JPH07160295A - Voice encoding device - Google Patents

Abstract

PURPOSE:To realize a voice encoding system in which degradation of voice quality is small even if a voice signal is encoded with a low bit rate. CONSTITUTION:A voice signal is inputted, stored in a buffer memory 110 for each frame, divided into a sub-frame by a sub-frame dividing circuit 150. A masking threshold value of hearing sense is calculated by a masking threshold value calculating circuit 205, further, this is converted to a filter coefficient of a filter weighing hearing sense. In a hearing sense weighing circuit 202, hearing sense weighing is performed using this filter coefficient, and in a voice source code book searching circuit 230, a code book or a multi-path is searched.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding apparatus for coding a speech signal with a low bit rate, especially at a high quality of 8 to 4.8 kb / s.

[0002]

2. Description of the Related Art As a method for encoding a voice signal at a low bit rate of about 8 to 4.8 kb / s, for example,
M. Schroeder and B.I. "Code-excited linear pre by Atal
Diction: High quality speed
ch at very low bit rates "
(Proc. ICASSP, pp. 937-940, 1
985) (reference 1) and Kleijn et al., “Improved speech qual”.
ity and efficient vector
"Quantization in SELP" (ICASSP, pp.155-158, 198).
8 years) (reference 2) and other CELP (Cod
e Excited LPC Coding) method,
B. “A new model of by Atal et al.
LPC excitement for product
cing natural-sounding spe
ech at low bit rates ", (Pr
oc. ICASSP, pp. 614-617, 198
The multi-pulse coding method described in a paper (reference 3) and the like entitled 2) is known.

According to the methods of References 1 and 2, on the transmitting side, a spectrum parameter representing a spectrum characteristic of a voice signal is extracted from the voice signal for each frame (for example, 20 ms), and the frame is further divided into sub-interval subframes (for example, 5 ms). An adaptive codebook that represents a long-term correlation (pitch correlation) so as to minimize a weighted squared error between a reproduction signal that is divided and reproduced based on a past excitation signal for each subframe and the excitation signal. A pitch parameter is extracted, a long-term prediction of the voice signal of the subframe is performed based on the pitch parameter, and a residual signal obtained by the long-term prediction is selected, and a signal selected from a codebook including a noise signal of a predetermined type. Assuming that one type of noise signal is selected so as to minimize the weighted squared error between the signal synthesized by the above method and the voice signal, To calculate the optimal gain. Then, the index and the gain indicating the type of the selected noise signal, and the spectrum parameter and the pitch parameter are transmitted. A description of the receiving side is omitted.

[0004]

In the conventional method of the above-mentioned document 1, when searching for a codebook composed of multipulses, adaptive codebooks, and noise signals, the input speech signal and the code are used as error evaluation measures. A weighted squared error with a signal reproduced by a book or multi-pulse is used. Here, such a weighting scale is shown by the following equation.

[0005]

[Equation 1]

W (z) is the transfer characteristic of the weighting filter. a _i is a linear prediction coefficient obtained from the spectrum parameter. γ ₁ and γ ₂ ⁱ are constants for controlling the weighting amount, and are generally 0 <γ ₂ <γ ₁ <1.

However, since this evaluation scale does not always match the auditory perception, the sound quality of the reproduced voice reproduced using the code vector selected by this scale or the obtained multi-pulse is not necessarily sufficient for hearing. There was a problem. In addition, this problem was particularly noticeable when the bit rate was reduced and the codebook size was reduced.

[0008]

According to the first aspect of the invention, an input discrete voice signal is divided into frames of a predetermined time length, and a spectrum parameter representing a spectrum envelope of the voice signal is obtained. A spectrum parameter calculation unit for outputting and dividing the frame into small sections of a predetermined time length, and pitching so that a signal reproduced based on an adaptive codebook composed of past sound source signals becomes close to the voice signal. An adaptive codebook section for finding parameters,
In a speech coding apparatus having a sound source search unit that outputs a sound source signal of the speech signal by expressing it as a codebook or a multi-pulse composed of a plurality of types of pre-configured code vectors, based on the auditory masking characteristic from the speech signal. A masking calculation unit for obtaining a masking threshold value, and a coefficient of the auditory weighting filter based on the threshold value, and a signal reproduced by the adaptive codebook, the codebook, or the multipulse, and the voice signal. An auditory weighting unit that weights the error signal with the weighting filter, and a search unit that obtains an optimal code vector from the codebook so as to reduce the power of the weighting value, or obtains and outputs multiple pulses. A speech coding apparatus characterized by having .

According to a second aspect of the invention, in the speech coding apparatus of the first aspect of the invention, a band dividing unit for band-dividing an input speech signal and the auditory weighting filter for the band-divided signal are used. A speech coding apparatus is provided which has an auditory weighting unit for weighting, and a search unit for performing codebook search or multipulse search so as to reduce the power of the weighting value.

Further, according to the third invention, in the speech coding apparatus according to the second invention, a bit allocation section for allocating a quantization bit to a band-divided signal, and a bit of a codebook according to the allocation bit. It is possible to obtain a speech coder having a switching unit for switching the number or the number of multi-pulses.

[0011]

The operation of the speech coder according to the present invention will be described.

According to the present invention, in each subframe obtained by dividing a frame, in the adaptive codebook search, the excitation codec search or the multipulse calculation, the signal reproduced by the adaptive codebook, the codebook or the multipulse. And the error signal with respect to the input audio signal are weighted by the auditory sense calculated based on the auditory masking characteristics. A masking threshold value is obtained based on the auditory masking characteristic, and is converted into a power spectrum based on the threshold value on the frequency axis, and further converted into a filter coefficient, and the error signal is converted into the filter coefficient. It is characterized in that a weighting error measure is obtained by weighting with a perceptual weighting filter composed of, and an optimum code vector is obtained from the codebook so as to reduce the error measure. That is, the error power weighted by the following equation is minimized.

Explaining the search of the sound source codebook as an example, the optimum sound source code vector is searched from the sound source codebook so as to minimize the following equation.

[0014]

[Equation 2]

Here, x (n) is the speech signal after the long-term prediction signal by the adaptive codebook is removed, and c _j (n) is the j-th code vector in the excitation codebook (however,
j = 0 to 2 ^B −1: B is the number of bits in the codebook), γ
_j is the optimum gain. h (n) is the impulse response of the synthesis filter composed of spectral parameters. Further, w _m (n) is an impulse response of an auditory weighting filter that performs weighting with a filter coefficient obtained by conversion from an auditory masking threshold value.

Here, to obtain the masking threshold value,
For example, FFT an audio signal to obtain a power spectrum | X
(K) | ² asking, which was analyzed by a critical band filter or auditory model, the power or RMS of each critical band is calculated to determine the masking threshold in each critical band from these values. As a method of obtaining the masking threshold value, for example, a method using a value obtained by an auditory psychology experiment is known, and details are described in "Transform coding offaud" by Johnston.
io signals using perceptu
al noisecriteria "(IEEE J.
Sel. Areas on Commun. , Pp31
4-323, 1988) and a paper (reference 4),
R. "Ve by Drgo de Iacovo and others
center quantitation and perce
ptu terrier in SVD bas
ed CELP ceders ”(ICA
SSP, pp. 33-36, 1990) (Reference 5) and the like. For the critical band filter or the critical band analysis, see, for example, J. "Foundation of modern au" by Tobias
You can refer to Chapter 5 (reference 6) of the book titled "Theory theory." For the auditory model, see "A computer" by Seneff.
regional model for the peri
general audio system: Ap
application to speech recog
"state research" (Pr
oc. ICASSP, pp. 1983-1986, 19
1986) (reference 7) and the like.

Next, the masking threshold value obtained for each critical band is converted into power, and the autocorrelation function is obtained by inverse FFT of the set of power.

The filter coefficients are calculated from the autocorrelation function by using the well-known linear prediction analysis. Then, the impulse response w _m (n) of the perceptual weighting filter composed of these filter coefficients is calculated and applied to the equation (2) to select the code vector that minimizes the equation (2).

According to a second aspect of the present invention, in the first aspect of the present invention, there is provided a band division section for preliminarily band-dividing the input voice. Assuming that the number of band divisions is 2 and the frequency bandwidth of the input voice is f _w (Hz) for simplification, the band division in the band division unit is 0-f _w / 2 on the reduction side and is high. the frequency band may be divided f _w / 2-f _w. For band division, a method is known in which a signal is once FFT, frequency-divided on the FFT, and then inverse FFT, or a method using a QMF filter for band division. The latter method is described, for example, in Crochiere et al. , By them ”
Multirate Signal Processi
ng "(Prentice-Hall,
1983) (reference 8) and the like.

With respect to the voice signal of each band divided into bands,
After the auditory weighting filter coefficient is calculated by the method of the first aspect of the present invention and the auditory weighting is performed, a codebook or multi-pulse search is performed.

In a third invention, in the second invention, there is provided a bit allocation section for allocating a quantization bit to the band-divided audio signal of each band. As a method of bit allocation in the bit allocation unit, for example, a method of performing bit allocation using a power ratio between a band-divided reduction-side signal and a high-frequency side signal, or a masking threshold calculation unit When calculating the threshold value, bit allocation can be performed by using the ratio of the average value or minimum value of the masking threshold value on the reduction side to the average value or minimum value of the masking threshold value on the high frequency side. .

[0022]

1 is a block diagram showing an embodiment of a speech coder according to the present invention. Here for simplicity,
An example is shown in which a perceptual weighting filter based on a masking threshold is used in the search of the excitation codebook, but this weighting filter can also be used in the search of the adaptive codebook.

In the figure, on the transmitting side, the input terminal 100
Input the audio signal from 1 frame (for example, 20m
The audio signal of s) is stored in the buffer memory 110.

The LPC analysis circuit 130 uses the LSP as a parameter indicating the spectral characteristic of the voice signal of the frame.
Parameters from the voice signal of the frame to the known LPC
An analysis is performed and only a predetermined order L is calculated.

Next, the LSP quantization circuit 140 quantizes the LSP parameter with a predetermined number of quantization bits,
The obtained code l _k is output to the multiplexer 260, and the code is decoded to further obtain the linear prediction coefficient a _i ′ (i =
1 to L) and outputs to the weighting circuit 200, the impulse response calculation circuit 170, and the synthesis filter 281. L
For the method of encoding the SP parameter and converting the LSP parameter and the linear prediction coefficient, see "Quantizer design in LS" by Sugamura et al.
P speech analysis-synthes
a paper entitled "is" (IEEE J. Sel. Area
s Common. , Pp. 432-440, 1988
(Year) (Reference 9) and the like can be referred to. Also LSP
In order to quantize the parameters more efficiently, vector scalar quantization or other popular vector quantization methods can be used. For vector scalar scalar quantization of LSP, see "Transfo by Moriya et al.
rm Coding of Speech using
a Weighted Vector Quantiz
er, ”(IEEE J. Sel. Are
as, Commun. , Pp. 425-431,198
8 years) (Reference 10) and the like can be referred to.

The subframe division circuit 150 divides the audio signal of the frame into subframes. Here, for example, the subframe length is 5 ms.

The subtracter 190 uses the signal to synthesize the filter 2
The output of 81 is subtracted and output.

The adaptive codebook 210 is the synthesis filter 2
81 input signal v (n) via delay circuit 206, weighted impulse response h _w (n) from impulse response output circuit 170, and weighted signal from subtractor 190, and pitch prediction based on long-term correlation And then
The delay M and the gain β are calculated as pitch parameters.
In the following description, the prediction order of the adaptive codebook is 1, but it may be higher than 2nd order. For the calculation of the delay M in the adaptive codebook, it is possible to refer to the documents 1 and 2 mentioned above. Further, the gain β is calculated, the adaptive code vector x _z (n) is _calculated by the following equation, and the adaptive code vector x _z (n) is subtracted from the output of 190.

X _z = x (n) −β · v (n−M) * h (n) (3) where x (n) is the output signal of the subtractor 190 and v (n)
Is a past synthesis filter drive signal, and h (n) is an impulse response of the synthesis filter.

Next, the masking threshold value calculation circuit 205
Is an N-point FFT transform on the input speech signal x (n) to obtain a spectrum X (k) (k = 0 to N-1), and a power spectrum | X (k) | ^2. Is analyzed by a critical band filter or an auditory model to calculate power or RMS for each critical band. To calculate the power here, follow the formula below.

[0031]

[Equation 3]

Here, bl _i and hb _i indicate the lower limit frequency and the upper limit frequency of the i-th critical band, respectively. R is the number of critical bands included in the voice signal band, and the masking threshold value C (i) in each critical band is calculated and output from the value of equation (4). Reference 4 and the like can be referred to for obtaining the masking threshold value. Regarding the auditory model, reference can be made to Document 7 and the like. For the critical band, reference can be made to Document 6 and the like. Further, the masking threshold c (i) is converted into power, a power spectrum is obtained, and these are subjected to inverse FFT.
By doing so, the autocorrelation function r (j) (j = 0.
N-1) is calculated. Next, the filter coefficient b _i (i = 1 ... P) is calculated by performing a known linear prediction analysis on the P + 1 autocorrelation functions.

The perceptual weighting circuit 220 has an error signal x _z obtained by the adaptive codebook 210 according to the equation (2).
The weighting signal x _zm (n) is obtained by weighting (n) by the following equation using the filter coefficient b _i .

X _zm (n) = x _z (n) * w _m (n) (4) where w _m (n) is the impulse response of the auditory weighting filter formed by the filter coefficients b _i .

Here, as the auditory weighting filter, one having a transfer function of the following equation can be used.

[0036]

[Equation 4]

Here, the constant is 0 <r ₂ <r ₁ <1.

The source codebook search circuit 230 selects a source code vector so as to minimize the following equation.

[0039]

[Equation 5]

Where γ _j is the code vector c
_{j (n) (j = 0} .... 2 B -1; B is the number of bits of the excitation code book) is optimal gain for. The sound source codebook is, for example, learned and configured in advance.
A method of constructing a codebook by learning is, for example, Linde
Et al. “An Algorithmic Vect
or Quantization Design ”(IEEE Trans. COM-28, pp.
84-95, 1980) (Reference 11) and the like.

The gain quantization circuit 282 quantizes the gain of the adaptive codebook excitation codebook using the gain codebook 285.

The adder 290 is used by the adaptive codebook 210.
And the output sound source of the sound source codebook search circuit 230 are added by the following equation and output.

V (n) = β ′ · v (n−M) + γ ′ _j c _j (n) (7) The synthesizing filter 281 inputs the output v (n) of the adder 290 and synthesizes by the following equation. For one frame of voice, a sequence of 0 is input to the filter for another frame to obtain a response signal sequence, and the response signal sequence for one frame is output to the subtractor 190.

[0044]

[Equation 6]

Multiplexer 260 combines and outputs the output code sequences of LSP quantizer 140, adaptive codebook 210 and excitation codebook search circuit 230.

This is the end of the description of the embodiment of the present invention.

FIG. 2 is a block diagram showing an embodiment of the second invention. In the figure, the components having the same numbers as in FIG. 1 perform the same operations as in FIG.

In FIG. 2, the band dividing circuit 300 divides the band of the input audio signal. Here, for the sake of simplicity, the number of divisions is 2. The method of division uses the method using the QMF filter among the methods described in the section of action. Then, the low-frequency side signal and the high-frequency side signal are output.

The switch 310 is tilted upward when processing a low-frequency signal, and is tilted downward when processing a high-frequency signal.

This is the end of the description of the second embodiment of the invention.

FIG. 3 is a block diagram showing an embodiment of the third invention. In the figure, the constituent elements denoted by the same numbers as in FIG. 1 and FIG. 2 perform the same operations as in FIG. 1 and FIG.

In FIG. 3, the switches 320-1, 32 are shown.
0-2 switches the switch according to the low frequency side and the high frequency side, and outputs the low frequency signal and the high frequency signal. Also, switch 3
20-2 outputs information to the codebook switching circuit 350 whether the output signal is a low frequency band or a high frequency band.

The masking threshold value calculation circuit 360 calculates the masking threshold value in the entire band for the signal before band division, and divides them into the low band side and the high band side. Next, for each of the low band side and the high band side, the coefficient of the perceptual weighting filter is calculated by the same method as in the first aspect of the invention, and is output to the perceptual weighting circuit 220.

The bit allocation calculation circuit 340 uses the output of the masking threshold value calculation circuit 360 to allocate the quantization bit numbers on the low band side and the high band side by the method described in the operation, and switches the codebook. Output to the circuit 350.

The codebook switching circuit 350 inputs the quantization bit number from the bit allocation circuit 340,
The high frequency information is input from the switch 320-2 to switch between the sound source codebook and the gain codebook. here,
The codebook may be preliminarily constructed by learning using training data, or may be a random number codebook having predetermined statistical properties.

This is the end of the description of the third embodiment of the present invention.

In the second invention, the codebook can be switched between the low frequency side and the high frequency side.

As the critical band analysis filter, another well-known filter having an equivalent operation can be used.

As the method of calculating the masking threshold value, other known methods can be used.

Further, as the tone generator codebook, another well-known structure can be used. Regarding the method of constructing the sound source codebook, for example, C.I. "On reducing computerization" by Laflamme et al.
al complexity of codebook
search in CELP coder thro
uugh the use of algebraic
Codes ”(Proc. ICASSP,
pp. 177-180, 1990) (Reference 12),
I. "CELP: A ca by Mr. Trancoso
ndatefor GSM half-rate
Coding ”(Proc. ICAS
P, pp. 469-472, 1990) (Reference 13)
Etc. can be referred to.

As a sound source codebook, a more efficient codebook, for example, matrix quantization, finite state vector quantization, trellis quantization, Delayed De
The characteristics can be further improved by using a codebook based on the decision quantization. For more information on these methods, see, for example, Gray's "Vector Quantiza."
section "(IEEE ASSP Magazine,
pp. 4-29, 1984) (Reference 1
4) etc. can be referred to.

In the above embodiment, the case where the tone generator codebook has one stage has been described, but the tone generator codebook may have multiple stages. For example, the number of stages can be two. By doing so, the calculation amount required for the codebook search can be reduced. Also, a multi-pulse can be used as the sound source codebook.

Although the adaptive codebook is the primary,
The sound quality can be further improved by setting the second or higher order or the delay to a decimal value instead of an integer value. Detail is,
P. "Pitch predict" by Mr. Kroon and others
tors with high temporal r
"Esolution" (Proc. ICASSP, p
p. 661-664, 1990) and the like (Reference 15) can be referred to.

In the above embodiment, the LSP parameter is coded as the spectral parameter and the LPC analysis is used as the analysis method. However, as the spectral parameter, other well-known parameters such as LPC cepstrum, cepstrum, improved cepstrum, Generalized cepstrum, mel cepstrum, etc. can also be used. In addition, an optimal analysis method can be used for each parameter.

In the vector quantization of the LSP parameter, the LSP parameter may be vector-quantized after the nonlinear transformation corresponding to the auditory characteristic. As the non-linear transformation, for example, Mel transformation is known.

Further, the LPC coefficient obtained in the frame is set to LS.
The configuration may be such that interpolation is performed for each subframe on P or a linear prediction coefficient, and the adaptive codebook and excitation codebook are searched using the interpolated coefficient. With such a configuration, the sound quality is further improved.

The perceptual weighting based on the masking threshold shown in the embodiment can be used for searching at least one of adaptive codebook source codebook and gain codebook. It can also be used when quantizing spectral parameters, eg LSP.

In the embodiment, the case where the auditory weighting filter obtained from the masking threshold value is used for the search of the codebook has been described, but it can be applied to the calculation of multipulse.

When obtaining the perceptual weighting filter, not only the masking threshold value by the simultaneous masking but also the masking threshold value by the successive masking can be used together.

Further, it is also possible to multiply the masking threshold value by the weighting coefficient and then convert it into the coefficient of the auditory weighting filter without directly obtaining the coefficient of the auditory weighting filter from the masking threshold value. Further, the auditory weighting filter can use another known configuration.

In the third invention, other well-known methods for bit allocation, for example, power ratio between low frequency band and high frequency band, can be used.

[0072]

As described above, according to the present invention, in the adaptive codebook search, the excitation codebook search, or the multipulse calculation, the adaptive codebook, the excitation codebook, or the For the error signal between the signal reproduced by multi-pulse and the input voice signal, a masking threshold value is calculated for each critical band of the input voice, and the filter coefficient of the auditory weighting filter is calculated based on the masking threshold value. Is calculated and the weighted error scale is used by weighting the error signal using the weighting filter. There is a great effect that the rate can be reduced.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a first invention.

FIG. 2 is a block diagram showing an embodiment of the second invention.

FIG. 3 is a block diagram showing an embodiment of a third invention.

[Explanation of symbols]

 110 buffer memory 130 LPC calculation circuit 140 LSP quantization circuit 150 sub-frame division circuit 190, 195 subtractor 220 auditory weighting circuit 205, 360 masking threshold calculation circuit 210 adaptive codebook 230 excitation codebook search circuit 235 excitation codebook 282 Gain quantization circuit 260 multiplexer 281 synthesis filter 285 gain codebook 340 bit allocation circuit 350 codebook switching circuit

Claims (3)

[Claims] 1. A spectrum parameter calculation unit that divides an input discrete voice signal into predetermined time length frames, obtains and outputs a spectrum parameter that represents a spectrum envelope of the voice signal, and the frame is predetermined. An adaptive codebook section for obtaining a pitch parameter so that a signal reproduced based on an adaptive codebook consisting of past sound source signals is divided into small sections of a given time length, and the speech signal. In a speech coding apparatus having a sound source search unit that outputs a sound source signal by expressing it as a codebook or a multi-pulse that is composed of a plurality of types of pre-configured code vectors, a masking threshold based on auditory masking characteristics of the sound signal. A masking calculation unit for obtaining a value and a coefficient for the auditory weighting filter are obtained based on the threshold value. Therefore, an auditory weighting unit that weights the error signal between the audio signal and the signal reproduced by the adaptive codebook or the codebook or the multipulse by the weighting filter, and reduces the power of the weighting value. Thus, the speech coding apparatus is characterized in that it has a search unit for obtaining an optimum code vector from the codebook or obtaining and outputting a multipulse. 2. A band dividing unit for band-dividing an input audio signal, an auditory weighting unit for weighting the band-divided signal using the auditory weighting filter, and a code for reducing the power of the weighting value. The speech coding apparatus according to claim 1, further comprising a search unit that searches for a book or searches for multiple pulses. 3. A bit allocation section for allocating a quantization bit to a band-divided signal, and a switching section for switching the number of bits of a codebook or the number of multipulses according to the allocation bit. The speech encoding apparatus according to claim 2.