JPH02160300A - Voice encoding system - Google Patents

Abstract

PURPOSE:To obtain a voice encoding and decoding device of good voice quality with a relatively small arithmetic quantity by weighting a spectrum parameter when the impulse response of a synthesizing filter has high periodicity. CONSTITUTION:A voice signal is classified according to phonetic features and a sound source signal which matches the classifications is used. The vowellike section of the sound source signal is represented by using multipulses in a representative section and at least either of an amplitude correction coefficient and a phase correction coefficient in other pitch sections among pitch sections obtained by dividing a frame and the frictional section is represented by using the combination of a small number of multipulses and a noise signal. Further, the intensity of the periodicity of the impulses response of the synthesizing filter is decided and when the periodicity is high, the spectrum parameter is weighted so as to widen the band width of the synthesizing filter. Therefore, even a female voice, specially, 'i' and 'u' are generated while the underestimation of the band width of the synthesizing filter is prevented. Consequently, an excellent synthetic voice is obtained.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to an audio encoding method, in particular, to encoding audio signals with high quality at a low bit rate, that is, about 4.8 kb/s, and with a relatively small amount of calculation. This paper relates to a speech encoding method for

[Conventional technology]

As a method for encoding audio signals at a low bit rate of about 4.8 kb/s, for example, Japanese Patent Application Laid-open No. 15000/1983
Pitch interpolation multi-pulse methods are known, such as those described in JP-A-62-038500 (Document 2) and the like. In these methods, the transmitting side extracts a spectral parameter representing the spectral characteristics of the audio signal and a pitch parameter representing the pitch from the audio signal for each frame, and classifies the audio signal into two types: voiced sections and unvoiced sections. In a voiced section, the sound source signal of one frame is expressed as a multipulse for one pitch section (representative section) out of a plurality of pitch sections obtained by dividing one frame into pitch sections, and the amplitude and position of the multipulse in the representative section are expressed as multipulses. and spectrum and pitch parameters. Furthermore, in the silent section, the sound source of one frame is represented by a small number of multipulses and a noise signal, and the amplitude and position of the multipulses and the gain and index of the noise signal are transmitted.

In the voiced section, the receiving side interpolates the amplitude and position of the multipulses using the multipulse in the representative section of the current frame and the multipulse in the representative section of the adjacent frame, and then interpolates the multipulse in the pitch section other than the representative section. and restore the driving sound source signal of the frame. Furthermore, in the silent section, the drive excitation signal of the frame is restored using the multi-pulse and the index and gain of the noise signal. Furthermore, the restored drive sound source signal is input to a synthesis filter using spectral parameters to output a synthesized speech signal.

[Problem to be solved by the invention]

However, in the conventional method described above, in voiced sections, the driving sound source signal of the frame is restored by interpolation between multi-pulses in the representative section, so it is not possible to restore the driving sound source signal of the frame by interpolating the multi-pulses of the representative section. For frames in which the characteristics of the sound source signal change, such as, the driving sound source signal restored by interpolation is significantly different from the actual sound source signal, and as a result, the sound quality of the synthesized voice is degraded. Furthermore, in the nasal section of the voiced section, no clear periodicity appears in the sound source signal, so the pitch interpolation method cannot adequately represent the sound source signal. On the other hand, it is known from perceptual experiments that parts of the voice where the characteristics change greatly are extremely important for phonological perception and naturalness perception, but conventional methods do not provide sufficient information on these parts. There is a big problem that the sound quality deteriorates because it cannot be restored. In addition, in unvoiced sections, the sound source signal is expressed using multipulses and noise, but even in consonant sections, the sound source of fricatives is noisy, but in plosives there are many pulse-like parts, so the conventional method Thus, there is a problem in that good synthesized speech cannot be obtained by simply classifying and representing speech signals into two types, voiced and unvoiced.

