JPH0782360B2 - Speech analysis and synthesis method - Google Patents

Description

The present invention relates to a voice analysis / synthesis method of synthesizing a voice signal by driving a linear filter representing a voice spectrum envelope characteristic with a sound source signal.

"Prior Art" As a prior art related to the present invention, there are a linear predictive vocoder and multi-pulse predictive coding. The linear prediction vocoder has been widely used as a speech coding method in a low bit rate region of 4.8 kb / s or less, and there are methods such as the Percoll method and the line spectrum pair (LSP) method. Details of these methods are described in, for example, "Basics of Speech Information Processing" by Saito and Nakata (published by Ohmsha). The linear predictive vocoder is composed of an all-pole filter that represents the spectral envelope characteristic of speech and a sound source signal generator that drives the filter. As the driving sound source signal, a pitch period pulse train is used for voiced sound and white noise is used for unvoiced sound. The sound source parameters are voiced / unvoiced discrimination, the pitch period, and the amplitude of the sound source signal, and these parameters are extracted as average features of the speech signal in the analysis period of about 30 milliseconds. The linear predictive vocoder synthesizes a voice by temporally interpolating the feature parameters of the voice extracted in such a fixed analysis interval in this way, and thus when the pitch period, amplitude, and spectral characteristics of the voice change rapidly. Cannot reproduce the characteristics of the voice waveform with sufficient accuracy. Furthermore, it is difficult to obtain highly natural synthetic speech because the driving sound source signal consisting of a periodic pulse train and white noise is not sufficient to reproduce the characteristics of various speech waveforms. As described above, in order to improve the quality of synthesized speech in the linear predictive vocoder, a driving sound source that can more reproduce the characteristics of the speech waveform has been required.

On the other hand, multi-pulse predictive coding is a method that uses a driving sound source with higher reproduction capability than the conventional vocoder (Patent
1234567). In this method, a driving sound source signal is expressed by a plurality of pulses, and two all-pole filters representing the proximity correlation and pitch correlation characteristics of the sound are driven to synthesize the sound. The temporal position and amplitude of the pulse are determined so as to minimize the error between the input speech waveform and the synthesized speech waveform. For details, refer to the literature (BSAtal, “A New model of LP
C excitation for producing natural-sounding speech
at low bit rates ", IEEE Int.Conf.on ASSP, pp614-61
7, 1982). In multi-pulse predictive coding, speech quality can be improved by increasing the number of pulses, but conversely, when the bit rate is low, the number of pulses is limited and the reproduction accuracy of speech waveform deteriorates. Voice quality cannot be obtained. The amount of information of about 8 kb / s was required to obtain good voice quality.

In the multi-pulse predictive coding, the driving sound source is determined so as to reproduce the input speech waveform itself, while as shown in the example of Japanese Patent Application No. 59-53757 "Speech Signal Processing Method", the speech waveform is There has been proposed a method for performing multi-pulse predictive coding on a phase-equalized speech signal after equalizing the phase component of P to a constant phase. In this method, the driving sound source signal is reproduced with a smaller number of pulses by removing the phase component of the auditory insensitive speech from the speech waveform, so that the speech quality at a low bit rate can be improved. However, even with this method, when the bit rate was reduced to about 4.8 kb / s, the number of pulses was insufficient and the characteristics of the speech waveform could not be sufficiently reproduced, so that high quality speech could not be obtained.

An object of the present invention is to provide a high-quality speech analysis / synthesis method in a boundary region (2.4-4.8 kb / s) between a linear prediction vocoder and waveform coding.

