JP3102017B2 - Audio coding method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION INDUSTRIAL APPLICATION The present invention relates to an audio signal having a low bit rate, especially 8 to 4.
The present invention relates to a voice coding method for performing high-quality coding at about 8 kb / s.

[Conventional technology]

As a method of encoding an audio signal at a low bit rate of about 8 to 4.8 kb / s, for example, M. Schroeder and B. At
al says, “Code-excited linear prediction: High
quality speech at very low bit rates "(Proc.ICASS
P, pp. 937-940, 1985), and a CELP (Code Excited LPC Coding) described in a paper (Reference 1) and the like are known. In this method, the transmitting side extracts a spectrum parameter representing a spectrum characteristic of a speech signal from a speech signal for each frame (for example, 20 ms), further divides the frame into small-section subframes (for example, 5 ms), and Extracts the pitch parameter of the adaptive codebook representing the long-term correlation (pitch correlation) based on the past sound source signal. A single noise signal is selected so as to minimize the error power between the signal synthesized from the signal selected from the codebook consisting of a predetermined type of noise signal and the audio signal, and an optimal gain is selected. calculate. Then, an index and a gain indicating the type of the selected noise signal, and the spectrum parameter and the pitch parameter are transmitted.

[Problems to be solved by the invention]

In the conventional method of the above-mentioned document 1, in order to obtain high sound quality, it is generally necessary to extremely increase the bit size of a codebook composed of a noise signal to 10 bits or more. There is a problem that an enormous amount of calculation is required to obtain a suitable noise signal (codeword). Further, since the codebook is basically composed of a noise signal, there is a problem that the sound quality of the reproduced sound reproduced by the sound source signal selected from the codebook is accompanied by noise. If the size of the codebook is reduced to further reduce the bit rate, there is a problem that the sound quality is rapidly deteriorated.

SUMMARY OF THE INVENTION It is an object of the present invention to solve the above-mentioned problems and to provide a speech encoding method with a good sound quality at about 8 to 4.8 kb / s with a relatively small amount of calculation and memory.

[Means for solving the problem]

A first invention divides an input discrete audio signal into frames of a predetermined time length, obtains and outputs a spectrum parameter representing a spectrum envelope of the audio signal, and outputs the frame at a predetermined time. A long section is divided, a pitch parameter is determined so that a signal reproduced based on a past sound source signal is close to the sound signal, and a sound source signal of the sound signal is selected from a first codebook and a signal selected from a first codebook. A speech encoding method represented by a linear combination with a signal selected from a second codebook, wherein the first codebook is modified based on a signal selected from the second codebook.

The second invention divides the input discrete audio signal into frames of a predetermined time length, obtains and outputs a spectrum parameter representing a spectrum envelope of the audio signal,
The frame is divided into small sections of a predetermined time length, and a pitch parameter is determined so that a signal reproduced based on a past sound source signal is close to the sound signal. A speech encoding method represented by a linear combination of a signal selected from the codebook and a signal selected from the second codebook, wherein the second codebook is generated based on the signal selected from the first codebook. It is characterized by being modified.

A third invention divides an input discrete audio signal into frames of a predetermined time length, obtains and outputs a spectrum parameter representing a spectrum envelope of the audio signal, and outputs the frame at a predetermined time. Speech coding that represents a sound source signal of the sound signal by dividing the sound signal into long sub-intervals, obtaining a pitch parameter for each of the sub-intervals to obtain a pitch prediction excitation signal, and using the pitch prediction excitation signal and a signal selected from a codebook. In the method, the codebook is modified based on the pitch prediction excitation signal, or the prediction excitation signal is modified by a signal selected from the codebook.

[Action]

 4 illustrates the operation of the speech encoding method according to the present invention.

In the first invention, a sound source signal is obtained so as to minimize the following equation for each subframe obtained by dividing a frame.

Here, β and M are pitch parameters of pitch prediction (adaptive codebook) based on long-term correlation, that is, gain and delay, and v (n) is a past sound source signal. h
(N) indicates the impulse response of the synthesis filter composed of the spectrum parameters, and w (n) indicates the impulse response of the auditory weighting filter. The signal * indicates a convolution operation. For details of w (n), reference can be made to Reference 1.

Further, d (n) represents the excitation signal represented by codebook, the following equation was first code code selected from book vectors c ₁ (n) and selected from the second codebook It is represented by a linear combination with the code vector c ₂ (n).

Here, γ ₁ and γ ₂ are the selected code words c
_1j (n) and c _2i (n) are shown. Therefore, in the present invention, since the sound source signal is represented by decomposing the two types of codebooks, each codebook may be half the number of bits of the entire codebook. For example, assuming that the number of bits of the entire codebook is 10 bits, the first and second codebooks may be, for example, 5 bits each, and the amount of calculation for codebook search can be greatly reduced.

When a noise codebook as described in Reference 1 is used as each codebook, if the codebook is divided as shown in the equation (2), the codebook is characteristically degraded more than a 10-bit codebook and has a total performance of 7 to 8 bits. I can only give it.

Therefore, in order to obtain high performance, the first codebook is configured by learning in advance using training data. As a method of constructing a codebook by learning,
For example, Linde et al., “An Algorithm for Vector Qua
ntization De-sign "(IEEE Trans.COM-2
8, pp. 84-95, 1980) (Literature 2).

As a distance measure at the time of learning, a square distance (Euclidean distance) is usually used, but in the present invention, an auditory weighting distance scale that has better performance than the square distance and is represented by the following equation is used.

Where t _j (n) is the j-th training data, c
_l (n) is the code word of cluster 1. Cluster 1 centroid (representative codeworder) is cluster 1
Equation (4) or Equation (5) is determined to be minimized using the training data in.

In the equation (5), g indicates an optimum gain.

Next, the second codebook rewrites the first codebook to relieve the training data dependency due to the first codebook.
For the remaining signal obtained by subtracting the contribution of the codebook from the sound source signal, the statistical characteristics such as the codebook obtained by learning according to the method of Reference 2 and the Gaussian noise signal of Reference 1 are determined in advance. A codebook including a noise signal or a random number signal obtained as described above and a codebook having other characteristics are used. The characteristics are further improved by learning the noise codebook under a certain distance scale. See “Transfor by T. Moriya et al.
m Coding of Speech using a Weighted Vector Quantiz
er, "(IEEE J. Sel. Areas Commun., pp. 425
-431, 1988) (Reference 3).

The first invention is characterized in that the first code vector is corrected using the code vector selected by the second code book. This is performed in order to adapt the first codebook to the short-time characteristics of the input signal and to obtain better characteristics with a small codebook size. The first codebook is modified according to the following formula.

c ′ _1j (n) = c _1j (n) + sign (γ ₁ ) · sign (γ ₂ ) · B · C _2i (n) (6) where B is a small positive amount that determines convergence. Also si
gn (γ) represents the sign of γ. According to the equation (6), the first codebook is modified on both the transmitting side and the receiving side.

Also, the following equation can be used in place of equation (6) in order to make the channel resistant to transmission line errors and to prevent the influence of erroneous correction.

c ′ _1j (n) = (1−δ) · c _1j (n) + A · δ · C _2i (n) (7) where δ is not only an erroneous correction but also an effect of a transmission path error. Small amount (eg, 10 ^-2 to 1
0 ^-3 ), and A is a convergence coefficient determined by the following equation.

A = γ ₂ / γ ₁ (8) Here, if the error between the input vector and the vector selected in the first stage is small, the ratio γ of the gain in the first and second stages
_{Since 2} / γ ₁ becomes smaller, the correction becomes difficult to proceed.