On the other hand, linear predictive analysis (LPC analysis) is conventionally well known for analyzing spectral parameters representing the spectral envelope of speech. However, since LPG analysis is influenced by the fundamental and harmonic components of the pitch frequency for female voices, especially i, tsu, etc., the characteristics of a synthesis filter using spectral parameters obtained by LPC analysis are There is a problem in that the bandwidth especially at the frequency corresponding to the first formant of the voice becomes extremely narrow compared to the spectral envelope of the voice. Therefore, if such spectral parameters are used when determining the sound source signal, the pitch periodicity will not appear in the sound source signal, and the sound source signal will be represented by multipulses using the pitch interpolation described above assuming periodicity of the sound source. In this case, there is a problem that the sound quality of the synthesized speech deteriorates significantly.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems and provide a speech encoding/decoding device that achieves good sound quality at approximately 4.8 kb/s with a relatively small amount of calculation.

[Means to solve the problem]

The audio encoding method according to the present invention calculates a spectral parameter representing a spectral envelope and a pitch parameter representing a pitch from an input discrete audio signal for each frame,
When the impulse response of a filter configured using the spectral parameters has strong periodicity, weighting is applied to the spectral parameters, a discrimination parameter representing the characteristics of the audio signal is extracted, and the audio signal is divided into a plurality of types. The frame section is divided into small sections according to the pitch parameter as the sound source signal of the audio signal for each frame according to the type, and one of the small sections is divided into subsections according to the pitch parameter.
It is configured by determining and outputting the multi-pulse obtained in one interval, correction information for correcting at least one of the amplitude or phase of the multi-pulse, or at least one of a codebook and the multi-pulse.

[Effect]

The first feature of the audio encoding method according to the present invention is to classify audio signals of frames into predetermined types. Below, we will discuss an example in which the sound source is classified into four types: vowel, nasal, fricative, and plosive, but these types are appropriate to express the sound source signal well according to the differences in sound production of the sound source. need to be selected. Configurations other than the four types are also possible.

The voice classification method involves extracting the discrimination parameters that represent the characteristics of the voice signal and classifying the voice signal into the above-mentioned 4 types as shown in Figure 2.
Classify into species. As this parameter, for example, the power of the signal or its RMS (Root Mean 5q
uare), power change or change rate every short time (for example, 5 ms), spectral change or change rate every short time, pitch gain, etc. can be used.

The second feature is that when the impulse response of the synthesis filter composed of the spectral parameters has strong periodicity in the spectral parameters representing the spectral envelope of the audio signal extracted using LPC analysis for each frame, the first It is determined that the bandwidth of the filter is underestimated in the band corresponding to the formant, and appropriate weighting is applied to the spectral parameters. Here, the transfer characteristic H(
z) can be written as follows.

Here, al is a spectral parameter, and P is the order of the filter. Impulse response h(n) of this synthesis filter
is determined by the following formula.

h(n)4a+h(n-i))Gδ(n)
(n≧0) −(2) where G is the amplitude of the excitation source. Then, if the pitch gain PCI obtained from h(n) is larger than a predetermined threshold,
It is determined that the impulse response has strong periodicity. Here, the pitch gain can be determined by calculating the autocorrelation function of h(n) by a predetermined time lag, and using the value of the autocorrelation coefficient at the time lag that takes the maximum value. Next, when the impulse response has strong periodicity, the spectral parameters are weighted as shown in the following equation.

aI: aIrl (1≦i≦P),
(3) Here, r takes a positive value smaller than 1. Depending on the value of r, the bandwidth of the synthesis filter is determined by the amount B(H
It spreads by z,).

B=Fs/π1. (r) (Hz)
-(4)-As an example, r is 0.98 and Fs is 8kHz.
If selected, B will be approximately 50 Hz.

For multi-pulse calculations explained below, when it is determined that the impulse response of the synthesis filter has strong periodicity,
Spectral parameters weighted using equation (3) are used. When the periodicity is not strong, the weighting in equation (3) is not performed. The above is the second feature of the present invention.

Next, how to obtain the sound source signal will be explained.

It is determined whether the frame is a vowel section or not using the signal power or its RmS and pitch gain. In the vowel interval, as shown in Figure 3, the frame interval is divided into a plurality of pitch intervals for each pitch period determined in advance, and a multipulse is applied to one pitch interval (representative interval) among these pitch intervals. demand. Next, for other pitch sections within the same frame, the amplitude correction coefficient Ck and phase correction coefficient dk for the multi-pulse are determined.