"Means for Solving the Problem" The present invention relates to a speech prediction residual obtained by phase-equalizing a driving sound source signal for a voiced sound used for speech analysis and synthesis with a quasi-periodic pulse train in which the magnitude of pitch period fluctuation is limited. A zero-type filter that characterizes the difference, and a parameter that constitutes the sound source signal, that is, the error between the sound waveform synthesized by this driving sound source signal and the phase-equalized input sound waveform is minimized, that is, It is characterized by determining the temporal position of the pulse, the amplitude and the coefficient of the zero filter. In the conventional vocoder, the periodic pulse train generated from the pitch period and the amplitude obtained for each constant analysis section is used as the driving sound source signal, whereas in the present invention, the position and the amplitude of the pulse are determined for each pitch period, Furthermore, the reproducibility of the voice waveform is improved by introducing a new zero filter. Further, in the conventional multi-pulse predictive coding, the driving excitation signal of one pitch period is represented by a plurality of pulses, whereas in the present invention, one pulse per pitch and zero set for each constant analysis section. The driving sound source signal is represented by a shape filter, and the amount of information of the driving sound source signal is reduced. Further, as the evaluation criterion for determining the sound source parameter, the error with the input speech waveform is used in the conventional method, whereas the error with the phase equalized speech waveform is used in the present invention. By using the error evaluation scale for the phase equalized speech waveform, the degree of matching between the speech waveform synthesized from the driving sound source signal used in the present invention and the input speech waveform can be improved. Since the phase equalized speech waveform and the synthesized speech waveform are close to each other, the number of sound source parameters can be reduced by comparing them to determine the sound source parameter. Finally, the difference between the conventional method of combining phase equalization and multi-pulse predictive coding is the method of determining the driving excitation signal and the excitation parameter to be used.

[Embodiment] FIG. 1 shows a configuration of a speech analysis and synthesis method according to the present invention. A sampled digital audio signal s (t) is input from the input terminal 1. In the linear prediction analysis unit 2, after storing N audio signal samples in a data buffer once, a linear prediction analysis is performed on these samples to predict coefficients a _i (i = 1, 2, ..., p) is calculated, and the prediction coefficient a _i is quantized by the quantizer 3. In addition, the prediction residual signal is obtained using an inverse filter that uses the prediction coefficient as a filter coefficient, and the voiced / unvoiced VU of the voice is determined based on the level determination with respect to the maximum value of the autocorrelation coefficient of the prediction residual signal. . Details of these processing methods are described in the above-mentioned book by Saito et al.

The phase equalization analysis unit 4 calculates the coefficient of a phase equalization filter that zero-phases the phase characteristic of voice and a reference time point for phase equalization. FIG. 2 shows a detailed configuration of the phase equalization analysis unit 4. The speech signal s (t) is input to the inverse filter 31 to obtain the prediction residual e (t). The prediction residual is supplied to the maximum amplitude position detector 32 and the phase equalization filter 37. The switch 33 is normally set on the output side of the amplitude comparison section 38, and is set on the output side of the maximum amplitude position detection section 32 only when the analysis frame is voiced and the previous analysis frame is unvoiced.
In this case, the amplitude of the prediction residual at the maximum amplitude position detector 32 is detected when t _'0 is maximized, this coefficient of the phase equalizing filter is input to the filter coefficient calculation unit 34 is calculated by the following formula .

After that, the switch 33 is switched to the output side of the amplitude comparison unit 38, and the output of the amplitude comparison unit 38 is input to the filter coefficient calculation unit 34.

In the case where the frame is voiced, the filter coefficient calculation unit 34 calculates the reference time point t _{i according} to the following equation similar to the above equation.

If the frame is unvoiced, it is set as follows.

The output of the filter coefficient calculation unit 34 is supplied to the smoothing unit 35, and the coefficient h ^* (m) of the phase equalization filter is temporally smoothed by using, for example, a primary filter such as the following equation.

_{h t (m) = bh t} -1 (m) + (1-b) h * (m) t i-1 <tt i where the coefficient b is set to a value of about .97. The filter coefficient holding unit 36 holds the smoothed filter coefficient h _t (m) at the value h _ti (m) at each reference time, and the phase equalization filter 37
To control. The prediction residual e (t) is sent to the phase equalization filter 37.
Is input, and the phase equalization prediction residual e _p (t) is output by the following equation.

In the amplitude comparison unit 38, the amplitude level of the phase equalization prediction residual e _p (t) is compared with the threshold value, and when it exceeds the threshold value, that time point is detected as the next reference time point t ′ _i .

As shown in FIG. 1, the filter coefficient h _t (m) obtained by the phase equalization analysis unit 4 controls the phase equalization filter 5. By inputting the audio signal s (t) into the phase equalizing filter 5, the phase equalized audio signal s _p (t) is obtained as its output.

Next, the sound source parameter analysis unit 30 will be described. In this analysis and synthesis method, separate voice sources are used for voiced sound and unvoiced sound, and the switch 17 is switched by the voiced / unvoiced parameter VU. The drive source of voiced sound is composed of a pulse sequence generator 7 and a zero filter 10.

