JP3065638B2 - Audio coding method - Google Patents

Description

Description: Object of the Invention (Field of Industrial Application) The present invention relates to an audio encoding system for compressing an audio signal or the like with high efficiency, and more particularly to an audio encoding system at a low bit rate. .

(Prior Art) When an audio signal is transmitted at a low bit rate, a multi-mode CELP (Code Excited) is an effective method for encoding at a transmission rate of about 10 kb / s or less.
Linear Prediction) is known. The details are described in the ICASSP paper (first paper) “Multimode coding: Application to CELP T” in Glasgow, 1989.
omohiko Taniguchi, Shigeyuki Unagami and Robert MG
ray ". This will be described briefly.
FIG. 6 is a diagram for explaining the principle of multi-mode coding described in the above-mentioned paper, and FIG. 7 is a block diagram showing processing of the multi-mode CELP coder.

In FIG. 6, the code side includes m encoders 510, 520,
530 (encoders # 1 to #m), and each encoder is set in advance to assign different bits to the drive signal parameters and the spectrum parameters.

Each encoder processes the input speech signal in parallel in the evaluation and optimal encoder decision unit 550 in frame units, and uses the input speech signal to generate a synthesized speech signal (decoded speech signal) given by each encoder. , And the selector 540 selects and transmits the drive signal parameter and the spectrum parameter to be transmitted by using the index n (n is one of 1, 2,... The information of the index n is also transmitted to the decoding side. On the decoding side, the decoder 560 corresponding to the encoder #n is based on the index n of the encoder.
(Decoder #n) to output a synthesized speech signal.

The above is the outline of the multi-mode coding shown in the above-mentioned paper. A multi-mode CELP encoder shown in FIG. 7 applies the concept of the multi-mode encoding to the CELP system.

The CELP method is a speech coding method that performs vector quantization of a drive signal at the level of a synthesized sound, and is a known technique.
For details on the CELP method, see `` MRSchroeder and B.
S. Atal, “Code-excited linear predection CELP): Hi
gh quality speech at very low bit rates, 2Proc.ICAS
SP'85, pp. 937-940 ".

FIG. 7 shows a multi-mode CELP system in which the above-described multi-mode coding system is CELP in the simplest form of two modes.
It is applied to That is, in the A mode, a drive signal parameter and a spectrum parameter (LPC parameter) are transmitted by a known CELP method, and 1-bit mode information is transmitted for each frame.

On the other hand, the B mode has a configuration in which the number of quantization bits allocated to the drive signal parameters is increased by using the same spectral parameters as the previous frame without transmitting the spectral parameters. In each frame, the A / B mode is determined based on the quality evaluation (using SNR or the like) of the synthesized speech signal in each mode,
Allocation of transmission information is dynamically controlled by switching between the two modes. In FIG. 7, in the A mode, the LPC analyzer 100 extracts a spectrum parameter (LPC parameter) from the input audio signal and outputs the extracted parameter to the switching terminal A and the short-time synthesis filter 110. The parameters of the long-term synthesis filter 150 and the waveform of the vector selected from the codebook (small) 170 (index + sign attached to the vector in the codebook) and the gain are the input voice and the short-time synthesis filter 110 (synthesis filter). An error signal with the synthesized signal synthesized in the above is obtained in a closed loop so that the power of the weighted error signal weighted by the weight filter 120 is minimized.

On the other hand, in the B mode, the spectrum parameter memory 24
Only when 0 is determined to be the A mode, it is connected to the terminal A to update the spectrum parameters. The spectrum parameters stored in the spectrum parameter memory 240 are not updated during the B mode, and are not updated. used. The parameters of the long-time synthesis filter 160 and the waveform and gain of the codebook (large) 180 are determined in the same manner as in the A mode. Mode determination unit 230
Inputs the minimum value of the error power of each mode calculated in the A mode and the B mode, and outputs the mode with the smaller error power as the determined mode.

The above is the description of the multi-mode CELP system (conventional system) shown in FIG.

This method is 4.8 kbit / s higher than the conventional CELP method.
Approximately 2dB segmental SNR at kbit / s transmission rate
Is also shown in the above first paper.

This audio coding method uses A mode B
By changing the mode, the bit assignment of the drive signal and the spectrum parameter was variable for each frame.

