JPH0455899A - Voice signal coding system - Google Patents

Abstract

PURPOSE:To allow extremely efficient quantization and to provide the voice coding system which is good in sound quality even at a low bit rate by subjecting the spectral parameter of a voice signal to a nonlinear transformation corresponding to the characteristic of hearing sensation to make vector quantization. CONSTITUTION:An LPC analysis circuit 130, an LSP quantization circuit 140, an impulse response calculating circuit 170, a weighting circuit 200, and adaptive code book 210, a sound source code book retrieval circuit 230, a sound source code book 235, etc. are provided and the spectral parameter is determined by approximating the inputted discrete input signal by an all Polar type model. The frame is divided to small sections and the pitch parameter is so determined that the signal reproduced in accordance with the past sound source signal is approximated to the voice signal. The spectral parameter is subjected to the nonlinear transformed so as to correspond to the characteristic of hearing sensation by the voice coding system which outputs the sound source signal of the voice signal by expressing the same by the code book. This parameter is expressed by the previously constituted code book 210 and is outputted. The voice coding system which is good in sound quality at <=4.8kb/s is obtd. by the relatively small computation quantity and memory quantity in this way.

Description

DETAILED DESCRIPTION OF THE INVENTION C. Industrial Application Fields The present invention provides a method for processing audio signals at low bit rates, in particular from 8 to 4.
The present invention relates to an audio signal encoding method for high-quality encoding at approximately 8 kb/s.

C. Prior Art] As a method for encoding an audio signal at a bit rate of about 8 to 4.8 kb/s, for example, M, 5 chroed
erand B, “Code-exc” by Mr. Ata1
ited 1inear predi-ction:
High quality 5peech at ve
ry low bitrates” (Proc, I
CELP (C
Excited LPC Coding) encoding method and A new mode by Mr. B.Ata1 et al.
l of LPCexcitation pro-du
singing natural-sounding 5pe
A multi-pulse encoding method is known, which is described in a paper titled "Ech at Low Bit Ra-tes" (Reference 2).

In the former method, on the transmitting side, the spectral characteristics of the audio signal are calculated from the audio signal for every frame (for example, 2 S).
All-pole) model is used to extract spectral parameters representing spectral characteristics (for example, PARCOR coefficients, I, and SP coefficients). Next, the frame is further divided into small interval subframes (for example, 5m5) 4::, and a pitch parameter (also called an adaptive codebook) representing a long-term correlation (pitch correlation) is calculated based on the past sound source signal for each subframe. The subframe audio signal is extracted, long-term prediction is performed using the pitch parameter, and the residual signal obtained through long-term prediction is synthesized with a signal selected from a codebook consisting of predetermined types of noise signals. One type of noise signal is selected so as to minimize the error power between the signal and the audio signal, and the optimal gain is calculated. Then, the index and gain indicating the type of the selected noise signal, as well as the spectral parameter and pitch parameter are transmitted.

[Problem to be solved by the invention]

In the conventional method described in the above-mentioned document 1, in order to obtain high sound quality, it is necessary to express the spectral characteristics of the audio signal well by spectral parameters. The spectral parameters are quantized for transmission from the transmitting side. As an efficient quantization method, for example, a vector-scalar quantization method using LSPSP coefficients is known. A specific method is, for example, “Tra
nsform CodjBof 5peech u
sing a Weighted Vector
Quantizer.

A paper entitled (IEEE J, Set, ^reas,
Comn+un,.

pp. 425-431.1988) (Reference 3).

In this method, after quantization and decoding, the original LSPSP coefficients obtained for each frame are used as spectral parameters.
The error signal between the SP and the quantized and decoded LSP is scalar quantized. Here, the vector quantization codebook is constructed by training a codebook consisting of 2B types of code vectors (B is the number of bits for spectral parameter quantization) against a large amount of spectral parameter database. . The codebook training method is, for example, “An
Algorithm for Vector Q
A paper entitled ``Antification Design'' (IEEE Trans, C0M-28, pp. 84
-95°1980) (Reference 4), etc.

In order to realize an audio encoding system of 4.8 kb/s or less, the number of quantized bits of spectral parameters must be reduced to 20 bits or less per frame while keeping the distortion caused by spectral parameter quantization below the auditory perceptible limit. There is a need to. Conventional methods were insufficient for this purpose, and when the number of quantization points was reduced to 20 bits or less, the sound quality deteriorated significantly.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems and provide an audio signal encoding method that achieves good sound quality at 4.8 kb/s or less with a relatively small amount of calculation and memory.