The pulse sequence generator 7 generates a quasi-periodic pulse train as shown in FIG. The quasi-periodic pulse train is represented by parameters of temporal position (pulse position) t _i and amplitude m _i of each pulse. The pulse position is controlled by the pulse position generator 6, and the pulse amplitude is controlled by the pulse amplitude calculator 8. The pulse positions are limited so that the position spacing is quasi-periodic. That is, the pulse position interval T _i =
t _i −t _i−1 is limited by the following equation so that the difference between successive pulse position intervals is equal to or less than a certain value, and the sum of the differences in the analysis frame is less than or equal to a certain value.

Condition 1 ΔT _i = | T _i −T _i-1 | J Condition 2 Where n _p is the number of pulses in the analysis frame, J and J
_sum is a constant. The pulse position generation unit 6 generates a sequence of pulse positions satisfying the above-mentioned restrictions based on the reference time point t ′ _i obtained by the phase equalization analysis unit 4. FIG. 4 shows a processing procedure for generating a pulse position sequence from a reference time point. In this process, first, the determination regarding the condition 1 is performed regarding the difference in the position interval obtained from the reference time point, and the condition 1 is determined.
If the condition is not satisfied, the pulse position is inserted, removed, and corrected according to the procedure shown in FIG. As a result, if all the reference time points satisfy the condition 1, the condition 2 is judged, and the condition 2
When the condition is satisfied, the reference time point is set as the pulse position. If the condition 2 is not satisfied, all pulse positions that satisfy the condition 2 in the vicinity of the reference time point are generated as candidates. Also, condition 1
If is not satisfied, the number of reference time points is compared with the maximum possible number N _P, and when the number is less than the maximum pulse number, the reference time point is used as it is as a pulse position. When the number of reference time points is greater than the maximum number of pulses, all combinations of pulse positions having the maximum number of pulses are generated from the reference time points. When there are a plurality of pulse position candidates to be generated, an error between the voice waveform synthesized for each pulse position and the input voice waveform after phase equalization is calculated by the waveform distortion calculation unit 19, and the distortion determination unit 20 Select the pulse position that minimizes the error in.

The pulse amplitude calculator 8 determines the amplitude of each pulse so that the frequency weighted mean square error between the synthesized speech waveform and the input speech waveform after phase equalization is minimized. FIG. 5 shows an internal configuration of the pulse amplitude calculating section 8. The input speech waveform s _p (t) of the phase equalized signal is supplied to the frequency weighting filter 39, and this filter 39 has a function of suppressing a strong frequency component of the speech spectrum, and its transfer characteristic is expressed as follows. .

However, Where a _i is the linear prediction coefficient and z ⁻¹ represents the sampling delay. γ is a parameter that controls the degree of suppression, and is 0
The value is in the range of <γ1, and the smaller the value, the greater the degree of suppression. Values of 0.7-0.9 are usually used.
The frequency-weighted filter 39 uses the output signal obtained by passing the phase-equalized speech signal through the frequency-weighted filter as the initial value of the synthesized speech of the immediately preceding analysis frame, and filters 1 / A (γ
The signal s _w (t) is obtained by subtracting the initial value response when z) is driven with zero input. On the other hand, the linear prediction coefficient a _i is supplied to the impulse response calculation unit 40, and the impulse response f (t) of the filter having the transfer characteristic of 1 / A (γz) is calculated. The correlator 41 calculates the mutual covariance ψ (i) between the impulse response f (t−t _i ) and the frequency signal S _w (t) for each pulse position t _i by the following equation.

Further, the correlator 42 calculates the self-covariance φ (i, j) of the impulse response for each set of pulse positions t _i , t _j by the following equation.

The pulse amplitude calculator 43 obtains the pulse amplitude from ψ (t) and φ (i, j) by solving the following simultaneous equations.

The pulse amplitude in FIG. 1 is quantized in the quantizer 9 using, for example, a vector quantization method. In the case of using vector quantization, a vector (amplitude pattern) having pulse amplitude as an element is compared with a plurality of pulse amplitude standard patterns and quantized to a standard pattern in which the distance between patterns is minimized. As the distance measure of the amplitude pattern, the mean square error between the speech waveform synthesized from the pulse amplitude standard pattern without using the zero-shaped film and the input speech waveform after phase equalization is used. The amplitude pattern vector is m = (m ₁ , m ₂ , ..., m _np ) (t
Represents the transposed value of the matrix), and the standard pattern vector is m _ci (i
= 1,2, ..., Nc), the mean square error is expressed by the following equation.

d (m, m _c ) = (m−m _ci ) ^t Φ (m−m _ci ), where Φ is a matrix whose elements are the autocovariance φ (i, j) of the impulse response. At this time, the quantized value of the amplitude pattern is obtained by the following equation as a standard pattern that minimizes the mean square error.

The zero filter 10 is a filter that characterizes the prediction residual waveform after phase equalization, and the filter coefficient is controlled by the zero filter coefficient calculation unit 11. FIG. 6 shows an example of a prediction residual waveform after phase equalization and an impulse response waveform of the zero filter 10 corresponding to it. The prediction residual after phase equalization is impulse-like because the spectrum envelope characteristic is flat and the phase is close to zero phase, and shows large amplitude at each pulse position, and becomes relatively small in other sections. . In addition, the waveform has a waveform that is almost symmetrical with respect to the middle point between the pulse positions and the adjacent pulse positions. As shown in FIG. 6, the amplitude at the intermediate point of the pulse position often has a relatively large amplitude as compared with other sections, and this tendency becomes stronger especially for a voice with a long pitch period. . As shown in FIG. 6, the zero filter 10 has an impulse response that takes values at q time points on the left and right centering on the pulse position and at r time points on the left and right centering on the intermediate point of the pulse position. Is set. At this time, the transfer characteristic of the zero filter 10 is expressed as follows.

The zero-type filter coefficient calculation unit 11 uses the filter coefficient v _k for a given pitch position and pulse amplitude so that the frequency-weighted mean square error between the synthesized speech waveform and the input speech waveform after phase equalization is minimized. Calculate to. FIG. 7 shows the configuration of the filter coefficient calculation unit 11. The frequency weighted filter 44 and the impulse response calculation unit 45 are respectively the fifth
Frequency-weighted filter 39 and impulse response calculator in the figure
It has the same structure as 40. The adder 46 adds the impulse response f (t) according to the following equation.

The correlator 47 calculates the mutual covariance ψ of the signals _sw (t) and u _i (t).
(I) is calculated, and the correlator 48 calculates the signals ui (t) and uj (t).
Compute the autocovariance φ (i, J) with and. The filter coefficient calculation unit 49 calculates the coefficient v _i of the zero filter 10 by solving the following simultaneous equations from ψ (i) and φ (i, J).

The filter coefficient v _i is quantized in the quantizer 12 in FIG. 1 by using, for example, a vector quantization method. When vector quantization is used, a vector (amplitude pattern) having a filter coefficient as an element is compared with a plurality of pulse amplitude standard patterns and quantized to a standard pattern in which the distance between patterns is minimized. Similar to vector quantization of pulse amplitude,
When the mean square error between the synthesized speech waveform and the input speech waveform after phase equalization is used as the distance measure, the quantized value of the filter coefficient is calculated by the following equation.

d (v, v _c ) = (v−v _ci ) ^t Φ (v−v _ci ), where v is a vector whose elements are filter coefficients, and v _ci
Is the standard pattern vector. Further, Φ is a matrix whose elements are the autocovariance Φ (i, j) of the impulse response u _i (t).

In summary, in the voice sound section, the quasi-periodic pulse train determined by the amplitude of the pulse position is passed through the zero-type filter 10 as a driving sound source signal, and the all-pole filter 18 that characterizes the voice spectrum envelope characteristic is driven. To synthesize the voice. As for the sound source parameter, with respect to the pulse amplitude and the coefficient of the zero filter, the optimum value that minimizes the error between the synthesized speech waveform and the input speech waveform after phase equalization is determined for the pulse position. When there are a plurality of pulse position candidates, the above error is obtained for each candidate, and the optimum pulse position that minimizes the error is determined by a full search.

Next, the driving sound source in the unvoiced sound section will be described.
In the unvoiced interval, code-excited predictive coding (Reference Schroede
r et al., “Code excited Iinearprediction (CELP)”, IEEE
Int.Conf.on ASSP, pp937-940, 1985), a random number pattern is used as a driving sound source signal. The random number pattern generation unit 13 in FIG. 1 stores a plurality of patterns in which a plurality of normal random numbers having an average of 0 and a variance of 1 are collected. The random number amplitude calculation unit 15 calculates, for each random number pattern, a gain optimum value that minimizes the error between the synthesized speech waveform and the input speech waveform after phase equalization for the random number pattern, and is quantized by the quantizer 16. The gain is used to control the gain amplifier 14. next,
For each random number pattern, the error between the synthesized voice and the phase equalized voice is found, the optimal random number pattern that minimizes it is found by full search, and the sequence of this random number pattern is passed through the gain amplifier 14 as the driving sound source signal to all poles. Supply to the filter 18.

