JPH0380300A - Voice synthesizing system - Google Patents

Abstract

PURPOSE:To obviate the deterioration in sound quality even if the pitch period at the time of synthesis is largely changed with respect to the pitch period of a unit voice by determining the change rate of the pitch period in accordance with delay time and rhythm information and synthesizing the voice by changing the pitch period of a residual signal by as much as the change rate of the pitch period. CONSTITUTION:Unit voice connection information is inputted from a terminal 100 and the rhythm information is inputted from a terminal 150. The unit voice connection information is supplied to a sound source signal storage section 200 and a spectral parameter storage section 210 and the rhythm information is supplied to a pitch control section 230 and a a time length control section 240. A pitch change rate calculating section 220 calculates the exact pitch period and delay time with respect to the residual waveforms between the adjacent pitch sections and determines the pitch change rate in the vocal section of the residual signal. The pitch control section 230 changes the pitch period of the residual signal by as much as the pitch change rate in the vocal section of the residual signal by using the pitch change rate. The good synthesized voices which are hardly deteriorated in sound quality when the pitch is changed are obtd. in this way.

Description

Detailed Description of the Invention (Industrial Field of Application) The present invention relates to a speech synthesis method, and in particular, the present invention relates to a speech synthesis method, and in particular, connects unit speech by changing the prosody of residual signals of unit speech according to unit speech connection information and prosody information. This invention relates to a speech synthesis method that obtains synthesized speech by driving a synthesis filter configured based on spectral parameters.

(Prior art) As a rule synthesis method that can obtain relatively good sound quality, a residual signal representing a sound source signal and a spectral parameter representing a spectral envelope are prepared in advance in the entire interval of a unit sound (e.g., CV, VC, etc.). The prosodic information (pitch period, A residual control type speech synthesis method that controls the amplitude and duration (amplitude and duration) to desired values, connects them, and synthesizes speech by passing the connected residual signal through a synthesis filter configured with the corresponding spectral parameters. Are known.

For details of this method, see Japanese Patent Application No. 63-136969.
specification (Reference 1), Japanese Patent Application No. 63-133478 (Reference 2), and “Study of speech synthesis system using residual control” by Hiro Iwata et al. (Acoustical Society of Japan M Theory, 3-2-7.19)
(October 1988) (Reference 3).

According to this method, the residual signal obtained by analyzing the unit speech in advance is used as the sound source signal for the entire section of the unit speech, so the sound source signal is an impulse train in the voiced section, and a noise signal in the unvoiced section. The sound quality of the synthesized speech is much better than the method using .

(Problem to be Solved by the Invention) However, in the residual control type speech synthesis method described above, synthesized speech with good quality can be obtained when the range of changing the pitch period of the residual signal during synthesis is small, but the pitch There was a problem in that the sound quality deteriorated if the cycle was changed significantly.

The pitch period in the residual control type speech synthesis method is changed as follows. That is, in the voiced section, the pitch period of the residual signal obtained from the unit voice is calculated in advance, and the residual signal is divided in advance into pitch sections having a length equal to the pitch period. The pitch change amount of the residual signal is calculated using the desired pitch period and the section length or pitch period, and this is used to change the pitch period of the residual signal by the pitch change amount.

However, when the residual signal is divided into pitch sections in advance, a pitch period extraction error of several percent usually occurs. This is no problem at all when the pitch period of the unit voice is not changed or when the change range is small even if it is changed, but if the pitch period during synthesis is changed greatly compared to the pitch period of the unit voice, a few percent of the pitch will be extracted. As errors accumulate, the pitch period after changing the pitch fluctuates with respect to the pitch period specified by the prosody information. This fluctuation caused deterioration in sound quality.

Furthermore, when dividing the pitch interval, it is difficult to align the peak position of the pitch waveform of the residual signal for each pitch interval. The sound quality deteriorated due to the phase shift of the pitch waveform.

SUMMARY OF THE INVENTION An object of the present invention is to provide a speech synthesis method that does not cause deterioration in sound quality even when the pitch period during synthesis is greatly changed with respect to the pitch period of a unit speech.

