JPH04346400A - Voice analysis/synthesis method - Google Patents

Abstract

PURPOSE:To provide a voice analysis/synthesis method which can analyze/ synthesize voices having a high voice quality at a low bit rate and has excellent ambient noise resistance. CONSTITUTION:A quasi-periodic pulse train, on which fluctuation size of a pitch period of an input voice is controlled, is created by means of a pulse series creating unit 10, and a noise code selected from a noise code book 12 is formed as a periodic noise series through a pitch correlation filter 13, and a result obtained by adding these quasi-periodic pulse train and noise series together is used as a sound source signal, and a voice signal (s'(t)) is synthesized by driving a total pole type filter 19 having voice spectrum envelope characteristics. Respective amplitudes of the quasi-periodic pulse train and the noise series and patterns of the noise series are determined, so that errors between a voice signal (sp(t)) obtained by transforming the input voice pitch- periodically into a zero phase by means of a phase equalization filter 4 and the synthesized voice signal (s'(t)) become minimal.

Description

[Detailed description of the invention]

0001

[Industrial Application Field] This invention relates to high-efficiency speech coding for providing high-quality speech with a small amount of information, in particular,
The present invention relates to a speech analysis and synthesis method that realizes high-quality speech encoding at a bit rate of 2.4 to 4.8 kb/s, which is the boundary between a conventional speech analysis and synthesis system called a vocoder and waveform encoding.

0002

[Prior Art] As prior art related to the present invention, a linear predictive vocoder and a code excitation predictive coding (CELP)
ode Excited Linear Pre
diction). The linear predictive vocoder consists of 4.
Audio encoding methods have been widely used in the low bit rate region of 8 kb/s or less, and include methods such as the Percoll method and the Line Spectral Pair (LSP) method. Details of these methods are described, for example, in Saito and Nakata's "Fundamentals of Audio Information Processing" (Ohmsha Publishing). A linear predictive vocoder is composed of an all-pole filter representing the spectral envelope characteristics of speech and a sound source signal generator that drives the filter. As the driving sound source signal, a pitch periodic pulse train is used for voiced sounds, and white noise is used for unvoiced sounds. The parameters of the sound source signal include voiced/unvoiced distinction, pitch period, and amplitude, and these parameters are extracted as average features of the sound signal in an analysis interval of about 30 milliseconds. In this way, linear predictive vocoders synthesize speech by temporally interpolating the speech feature parameters extracted for each fixed analysis interval. , the features of the audio waveform cannot be reproduced with sufficient accuracy. Furthermore, since the drive sound source consisting of a periodic pulse train and white noise is insufficient to reproduce the characteristics of various speech waveforms, it has been difficult to obtain highly natural synthesized speech. As described above, in order to improve the quality of synthesized speech in a linear predictive vocoder, a driving sound source that can better reproduce the characteristics of the speech waveform has been required.

On the other hand, in code-excited predictive coding, speech is synthesized by using a noise sequence as a driving sound source to drive two all-pole filters representing the proximity correlation and pitch correlation characteristics of speech. The noise sequence is prepared in advance as a plurality of code patterns, from which a code pattern that minimizes the error between the input speech waveform and the synthesized speech waveform is selected. For details, see Schroeder et al., “Code
excited linear predictor
ion (CELP)”, IEEE Int. Conf
．． on ASSP, pp937-940, 1985. In code-excited predictive coding, there is a proportional relationship between the number of code patterns and the reproduction accuracy of the encoded speech waveform. Therefore, by preparing a large number of sequence patterns, the reproduction accuracy of the audio waveform can be improved, and the quality can be improved accordingly. However, the audio encoding bit rate is 4kb.
If it is less than /s, the number of code patterns will be limited, and as a result, sufficient voice quality will not be obtained. To obtain good voice quality, an amount of information of about 4.8 kb/s was required.