Further, simplified equations (9) and (10) can also be used.

c ′ _1j (n) = (1−δ) · c _1j (n) + A · δ (9) c ′ _1j (n) = (1−δ) · c _1j (n) + sign (γ ₁ ) · sign ( γ ₂ ) · δ (10) Next, in the second invention, the second codebook is modified using a code vector selected from the first codebook. Equation (11) or (12) is used for the correction.

c ′ _2i (n) = sign (γ ₁ ) · sign (γ ₂ ) · B · C _1j (n) + C _2i (n) (11) c ′ _2i (n) = A · δ · c _1j (n) + (1−δ) · C _2i (n) (12) Further, simplified equations (13) and (14) can also be used.

c ′ _2i (n) = A · δ + (1−δ) · C _2i (n) (13) c ′ _2i (n) = sign (γ ₁ ) · sign (γ ₂ ) · δ + (1) −δ) · C _2i (n) (14) Further, in the third invention, the sound source signal of the audio signal is represented by an adaptive codebook and a codebook as in the above-mentioned document 1. The code vector selected in the codebook is corrected using the pitch prediction excitation signal obtained based on the past excitation signal by the adaptive codebook. The correction accompanies equation (15) or equation (16).

c ′ _j (n) = sign (β) · sign (γ ₁ ) B · v (n−M) + C _j (n) (15) c ′ _j (n) = A · δ · v (n−M) + (1−δ) · C _j (n) (16) Here, v (n−M) is a pitch prediction excitation signal obtained by the adaptive codebook. Also, c _j (n) is the j-th code vector selected from the codebook.

Further, simplified equations (17) and (18) can also be used.

c ′ _j (n) = A · δ + (1−δ) · C _j (n) (17) c ′ _j (n) = sign (β) · sign (γ ₁ ) · δ + (1−δ) · C _j (n) (18) Also, the code vector C _j selected from the codebook
(N) may be used to modify the pitch prediction excitation signal.

〔Example〕

FIG. 1 is a block diagram showing a speech encoding apparatus for implementing a speech encoding method according to the first invention.

On the transmitting side, an audio signal is input from the input terminal 110 and
Buffer memory for audio signals for frames (for example, 20 ms)
Store in 120.

The LPC calculation circuit 130 performs a well-known LPC analysis on the LSP parameter from the audio signal of the frame as a parameter representing the spectral characteristic of the audio signal of the frame, and calculates only a predetermined order L. Reference 1 can be referred to for the specific calculation method.

Next, the LSP quantization circuit 140 quantizes the LSP parameter with a predetermined number of quantization bits, outputs the obtained code 1 _k to the multiplexer 260, decodes it, and further decodes the linear prediction coefficient a _i ′ ( i = 1 to L), the weighting circuit 200, the impulse response calculation circuit 170, the synthesis filter
Output to 281. For information on how to encode LSP parameters and how to convert LSP parameters to linear prediction coefficients, see Sugamura
“Quantizer design in LSP speech analysi
s-synthesis "(IEEE J. Sel. Areas Commu
n., pp. 432-440, 1988) (Reference 4).

The subframe division circuit 150 divides the audio signal of the frame into subframes. Here, for example, the frame length is 20
ms, and the subframe length is 5 ms.

The weighting circuit 200 performs well-known auditory weighting on the subtracted signal. Reference 1 can be referred to for details of the auditory sensation weighting function.

The subtractor 190 subtracts the output of the synthesis filter 281 from the input signal divided into subframes and outputs the result.

The adaptive codebook 210 inputs the input signal v (n) of the synthesis filter 281 via the delay circuit 206, and further outputs the weighted impulse response h from the impulse response output circuit 170.
_w (n), input a heavy weighting signal from the weighting circuit 200,
Pitch prediction is performed based on the long-term correlation, and delay M and gain β are calculated as pitch parameters. In the following description, the prediction order of the adaptive codebook is assumed to be 1, but may be higher than or equal to second order. The calculation method of the delay M and the gain β in the first-order adaptive codebook is described in Kleijin “Improved sp
eech quality and efficient vector quantization in
SELP "(Proc. ICASSP, pp. 155-158, 1988
Year) (Reference 5). Further, the obtained gain β is quantized and decoded by a predetermined number of quantization bits by a gain quantizer to obtain a gain β ′, which is used to calculate a prediction signal _w (n) according to the following equation. Output to The delay M is output to the multiplexer 260. _w (n) = β ′ · v (n−M) * h _w (n) (19) In the above equation, v (n−M) is a past sound source signal, which is an input signal of the synthesis filter 281. h _w (n) is a weighted impulse response obtained by the impulse response calculation circuit 170.

The delay circuit 206 sets the synthesis filter input signal v (n) to 1
Output to adaptive codebook 210 after delaying by a subframe.

The subtractor 205 subtracts the output of the adaptive codebook 210 from the output signal of the weighting circuit 200 according to the following equation, and generates a residual signal e _w
(N) is output to the first codebook search circuit 230.

e _w (n) = x _w (n) _−w (n) (20) The impulse response calculation circuit 170 calculates the impulse response h _w (n) of the synthesis filter weighted by the auditory sense for a predetermined number L of samples. I do. For a specific calculation method, reference can be made to Document 1 and the like.

The first codebook search circuit 230 uses the first codebook 235 to search for an optimal codeword c _1j (n). Here, as described in the operation section, the first codebook is learned in advance using a training signal. The method of searching for the optimal code vector C _1j (n) is disclosed in Japanese Patent Application No. Hei.
2956 (Reference 6) and the like can be referred to. Then, the optimum gain γ ₁ is obtained, and using this and C _1j (n), the weighted reproduction signal y _w (n) is obtained and output by the method of Reference 6.

The subtractor 255 subtracts y _w (n) from e _w (n) to obtain a second
To the codebook search circuit 270.

The second codebook search circuit 270 calculates an optimal codeword from the second codebook 275. The configuration of the second codebook search circuit can be basically the same as the configuration of the first codebook search circuit. Also, the same method as the method of searching the first codebook 235 can be used as a codeword search method. As described in the section of the operation, the second codebook is composed of a random number sequence codebook in order to relieve training data dependency while maintaining high efficiency of the learning codebook. For the method of constructing a codebook composed of a random number sequence, reference can be made to the above-mentioned document 1.

In addition, in order to reduce the amount of calculation for codebook search,
As the second codebook 275, a superimposed random number codebook can be used. For the construction method of the superimposed random number codebook and the codeword search method, reference can be made to the above-mentioned Document 5. Further, it is also possible to configure by learning in advance similarly to the first codebook.

The gain quantizer 286 performs vector quantization on the gains γ ₁ and γ ₂ by using the gain codebook 287 created in advance by learning using the equations (12) and (13) by the method described in the operation. Reference can be made to the above-mentioned reference 6 and the like.

The correction circuit 280 corrects the code vector c _1j (n) selected in the first codebook search circuit 230 using the equations (6) to (10) described in the section of the operation.

The adder 290 adds the predicted excitation signal of the adaptive codebook 210, the output excitation signal of the first codebook search circuit 230, and the output excitation signal of the second codebook search circuit 270 to the synthesis filter 281. Output.

The synthesis filter 281 receives the output v (n) of the adder 290, obtains a synthesized speech for one frame by the following equation, and inputs a 0 series filter for another frame to obtain a response signal sequence. One frame of the response signal sequence is subtracted by 19
Output to 0.

The multiplexer 260 combines and outputs the output code sequences of the LSP quantization circuit 140, the adaptive codebook 210, the first codebook search circuit 230, the second codebook search circuit 270, and the gain quantizer 286.

 This concludes the description of the embodiment of the first invention.

FIG. 2 is a block diagram showing a speech encoding apparatus for implementing the speech encoding method according to the second invention. In the figure, components having the same reference numerals as those in FIG. 1 perform the same operations as those in FIG.

The correction circuit 380 uses the code vector c _1j (n) selected in the first codebook search circuit 230 to calculate the second codebook based on the expressions (11) to (14) described in the section of operation. The search circuit 270 corrects the selected code vector c _2i (n).

 This concludes the description of the second invention.