<Means for solving the problem> The speech synthesis method according to the present invention stores a residual signal representing the sound source of the entire unit speech and a spectral parameter representing the spectral envelope, and combines the unit speech connection information and the prosody information. In a speech synthesis method that obtains synthesized speech by changing the prosody of the residual signals according to the spectral parameters and driving a synthesis filter configured based on the spectral parameters, the cross-correlation between the residual signals of the adjacent pitch sections is determined. A delay time that gives a maximum value of the function is determined, a pitch period variable amount is determined based on the delay time and the prosody information, and the pitch period of the residual signal is changed by the pitch period variable amount.

Further, the synthesis filter is driven by the residual signal of the adjacent pitch section to obtain a synthesized speech, a delay time that gives a maximum value of a cross-correlation function between the synthesized speech of the adjacent pitch sections is obtained, and the delay time and the A pitch period change amount is determined based on the prosody information, and the pitch period of the residual signal is changed by the pitch period change amount to synthesize speech.

(composition) The operation of the first invention will be explained with reference to FIG.

Here, as unit sounds, for example, Japanese cv, v
It is assumed that 300 to 400 types of c are used.

In the voiced section, the pitch period is extracted from the unit voice in advance,
For pitch extraction, in which a boundary is set for a unit voice for each pitch section whose length is equal to the pitch period, a method obtained from autocorrelation of the voice signal or other known methods can be used. Further, as a method for determining the pitch division boundary, for example, the technique proposed in Japanese Patent Application No. 1983-210690 (Reference 4) is used.

In voiced sections, the unit speech is analyzed for each pitch section to obtain spectral parameters and residual signals representing the spectral envelope of the speech signal.In unvoiced sections, the analysis is performed for each predetermined interval (for example, 5m5). The modified Cepnutrum analysis is used for the 0 analysis as in References 2 and 3, but other known good spectral analysis methods can be used. The residual signal is obtained for the entire section of the unit speech in the same manner as in References 1 and 2.3.

The above processing is performed in advance, and the sound source signal storage unit 200 stores the residual signal for the entire unit voice for each unit voice, and the spectral parameter storage unit 210 stores the residual signal for each pitch interval in the voiced interval. In addition, in the silent section, spectral parameters obtained at predetermined intervals are stored.

Next, unit audio connection information is transferred from the terminal 100 to the terminal 150.
prosodic information (pitch period, phonological duration, amplitude)
Enter. The unit audio connection information is stored in the audio source signal storage section 200.
and spectral parameter storage section 210, and prosody information is supplied to pitch control section 230 and time length control section 240.

The pitch change calculation unit 220 calculates an accurate pitch period T, with respect to the residual waveform of the adjacent pitch section in the voiced section of the residual signal,
Delay time τ11. Calculate and find the amount of pitch change. As shown in Fig. 4, the section length of the divided pitch section ■ is LL, the residual waveform is e+ (n), the section length of the pitch section ■ is L2, and the TA waveform is e2 (n), el ( The cross-correlation function of n) and e2(n) is calculated using the following equation.

Φ (τ)=Σet (n+L+ +τ)e2 (
n) (1) Calculate equation (1) for various delay times, and set the delay time that maximizes equation (1) to τ6.8. At this time, the accurate pitch period T between adjacent pitches can be determined by the following equation.

T=L1+τ,,, (2) Therefore, in order to make the pitch period T of the unit speech to the pitch period T° that is inputted as prosody information and which is desired to be synthesized,
The pitch change scale can be obtained from the following formula.

D=T'-LL-τ,,,=T'-T (3
) The determined pitch change scale is output to the pitch control section 230.

The pitch control unit 230 changes the pitch period of the residual signal by the amount of pitch change in the voiced section of the residual signal using a pitch change measure. Specifically, as in the above document 1.2,
When the pitch period is lengthened by D, O is inserted at the rear of the pitch section by D samples. On the other hand, when the pitch period is shortened by D, the residual signal is truncated by D samples from the rear of the pitch section. Of course, other known methods can also be used to change the pitch period.

The duration control unit 240 uses the duration length of the input prosody information to control the duration length of a phoneme obtained by connecting unit speech. Specifically, reference can be made to the time length control section in Document 1.2.