[0004] In code excitation predictive coding, the driving sound source is determined so as to reproduce the speech waveform itself.
An encoding method has been proposed in which a driving sound source is determined so as to reproduce a waveform after removing the phase component of an audio waveform that is auditory insensitive, that is, a waveform with zero phase. The details are described in Japanese Patent Application No. 59-53757 "Audio Signal Processing Method" and Japanese Patent Application No. 1-257503 "Speech Analysis and Synthesis Method". In this method, the short-time phase of the speech predicted residual waveform corresponding to the driving sound source signal is approximately equalized to zero phase, so the zero-phase speech waveform is a waveform that shows a large peak at the pitch drive point. is converted to As a result, it has become possible to encode the zero-phase predicted residual waveform with a smaller amount of information than the original waveform. In the aforementioned patent application "Speech analysis and synthesis method" (Japanese Patent Application No. 1-257503), a zero-phase predicted residual waveform (drive sound source) for a voiced sound is generated using a quasi-periodic pulse train and a filter coefficient of a zero-type filter. It shows how to express it. This method can provide high audio quality at a bit rate of 4 kb/s or less if there is no mixed noise in the input audio. However, when ambient noise is mixed into the input speech, this driving sound source signal cannot express the noise component superimposed on the voiced speech, so the mixed noise becomes another distortion that is not noise-like, resulting in the quality of the encoded speech. There was a problem that caused deterioration.

An object of the present invention is to provide speech analysis and synthesis that has high speech quality and excellent resistance to ambient noise in the boundary region between linear predictive vocoder and waveform coding (2.4-4.8 kb/s). The purpose is to provide a method.

[0006]

[Means for Solving the Problems] According to the present invention, a quasi-periodic pulse train and a noise sequence are added as driving sound source signals for voiced sounds used in speech analysis and synthesis, in which the magnitude of fluctuation in the pitch period of input speech is limited. Using the combined signal,
This driving sound source signal drives a linear filter representing the sound spectrum envelope characteristic to synthesize a sound waveform, and the sound source signal is parameters, that is, the temporal position and amplitude of the pulse sequence, and the pattern and amplitude of the noise sequence.

While conventional vocoders use a periodic pulse train generated from the average pitch period and amplitude determined for each fixed analysis interval as the driving sound source signal, in the present invention, the position of the pulse is determined for each pitch period. The driving sound source signal is composed of the sum of a quasi-periodic pulse train given an amplitude and a noise sequence having a constant block length. In addition, in conventional code excitation predictive coding, the driving excitation signal is composed of only a noise sequence, whereas in this invention, one
The driving sound source signal is composed of the sum of the pulse sequence and the noise sequence having a constant block length. Furthermore, in the conventional multi-pulse predictive coding method, the drive sound source signal is composed of a plurality of pulses that are determined independently of the pitch period, whereas in this invention, the drive sound source signal is composed of a plurality of pulses that are determined independently of the pitch period. The driving sound source signal is composed of the sum of noise sequences having a block length. Furthermore, in the above-mentioned code excitation predictive coding and multipulse coding, the square error between the input speech waveform and the synthesized speech waveform is conventionally used as an evaluation criterion for determining the sound source parameters. The squared error between the equalized speech waveform and the synthesized speech waveform is used. Finally, in the patent application "Speech analysis and synthesis method" (Japanese Patent Application No. 1-257503), the drive sound source signal is composed of a quasi-periodic pulse train and the filter coefficients of a zero-type filter, whereas in this invention, A noise sequence is used instead of a zero-type filter.

[0008]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows the configuration of an apparatus to which the speech analysis and synthesis method according to the present invention is applied. A sampled audio signal s(t) is input from the input terminal 1. In the linear prediction analysis unit 2, after once storing N audio signal samples in a data buffer, linear prediction analysis is performed on these samples to obtain prediction coefficients ai (i=1, 2,..., p). calculate. Further, a prediction residual signal e(t) is obtained by the following equation using an inverse filter using the prediction coefficient as a filter coefficient. Σ is i
=1 to p.