Then, for each frame, as sound source information, the pitch position in the frame of the representative section, the amplitude and position of the multipulse in the representative section, and the amplitude correction coefficient ck and phase correction coefficient dk in other pitch sections of the same frame are used as correction information. Transmit. The representative section may be found by searching for a section where the best synthesized speech signal is found, or may be fixed within the frame. The former has better sound quality, but requires more calculations.

The following describes how to obtain the amplitude correction coefficient Ck and the phase correction coefficient dk, and how to search for a representative section. Now, let T be the average pitch period found for each frame. FIG. 3(b) shows how the frame is divided into T subframe sections. Here, a case will be described in which a representative section is searched. For example, a subframe that is a candidate for the representative section is subframe ■. The amplitudes and positions of a predetermined number of multipulses are determined for subframe ■. As for how to obtain multipulses, a method is known that uses a cross-correlation function Φ□ and an autocorrelation function Rhb, and this method is described, for example, in the above-mentioned document 1.2 and Araseki.

°Mult by Ozawa, Ono, Ochiai
i-pulse ExcitedSpeech Cod
er Ba5ed on Maximum Cross
-correla ion 5earch Alg
orithm,” (GLOBECO) 4 83. I
EEEGlobal Tele-communicat
ions Conference, lecture number 23.3
．． 1983) (Reference 3), so the explanation will be omitted here.

The amplitude and position of the multipulse in the representative section are expressed as g.
, and m+H=t~L). This is shown in Figure 3(c)
Shown below. Amplitude correction coefficient c in section k other than the representative section
k, phase correction coefficient dk, and synthesized speech x'k(n) synthesized for section k using these and a synthesis filter,
It can be determined to minimize the weighted difference power 9 with respect to the sound Xk(n) in the corresponding section. The weighted error power Ek is Ek=ΣI[xk(n)-x' h(n)]*w(n)
l 2- (5) However, X' k(n)=ckΣg+ ・h(n-m+-T-
d k) ・+ ・ (6
) Here, w(n) represents the impulse response of the perceptual weighting filter. However, this filter is not necessary. Further, h(n) represents an impulse response of a synthesis filter for synthesizing speech. ck and dy can be obtained by minimizing equation (5). To do this, for example, first fix di, partially differentiate equation (5) with respect to ck, and set it as O to obtain the following equation.

ΣXwk(n)x'wi+(n) where x, k (n)=xk(n)new(n)
−(8a)x'wk(n)"Σg+
h(n-m+-T-dk)1w(n)-(8b)
Therefore, the value of equation (7) is found for various values of dk, and (
7) By finding a combination of d, ck that minimizes ck in equation (5), EK in equation (5) is minimized. In this way, Ck, d for pitch sections other than the representative section
k is determined, and the weighted offset error power E defined by the following equation is determined for the entire frame.

E・ΣEk...(
9) Here, N is the number of subframes included in the corresponding frame. However, the weighted error power E2 of the representative pitch section (subframe section ■ in the example of FIG. 3) is determined by the following equation.

E2=Σf[x(n)=Σg+・h(n-m+)]*w
(n)12-(10) Search for a representative pitch section calculates the values of equations (5) to (10) for all representative pitch section candidates, and minimizes the value of the error power of equation (9). The section where the pitch is played can be set as the representative pitch section. Figure 3 (c
), when the representative pitch section after searching is subframe ■, the multi-pulse of the representative section and the second section other than the representative section (k・■, ■, ■ in Fig. 3(c))
An example will be shown in which the sound source vk(n) is generated according to the following equation using amplitude and phase correction coefficients.

VK(n)"CJgl ・δ(n-m+-T-dk)
-(11) Next, in the nasal interval, it is expected that the periodicity of each pitch of the sound source is not strong in the vowel interval, so
Rather than using the method described above, the sound source is represented by pitch prediction multipulses or multipulses. Here, for how to obtain the pitch prediction multi-pulse, reference can be made to JP-A-6O-051900 (Reference 4). Further, reference can be made to the above-mentioned document 3 for how to obtain multi-pulses. Note that the nasal interval can be determined using, for example, power or its RMS, pitch gain, and first-order logarithmic cross-sectional area ratio r1 defined by the following equation. Particularly in the nasal interval, rl is characteristically large.

Here, is a parameter (also called PARCOR) for the first missing point.

On the other hand, in the consonant section, the sound source is represented by multipulses or a combination of multipulses and noise.

