then, the autocorrelation function R is obtained by inverse Fourier transform of the formula <3)
(m), and calculate the correction spectral parameter b, (order 3) from R(m) by LPC analysis.
Equations (1) and (2) can be calculated using Fourier transform.

The correction filter 250 has the following transfer characteristic Q(z), ] Q(Z) −−□ ・・・ ・・・
5 > Synthesized speech x (n>) is input, and the synthesized speech x'(n,) whose spectral distortion is corrected using the correction spectral parameter bI is output to the terminal 360.

〔Example〕

Next, the present invention will be explained in detail with reference to FIG.

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention. The control circuit 5]0 inputs prosody control (pitch, duration, amplitude) information and unit voice connection information from the terminal 500,
Sound source storage circuit 550. Spectral parameter storage circuit 5
80. Pitch control circuit 560. Output to amplitude control circuit 570. The sound source storage circuit 550 receives connection information of a unit sound and outputs a prediction residual signal corresponding to the unit sound.

The pitch control circuit 560 receives pitch control information and changes the pitch of the prediction residual signal using pitch division positions specified in advance in the vowel section. As a specific method for changing the pitch, the method explained in the section of the above-mentioned operation or other known methods can be used.

Next, the amplitude control circuit 570 inputs the amplitude control information, controls the amplitude of the prediction residual signal according to the input, and outputs the prediction residual signal e (n). The spectral parameter storage circuit 580 receives connection information of a unit voice and outputs a spectral parameter series corresponding to the unit voice. Here, the LPG coefficient aI is used as the spectrum parameter as in the effect section, but other known parameters may be used.

The synthesis filter circuit 600 has a characteristic expressed by equation (1), and receives the pitch-changed prediction residual signal and calculates synthesized speech x(n) using the LPG coefficient a according to the following equation.

(6) The amplitude control circuit 710 multiplies the synthesized speech x(n) by a gain G and outputs the result. The gain G is input from the gain calculation circuit 700. Note that the operation of the gain calculation circuit 700 will be described later.

The FFT calculation circuit 610 inputs the LPC coefficient a,
FFT with a predetermined number of points (for example, 256 points)
(fast Fourier transform) is performed to calculate and output the power spectrum H2(z) defined by equation (2) above. In addition,
The FFT calculation method is described, for example, in 'Digital Signal Processing' by Oppenheim et al.
ng” (Prentice-Hall, 1975
The explanation is omitted here because it is described in Chapter 6 (Reference 3) of the book titled 2007.

The LPG analysis circuit 640 performs LPC analysis on the vowel section of the synthesized speech x(n), which is advantageous by changing the pitch period.
Calculate the PC coefficient al'. At this time, as mentioned in the section on effects, LPC analysis may be performed in pitch synchronization, or
It may be performed for each fixed length frame section. The FFT calculation circuit 630 inputs the coefficient a+', calculates and outputs the power spectrum F2 (z) determined by the above equation (3).

The correction spectral parameter calculation circuit 620 uses the power spectra H2(z) and F2(z) to calculate the ratio G2(z) according to equation (4). Furthermore, this is inverted F
F T to obtain the autocorrelation function R(m), and LPG analysis to obtain the LPG coefficient b1.

The correction filter 650 uses the coefficient b+ and the amplitude control circuit 7
By inputting the output of 10, the synthesized speech x' (n) with the spectral distortion corrected is calculated according to the following formula.

x′ (n)=G−x(n) In the equation (7), G−x(n) represents the input signal of the correction filter 650.

The gain calculation circuit 700 calculates a gain G to equalize the average power for each pitch of the synthesized speech x (n) and x' (n) in the interval where the pitch is changed. This is because the gain of the correction filter 650 is not 1.

Specifically, in the interval where the pitch is changed, the average power of the synthesized speech x(n) and x'(n>) is calculated for each pitch according to the following formula.

P2-1/'N・Σx'" (n) --
--(8b)n+1 Here, N indicates the number of samples in the pitch section. Then, obtain the gain G from the following formula.

G = 5 shoulders - stone T7 ・・・・・・
(9) The final synthesized speech signal x' (n) multiplied by this gain G is outputted through the terminal 660.