[0009] e(t)=st −Σai s(t−i)
Next, the autocorrelation coefficient of the prediction residual is calculated, the level is judged based on the maximum value, and the voiced and
Determine silent VUV. Details of these processing methods are described in the aforementioned book by Saito et al. The phase equalization analysis unit 3 calculates the coefficient ct (n) of the phase equalization filter 4 that zero-phases the phase characteristic of the audio and the reference time t'i for phase equalization.
Calculate. The details of its configuration are described in the aforementioned patent application "Speech analysis and synthesis method" (Japanese Patent Application No. 1-257503). The phase equalization filter 4 is controlled on a sample-by-sample basis using the filter coefficients determined by the phase equalization analysis section 3. By inputting the audio signal from the terminal 1 to the phase equalization filter 4, a phase equalized audio signal sp (t) is obtained as its output by the following equation. Σ is from i=-M/2 to M/2.

[0010] sp (t)=Σct (i) s(t-i
) In this analysis and synthesis method, separate driving sound sources are used for voiced sounds and unvoiced sounds, and the switch 18 is switched depending on the voiced/unvoiced parameter VUV. First, the configuration of the driving sound source for voiced sounds will be explained below. The voiced sound driving sound source is composed of a pulse sequence generation section 10 and a noise codebook 12. The pulse sequence generation section 10 generates a quasi-periodic pulse sequence by giving pulse time points ti. The amplitude of each pulse is determined by a gain m in the pulse amplitude control section 11.
controlled by multiplying by i.

On the other hand, the noise codebook 12 stores a plurality of noise sequences (noise vectors) having a fixed block length. As the noise sequence, for example, a normal random number sequence with an average of 0 and a variance of 1 is used. Noise code book 1
The noise sequence output from 2 is passed through the pitch correlation filter 13
and is converted into periodic noise. The pitch correlation filter 13 is realized by a digital filter having the following transfer characteristics.

B(z)=1/(1-γb bz-Tp
) Here, b is the pitch gain, Tp is the pitch period, and γb is the periodicity emphasis coefficient. pitch gain b
is calculated by the pitch gain calculation unit 5 as an autocorrelation coefficient of the phase equalization prediction residual signal with respect to the time delay of the pitch period. The pitch gain calculated by the pitch gain calculating section 5 is quantized by the quantizing section 5a, and is given to the pitch correlation filter 13 as a pitch gain b. Pitch period T
p is given as the average spacing of adjacent pulse instants of the quasi-periodic pulse train. The amplitude of the output signal of the pitch correlation filter 13 is controlled by multiplying it by a gain GV in an amplitude control section 14.

The thus obtained quasi-periodic pulse sequence p(t) from the pulse characteristic control section 11 and the amplitude control section 1
The noise sequence v(t) from 4 is added to the adder 15 for each sample.
and the drive excitation signal ev (t) is generated. Next, the configuration of the unvoiced sound driving sound source will be explained. For unvoiced sounds, a noise sequence is used as the driving sound source signal. Similar to noise codebook 12 for voiced sounds, noise codebook 1
6 stores a plurality of noise sequences having a constant block length. The amplitude of the noise sequence output from the noise codebook 16 is controlled by multiplying it by a gain GU in the amplitude control unit 17, and the drive excitation signal eu (t
) is generated. Furthermore, in the case of unvoiced sounds, the pitch correlation filter is not included in the configuration because the periodicity of the speech is weak. Generally, the block length of an unvoiced noise sequence is different from the block length of a voiced noise sequence.

For voice synthesis, the switch 18 selects the driving excitation signal ev (t) or eu (t) according to the voiced/unvoiced parameter VUV, and an all-pole (linear) filter is applied to characterize the spectral envelope characteristic of the voice. This is done by driving 19. The all-pole filter 19 is realized by a digital filter having the following transfer characteristic A(z). A(z)=1/(1+a1 z-1+...+ap z-p
) where ai are the linear prediction coefficients, z-1 is the sampling delay, and p is the order of the filter. The linear prediction coefficient ai used at the time of synthesis is calculated by the linear prediction analysis unit 6,
It is obtained by linear predictive analysis of the output phase equalized audio of the phase equalization filter 4, and is quantized by the quantization section 6a.