In the consonant interval, it is determined whether the sound is fricative or plosive, and if it is fricative, the sound source is represented by multipulses and noise or a codebook. For the specific method, reference can be made to the above-mentioned document 2 and the like. In addition, in the case of rupture, the sound source is represented by multipulses. As a method for determining the frictional property and the rupturable property, parameters such as power every short period of time (for example, 5 m5) or a change or rate of change in the RMS thereof can be used.

〔Example〕

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of a speech encoding method according to the present invention. In the figure, an audio signal is input from an input terminal 100, and the audio signal for one frame (for example, 20 m5) is stored in a buffer memory 110. Pitch analysis circuit 130 calculates an average pitch period T from the audio signal of the frame.

As this method, for example, a method based on an autocorrelation method is known, and for details, refer to the pitch extraction circuit in Document 1.2. In addition to this method, other well-known methods (for example, cepstrum method, 5IFT method, phase change open method, etc.) can be used. Pitch encoding circuit 1
50 outputs a code obtained by quantizing the average pitch period T with a predetermined number of bits to the multiplexer 260, and also outputs a code obtained by decoding this to the decoded pitch period T.
' is output to the sound source signal calculation circuit 220, the interpolation circuit 282, and the drive signal restoration circuit 283.

The K parameter calculation circuit 14.0 calculates the following as a spectral parameter representing the spectral characteristics of the audio signal of the frame.
The K parameter is calculated from the audio signal of the frame by the well-known LP
Perform C analysis and calculate only a predetermined order M. Regarding this specific calculation method, reference can be made to the above-mentioned document 1.2 regarding the parameter calculation circuit. Note that the K parameter is the same as the PARCOR coefficient.

The periodicity determination circuit 145 once converts the K parameter into a linear prediction coefficient a using a well-known method, and calculates the impulse response of the synthesis filter constituted by the linear prediction coefficient. Here, the transfer characteristic of the synthesis filter is expressed by the above equation (1). Further, the above equation (2) is used to calculate the impulse response. Next, determine the periodicity of the impulse response.

Specifically, we calculate the pitch gain of the impulse response,
This is compared with a predetermined threshold, and if it is larger than the predetermined threshold, it is determined that the periodicity is strong. Here, in calculating the pitch gain, the autocorrelation function of the impulse response is calculated for a predetermined delay time, and the value of the autocorrelation coefficient at the delay time that takes the maximum value can be taken as the pitch gain. For details, refer to the pitch extraction circuit in Document 1.2. When the pitch gain is larger than a predetermined threshold, a predetermined weight r is applied to the linear prediction coefficient a according to equation (3). Here, the value of r is a positive value smaller than 1. The linear prediction coefficients weighted in this way are inversely transformed into parameters again and sent to the K-parameter encoding circuit 160.
Output to. The conversion between parameters and linear prediction coefficients is described in 'Linear Pr' by J. Makhou et al.
You can refer to the book titled "Edict, ion of 5peech" (Reference 5).

The K parameter encoding circuit 160 generates a code ρk obtained by quantizing the parameters with a predetermined number of quantization bits.
is output to the multiplexer 260, and is also decoded and further converted into linear prediction coefficients a,' (i=1 to M), which are output to the weighting circuit 200 and the interpolation circuit 282. Regarding the method of converting the sign (hi, ni) parameter of the K parameter into a linear prediction coefficient, reference can be made to the above-mentioned document 1.2.

The impulse response calculation circuit 170 calculates the linear prediction coefficient a.
The impulse response (n) of the synthesis filter subjected to perceptual weighting using 1' is calculated and outputted to the autocorrelation function calculation circuit 180. Autocorrelation function calculation circuit 18
0 calculates and outputs the autocorrelation function Rhh(n) of the impulse response up to a predetermined delay time. Impulse response calculation circuit 170 and autocorrelation function calculation circuit 1
Regarding the operation of 80, reference can be made to the above-mentioned documents 1.2, etc.

The subtracter 190 subtracts the output of the synthesis filter 281 by one frame from the frame audio signal x(n), and outputs the subtraction result to the weighting circuit 200. For weighting, the circuit 200 passes the subtraction result through an auditory weighting filter whose impulse response is represented by w(n) to obtain a weighting signal x,(n) and output it. The weighting method is described in the above document 1.2.
etc. can be referred to.

The cross-correlation function calculation circuit 210 calculates the weighted scale signal x, (n
) and the impulse response hW(n) to obtain the cross-correlation function Φ. is calculated and output up to a predetermined delay time.

For this calculation method, reference can be made to the above-mentioned documents 1 and 2.

The determination circuit 215 determines the type of audio signal of the frame. Here, as an example, as mentioned in the section on action, we will classify it into four types: vowel, nasal, fricative, and plosive, but the number of classifications is not limited to four, and there are other types as well. You can also use the law. For these determinations, as mentioned in the section on effects, the power of the frame audio signal and its RM are used.
S, pitch gain, power per short time or its RM
A change in S, a change in spectrum between frames, etc. can be used. The type determined using these parameters is determined by the sound source signal calculation circuit 220 and the multiplexer 260.
Output to.

In the sound source signal calculation circuit 220, vowel character is determined when the power or its RMS is equal to or higher than a predetermined threshold;
The determination is made based on whether the pitch gain is greater than or equal to a predetermined threshold. In this case, as explained in the operation section, the frame is divided into subframes (pitch sections) for each pitch period in advance using the decoded average pitch period T'.
The position ll11 and the amplitude g+ of the multi-pulse are determined for a pitch section that is a candidate for a representative one pitch section (representative section) as a sound source signal.

Next, the amplitude/phase correction circuit 270 performs (3
), according to equation (4), multi-pulse amplitude correction coefficient ck for generating sound source signal in other pitch section k,
Calculate the phase correction coefficient d1. Furthermore, these values are output to the sound source signal calculation circuit 220, and the sound source signal calculation circuit 220
Now, (1), (5), (
6) Based on the formula, calculate the error power E of the entire frame for several candidate sections, select the pitch section that minimizes E as the representative section, and calculate the information PI indicating the subframe number of the representative section, the information PI indicating the subframe number of the representative section, Multipulse amplitude g
1, the amplitude correction coefficient ck and the phase correction coefficient d of the position n+(t-t~L) and other pitch sections are output.

Next, nasality is determined based on whether the pitch gain is larger than a predetermined threshold value and the logarithmic cross-sectional area ratio of one mouthpiece is larger than a predetermined threshold value. In this case, for example, multi-pulses are determined for the entire frame section.

On the other hand, in consonant intervals, the discrimination between fricative and plosive is, for example,
Spectrum changes every short time (for example, 5ms), power every short time (for example, about 5ms), or its RMS
If the change in is larger than a predetermined threshold value, it is determined that there is rupture, and otherwise, it is determined that there is friction. To determine the friction property, low range (e.g. 1 kHz or less) and high range (e.g. 2 kHz) are used.
It is also possible to use the ratio of the power or its RMS to (Hz or higher).

In the case of friction, the sound source signal is represented by a predetermined number of multipulses and a noise signal or codebook. For the specific method, reference can be made to the above-mentioned document 1.2.

First, a predetermined number of multipulses are determined, and then the index and gain representing the types of noise signals or codebooks stored in the noise memory are determined. These calculations are performed for each subframe obtained by dividing the frame into predetermined section lengths. In this case, what is transmitted as the sound source signal are the amplitude and position of the multipulse and the index and gain of the noise signal.

Furthermore, in the case of rupture, the amplitude and position of a predetermined number of multipulses are determined over the entire frame.

In the case of vowel character, the encoding circuit 230 encodes the amplitude g1 and position m1 of the multi-pulse in the representative section using a predetermined number of bits and outputs the encoded signal. Further, the information P+ indicating the subframe of the representative section, the amplitude correction coefficient Ck, and the phase correction coefficient dk are encoded with a predetermined number of bits and output to the multiplexer 260. Furthermore, these are decoded and output to the drive signal restoration circuit 283. In the case of nasality or plosiveness, the amplitude and position of the multipulse are encoded and output to the multiplexer 260, and also decoded and output to the driving sound source restoration circuit 283. In the case of friction, the amplitude and position of the multi-pulse are encoded, and the gain and index of the noise signal are encoded and output to the multiplexer 260, and at the same time, these are decoded and output to the drive sound source restoration circuit 283.

In the vowel interval, the drive sound source restoration circuit 283 converts the frame into the sound source signal calculation circuit 2 using the average pitch period T''.
The information P indicating the subframe of the representative section and the decoded amplitude and position of the multi-pulse in the representative section are used to generate multi-pulses in the representative section. In other pitch sections, the sound source signal Vk(n) is restored using the multipulse of the representative section, the decoded amplitude correction coefficient, and the decoded phase correction coefficient according to the equation (7).

On the other hand, multi-pulses are generated in the nasal, plosive, and fricative sections. In the frictional section, the noise signal is further accessed from the noise memory 225 using the index of the noise signal, and is multiplied by a gain to restore the drive sound source signal. Details of drive sound source signal restoration in a frictional section can be found in the above-mentioned document 2.
can be referred to.

In the vowel interval, the interpolation circuit 282 converts the linear prediction coefficients into parameters at once, interpolates on the parameters for each subframe interval of pitch period T', and inversely converts them into linear prediction coefficients for output. Note that the interpolation can use not only the parameters but also other well-known parameters such as the logarithmic cross-sectional area ratio. Interpolation is not performed in nasal or consonant intervals.

The synthesis filter 281 inputs the restored driving sound source signal and the linear prediction coefficient al' to obtain a synthesized speech signal for one frame, and also calculates the influence signal for the next frame for one frame. It is calculated and output to the subtracter 190. The method for calculating the influence signal is based on Japanese Patent Application Laid-open No. 59-11.
6794 (Reference 7) etc.

The multiplexer 260 receives a code representing the sound source signal, a code representing the type of audio of the frame, a code representing the subframe position of the representative interval in the vowel interval, a code representing the average pitch period,
and the code representing the parameter are combined and output.

Although the above is an explanation of one embodiment of the present invention, it is only one configuration of the present invention, and various modifications thereof can be considered.

For example, the strength of the periodicity of the impulse response of the synthesis filter may be determined only in the vowel section of the speech. Further, in the above embodiment, the frictional section is represented by a sound source signal with a small number of multipulses and a noise signal, but this can also be represented by the well-known method of 5 tochasjic codiB. For details on this method, see, for example, “Code-e
xcited 1inear prediction (
CELP): High quality 5peec
h at very low bit, r
aes,” (ICASSP, 937-940.1
985) (Reference 8).

Further, as the noise signal stored in the noise memory 225, a white noise signal having a predetermined probability density characteristic (for example, Gaussian distribution, etc.) may be stored, or a large amount of audio signals may be predicted in advance. It may be a value calculated by learning from the prediction residual signal obtained by the calculation. For the former method, reference can be made to the above-mentioned document 6. Regarding the latter method, for example, °Vec by Makhou et al.
tor Quantization in 5peec
h Coding,” (Proc, IEEE, vol.
73, 11.1551-1588.1985) (Reference 9).

Furthermore, in the embodiment, the frame audio signals are classified into four types: vowel, nasal, fricative, and plosive, and different sound source signals are used, but the number of classifications may be changed.

In addition, in the example, the parameters were encoded as spectral parameters and LPG analysis was used as the analysis method, but the spectral parameters may be the well-known parameters such as LSP, LPC cepstrum, cepstrum, improved cepstrum, generalized cepstrum. , melkebstrum, etc. can also be used. Furthermore, it is possible to use the optimal analysis method for each parameter.

Further, other well-known methods can be used as the parameters to be interpolated in the interpolation circuit 282 and the interpolation method thereof. A specific interpolation method is, for example, '5peech Analysis and Syn' by A Simi et al.
thesis by Linear Predicti
on of 5peechWave” (J
, Acoust, Soc, An+, , pp, 637
-655.1971) (Reference 10) etc.

Furthermore, in the vowel section, the amplitude correction coefficient ck and the phase correction coefficient dk were determined and transmitted for pitch sections other than the representative section, but the decoded average pitch period T' was calculated for each pitch section using the adjacent pitch period. It is also possible to adopt a configuration in which the phase correction coefficient is not transmitted by interpolation. Also, the amplitude correction coefficient is not transmitted for each pitch section, but the value of the amplitude correction coefficient obtained for each pitch section is shaped like a least squares curve or a least squares straight line, and the coefficient of the curve or straight line is transmitted. These methods can be used in any combination. These configurations can reduce the amount of information for transmitting correction information.

In addition, as a phase correction coefficient, for example, "2.4 kbps Pitch Predic" by Ono, O□awa et al.
tion Multi-p (11se 5peech
Coding” (Proc, ICASS
PS4.9.1988) As described in <Reference 11>, a configuration in which the linear phase term τ is determined at the edge of the frame, this is distributed to each pitch section, and the phase correction coefficient is not determined for each pitch section. It is also possible to do this.

In addition, in order to significantly reduce the amount of calculation, in the vowel interval,
The representative section may be fixed to a predetermined section within the frame (for example, the pitch section approximately in the center of the frame, the pitch section with the largest power within the frame, etc.), or the representative section may not be searched. In this case, calculation of equations (9) and (10) for the candidate section of the representative section becomes unnecessary, making it possible to significantly reduce the amount of calculation, but the sound quality deteriorates.

Furthermore, in order to further reduce the amount of calculation, calculation of the influence signal can be omitted. As a result, the drive signal 1 summer source circuit 283, the interpolation circuit 282, the synthesis filter 281,
The subtracter 190 becomes unnecessary and the amount of calculation can be reduced, but
Again, the sound quality deteriorates.

Note that, as is well known in the field of digital signal processing, the autocorrelation function corresponds to the power spectrum on the frequency axis, and the cross-correlation function corresponds to the cross-power spectrum, so it can also be calculated from these. These calculation methods are described in “Di
Digital Signal Processing”
(Prentice-Hall, 1975).

〔Effect of the invention〕

As described above, according to the present invention, audio signals are classified into several types based on phonetic characteristics, and an fS sound source signal suitable for the classification is used. Among the pitch sections divided into pitch periods, the multi-pulse of one pitch section (representative section) and the other pitch sections are expressed using at least one of the amplitude correction coefficient and the phase correction coefficient, and the frictional section. So,
It is determined that the sound source signal is represented by a combination of a small number of multi-pulses and a noise signal, and the strength of the periodicity of the impulse response of the synthesis filter composed of the spectral parameters.If the periodicity is strong, Since the spectral parameters are weighted to widen the bandwidth of the synthesis filter, it is possible to prevent underestimation of the bandwidth of the synthesis filter even for female voices that are particularly ugly or soft, resulting in good synthesized speech. It has the effect of being able to obtain As a result, the characteristics of the voice that are important for phonological perception and perception of naturalness change (voiced transient parts, vowel This has the great effect of making it possible to obtain synthesized speech with almost no deterioration in sound quality, even in the case of changes between the two.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of a voice classification method, and FIG. 3 is an explanatory diagram showing a representative section in a voiced frame and multipulses in the representative section. It is. 110...Buffer memory, 130...Pitch analysis circuit, 140...K parameter calculation circuit, 145...
- Periodicity discrimination circuit, 150... pitch encoding circuit, 1
60...K parameter encoding circuit, 170...Impulse response calculation circuit, 180...Autocorrelation function calculation circuit, 200...Weighting circuit, 210...Cross correlation function calculation circuit, 215. ... Discrimination circuit, 220 ... Sound source signal calculation circuit, 225 ... Noise memory, 230 ...
・Encoding circuit, 260...Multiplexer, 270・
... Amplitude/phase correction coefficient calculation circuit, 281 ... Synthesis filter, 282 ... Interpolation circuit, 283 ... Drive signal restoration circuit.

Claims (1)

[Claims] A spectral parameter representing the spectral envelope and a pitch parameter representing the pitch are obtained for each frame from the input discrete audio signal, and when the impulse response of a filter configured using the spectral parameters has strong periodicity, the spectral parameters are The audio signal is classified into a plurality of types by weighting and discriminating parameters representing the characteristics of the audio signal are extracted, and the frame section is used as the source signal of the audio signal for each frame according to the type. The multi-pulse is divided into small sections according to the pitch parameter, and the multi-pulse obtained in one of the small sections, correction information or a codebook for correcting at least one of the amplitude or phase of the multi-pulse, and the multi-pulse are divided into small sections according to the pitch parameter. A voice encoding method characterized by determining and outputting at least one of a pulse and a pulse.