The above embodiment is only one configuration example of the present invention,
Various variations are also possible.

That is, in this example, the prediction residual signal obtained by linear predictive analysis was used as the sound source signal for the entire section of the unit speech, but in order to reduce the amount of calculation and memory, the voiced section, especially the vowel section Then, the prediction residual signal of a representative one pitch section may be used repeatedly while controlling the vibration and pitch 3.

Furthermore, as the sound source signal, not only the prediction residual signal obtained by linear prediction analysis but also other good sound source signals such as a zero-phase signal, a 9-phase equalized signal, a multi-pulse sound source, etc. can be used.

In addition to LPC, other good spectral parameters such as L S P ,
Formant, cepstrum, etc. can be used.

In addition, the spectral parameters of the correction filter are also LPC
Besides, other good parameters, such as L S r-'
, pormant, cepstrum, etc. can be used.

Further, as the configuration of the correction filter, a fully horizontal filter as shown in equation (5) is used, but a configuration using a polar-cloud filter or an FIR filter may be used. However, if this is done, the amount of calculation will increase considerably.

Further, in order to reduce the amount of calculation, the amplitude control circuit 710 and the gain calculation circuit 700 can be omitted. However, if this is done, there is a possibility that the level of the synthesized speech x' (n) will change somewhat.

Further, the amplitude control circuit 570 does not control the power of the residual signal, but the gain calculation circuit 700. Amplitude control circuit 71
The configuration may be the same as that of 0, and the power of the synthesized speech x (n) may be controlled. However, in this case, the control circuit 5
The control signal input from 10 needs to be a unit power for each pitch of the synthesized speech, not a unit power for each pitch of the residual signal.

Further, in this embodiment, the prosody control information is input through the terminal 500, but for prosody control, accent information and intonation information may be input and the prosody control information may be generated according to rules.

Further, in order to reduce the amount of calculation, the correction Lylter calculation may be performed in the pitch control circuit 560 only when the pitch change (H) is large.

〔Effect of the invention〕

As explained above, according to the present invention, the sound source signal and spectral parameters are provided for all sections of unit speech, and these are used to synthesize speech, so that not only the consonant section but also the conventional sound quality A great effect can be obtained in that a synthesized sound with good sound quality can be obtained even in vowel sections that have deteriorated.

Furthermore, according to the present invention, even if the pitch period of the sound source (No. B is analyzed and stored in advance) and synthesized by changing it greatly compared to the pitch period of the sound source signal, the resulting spectral distortion can be corrected. Therefore, it is possible to synthesize speech with almost no deterioration in sound quality.This effect is particularly noticeable for female speakers with short pitch periods.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention, and FIG. 2 is a block diagram showing the operation of the present invention. 110... Sound source signal storage section, 150... Pitch control section, 200.600... Synthesis filter, 250, 65
0...Filter for correction, 300...Spec for correction 1
Hell parameter calculation section, 350... Spec I/Le parameter storage section, 510... Control circuit, 550... Sound source storage circuit, 560... Pitch control circuit, 570°7
DESCRIPTION OF SYMBOLS 10... Amplitude control circuit, 580... Spectrum parameter storage circuit, 610, 630... FFT calculation circuit, 620... Spectral parameter calculation circuit for correction,
640...LPC analysis circuit, 700...gain calculation circuit.

Claims (2)

[Claims] (1) A sound source signal and a spectral parameter are stored for each unit speech, a speech is synthesized using the spectral parameter while controlling the prosody of the sound source signal, and the spectrum of the synthesized speech is corrected by a filter. Speech synthesis method. (2) a sound source signal storage circuit that stores a sound source signal for each unit sound; a spectral parameter storage circuit that stores spectral parameters representing spectral characteristics for each unit sound; and a prosody control circuit that controls the prosody of the sound source signal; , a synthesis circuit that synthesizes speech using the prosody-controlled sound source signal and the spectral parameters, and a filter circuit that corrects the spectrum of the synthesized speech using the spectral parameters and spectral parameters obtained from the synthesized speech. A speech synthesis device characterized by having: