JP3050978B2 - Audio coding method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for compressing and encoding an audio signal.

[0002]

2. Description of the Related Art A high-efficiency encoding method for audio signals at a compression rate of 8 kbits / sec or less is a code-excited linear prediction encoding method (hereinafter referred to as CELP) by Atal et al. This expresses a voice signal by parameters of a vocal tract and parameters of an excitation source. Regarding vector quantization of parameters of an excitation source by two of a statistical codebook and an adaptive codebook,
It is disclosed in the following document.

Reference: NSJayant & JHChen, “Speech
Coding with Time-Varying Bit Allocations to Excit
ation and LPC parameters ", Proc, ICASSP-89, (1989)

[0004]

A speech signal can be considered to be divided into a stationary voiced section and a transient unvoiced section, and both have significantly different statistical properties. Concerning the compression coding by CELP, the contribution to the sound quality by the adaptive codebook is very large for voiced sounds, but not so much for unvoiced sounds, and the accuracy of vocal tract parameters is rather important. Therefore, for effective compression encoding, it is desirable to separately prepare an encoding method for voiced sound sections and an encoding method for unvoiced sounds.

Accordingly, the present invention distributes information mainly to the adaptive excitation code in the case of voiced sounds and distributes information mainly to the vocal tract parameters in the case of unvoiced sounds, thereby achieving high quality and high efficiency. It is intended to provide a method for compression-encoding the speech.

[0006]

The present invention is an improvement on CELP specified by the following items.

That is, the present invention has means for obtaining a vocal tract parameter (vector) by performing linear prediction analysis on an input speech signal (vector). Further, it has a vocal tract parameter quantization means for quantizing the vocal tract parameters of the input voice signal and outputting the quantized vocal tract parameters (vocal tract main parameters) and vocal tract parameter codes corresponding thereto. Further, the vocal tract parameter of the input voice signal is quantized with respect to an error between the vocal tract parameter after interpolation or the quantized vocal tract parameter without interpolation, and the quantized vocal tract correction parameter and the voice corresponding thereto are quantized. Vocal tract correction parameter quantization means for outputting a tract parameter correction code. Also,
It has an adaptive codebook that stores an adaptive excitation vector representing an excitation source parameter of a past input speech signal, and a statistical codebook that stores a statistical excitation vector that is a predetermined excitation source parameter. Further, a synthesized speech signal is created based on the quantized vocal tract parameters and excitation source parameters, and an error between the input speech signal and the synthesized speech signal (vector) is evaluated, whereby an excitation source output code is calculated. decide.

According to the present invention, a voiced sound or an unvoiced sound is discriminated by performing a long-period analysis of an input voice signal. In the case of a voiced sound, the quantized vocal tract parameters and the adaptive excitation vector are determined. The adaptive excitation code and the statistical excitation code are determined by creating the synthesized speech signal based on an addition vector of the statistical excitation vector and evaluating an error between the input speech signal and the synthesized speech signal. I do.

In the case of unvoiced sound, the synthesized speech signal is created based on the sum of the quantized vocal tract parameters and the quantized vocal tract correction parameters and the statistical excitation vector. A statistical excitation code is determined by evaluating an error between the input audio signal and the synthesized audio signal, and one of the adaptive excitation code and the vocal tract parameter correction code is replaced with another according to whether a voiced sound or an unvoiced sound is present. The output word is multiplexed with the code.

In the present invention, a vocal tract parameter code and a voiced / unvoiced identification code as a result of discrimination between a voiced voice and an unvoiced voice, which are multiplexed as output words, are information updated for each frame, and vocal tract parameter correction is performed. The code, the adaptive excitation code, and the statistical excitation code can be information updated for each subframe.

[0011]

According to the coding method of the present invention, first, a long-period correlation of a voice is analyzed to determine whether it is a voiced sound or an unvoiced sound.

In the case of a voiced sound, an excitation signal vector having a long-period correlation is encoded using an adaptive codebook.

In the case of unvoiced sound, an adaptive codebook is not used. Instead, a codebook for correcting vocal tract parameters is used to code errors due to quantization and interpolation of vocal tract parameters. Become

The voiced / unvoiced identification code may be updated for each frame or for each subframe. In the former case, the interpolation processing is brought to the subsequent stage, so that the vocal tract parameters can be easily viewed. There is an advantage that a conventional one can be used for creating the codebook.

[0015]

FIG. 1 is a block diagram showing an encoder to which the present invention is applied.

In FIG. 1, an A / D-converted input speech signal sequence is inputted in units of a specific frame length, and the vocal tract analyzer 101 analyzes the input speech signal to obtain vocal tract parameters. The vocal tract parameters of the input voice signal are quantized by a vocal tract parameter quantizer 102 using a built-in quantization table.

The vocal tract parameter code C1 corresponding to the quantized vocal tract parameter P1 is sent to the multiplexer 113 once in each frame.

The quantized vocal tract parameter P1 is used by the interpolator 103 after being interpolated in subframe units obtained by further dividing the frame (the vocal tract interpolation parameter P1).
h).

The vocal tract parameters used in the current subframe include errors due to quantization and interpolation. Therefore, the vocal tract analysis is performed again for each subframe, and the difference is obtained by the subtractor 114 from the difference from the vocal tract parameter containing no error obtained as a result of the analysis in the current subframe. The vocal tract correction parameter quantizer 104 quantizes the vocal tract using a built-in quantization table to obtain a vocal tract correction parameter P2.

In the case of an unvoiced sound, a vocal tract parameter correction code C2 corresponding to the vocal tract correction parameter P2 is sent to the multiplexer 113 once for each subframe.

The long-period analyzer 105 calculates the long-period correlation of the input speech signal, and determines whether the section of the input speech signal at the current time is a voiced sound or an unvoiced sound for each subframe. It discriminates and sends the voiced / unvoiced identification code C3 to the multiplexer 113.

If it is determined that the sound is a voiced sound, the switch 106 is closed, the switch 107 is opened, and the adaptive excitation vector from the adaptive codebook 108 and the statistical excitation vector from the statistical codebook 109 are added by the adder 115. To synthesize an synthesized speech signal from the excitation vector by the synthesis filter 110 using the vocal tract interpolation parameter Ph which is not corrected, and subtracts an error between the synthesized speech signal and the input speech signal. Is calculated by the calculator 116 and the error calculator 111.

From the errors obtained by the error calculator 111, an optimal adaptive excitation vector and an optimal statistical excitation vector are selected by a minimum error selector 112.

The statistical excitation code C4 corresponding to the statistical excitation vector and the adaptive excitation code C corresponding to the adaptive excitation vector
5 is sent to the multiplexer 113 for each subframe.

When it is determined that the sound is unvoiced, the switch 106 is opened, the switch 107 is closed, and the excitation vector is constituted only by the statistical excitation vector from the statistical codebook 109. The vocal tract parameter P1 and the vocal tract parameter The correction parameter P2 is added by the adder 117 to generate a corrected vocal tract parameter, and the synthesized speech signal is synthesized by the synthesis filter 110 using the excitation vector of only the statistical excitation and the corrected vocal tract parameter, and the error The calculator 111 calculates an error from the input voice signal.

In the case of unvoiced sound, the minimum error selector 112
Selects the optimal excitation vector only for the statistical codebook 109.

The multiplexer 113 is obtained as described above.
Vocal tract parameter code P1 and voiced unvoiced identification code C3
, A statistical excitation code C5, an adaptive excitation code C2 for voiced sound, and a vocal tract parameter correction code C2 for unvoiced sound, and transmit them to the communication line.

In this example, only the vocal tract parameter code C1 is information for each frame, and the other codes C2 to C
5 is information for each subframe.

FIG. 2 shows a block diagram of a decoder corresponding to the encoder of FIG.

In FIG. 2, a demultiplexer 201 converts a codeword received from a communication line into a vocal tract parameter C1, a voiced unvoiced identification code C3, a statistical excitation code C5, an adaptive excitation code C4, or a vocal tract correction parameter code C4.
2 and sent to each part of the decoder. At this time, the switch 202 is connected to the adaptive codebook 205 if voiced, and to the vocal tract correction parameter inverse quantizer 204 if unvoiced.

The vocal tract parameter code C1 is inversely quantized by the vocal tract parameter inverse quantizer 203, and the vocal tract parameter P
It becomes 1. Further, the vocal tract parameter P1 is
Is interpolated for each subframe.

When the voiced / unvoiced identification code C3 indicates a voiced sound, the switch 210 is closed and the switch 209 is closed.
And add the optimal adaptive excitation vector corresponding to the adaptive excitation code C4 from the adaptive codebook 205 and the optimal statistical excitation vector corresponding to the statistical excitation code C5 from the statistical codebook 206 by the adder 211. In addition, an excitation vector is formed, and the reproduced voice output is synthesized by the synthesis filter 208 using the vocal tract interpolation parameter Ph that is not corrected.

Conversely, when the voiced unvoiced identification code C3 indicates an unvoiced sound, the switch 210 is heard, the switch 209 is closed, and the vocal tract correction parameter inverse quantizer 204
Then, the vocal tract correction parameter code C2 is inversely quantized to obtain a vocal tract correction parameter P2, and the vocal tract parameter Ph is corrected by the adder 212. And the statistical code book 206
, An excitation vector is formed only from the optimal statistical excitation vector corresponding to the statistical excitation code C5, and the reproduced sound output is synthesized by the synthesis filter 208 using the corrected vocal tract parameters.

FIG. 3 is a block diagram showing another encoder to which the present invention is applied. The difference from the example of FIG. 1 is that voiced / unvoiced identification is performed in units of frames and not in units of subframes.

In FIG. 3, the input audio signal sequence that has been subjected to A / D conversion is input in units of a specific frame length.

First, the input voice signal is analyzed by the vocal tract analyzer 101 to obtain vocal tract parameters. The long-period analyzer 105 calculates a long-period correlation of the input voice signal, and determines whether the section of the input voice signal at the current time is a voiced sound or an unvoiced sound. The determination (C3) is performed in frame units.

The vocal tract parameters of the input speech signal are quantized by a vocal tract parameter quantizer 102. The quantized vocal tract parameters are P1 once in each frame,
Sent to 3.

If the result of the voiced / unvoiced sound discrimination C3 is unvoiced, the difference between the vocal tract parameter before quantization and the vocal tract parameter P1 after quantization is subtracted by a subtractor 314 from the difference between the vocal tract parameter before quantization and the vocal tract parameter P1 after quantization. The error is quantized by a vocal tract correction parameter quantizer 304, and a vocal tract correction parameter P2 is obtained for each subframe.

In the case of voiced sound, the switch 307 is opened, and the quantized vocal tract parameters are interpolated by the interpolator 303 in subframe units obtained by further dividing the frame, and in the case of unvoiced sound, the switch 307 is opened. Is closed, and the sum of the quantized vocal tract parameter P1 and the vocal tract correction parameter Ph is used by interpolation by the interpolator 303.

Regarding the encoding of the excitation source, in the case of voiced sound, the switch 106 is closed and the adaptive code book 108
, And the statistical excitation vector from the statistical codebook 109 are added by an adder 115 to form an excitation vector, a synthesis filter 110 synthesizes a synthesized speech signal, and the synthesized speech signal and the input speech signal are combined. Error of
The calculation is performed by the subtractor 116 and the error calculator 111. From the error obtained by the error calculator 111, a minimum error selector 112
Select the optimal adaptive excitation vector and the optimal statistical excitation vector with.

If it is determined that the sound is unvoiced, the switch 106 is opened, an excitation vector is composed only of the statistical excitation vector from the statistical codebook 109, the synthesized filter 110 synthesizes the synthesized speech signal, and the subtracter 116 An error calculator 111 calculates an error from the input voice signal. In the case of unvoiced sound, the minimum error selector 112 selects the statistical codebook 1
An optimal excitation vector is selected only for 09.

The multiplexer 113 is obtained by the above-mentioned device.
Vocal tract parameter code C1, voiced unvoiced identification code C3,
The statistical excitation code C5 and one of the adaptive excitation code C4 and the vocal tract correction parameter code C2 are multiplexed and transmitted to the communication line.

The codes C1 and C3 are information for each frame, and the codes C2, C4 and C5 are information for each frame.

FIG. 4 is a block diagram of a decoder corresponding to the encoder of FIG.

In FIG. 4, a demultiplexer 201 converts a codeword received from a communication line into a vocal tract parameter code C
1. It is separated into a voiced / unvoiced identification code C3, a statistical excitation code C5, an adaptive excitation code C4, or a vocal tract correction parameter code C2 and sent to each device of the decoder. At this time, the switch 202 is connected to the adaptive codebook 205 if voiced, and to the vocal tract correction parameter inverse quantizer 204 if unvoiced.