Next, a method of analyzing sound source parameters will be explained. The pulse time calculation unit 7 calculates the pulse time of the quasi-periodic pulse system. The pulse instants are limited so that their position spacing is quasi-periodic. That is, the pulse time interval Ti=ti-ti-1 in FIG. 2 is limited by the following equation so that the difference between successive pulse time intervals is equal to or less than a certain value.

ΔTi =|Ti −Ti−1 |≦J where J is the tolerance for the difference in pulse time intervals. The pulse time point ti is determined by using the reference time point t'i obtained by the phase equalization analysis section 3 as an initial value, and determines a series of pulse time points that satisfy the above restrictions, quantizes it in the quantization section 7a, and generates the pulse sequence generation section. Supply to 10. FIG. 3 shows the reference time t'i
This figure shows a processing procedure for generating a sequence of pulse time points t'i from . In this process, the reference time point ti is input as the initial value ti of the pulse time point, the number of reference time points is first determined (S1), and if the number of reference time points is 2 or less, the reference time point is set as the pulse time point. If the number of reference points is 3 or more,
Calculate the difference ΔTi between the time intervals of adjacent reference points (S
2), first determine whether ΔTi is less than the allowable value J (
S3), if it is less than the tolerance value, move on to step S4, and if it is not less than the tolerance value J, it is determined whether 1/2 of ΔTi is less than J (S5), and if it is less than J, the pulse time interval is too long. A pulse position is inserted at an intermediate point from then to step S4 (S6). The interval ti+1 -ti-1 between the next reference time and the interval t between the previous reference time
Find the difference ΔTi from i-1 -ti-2 (S7)
, it is determined whether this is less than or equal to J (S8), and if it is less than J, the pulse time interval is too narrow, the pulse time ti is removed, and the process moves to step S4 (S9). If it is not less than J in step S8, it is determined whether 1/2 of ΔTi is less than J (S10), and if it is less than J, the pulse position ti is corrected to the midpoint between ti+1 and ti-1, and step S4
The process moves to (S11). If it is not J or less in step S10, the amplitude of each pulse is calculated using the pulse amplitude calculation method described later with respect to the reference time (S12), and the time point with the minimum pulse amplitude is deleted from the reference time, and step Return to S1 (S13
). In step S4, it is checked whether all pulse positions (time points) have been determined, and if the determination has not been completed, the process returns to step S1, and once the determination has been completed, the process ends. As described above, the pulse time points are determined by repeating the insertion, removal, and modification of the pulse time points.

The pulse amplitude calculation unit 8 calculates the amplitude of each pulse of the quasi-periodic pulse train. The amplitude of each pulse is determined so that the frequency-weighted mean square error between the speech waveform synthesized using the quasi-periodic pulse sequence and the phase-equalized input speech waveform is minimized. The frequency weighted mean square error is expressed by the following equation. The first Σ is from t=0 to N-1, the next Σ
is from j=1 to nP. np is the number of pulses within the analysis frame.

d=Σ{(sp (t) −fz (t)
-f(t) *Σmj δ(t-tj ))*w(t
)}2 Here, δ(.) represents a delta function and * represents convolution. f(t) is the impulse response of the all-pole filter 19. fz (t) is the transfer characteristic A(z
) is the initial value response when the filter is driven with zero input. w(t) is the impulse response of the frequency weighting filter, and the transfer characteristic is expressed as follows.

W(z)=A(z)/A(γz) where,
γ is a parameter that controls the degree of frequency weighting, and takes a value in the range 0<γ≦1, usually 0.7-0.9
The value of is used. FIG. 4 shows the internal configuration of the pulse amplitude calculation unit 8 for determining the pulse amplitude that minimizes the above-mentioned mean square error. Phase equalized audio sp (
t) as input, the filter 41 inputs sp (t)*w(
t), the filter 42 calculates fz(t)*w(t), and the adder 43 subtracts the output of the filter 42 from the output of the filter 41 to obtain sw (t). The impulse response calculation unit 44 calculates the impulse response fw (t) of a filter having a transfer characteristic of 1/A(γz).
Calculate. The correlator 45 calculates the mutual covariance ψ(i) between the impulse response fw (t) and the signal sw (t) at each pulse time ti using the following equation. Σ is from t=0 to N-1.

[0020]ψ(i)=Σfw (t-ti)sw
(t) Also, in the correlator 46, each pulse time point ti, t
The autocovariance of the impulse response φ(i,
j) is calculated using the following formula. Σ is from t=0 to N-1. φ(i,j)=Σfw(t-ti)fw(t-t
j) The pulse amplitude calculating section 47 calculates the pulse amplitude by solving the following simultaneous equations.

[0021]

[Math 1]

These pulse amplitudes mi are quantized by a quantizer 8a and provided to an amplitude controller 11. The noise sequence/noise gain calculation unit 9 determines the noise sequence and its gain (noise gain) in the voiced sound. The noise gain is determined so that the frequency-weighted mean square error between the phase-equalized input audio waveform and the audio waveform obtained by combining the sum of the quasi-periodic pulse sequence and the noise sequence as a driving sound source signal is minimized. If the i-th noise sequence CVi (t) in the noise codebook 12 and the impulse response of the combined synthesis filters 1/A(z) and 1/B(z) are h(t), then the frequency-weighted average of the synthesized speech is The squared error is given by the following equation. Σ is from t=0 to N-1.

di = Σ{(sp (t) − fz (
t)-p(t)*f(t)-hz(t)-GVi
CVi *h(t))*w(t) }2 where p(
t) is the quasi-periodic pulse sequence determined by the method described above, hz
(t) is the zero-input initial value response of the composite synthesis filter. At this time, the optimal gain that minimizes the squared error is calculated using the following equation. Each Σ is from t=0 to n-1. GVi =Σz(t)y(t)/Σy2(t)where z(t)=(sp(t)-fz(t)-p(t)
*f(t)-hz (t))*w(t),y(t)=C
Vi (t)*h(t)*w(t).

FIG. 5 shows the internal configuration of the noise sequence/noise gain calculation section 9. The signal sw (t) obtained by the pulse amplitude calculation unit 8 in FIG. 4 is input. The filter 51 calculates p(t)*f(t)*w(t),
The filter 52 calculates hz (t)*w(t), and the adders 53 and 54 calculate the difference between the signals for each sample to obtain the signal z(t). The filter 55 receives the noise sequence CVi as input, and y(t)=CVi
(t)*h(t)*w(t), y(t)
seek. In the correlator 56, the correlation function between the signals z(t) and y(t) is obtained as czy=Σz(t)y(t) (Σ is from t=0 to N-1), and the correlator 57, the power of the signal y(t) is determined as cyy=Σy2 (t) (Σ is from t=0 to N-1). The divider 59 performs the calculation GVi=czy/cyy,
The optimal gain is calculated. In the correlator 58, the signal z(t)
The multiplier 60 calculates the power czz of GVi and cz
The adder 61 multiplies GV from czz.
By subtracting i cz y , the mean square error di of the synthesized speech is determined. In the minimum value selection section 62,
From among the noise sequences included in the noise codebook 12, select the noise sequence that minimizes the mean squared error di of synthesized speech, and calculate the number ICV and optimal noise gain GV of that noise sequence.
Output. The gain GV is quantized by the quantization section 9a and given to the amplitude control section 14.