[Means to solve the problem]

The first invention divides an input discrete audio signal into frames of a predetermined time length, approximates the audio signal using an all-pole model, and obtains spectral parameters representing the spectral envelope of the audio signal. , the frame is divided into small sections of a predetermined time length, a pitch parameter is determined so that the signal reproduced based on the past sound source signal is close to the sound signal, and the sound source signal of the sound signal is determined in advance. The speech encoding method is characterized in that the spectral parameters are non-linearly transformed so as to correspond to auditory characteristics, and the spectral parameters are represented and output using a pre-configured codebook or multi-pulse. .

The second invention divides an input discrete audio signal into frames of a predetermined time length, and calculates spectral parameters representing the spectral envelope of the audio signal by dividing the audio signal using an all-pole model. The frame is divided into small sections of a predetermined time length, the pinch parameter is determined so that the signal reproduced based on the past sound source signal is close to the sound signal, and the sound source signal of the sound signal is determined. In a speech encoding method that outputs a pre-configured codebook or multi-pulse representation, the spectral parameters are non-linearly transformed to correspond to auditory characteristics, and the difference or prediction error between frames is calculated for the non-linearly transformed spectral parameters. , or the difference or prediction error between parameters in the same frame is represented by a pre-configured codebook and output.

[Effect]

The operation of the audio signal encoding method according to the present invention will be explained.

In the following explanation, LSP
However, other well-known parameters may be used, such as PARCOR, cepstrum.

Mel cepstrum etc. can be used in the same way. For information on how to obtain LSP, see “Quantizer design in LSP” by Sugat*ura et al.
5peech analysis-synthesis
” (IEEE, J, Se1. Are
as.

Common, pp. 432-440.1988) (Reference 5).

In the first invention, LSP parameters are obtained by nonlinear transformation so as to correspond to auditory characteristics. It is known that the characteristics of hearing are nonlinear on the frequency axis, with high resolution in the low range and low resolution in the high range. Mel transformation is known as a nonlinear transformation that meets these characteristics. Regarding Mel transformation of spectral parameters, a method of converting from a power spectrum and a method of converting from an autocorrelation function are known. Details of these methods can be found, for example, in "Linear p" by Mr. 5trube.
redic-tion on a warped fr
The paper entitled “Equency 5cale” (J, A
coust, Soc, ^-,, pp, 1071-1
076, 1980) (Reference 6).

By the method described above, a vector quantization codebook is constructed in advance by training for the nonlinearly transformed LSP. For the method of configuring the vector quantization codebook, reference can be made to the above-mentioned document 4 and the like. The LSP is vector quantized using the following formula.

In other words, by searching the code vector for 2”[! class, (1)
Select the codevector that minimizes the distance of Eq.

Here, LSP' i j is the j-th codevector of the vector quantization codebook, and P is the order of the LSP.

In equation (1), the code vector was searched using the squared distance on the nonlinearly transformed LSP, but as another method,
For example, the nonlinearly transformed LSP is
(n) etc., and the code vector may be searched using the squared distance on the melkebstrum as shown in equation (2). Melkebstrum corresponds to a nonlinearly transformed logarithmic spectrum, so better sound quality can be obtained with this method.

Furthermore, instead of the square distance, a weighted square distance as shown in the following equation can also be used.

(j=1~2 m) (3
) Here, w and (n) can be weighted mel cepstrum or mel cepstrum. here,
δ is a weighting coefficient that determines the degree of auditory weighting,
It takes the value 0<δ<1.

The conversion method from the power spectrum to the mel-kebstrum is described, for example, in a paper entitled "Information compression of speech using the mel-kebstrum" by Mr. Kitamura et al. (Transactions of the Institute of Electronics and Communication Engineers, pp. 1092-1093, 1984) (Reference 7). )
etc. can be referred to.

Note that the calculation of the mel cepstrum corresponding to the code vector of the mel LSP may be performed when vector quantizing the mel LSP, or the mel cepstrum corresponding to the code vector of the mel LSP may be calculated in advance and the mel cepstrum code You may keep it as a book. The latter can significantly reduce the amount of calculation.

In the second invention, a difference is obtained between frames of nonlinearly transformed LSP as shown in the following equation, and a vector quantization codebook is created for this difference.

LSP, 1a-LSP, ILQLSP'-i'-'
(4) Here, LSP, , L, QLSP'
The L issue is the i-th Mel L in the L frame.
SP represents the i-th quantized and decoded mel LSP in the L-1th frame.