The synthesis filter 250 inputs the residual signal to which the unit speech is connected and in which the prosodic information such as the pitch period and duration of one hour is controlled, synthesizes the speech using the spectral parameters according to the following equation, and outputs the synthesized speech from the terminal 260. . Note that the spectral parameters here are as follows, considering ease of control:
A modified gepnutrum converted into a linear prediction coefficient a1 is used.

For the conversion from improved gepnutrum to linear prediction coefficients, reference can be made to the above-mentioned document 2, for example.

Next, in the second aspect of the invention, first, synthesized speech is obtained and a pitch change measurement is obtained based on the synthesized speech signal level.

The calculation method is shown below.6 In Figure 4, the synthesized speech in pitch interval ■ is x + (n), and the synthesized speech in pitch interval ■ is x2 (n). These synthesized voices are synthesized by the residual signals of the pitch interval. Once passed through the filter 250, a cross-correlation function is calculated according to the zero-order equation.

Φ (τ) = Σ XI (n+L + τ)
X2 (n) (5) Find τ that maximizes equation (5) as τ, 1 work, and find the accurate pitch period T and pitch change amount from equations (2) and (3).

(Example) FIG. 1 shows a block diagram showing a first embodiment.

The control circuit 510 inputs prosody control information (pitch, duration, amplitude) unit voice connection information from the terminal 500,
Sound source storage circuit 550, spectrum parameter storage port v?
1580, the pitch change amount calculation circuit 555, the amplitude control circuit 570, and the time length control circuit 590.

The sound source storage circuit 550 inputs the connection information of the unit sound,
A prediction residual signal corresponding to the unit speech is output.

The pitch change calculation circuit 555 inputs the pitch period T'' at the time of synthesis from the prosody control information.In addition, the pitch change calculation circuit 555 inputs the pitch period T'' at the time of synthesis from the prosody control information. τ16. is calculated to obtain the pitch change measure.The pitch change measure can be calculated according to formulas (1) to (3> above.The obtained D is calculated by the pitch control circuit 56.
Output to 0.

The pitch control circuit 560 inputs a pitch change measure and changes the pitch period of the residual signal using pitch division positions specified in advance in a voiced section. For this purpose, the method explained in the section of the above-mentioned operation or other known methods can be used.

The duration control circuit 590 receives duration information from the control circuit 510 and controls the duration so that the duration of the phoneme obtained by connecting the unit voices becomes a desired duration. For details, refer to the time length control circuit in Document 1.2.

Next, the amplitude control circuit 570 inputs amplitude control information,
Accordingly, the amplitude of the residual signal is controlled and e (n) is output.

The spectral parameter storage port F#1580 inputs the connection information of the unit voice and outputs the spectral parameter series corresponding to the unit voice. Here, the pc coefficient al obtained by converting the cepstral coefficients will be used as the spectral parameter in the same manner as in the above-mentioned action section, but other known parameters may be used.

The synthesis filter circuit 600 inputs the residual signal whose pitch period has been changed and uses the coefficient &l to generate synthesized speech x according to the following formula.
Calculate (n).

x (n)=e(n) ten Σat −x (ni
) (6) This concludes the description of the embodiment of the first invention.

FIG. 2 is a block diagram showing an example of a solid tumor according to the second invention. In FIG. 2, the components labeled with the same numbers as in FIG. 1 operate in the same manner as in FIG. 1, and therefore their explanations will be omitted here.

In FIG. 2, a pitch change amount calculation circuit 610 once drives a synchronization file using the residual signal in a voiced section of the residual signal output from the sound source storage circuit to obtain a synthesized speech signal. Here, the coefficients of the synthesis filter are read out from the spectrum parameter storage circuit 580 and used.

Then, for the synthesized speech of the adjacent pitch section, the above-mentioned <5)
Delay time τ1 according to the formula. Then, the pitch change amount is determined according to the above equation (3) and outputted to the pitch control circuit 560.

It should be noted that the above-mentioned embodiment is merely one example of the present invention, and various modifications are possible.

In this example, the prediction residual signal obtained by predictive analysis was used as the sound source signal for the entire section of the unit speech.
In order to reduce the amount of calculation and memory, in voiced sections, especially vowel sections, a prediction residual signal of a typical one pitch section may be used repeatedly while controlling the amplitude and pitch.