When a frame is transmitted with a fixed code amount, in the A mode, bits are frequently allocated to spectrum parameters, and bits cannot be allocated much to drive signal parameters. Therefore, in the A mode, the conventional CE
This is the same as the LP system, and in the speech section in which the B mode is used, more quantization bits can be allocated to the drive code signal parameters by using the same spectral parameters as in the previous frame. So in B mode
The speech quality in the CELP system is improved.

On the other hand, in the B mode, it is clear that the voice section is easily selected in a voice section in which the spectrum parameter of the previous frame can be used instead of the spectrum parameter of the current frame, that is, a vowel section having a small temporal change in spectrum. .

However, such a speech section generally has a high degree of redundancy due to the cyclic repetition of the drive signal, so that a synthesized speech having a high SN ratio can be obtained even with the normal CELP method. When B-mode coding is performed in such a voice section, it is expected that a synthesized voice having a higher SN ratio than that of the CELP system will be obtained, but the sound that clears the SN ratio to a certain extent is perceived. The difference is hard to understand.

In addition, A mode (normal CELP method) is easy to select in the voice section where the spectrum change other than the vowel is large.
In terms of hearing, there is a problem that the deterioration of voice quality by the normal CELP method is not improved.

(Problems to be Solved by the Invention) As described above, in the conventional speech coding method, the driving is performed by switching between two modes, a mode using the spectrum parameters of the current frame and a mode using the spectrum parameters of the previous frame. Although the bit assignment of the signal parameter and the spectrum parameter is variable for each frame, the mode using the spectrum parameter of the previous frame is hardly used in a speech section such as a consonant having a large temporal change in the spectrum. As a result, there is a problem that the deterioration of the voice quality in the non-stationary section in the CELP system which is the conventional voice coding system is not improved.

The present invention has been made in order to solve such a problem, and an object of the present invention is to provide a speech coding system capable of obtaining a high-quality synthesized speech at a low bit transmission rate. .

[Structure of the Invention] (Means for Solving the Problems) In order to achieve the above-mentioned object, a speech encoding method according to the present invention provides a synthesized speech signal by driving a synthesis filter including a pole filter and a zero filter with a drive signal. In the speech coding method for obtaining the above, there is provided means for storing coefficient information of the zero filter, and the synthesized speech signal is obtained using the coefficient information.

(Operation) According to the speech coding method of the present invention having the above-described configuration, of the synthesis filter composed of the pole filter and the zero filter, there is provided means for storing coefficient information of the zero filter,
Since the synthesized speech signal is obtained using the coefficient information, even in a speech section such as a consonant having a large spectrum change, a filter suitable for the speech in the section can be selected. Therefore, a high-quality and stable synthesized speech can be obtained.

(Embodiment) Hereinafter, the encoding system of the present invention will be described in detail with reference to the drawings.

FIG. 1 and FIG. 2 are block diagrams for performing the speech coding method of the present invention. In FIG. 1, the input audio signal is
Linear prediction and pitch extraction are performed by the LPC analysis unit 100,
The obtained filter parameters are output to the short-time synthesis filter 110 and the long-time synthesis filter 150.
Then, the waveform of the vector selected from the codebook A175 (the index + sign attached to the vector in the codebook A) and the gain are input to the long-time synthesis filter 150 via the multiplication circuit 190. Long-time synthesis filter 150
Adds pitch periodicity to the input signal. When this is input to a short-time synthesis filter (hereinafter referred to as a synthesis filter) 110, a synthesized speech signal is generated from prediction parameters (coefficient information of the synthesis filter (polar filter) 110) based on linear prediction of the LPC analysis unit 100. Here, according to the present invention, since the synthesis filter is constituted by a pole-zero filter, the filter has the zero filter 115. And zero filter 115 is codebook B1
76 has zero filter coefficient information. Therefore, the power of the weighted error signal obtained by weighting the error signal between the synthesized voice signal output from the synthesis filter 113 including the zero filter 115, the codebook Book A175
And the coefficients in the codebook B176 are changed in a closed loop. When the weighted error is minimized, the distortion comparator 210 determines that the codebook A1 at the time when the error is minimized is obtained.
The index of the coefficient in 75 and the index of the coefficient in the codebook B176 are output as a coded signal corresponding to the input audio signal. FIG. 2 is a block diagram of the present invention in which an error between an input signal and a synthesized speech signal is weighted to evaluate distortion. In FIG.
The parameter of 0 is configured to be obtained in a closed loop. Therefore, in this case, the LPC analysis unit 100 obtains only the parameters of the short-time synthesis filter. The first
B (Z) in FIG. 2 corresponding to the zero filter 115 in FIG.
When (Z) = 1, there is no coefficient information of the zero filter. Here, when transmission is performed at a fixed rate, the number of bits allocated to one frame is determined. However, the bit may be arbitrarily assigned to each parameter as long as the code amount is constant. Therefore, as described above, B (Z)
In the case of = 1, it is not necessary to send the parameters of the zero filter, and more bits can be allocated to the drive signal parameters. Conversely, when B (Z) ≠ 1, the coefficients of the zero filter must also be transmitted, so that the transmission rate can be made constant by reducing the bit allocation of the drive signal parameters.

Next, FIG. 3 is a block diagram showing a system using a plurality of speech coding systems shown in FIG. In FIG. 3, B
If (Z) ≠ 1, the zero filter 115 is a codebook B176
Codebook of drive signal parameters
170 is smaller than the codebook 180 of the drive signal parameters when B (Z) = 1 of the zero filter 116.

FIG. 4 is a block diagram showing a case where an encoding method according to an embodiment of the present invention is applied to an encoding device.

In FIG. 4, a series of A / D converted input audio signals is input from an input terminal 10. The frame buffer 11 is a circuit for accumulating an input audio signal for one frame. Each block in FIG. 4 performs the following processing on a frame basis or on a subframe basis obtained by dividing a frame into a plurality of frames.

The prediction parameter calculation circuit 12 calculates a prediction parameter using a known method. When the prediction filter is configured by cascading a long-term prediction filter 41 and a short-term prediction filter 42 as shown in FIG.
12 is a pitch period pitch prediction coefficient and a linear prediction coefficient (α
Parameters or K parameters: all referred to as LPC parameters) are calculated by a known method such as an autocorrelation method or a covariance method. The calculation method is described, for example, in (Digital Speech Processing by Sadateru Furui, published by Tokai University Press, 1985). The calculated prediction parameters are input to the prediction parameter coding circuit 13. Prediction parameter coding circuit
13 encodes the prediction parameter based on a predetermined number of quantization bits, outputs this encoding to the multiplexer 25, and calculates the gain calculation circuit 15, the synthesis filter 18,
Output to the weight filters 20 respectively.

The gain calculation circuit 15 calculates a zero filter coefficient from a zero filter coefficient codebook 14 described later, a coefficient update signal output from the coefficient search circuit 24, and a prediction parameter (polar filter coefficient information) from the encoding circuit 13. A pole-zero type synthesis filter H (Z) is constructed based on this. This inverse filter 1 / H
The input speech signal is predicted using (Z) as a prediction filter, and a prediction residual signal is created. Next, the gain calculation circuit 15 calculates the average power of the prediction residual signal, and outputs this to the encoding circuit 16 as a gain. As the average power of the prediction residual signal, for example, a standard deviation can be used. The encoding circuit 16 encodes the gain based on a predetermined number of quantization bits, and outputs this encoding to the multiplexer 25 and the multiplication circuit 17. The zero filter coefficient codebook 14 stores a predetermined order and filter coefficient information of zero filters of the number corresponding to the number M of quantization bits. If one of the zero filters B (Z) stored in the zero filter coefficient codebook 14 is stored in the filter information where B (Z) = 1, the all-pole synthesis filter H ( Z) can be automatically created with the same configuration.

In the present embodiment, the zero filter coefficient codebook 14 is 2 ^M
+1 types of zero filter coefficient information are stored, and a zero filter B (Z) created using the first code vector is stored.
It is assumed that the code book 14 has been created in advance so that B (Z) = 1.

The zero filter coefficient codebook 14 outputs the zero filter coefficients (code vectors) stored in the zero filter codebook 14 to the gain calculation circuit 15 and the synthesis filter 18 based on the code update signal input from the coefficient search circuit 24. And zero filter B (Z) is B (Z) = 1 or B (Z)
≠ Output one information PZ to the codebook 21.

The variance value of the codebook 21 is normalized in advance, and the codebook 21 outputs to the multiplication circuit 17 a limited number of codevectors set in advance according to the information PZ from the codebook 14. The output of the code vector at this time is controlled by the code update signal input from the code search circuit 23. The limitation on the search range of the code vector in the code book 21 can be determined, for example, as follows.

Information PZ from the codebook is zero filter B (Z) = 1
In the case of the information indicating the drive signal, there is no information of the zero filter coefficient, so that a larger number of bits are allocated to the drive signal to expand the search range of the code vector in the code book 21 representing the shape of the drive signal. it can.

Conversely, if the information PZ is information indicating a zero filter B (Z) ≠ 1, it is necessary to transmit information of a zero filter coefficient. It is assumed that the search range of the code vector in 21 is narrowed.

The multiplication circuit 17 multiplies the code vector output from the code book 21 by the encoded gain to generate a vector as a drive signal candidate, and inputs the vector to the synthesis filter 18.

The synthesis filter 18 receives the zero filter coefficient information and the pole filter coefficient information (collectively referred to as spectral parameters) from the zero filter coefficient codebook 14 and the encoding circuit 13, respectively.
(Z), and outputs a synthesized speech signal using the candidate vector of the drive signal from the multiplication circuit 17 as an input signal.

The subtraction circuit 19 receives the input voice signal and the above-described synthesized voice signal, and outputs an error signal.

The weight filter 20 adds the weight generated from the prediction parameter to the above-described error signal and outputs the weighted error signal. The weight filter 20 has a transfer function It is known that the filter represented by ## EQU1 ## has an effect of making the encoding noise included in the synthesized speech difficult to hear at the time of decoding using the auditory masking effect. In equation (1), A (Z) represents a prediction filter created from prediction parameters.

The square error calculation circuit 22 calculates the weighted error signal 2
The sum of squares is calculated for each code vector output from the codebook 21, the result is output to the code search circuit 23, and the value obtained by calculating the sum of squares of the error signal for one frame is output to the coefficient search circuit 24. .

The code search circuit 23 inputs the code number of the zero filter currently being searched output from the coefficient search circuit 24 described later,
2 for each subframe for each code number of the zero filter
The code that minimizes the power error is searched from the code book 21, and this code is retained. When the code number of the zero filter is finally determined by the coefficient search circuit 24, the code held in correspondence with the code number of the zero filter among the codes of the drive signal that has been input and held is input to the multiplexer 25. Output to

The coefficient search circuit 24 compares the sum of squares of the error signal calculated for each frame for each code number of each zero filter input from the square error calculation circuit 22, and selects the code number of the zero filter that minimizes this. The code number is output to the multiplexer 25 and the code search circuit 23. If the code number of the searched zero filter coefficient is 1, as described above, it is known that the zero filter is not used.
At this time, the code of the drive signal output from the code search circuit 23 is represented by a larger number of bits than when the zero filter is used. The coefficient search circuit 24 also outputs information on the use / non-use of the zero filter to the multiplexer 25 at the same time.
Table 1 shows an example of bit allocation between drive signals and spectral parameters in the present embodiment.

In Table 1, the synthesis filter used can be divided into an all-pole filter and a pole-zero filter depending on the case where the zero filter is B (Z) = 1 and B (Z) ≠. Now, when the number of bits per frame is R, the number of bits for the spectral parameter is only K, the number of bits of the polar filter, and the number of bits required for the drive signal is, of course, RK. Therefore, the number of bits per frame is always R constant. When a pole-zero filter is used, M bits are also assigned to the zero filter as the bits for the spectrum parameter, and the remainder is used as a drive signal. The multiplexer 25 multiplexes the input code information, and outputs
The code information is output to the transmission path.

As described above, according to the speech encoding of the present invention, not only does the filter representing the spectral envelope and the bit allocation of the parameters of the drive signal change in frame units in accordance with the change in the sound quality of the input speech signal. This filter is represented by a pole-zero form, and the quantization of the filter coefficient of the zero filter, that is, the selection of the codebook, is performed so that the perceptually weighted error between the input speech signal and the synthesized speech signal is minimized. For this reason, even in a voice section in which the temporal change of the spectrum is large, a filter suitable for the section can be selected, so that the quality of the synthesized voice can be stably improved.

The embodiment described here is one embodiment of the present invention, and various modifications are possible.