With the above procedure, the speech signal is linearly predicted, a _i ,
It is represented by the unvoiced parameter VU, the pulse position t _i , the pulse amplitude m _i , the zero-shaped filter coefficient v _i for voiced sound, and the random code pattern (number) c _i and the gain g _{i for} unvoiced sound. These audio parameters are transmitted or stored after being encoded by the encoding unit 21. In the speech synthesis unit, after decoding the speech parameters in the decoding unit 22, in the case of voiced sound, the pulse train generated in the pulse sequence generation unit 23 by the pulse position t _i and the pulse amplitude m _i to the zero filter 24. To generate a driving sound source signal, and in the case of an unvoiced sound, a random number pattern (signal) c _i is used to selectively generate a random number pattern from the random number pattern generation unit 25, and this is passed through an amplifier 26 controlled by a gain g _i to generate an amplitude. One of the two driving sound source signals is selected by the switch 27 which is controlled to generate a driving sound source signal, and which is switched between voiced and unvoiced, and by driving the all-pole filter 28, a synthetic voice is output to the output end 29 thereof. The filter coefficient of the zero filter 24 is v _i
And the filter coefficient of the all-pole filter 28 is controlled by a _i .

Modified Example A drive source is not distinguished by voiced and unvoiced, and a pulse drive source is used in both cases. In this case, the quality of the fricative consonants is slightly deteriorated, but the processing configuration is simple, the processing amount can be reduced, and the hardware scale can be small. Also, since it is not necessary to transmit voiced / unvoiced parameters, the bit rate is reduced by 60 bits per second.

A configuration that does not include the zero filter in the pulse-driven sound source.
With this method, the naturalness of synthesized speech is slightly degraded especially for male voice with a low pitch frequency, but the hardware scale is reduced by eliminating the zero filter, and 600 bits per second required for coding the filter coefficient is used. , The bit rate is reduced.

A configuration in which the processes of the pulse amplitude calculation unit 8 and the vector quantization unit 9 are integrated to calculate the quantized value of the pulse amplitude. The structure of this method is shown in FIG. Frequency weighted filter 50,
The impulse response calculator 51, the correlator 52, and the correlator 53 have the same configurations as those corresponding to those in FIG. 5 of the first embodiment. In the pulse amplitude quantizing unit 54, for each pulse amplitude standard pattern m _ci (i = 1,2, ..., N _c ) stored in the pattern code book 55, the speech when synthesized using that amplitude standard pattern Calculate the mean square error of the waveform and the input speech waveform after phase equalization,
The pulse amplitude standard pattern with the smallest error is obtained. The distance calculation is performed according to the following equation.

Where Φ is the autocovariance φ of the impulse response f (t)
A matrix having (i, j) as elements, and ψ is the mutual covariance ψ between the impulse response and the output of the frequency weighted filter s _w (t)
(I) A column vector having (i = 1, 2, ..., N _P ) as an element.

The processing amount required to obtain the optimum pulse amplitude is almost the same in FIGS. 8 and 5, but in FIG.
The solution of simultaneous equations included in the processing of the figure is not required, and the processing configuration is simplified. However, in FIG. 5, it is possible to perform scalar quantization after obtaining the optimum value of the pulse amplitude, whereas in FIG. 8, it is assumed that vector quantization is used as the quantization method. Become.

It is also possible to calculate the quantized value of the coefficient by integrating the calculation of the coefficient of the zero filter and the vector quantization in the same manner as in FIG.

"Effect of Invention" In order to investigate the effect of the speech analysis and synthesis method according to the present invention,
An analysis and synthesis speech experiment was conducted under the following conditions. 0-4kHz
After sampling the speech in the band at a sampling frequency of 8 kHz, multiply the speech signal by a Hamming window with an analysis window length of 30 ms, perform a linear prediction analysis by the autocorrelation method with the analysis order as 12th order, and perform 12 prediction coefficients and speech. • Find unvoiced parameters. The analysis frame length for encoding is 15 ms (120 audio samples). The prediction coefficient is quantized using the differential multistage vector quantization method. A frequency weighted cepstrum distance was used as a distance measure in vector quantization. Bit rate 4.8kb / s
, The number of bits per frame is 72 bits,
The breakdown is as follows.