Next, a method for determining the optimal noise sequence and optimal gain for unvoiced speech will be described. The noise sequence/noise gain calculation unit 20 determines the noise sequence and its gain (noise gain) in unvoiced sounds. The noise gain is determined so that the frequency-weighted mean square error between the speech waveform synthesized from the noise sequence as the driving sound source signal and the phase-equalized input speech waveform is minimized. The i-th noise sequence CUi (t) in the noise codebook 16, the synthesis filter 1/A(z
) is the impulse response of f(t), the frequency-weighted mean square error of the synthesized speech is given by the following equation. Σ is t
=0 to N-1.

di = Σ{(sp (t) − fz (
t) -GUi CUi *f(t))*w(t) }
2 Here, fz (t) is the 0-input initial value response of the aforementioned synthesis filter. At this time, the optimal gain that minimizes the squared error is calculated using the following equation. Each Σ is from t=0 to N
-1. GUi = Σsw (t)y(t)/Σy2 (t) where sw (t) = (sp (t) - fz (t))
*w(t),y(t)=CUi(t)*f(t)*w
(t).

FIG. 6 shows the internal configuration of the noise sequence and noise gain calculation section 20. phase equalized audio signal sp
In the filter 63 to which (t) is input, sp (t)
*w(t) is calculated, and the filter 64 calculates fz (t)*
w(t) is calculated, and the adder 65 calculates the difference between the signals for each sample to obtain the signal sw (t). The filter 66 inputs the noise sequence CUi and calculates y(t
)=CUi (t)*f(t)*w(t) is performed to obtain y(t). In the correlator 67, the signal sw (
t) and y(t) as csy=Σsw (t
)y(t) (Σ is from t=0 to N-1), and the correlator 68 calculates the power of the signal y(t) as cyy
= Σy2 (t) (Σ is from t=0 to N-1
up to). In the divider 70, GVi =czy/c
The optimum gain is calculated by calculating yy. The correlator 69 calculates the power css of the signal sw (t), the multiplier 71 multiplies GUi by csy, and the adder 72 subtracts GUi csy from css, so that the mean square error di of the synthesized speech is Desired. The minimum value selection unit 73 selects the noise sequence with the minimum mean squared error di of synthesized speech from among the noise sequences included in the noise codebook 16, and calculates the number ICU and optimal noise gain GU of the noise sequence. Output. The gain GU is quantized by the quantization section 20a, and the amplitude control section 17 is controlled.

[0028] Through the processing described above, the audio signal is
For unvoiced sounds, linear prediction coefficient ai, voiced/unvoiced parameter VUV, for voiced sounds, pulse time ti, pulse amplitude mi, noise sequence number ICV and amplitude GV, pitch correlation filter coefficient b, for unvoiced sounds, noise sequence number ICU and amplitude GU. Represented by These audio parameters are encoded by the encoder 21 and transmitted or stored.

In the speech synthesis section, as shown in FIG. 7, after decoding all speech parameters in the decoding section 22, voiced/
The driving sound source signal is decoded according to the unvoiced parameter VUV. In the case of a voiced sound, the pulse sequence generation unit 23
A quasi-periodic pulse train is generated from the pulse time point ti, and the amplitude of each pulse of the quasi-periodic pulse train is controlled by the amplitude controller 24.
to control. Also, using the same noise codebook 25 as the noise codebook 12 on the transmitting side, the noise sequence number I
Read out the noise series corresponding to the CV. After passing the noise sequence through a pitch correlation filter 26, the amplitude control unit 27 multiplies it by a noise gain GV to generate a noise drive sequence. This 2
The two signal sequences are added sample by sample in an adder 28 to generate a drive sound source signal. On the other hand, in the case of unvoiced sound, the noise codebook 29 that is the same as the noise codebook 16 on the transmitting side is used to read out the noise sequence corresponding to the noise sequence number ICU, and this noise sequence is multiplied by the noise gain GU in the amplitude control unit 30. to generate a driving sound source signal. The switch 31 selects a driving sound source signal according to the voiced/unvoiced parameter VUV, and uses the selected driving sound source signal to filter an all-pole filter 32 in which a linear prediction coefficient ai is set as a filter coefficient.
By driving the synthesized speech, synthesized speech is outputted to the output terminal 33 thereof.