FIG. 3 is a block diagram showing a speech encoding apparatus for implementing the speech encoding method according to the third invention. In the figure, components having the same reference numerals as those in FIG. 1 perform the same operations as those in FIG.

The sound source codebook search circuit 430 performs the same operation as the first codebook search circuit 230, and selects an optimum sound source signal from the sound source codebook 435.

The correction circuit 480 uses the pitch prediction excitation signal v (n−M) obtained by the adaptive codebook 210 to calculate (15) to (1)
Using the expression 8), the sound source signal c _1j (n) selected by the sound source codebook search circuit 430 is corrected.

 This concludes the description of the third embodiment of the present invention.

In the correction method in the correction circuit, the following method can be used in addition to the method described in the embodiment. For example, taking the first invention as an example, c ′ _1j (n) = (1−δ) · c _1j (n) + sign · | γ ₂ / γ ₁ | · δ · C _2i (n) (23) Here, sign indicates a positive or negative sign. The sign selects the method that minimizes:

Here, e _w (n) is the output signal of the subtractor 205. In order to minimize the above equation, the following equation obtained by knitting and differentiating the above equation with γ ₁ and setting it to 0 may be minimized.

Therefore, equation (25) is calculated for both the positive and negative signs of equation (23), and the sign of equation (25) or the one with a larger value is transmitted in one bit.

The second and third inventions can also have the same configuration as described above.

In the above-described embodiment, the adaptive codebook gain, the first and second codebook gains, or the adaptive codebook gain and the excitation codebook gain are not simultaneously optimized. Regarding the gain of the first codebook and the second codebook,
Perform simultaneous optimization to further improve the characteristics. In this simultaneous optimization, when calculating the code vectors of the first and second codebooks, the gain optimization is performed only at the time of searching for the code vectors of the first codebook in order to reduce the amount of calculation. A configuration in which the search is not performed when searching for the second codebook may be adopted.

Further, in order to further reduce the amount of calculation, the gain of the adaptive codebook and the first codebook are simultaneously optimized without performing the gain optimization when searching for the code vector of the codebook. A configuration in which the gains of the adaptive codebook and the first and second codebooks are simultaneously optimized may be used. For details, reference can be made to the aforementioned reference 5.

In order to further reduce the amount of calculation, after the code vectors of the first and second codebooks are selected, the gain β of the adaptive codebook and the gains γ ₁ and γ of the _first and second codebooks are selected. It is also possible to adopt a configuration in which the three types ₂ are simultaneously optimized. For details, reference can be made to the aforementioned reference 6.

Further, as the first codebook search method, other well-known methods can be used other than the method of the embodiment. For example, the method described in Document 1 or the orthogonal transform C ₁ (k) of each codeword c _1j (n) in the codebook is obtained and stored in advance, and the weighted impulse response h _w ( The orthogonal transform H _w (k) of n) and the orthogonal transform E _w (k) of the residual signal e _w (n) can be obtained by a predetermined number of points, and searched on the frequency axis. For details, reference can be made to the aforementioned reference 5.

As the second codebook search method, in addition to the method of the above-described embodiment, the method described above, the method described in the above-mentioned document 6, and other well-known good methods can be used.

As a method of constructing the second codebook, in addition to the method described in the above embodiment, for example, an enormous random number sequence is prepared in advance as a codebook, and the random number sequence is searched for training data by using them. And register it as a codeword from the one that is selected
Can be configured. This configuration method can also be applied to the configuration of the first codebook.

In the above embodiment, the gain of the adaptive codebook and the gains of the first and second codebooks are separately vector-quantized, but three types of gains β, γ ₁ , γ ₂ or β, γ ₁
Can be taken together and vector-quantized. For details, refer to the above-mentioned document 5 and “Ve by I. Gerson et al.
ctor sum excited linear prediction "(VSELP) speech
coding at 8kbp / s "(Proc.ICASSP, pp.461
-464, 1990) (Reference 7).