Alternatively, instead of the inter-frame difference, a difference between orders in the same frame may be determined, and a vector quantization codebook for this may be created.

DLSPIIH=LSP, fLLSP−ffi−+L(
However, LSPIIO=0)...(5) Furthermore, in the above equations (4) and (5), the prediction error may be calculated instead of the difference. The prediction coefficient for prediction is
It may be fixed, or a codebook of prediction coefficients may be configured in advance and an optimal one may be searched for.

〔Example〕

FIG. 1 is a block diagram showing an audio signal encoding device implementing the audio encoding method according to the first invention.

Input the audio signal from the input terminal 100, and input the audio signal for one frame (
For example, an audio signal of 2 hs) is stored in the buffer memory 110.

The LPG analysis circuit 130 performs well-known LPG analysis on the frame audio signal to calculate a linear prediction coefficient α1 of a predetermined order P as a parameter representing the spectral characteristics of the frame audio signal.

The LSP quantization circuit 140 performs vector-scalar quantization of the mel LSP. As shown in Figure 2, the process flow is as follows:
The linear prediction coefficient α is first converted into an LPC cepstrum (step 311), and the LPC cepstrum is further converted into a mel cepstrum C(n) (step 512).
). The conversion from the linear prediction coefficient α to the LPC cepstrum is described in “Effective~ness” by Mr. Ata1.
of 1inear prediction cha
racteristics of the 5peech
5ave for automatic 5peak
er 1 denti-fication and ver.
ification” (J, Acoust, Soc,
^m.

pp. 1304-1312.1974) (Reference 8). Further, for the conversion from cepstrum to mercebstrum cII(n), reference can be made to the above-mentioned document 7. Furthermore, the mel cebstrum is inversely transformed to the mel LPC cepstrum and then to the mel linear prediction coefficients (
Step 513). Furthermore, by converting this into LSP, mel LSP (LSP, . . . ) is calculated (step 514). The mel LSP is first subjected to vector quantization using the codebook 145 configured in advance (step 515). Here, the vector quantization codebook search uses the squared distance or the weighted distance on the norcepstrum, as described in the section on effects. In other words, the mel cepstrum CJ corresponding to the mel LSP obtained from the input voice, and the mel cepstral codebook 14 corresponding to each code vector of the mel LSP codebook 145.
The distance to the melkebstrum Cmj' (n) in 6 is determined by equation (2) or equation (3), and a code vector j that minimizes this distance is selected. Then, the mel LSP code vectors LSP', , corresponding to the selected code vectors c, j(n).

is selected from the LSP codebook 145.

Next, input mel LSP and vector quantized mel L
The SP error is calculated using the following equation (step 516).

ΔLSP,,=LSP,,-LSP',,,
(6) The error ΔLSP, , is scalar quantized using a scalar quantizer with a predetermined number of bits for each order i (step 517). Here, the quantization range (minimum value, maximum value) of the scalar quantizer is determined in advance using a large amount of training error signals ΔLSP, .

Furthermore, the range of scalar quantization is limited by utilizing the fact that the mel LSP coefficients have the following relationship.

For a specific method, reference may be made to Japanese Patent Application No. 2-42955 (Document 9).

LSP□〈・・・・・・＜LSpHr
(7) The code obtained by vector quantization and scalar quantization is output to the multiplexer 260, and the decoded mel LSP (QLSP□) is determined by the following equation.

QLSP,, = LSP'□, +Δt, sp'11.
(8) Here, ΔLSP'□ is an error signal decoded by scalar quantization. Then, the mouth LSP, , decodes the linear prediction coefficients, 1. '(1,=l~
P). The inverse transformation can be performed by retracing the aforementioned transformation from linear prediction coefficients to mel LSP. The decoded linear prediction coefficients a,' are output to the impulse response calculation circuit fWI70.

The subframe division circuit 150 divides the frame input audio signal into subframes. Here, for example, the frame length is 20Ims and the subframe length is 5a+s.

The weighting circuit 200 performs well-known perceptual weighting on the signal divided into subframes. For details of the perceptual weighting function, refer to the above-mentioned document 1.

The subtracter 190 performs the weighting from the signal to the synthesis filter 281.
Subtract and output the output of .

The adaptive codebook 210 inputs the input signal v(n) of the synthesis filter 281 via the delay circuit 206, and further receives the weighted impulse response signal v(n) from the impulse response output circuit 170, and the signal from the subtracter 190. manually,
Pitch prediction is performed based on long-term correlation.

Delay M and gain β are calculated as the other parameters.

In the following explanation, the prediction order of the adaptive codebook is assumed to be 1, but it can also be set to a higher order than 2.

The calculation method for the delay M and gain β in the first-order adaptive codebook is described by Kleijn "Improved 5pee
ch quality and effective v
ector quantization in 5EL
A paper entitled P' (ICASSP, pp, 155-1
58.1988) (Reference 10). Further, the obtained gain β is quantized and decoded using a predetermined number of quantization bits to obtain a gain β', which is used to calculate a prediction signal L(n) according to the following equation and output to the subtracter 205. The multiplexer 26
Output to 0.

L(n) = β'-v(n-M)*h,
(r+) (9)
In equation (9), ν(n-M) is a past sound source signal and is an input signal of the synthesis filter 281. h and (n) are the weighted impulse responses obtained by the impulse response calculation circuit 170.

The delay circuit 206 delays the synthesis filter input signal v (n) by one subframe and outputs the delayed signal to the adaptive codebook 210 .

The subtracter 205 subtracts the output of the adaptive codebook 210 from the output signal of the subtracter 190 according to the following formula, and obtains a residual signal e.
, 1(n) to the sound source codebook search circuit 23o.

e, (n) = x, (n) - (n) 0ω The impulse response calculation circuit 170 calculates the impulse response hw(n) of the auditory weighted synthesis filter for a predetermined number of samples. For a specific calculation method, reference can be made to the above-mentioned document 1, etc.

The sound source codebook search circuit 230 searches a sound source codebook 235 constructed by learning from Gaussian noise signals or a large amount of speech signals, and finds the optimal code vector c.
Search r
The above-mentioned document 4, document 9, etc. can be referred to. Further, for the optimal code vector search method and the optimal gain γ calculation method, reference can be made to the aforementioned document 1, document 9, and the like.

Furthermore, the codebook search circuit 230 quantizes the obtained gain with a predetermined number of quantization bits, and adds a quantization gain T' to the selected optimal code vector Cj (n).
is multiplied by the following formula to obtain the sound source signal q(n), and the adder 29
Output to 0.

q(n)-γ'cJ(n)
The 00 adder 290 adds the output excitation signal of the adaptive codebook 210 and the output excitation signal of the codebook search circuit according to the following formula, and outputs the result to the synthesis filter 281.

v(n) = β'v(n-M) + 7 cj(n)
02) The synthesis filter 281 is the adder 2
Input the output v (n) of 90 and obtain one frame of synthesized speech using the formula below, and for another frame, input the O sequence to the filter to obtain the response signal sequence, and obtain one frame of synthesized speech. The response signal sequence of is output to the subtracter 190.

However, in Equation 0, δ is an audible weighting coefficient that determines the degree of audible weighting in the circuit 200, and 0
〈δ〈takes the value of 1.

Multiplexer 260 includes LSP quantizers 140 . Adaptive codebook 21O1 sound source codebook search circuit 230
The output code sequences of are combined and output.

A block diagram showing an audio signal encoding device implementing the audio signal encoding method of the second invention is shown in FIG. In the figure, the components labeled with the same numbers as in FIG. 1 operate in the same manner as in FIG. 1, and therefore their explanation will be omitted. In Figure 3, LS
P@ child circuit 340. LSP codebook 345. The mel cepstrum codebook 346 is different from that in FIG.

FIG. 4 is a diagram showing the processing flow of the LSP quantization circuit 340.

In FIG. 4, the linear prediction coefficient α is first converted into an LPC cepstrum (step 521), and the LPC cepstrum is further converted into a mel cepstrum (step 522). Furthermore, the mel cepstrum is inversely transformed to the mel LPC cepstrum, and then inversely transformed to the mel linear prediction coefficient (
Step 523). Furthermore, by converting this to LSP, mel LSP (LSP, i) is calculated (
step 524).

For Mel LSP, the difference between frames or the Mel LSP between orders is calculated based on equation (4) or equation (5).
(step 525). In the following, a case will be described in which a difference signal is obtained between orders in the same frame. Then, vector quantization is performed on the difference signal using the Mel LSP codebook 345 that has been learned and configured in advance for a large amount of difference signals (step 5
26). Here, the vector quantization codebook search is
As mentioned in the section on the effect, the square distance on the Melkebstrum or the weighted square distance is used. In other words, the mel cepstrum C,(n) corresponding to the mel LSP obtained from the input voice and the mel LSP codebook 34
Melkebstrum CIIj' (n
) is determined by equation (2) or (3), and a code vector j that minimizes this distance is selected. Then, a mel LSP code vector LSP'□j corresponding to the selected code vector C+*j (n) is selected from the LSP code book 345. Then, by the following formula, Mel L
The SP is decoded (step 527).

LSP', 1=LSP', i-1+Δt, sp',...
Q5) Here, Δt, sp',...
This is a code vector selected in the Mel LSP codebook 345.

Next, the error between the input mel LSP and the vector quantized decoded mel LSP is determined (step 528), and the result is scalar quantized and output in the same way as the LSP quantization circuit 140 in FIG. 1 (step 529).

In each of the above-mentioned embodiments, an example applicable to the CELP method of Document 1 has been described, but it is also possible to apply to other well-known methods. For example, it can also be applied to improved CELP in which a sound source signal is represented by two different codebooks. For the improved CELP method, reference can be made to the above-mentioned document 9 and the like. Furthermore, it is also possible to apply the multi-pulse method described in Document 2 and the like.

Furthermore, in each embodiment, an example was explained in which spectral parameters are vector-scalar quantized in the auditory domain.
Vector quantize the spectral parameters in the auditory domain,
The difference between the original parameter and the vector-quantized decoded parameter can also be vector-quantized using, for example, a random number codebook having predetermined statistical properties. In this way, it becomes possible to further reduce the number of quantization bits.

In addition, quantization methods for Mel LSP include matrix quantization and trellis quantization, which are more efficient.

For details on these quantization methods, which can be applied to finite state vector quantization methods, etc., see @Vect by Mr. Gray.
or quantization” (J
EERASSPMag, pp, 4-29.1984
) (Reference 11) etc.

Furthermore, in order to reduce the amount of memory, the mel-kebstrum codebook 146.346 may be removed, and the mel-LSP code vector may be converted to a mel-cepstrum during vector quantization of the mel-LSP.

Furthermore, a codebook search may be performed on the Mel LSP.

〔Effect of the invention〕

As described above, according to the present invention, vector quantization is performed by subjecting the spectral parameters of an audio signal to nonlinear exchange corresponding to auditory characteristics, or parameters subjected to nonlinear transformation are changed between frames or at the same time. Since the difference between orders is determined within a frame and this difference is vector quantized, extremely efficient quantization is possible, and the effect is that it is possible to provide an audio encoding system with good sound quality even at a low bit rate.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an audio signal encoding device implementing the audio signal encoding method according to the first invention, and FIG.
FIG. 3 is a block diagram showing the processing flow of the LSP quantization circuit shown in FIG.
It is a figure which shows the flow of a process of the LSP quantization circuit of a figure. 110... 130... 140, 340... 145, 345... 146, 346... Buffer memory LPG calculation circuit LSPii childization circuit LSP code book Mel cepstrum code book Subframe division circuit Impulse response calculation @ path subtractor Weighting circuit Delay circuit Adaptive codebook Sound source codebook search circuit Sound source codebook Synthesis filter 190° 200 ・ 206 ・ 210 ・ 230 ・ 235 ・ 281 ・ 260 ・・・・・・Multiplexer

Claims (2)

[Claims] (1) Divide the input discrete audio signal into frames of a predetermined time length, approximate the audio signal using an all-pole model to obtain spectral parameters representing the spectral envelope of the audio signal, and is divided into small sections of a predetermined time length, and the pitch parameter is determined so that the signal reproduced based on the past sound source signal is close to the audio signal, and the sound source signal of the audio signal is An audio signal encoding system that represents and outputs a codebook or multipulse, characterized in that the spectral parameters are non-linearly transformed so as to correspond to auditory characteristics, and are represented and output using a preconfigured codebook. method. (2) Divide the input discrete audio signal into frames of a predetermined time length, approximate the audio signal using an all-pole model to obtain spectral parameters representing the spectral envelope of the audio signal, and is divided into small sections of a predetermined time length, and the pitch parameter is determined so that the signal reproduced based on the past sound source signal is close to the audio signal, and the sound source signal of the audio signal is In a speech encoding method that outputs a codebook or multi-pulse representation, the spectral parameters are nonlinearly transformed to correspond to auditory characteristics, and the nonlinearly transformed spectral parameters are calculated as differences or prediction errors between frames, or in the same frame. An audio signal encoding method characterized in that a difference between parameters or a prediction error is represented by a preconfigured codebook and output.