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

In addition, the accurate pitch period T is not calculated at the time of synthesis as in the above embodiment, but T is calculated in advance based on the above (1), (2), and (5) when analyzing the unit voice. It is also possible to control the pitch by storing the pitch and calculating the pitch change amount based on the equation (3) at the time of synthesis.

Furthermore, when dividing the pitch section, it is also possible to divide the pitch section so that the length of the divided section is equal to the exact pitch period T.

In addition to the actual example method, other spectral parameters such as formant, ARMA, PSE, LSP, PARCOR, mel cepstrum, -membrane cepstrum, mel generalized cepstrum, etc. can be used as the spectral parameters to be stored. can.

Further, although the LPCl number is stored as a spectral parameter in the spectral parameter storage circuit 580, it is also possible to directly store a cepstrum or an improved cepstrum and perform synthesis using the cepstrum or improved cepstrum.

Furthermore, when the pitch period is changed significantly, the spectral envelope of the synthesized speech may be deformed or distorted compared to the spectral envelope before the pitch period was changed, so a correction for correcting the spectral envelope is provided after the synthesis filter 600. As a specific method of configuring the correction filter, which may be connected with filters, the configuration disclosed in the above-mentioned document 1.2 can be used.

In addition, for each unit voice, the correction spectral parameters of the correction filter are coded according to the amount of change in pitch.
ook, or the changes in the spectral parameters themselves can be stored in advance as a codebook or table, and the optimal change in the spectral parameters can be referenced. In this way, in the former case, the correction In the latter case, calculation of the correction filter becomes unnecessary.

Additionally, amplitude control circuit 570 may be omitted for simplicity.

Furthermore, in this practical example, the configuration is such that the prosody information 1tIIfjRIr1 is input through the terminal 500, but for prosody control, accent information and intonation information are input, and prosody control information is generated according to rules. You can.

(Effects of the Invention) As explained above, according to the present invention, a residual signal and a spectrum parameter are provided for every section of a unit speech, and when changing the pitch period of the residual signal, adjacent pitch Since the pitch period is changed based on the amount of pitch change calculated from the calculation of the cross-correlation function between the residual signals of the intervals or between the synthesized speech signals, there is almost no deterioration in sound quality when changing the pitch compared to conventional methods. This has the great effect of making it possible to obtain good synthesized speech.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the first invention in the present invention, FIG. 2 is a block diagram showing an embodiment of the second invention, FIG. 3 is a block diagram showing the operation of the present invention, and FIG. The figure shows a residual waveform of a pitch section in a voiced section. 200.550...Sound source signal storage circuit, 210°58
0...spectral parameter storage circuit, 220°55
5.610...Pitch change calculation circuit, 230°56
0... Pitch control circuit, 240.590... Time length control circuit, 250,600... Synthesis filter, 570
...Amplitude control circuit.

Claims (2)

[Claims] (1) storing a residual signal representing the sound source of the entire unit voice and a spectral parameter representing the spectral envelope, and connecting the residual signal by changing the prosody of the residual signal according to input unit voice connection information and prosody information; In a speech synthesis method that obtains synthesized speech by driving a synthesis filter configured based on the spectral parameters, a delay time that gives the maximum value of the cross-correlation function between the residual signals of the adjacent pitch sections is determined, and the delay time is and the prosodic information, and synthesizes speech by changing the pitch period of the residual signal by the pitch period change amount. (2) Store a residual signal representing the sound source of the entire unit voice and a spectral parameter representing the spectral envelope, and connect the residual signal by changing the prosody of the residual signal according to input unit voice connection information and prosody information. In a speech synthesis method for obtaining synthesized speech by driving a synthesis filter configured based on spectral parameters, the synthesis filter is driven by a residual signal of an adjacent pitch section to obtain synthesized speech;
Determine the delay time that gives the maximum value of the cross-correlation function between the synthesized speech in the adjacent pitch sections, determine the amount of change in pitch period based on the delay time and the prosody information, and calculate the pitch period of the residual signal as described above. A speech synthesis method characterized by synthesizing speech by changing the amount of change in pitch period.