The vocal tract parameter code C1 is inversely quantized by the vocal tract parameter inverse quantizer 403 to become a vocal tract parameter.

If the voiced unvoiced identification code C3 indicates a voiceless sound, the switch 210 is opened and the switch 409 is opened.
Is closed, the vocal tract correction parameter inverse quantizer 204 inversely quantizes the vocal tract parameter correction code C2 to obtain the vocal tract correction parameter P2, and corrects the vocal tract parameter. Further, the data is interpolated by the interpolator 407 in each sub-frame unit (Ph).

Then, from the statistical code book 206,
An excitation signal vector is formed only from the optimal statistical excitation vector corresponding to the statistical excitation vector code, and the reproduced sound output is synthesized by the synthesis filter 408 using the corrected vocal tract parameters.

Conversely, when the voiced unvoiced identification code C3 indicates a voiced sound, the switch 210 is closed and the switch 409 is opened, and the optimum adaptation corresponding to the adaptive excitation code C4 from the adaptive codebook 205 is performed. The excitation vector,
An excitation vector is constructed from the optimal statistical excitation vector corresponding to the statistical excitation code C5 from the statistical codebook 206, and the reproduced voice output is synthesized by the synthesis filter 208 using the vocal tract interpolation parameter Ph which is not corrected.

[0050]

According to the present invention, it is possible to selectively use a coding method for effectively coding voiced sound and a coding method for effectively coding unvoiced sound. By doing so, a higher quality and more efficient voice compression encoding method can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram of an encoder to which the present invention is applied.

FIG. 2 is a block diagram of a decoder corresponding to FIG.

FIG. 3 is a block diagram of another encoder to which the present invention is applied;

FIG. 4 is a block diagram of a decoder corresponding to FIG. 3;

[Explanation of symbols]

 Reference Signs List 101 vocal tract analyzer 102 vocal tract parameter quantizer 103 interpolator 104 vocal tract correction parameter quantizer 105 long period analyzer 106 switch 107 switch 108 adaptive codebook 109 adaptive codebook 110 synthesis filter 111 error calculator 112 minimum error Selector 113 Multiplexer 114 Subtractor 115 Adder 116 Subtractor

──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Yoshihiro Ariyama 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (56) References JP-A-1-54497 (JP, A) Kaisho 59-172690 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 11/00-21/06 H03M 7/30 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A means for linearly predicting and analyzing an input speech signal to obtain a vocal tract parameter, a voice for quantizing the vocal tract parameter, and outputting a quantized vocal tract parameter and a vocal tract parameter code corresponding thereto. Vocal tract parameter quantizing means, and a quantized vocal tract correction parameter which is quantized with respect to an error between the vocal tract parameter of the input voice signal and the quantized vocal tract parameter after interpolation or without interpolation. And a vocal tract correction parameter quantizing means for outputting a vocal tract parameter correction code corresponding thereto and an adaptive codebook storing an adaptive excitation vector representing an excitation source parameter of the past input voice signal. A statistical codebook that stores a statistical excitation vector, which is an excitation source parameter. And generating a synthesized speech signal based on the excitation source parameter and evaluating an error between the input speech signal and the synthesized speech signal to determine an excitation source output code. Determine whether voiced or unvoiced by performing a long-period analysis, in the case of voiced sound, the quantized vocal tract parameters,
Based on the sum of the adaptive excitation vector and the statistical excitation vector, the synthesized speech signal is created, and the error between the input speech signal and the synthesized speech signal is evaluated, whereby the adaptive excitation code and the statistical An excitation code is determined, and in the case of unvoiced sound, the synthesized speech signal is calculated based on an added value of the quantized vocal tract parameter and the quantized vocal tract correction parameter, and the statistical excitation vector. A statistical excitation code is determined by evaluating an error between the input audio signal and the synthesized audio signal, and the adaptive excitation code and the vocal tract parameter correction code are respectively determined according to voiced sound or unvoiced sound. A speech encoding method characterized in that one is multiplexed with another code to produce an output word. 2. The voice coding method according to claim 1, wherein the vocal tract parameter code and the voiced / unvoiced identification code as a result of determining whether the voiced sound or unvoiced sound are multiplexed as output words are information updated for each frame. A speech coding method characterized in that there are vocal tract parameter correction codes, adaptive excitation codes, and statistical excitation codes, which are information updated for each subframe.