In order to simplify the illustration in FIG. 4, 5, 20, respectively.
As is clear from 6, specifically, the output sp (t),
s′(t) is calculated by the pulse amplitude calculation unit 8 and the noise gain calculation unit 9.
, 20, respectively.

[0031]

[Effects of the Invention] In order to examine the effects of the speech analysis and synthesis method according to the present invention, an analysis and synthesis speech experiment was conducted under the following conditions. Sampling frequency 8k for audio in the 0-4kHz band
After sampling at Hz, the speech signal is multiplied by a Hamming window with an analysis window length of 30 ms, the analysis order is set to 12th order, linear prediction analysis is performed using the autocorrelation method, and the 12 prediction coefficients and voiced/voiced
Find the silent parameters. The analysis frame length for encoding is 2
5ms (160 audio samples). The prediction coefficient is LS
Multistage vector quantization is performed using the Euclidean distance of the P parameter. Furthermore, a plurality of pulse amplitudes are collectively regarded as a vector and vector quantized. The noise gain is scalar quantized. The pulse point is the first pulse position in the frame,
The interval between the second pulse time and the first pulse time, 3
From the second pulse onward, the difference in interval between adjacent pulse points was encoded. The block length of the noise sequence is 20 ms (160 samples) for voiced sounds and 5 for unvoiced sounds.
ms (40 samples). Bitrate is 3.7k
In the case of b/s, the number of bits per frame is 74 bits, the breakdown of which is as follows.

Parameters
Number of bits/frame Prediction coefficient

24 Voiced/unvoiced parameters
1
Driving sound source (if voiced) pulse time
30
pulse amplitude
8
Number of noise sequences
6
noise gain
3
Pitch prediction filter coefficient 2 Number of driving sound source (unvoiced) noise sequences
36 (=9×4)
noise gain
12=(3×4) Speech encoded under the above conditions has much higher naturalness than a conventional vocoder, and high speech quality is achieved. Furthermore, the dependence of voice quality on the speaker is smaller than with conventional vocoders. It was also confirmed that the quality of encoded speech is higher than that of conventional multi-pulse predictive coding or code-excited predictive coding. The time delay caused by encoding is 60m
s, which is about the same level or lower than the conventional method in the low bit rate area. Additionally, even when ambient noise is mixed into the input audio, the audio mixed with the noise is reproduced as is, resulting in more natural audio and improved immunity to ambient noise than when using only a conventional quasi-periodic pulse sound source. has been done.

The advantage of the present invention is that it is possible to realize speech encoding with extremely natural speech quality and resistance to ambient noise at a low bit rate of 4 kb/s or less.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of a device to which a speech analysis and synthesis method according to the present invention is applied.

FIG. 2 is an explanatory diagram of a quasi-periodic pulse drive sound source signal.

FIG. 3 is a flowchart illustrating an example process for generating pulse instants.

FIG. 4 is a block diagram showing a specific example of the pulse amplitude calculation section 8.

[Fig. 5] Noise sequence/noise gain calculation unit 9 for voiced sounds
FIG. 2 is a block diagram showing a specific example.

[Figure 6] Noise sequence/noise gain calculation unit 2 for unvoiced sounds
1 is a block diagram showing a specific example of 0. FIG.

FIG. 7 is a block diagram showing the configuration of a synthesis device to which the speech analysis and synthesis method according to the present invention is applied.

Claims (1)

[Claims] [Claim 1] A speech analysis and synthesis method using a linear filter representing speech spectral envelope characteristics and a sound source signal generation section that generates a sound source signal for driving the linear filter, in which the magnitude of pitch period fluctuation is limited. A signal obtained by mixing a quasi-periodic pulse train and a noise sequence is used as the sound source signal, and the sound source signal drives the linear filter to synthesize a sound signal, and the phase of the input sound is pitch-periodically set to zero phase. A speech analysis and synthesis method, characterized in that a sound source signal generation parameter of the sound source signal generation section is determined so that an error between the phase-equalized speech signal and the synthesized speech signal is minimized.