(Effects of the Invention) As described in detail above, according to the speech encoding method of the present invention, a high-quality and stable synthesized speech can be obtained.

[Brief description of the drawings]

1 and 2 are block diagrams for performing the speech coding method of the present invention, FIG. 3 is a block diagram using the speech coding method of the present invention for a plurality of speech coding methods, and FIG. FIG. 5 is a block diagram showing a configuration in which a speech coding method according to one embodiment of the present invention is applied to a coding apparatus. FIG. 5 is a block diagram showing one configuration example of a prediction filter described in the embodiment using FIG. 6 and 7 are block diagrams showing the configuration of an encoding device according to the prior art. 110: Short-time synthesis filter (polar filter) 113: Synthesis filter 115: Zero filter 175, 176: Code book 195: Drive signal generator

Continuation of front page (56) References Proceedings of IE 1988 International Conference on Acoustics, Speech and Signal Processing, Vol. 1, "S.12.5 AM Multipulse Excited Pole-Zero Filtering Approach for Speech Enhancement" p. 1, "S.14.10 A Generalized Vocal Tract Model for Pole Zero Type Linear Prediction" p. 687-690 Proceedings of IEEE 1989 International Conference on Acoustics, Speech and Signal Processing, Vol. 1, "S4.10 Multimode coding: Application to CELP," p. 156-159, Proceedings of IEEE 1989, International Conference on Acoustics, Speech and Signal. 1, "E2.3 Discrete Pole-Zero Modeling and Application ions" p. 2162-2165 Proceedings of IE 1991 International Conference on Acoustics, Speech and Signal Processing, Vol. 1, "S4.8 Adaptive Bit-Allocaton Between the Pole-Zero Synthesis Filter and Excitation in CELP" p. 229-232 Proceedings of IE EE 1992 International Conference on Acoustics, Speech and Signal Processing, Vol. 1, "Pole-Zero code Excited Linear Prediction usin a Perceptually Weighed Error Criterion" p. I-637 to I-639 Proceedings of 6th International Conferencing on Digital Signals in Communications, IEEE Conference Publication No. 340, "Pole-Zero Code Excited Liner Prediction", p. 42-47, 2-6 September 1991 Transactions of the Institute of Electronics, Information and Communication Engineers, Vol. J 72-D- ▲ II ▼ No. 8, August 1989, “CELP Coding Scheme Applying Multi-Mode Coding-Optimization of Transmission of Stimulus Information and Vocal Tract Information-”, p. 1159-1165, (Published August 25, 1989) Proceedings of the 1991 IEICE Spring Conference, Volume 1, [A-216 pole-zero synthesis filter adaptive signal allocation between driving signals Low-rate speech coding Method ", p. 1-216, (issued on March 15, 1991) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 11/00-13/08 G10L 19/00-21/06 H04B 14/04 INSPEC (DIALOG) JICST file (JOIS)

Claims (5)

(57) [Claims] 1. A speech coding system for obtaining a synthesized speech signal by driving a synthesis filter comprising a pole filter and a zero filter with a drive signal, comprising: means for storing coefficient information of the zero filter; A speech encoding method for obtaining the synthesized speech signal by using the speech signal. 2. A speech coding system for obtaining a synthesized speech signal by driving with a synthesis filter comprising a polar filter and a zero filter and a drive signal, comprising means for storing coefficient information of the filter, and using the coefficient information. A speech coding method characterized in that a synthesized speech signal is generated by the above-mentioned method, and coefficient information of the zero filter is selected based on a distortion between the synthesized speech signal and an input speech signal. 3. A distortion of a synthesized speech signal and an input speech signal by each coding method is calculated from a plurality of coding methods in which bit assignments of drive signal parameters and parameters of a synthesis filter including a polar filter and a zero filter are different. And then select one encoding method,
At least one of the plurality of encoding schemes has means for storing coefficient information of the zero filter, and generates a synthesized speech signal using the coefficient information. A speech coding system characterized in that the coefficient of the zero filter is selected based on a distortion with a signal. 4. The method according to claim 2, wherein the bit assignment between the drive signal parameter and the parameter of the synthesis filter is determined depending on whether or not the zero filter is used in the synthesis filter. 3. The speech encoding method according to 3. 5. The polar filter in the synthesis filter, wherein:
4. A speech encoding system according to claim 2, wherein said encoding system is common to all encoding systems.