In the above embodiment, the order of the adaptive codebook is 1
The order is higher, but can be higher than second order. Further, it is possible to set the delay to a decimal value instead of an integer value while keeping the order. For details on these, see, for example, P.
“Pitch predictors with high tempor” by Kroon et al.
al resolution "(Proc.ICASSP, pp.661-66)
4, 1990) (Reference 8). As described above, the characteristics are improved, but the amount of information necessary for transmitting the gain or the delay is slightly increased.

In the above embodiment, the K parameter and the LSP parameter are encoded as the spectrum parameters, and the LPC analysis is used as the analysis method. However, other known parameters such as the LPC cepstrum,
Cepstrum, improved cepstrum, generalized cepstrum, mel cepstrum and the like can also be used. In addition, an optimal analysis method can be used for each parameter.

Further, the LPC coefficient obtained in the frame may be interpolated for each subframe on the LSP or linear prediction coefficient, and the adaptive codebook and the first and second codebooks may be searched using the interpolated coefficient. . With such a configuration, the sound quality is further improved.

Further, the LSP coefficient can be more efficiently encoded by performing vector quantization or vector-scalar quantization by a known method. For the method of vector-scalar quantization, reference can be made to the above-mentioned reference 3, for example.

Also, on the receiving side, an adaptive post filter that operates on at least one of the pitch and the spectral envelope may be added in order to make the quantization noise shaped to make it easier to hear audibly. For the configuration of the adaptive post film, see, for example, “A Class of Analys
is-by-synthesis Predictive Coders for High Quali
ty Speech Coding at Rates between 4.8 and 16kb / s, "
(IEEE JSAC, vol. 6, 2, 353-363, 1988) (Reference 9).

〔The invention's effect〕

As described above, according to the present invention, the code vector selected by the first or second codebook is
Correcting based on the code vector selected by the second or first codebook, or correcting the code vector selected by the excitation codebook based on the pitch prediction excitation signal selected by the adaptive codebook Or, since the pitch prediction excitation signal is modified by the code vector, it is possible to adapt the characteristics of the codebook to the characteristics of the input signal, and to reduce the codebook size at a low bit rate. There is a great effect that better characteristics can be obtained than in the conventional method.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a speech encoding device that implements the speech encoding method according to the first invention, FIG. 2 is a block diagram showing a speech encoding device that implements the speech encoding method according to the second invention, FIG. 3 is a block diagram showing a speech encoding apparatus for implementing the speech encoding method according to the third invention. 110 buffer memory 130 LPC calculation circuit 140 quantization circuit 150 subframe division circuit 170 impulse response calculation circuit 190, 205, 255 subtractor 200 weighting circuit 206 delay circuit 210 adaptation Codebook 230: First codebook search circuit 235: First codebook 281: Synthesis filter 270: Second codebook search circuit 275: Second codebook 280, 380, 480: Correction circuit 286: … Gain quantizer 287 …… Gain codebook 430 …… Sound source codebook search circuit 435 …… Soundsource codebook

Continuation of the front page (56) References Tokuhoku Hei 8-32033 (JP, B2) 1990 IEICE Spring National Convention Lecture Papers, Volume 1, SA-5-4, "8 kb / sC using Learning Codebook Improvement of ELP (LCELP) ", p. 1-427 to 1-428, (issued March 5, 1990) (58) Fields surveyed (Int. Cl. ⁷ , DB name) G10L 11/00-13/08 G10L 19/00-21/06 H03M 7/30 H04B 14/04 INSPEC (DIALOG) JICST file (JOIS) WPI (DIALOG)

Claims (1)

(57) [Claims] 1. An input discrete audio signal is divided into frames of a predetermined time length, and a spectrum parameter representing a spectrum envelope of the audio signal is obtained and output.
The frame is divided into small sections of a predetermined time length, a pitch parameter is obtained for each of the small sections to obtain a pitch prediction excitation signal, and the speech signal is obtained by the pitch prediction excitation signal and a signal selected from a codebook. In the speech encoding method representing the excitation signal, the codebook is modified based on the pitch prediction excitation signal, or the prediction excitation signal is modified by a signal selected from the codebook. Audio coding method.