The constants J and J _sum , which represent the allowable range of fluctuations in the pulse period in the pulse sound source, and the maximum number of pulses N _P when it does not fall within the allowable range are determined by the number of bits allocated for encoding the pulse position. When the pulse position is encoded with 29 bits / frame, the difference ΔT between adjacent pulse periods is 5 samples or less, and the total sum is 14 samples or less in the frame. Further, the maximum number of pulses is 5 when it is not within the allowable range. As the zero-type filter, a 7th-order (q = r = 1) filter was used. Random pattern vector is 40 samples (5m
s) and is selected from 512 types (9 bits) of patterns. Further, the random number amplitude is quantized by 6 bits including positive and negative signs.

The speech coded under the above conditions has much higher naturalness than the conventional vocoder, and its quality is close to that of the original sound. Moreover, the dependence of the voice quality on the speaker is smaller than that of the conventional vocoder. It was also confirmed that the quality of the coded speech is clearly higher than that of the conventional multi-pulse predictive coding and code-excited predictive coding. The spectral envelope distortion of speech coded at 4.8 kb / s is about 1 dB. The time delay caused by encoding is 45 ms, which is less than or equal to the conventional method in the low bit rate region.

The effect of the present invention is that, by expressing the driving sound source signal for voiced sound as a quasi-periodic pulse train, the reproducibility of the waveform information of the voice is higher than that of the conventional vocoder, and the amount of information is less than that of the conventional multi-pulse predictive coding. It is possible to express the sound source signal. In addition, as a method of estimating the parameters of this driving sound source signal from the input speech, since the error with respect to the speech waveform after phase equalization is used as an evaluation measure, compared with the conventional method that uses the error with respect to the input speech itself, The degree of matching between the voice waveform and the input voice waveform is improved, and the sound source parameter can be estimated more accurately. In addition, the zero-shaped filter has an effect of reproducing minute features of the speech spectrum, which improves the naturalness of the synthesized speech.

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of the analysis and synthesis method according to the present invention, FIG. 2 is a block diagram showing a configuration example of the phase equalization analysis unit 4, and FIG. 3 is an explanatory diagram of a quasi-periodic pulse drive source signal. Four
FIG. 5 is a flow chart of processing for generating a pulse position, FIG. 5 is a block diagram showing a configuration example of the pulse amplitude calculation unit 8, FIG. 6 is an explanatory diagram of a zero filter, and FIG. 7 is a zero filter coefficient calculation unit.
FIG. 8 is a block diagram showing a configuration example of 11, and FIG. 8 is a block diagram showing another configuration example of the pulse amplitude calculation unit 8.

Claims (3)

[Claims] 1. A speech analysis and synthesis system comprising a linear filter representing a speech spectrum envelope characteristic and a sound source signal generating section for driving the linear filter, wherein said sound source signal is quasi-limited to a size of fluctuation of pitch period. Expressed by a periodic pulse train, the parameters that make up the sound source signal are determined so as to minimize the error between the phase-equalized speech waveform and the synthesized speech waveform after the phase of the input speech is pitch-synchronized to zero phase, A voice analysis / synthesis method comprising synthesizing a voice signal by driving a linear filter representing the voice spectrum envelope characteristic with the sound source signal. 2. The sound source signal is used for voiced sound, and for unvoiced sound, a random number sequence selected from a plurality of random number patterns and having its average power set is used as the sound source signal, and for this unvoiced sound. 2. The speech analysis / synthesis method according to claim 1, wherein the parameters forming the sound source signal are determined so as to minimize an error between the phase equalized speech waveform and the synthesized speech waveform. 3. A source signal represented by a quasi-periodic pulse train in which the fluctuation of the pitch period is limited is supplied to the linear filter through a zero-shaped filter which characterizes the fine structure of the speech spectrum, and the zero filter is supplied. 3. The speech analysis / synthesis method according to claim 1, wherein the coefficient of the shape filter is determined so as to minimize an error between the phase equalized speech waveform and the synthesized speech waveform.