JP2658794B2 - 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 speech coding system for coding a speech signal at a low bit rate, particularly at a bit rate of 8 kb / s or less, with a relatively small amount of computation and high quality.

[0002]

2. Description of the Related Art Conventionally, as a speech coding system for vector-quantizing an excitation sound source signal by a sound source codebook, a "code-excited code" by a shredder (MR Shroeder) and an atal (BS Atal) is used.
Linear Prediction (CLP): High-Quality Speech at Berry Low Bit
RAITS (CODE-EXCITED LINEAR P
REDICTION (CELP): HIGH-QUAL
ITY SPEECH AT VERY LOWBIT
RATES), Proceedings of the International Conference on Sound, Speech and Signal Processing (Proc. ICASP), 198.
The CELP method described in a paper (Reference 1) of 937 to 940 pages for 5 years is known. As a CELP system having an adaptive codebook, "Improved に よ る" by Clayjin (WB Kleijn) et al.
Speech Quality and Efficient Vector Quantization in SELP (IMPROVED SPEECH QUALITYA)
ND EFFICIENT VECTOR QUANT
IZATION IN SELP) ", Proceedings of the International Conference on Sound, Speech and Signal Processing (Proc. IC
(ASSP), 1988, pp. 155-158 (Reference 2). These CELP schemes search the excitation codebook, the adaptive codebook, and the gain codebook so as to minimize the perceptual weighted square distance between the input audio signal and the encoded audio signal for each subframe length. However, since encoding is performed for each sub-frame, distortion at the block boundary of block encoding tends to occur, and sufficient sound quality cannot be obtained. Therefore, in order to reduce the distortion of the block boundary of the block coding, when searching the excitation codebook, the input audio signal of the next subframe is added to the input audio signal of the current subframe by a predetermined length called an overlap length. Sound source code so as to minimize the perceptual weighted square distance between the signal connected by the specified length and the signal connecting the coded voice signal and the influence signal of the coded voice signal having the overlap length. A speech coding method for searching for a book is described in "Structured Codebook Design in CLP (Structured Codebook)" by LeBlanc et al.
Design in CELP) ”, International Mobile Satellite Conference (International MobileS)
attellite Conference), 1990
Year, pp. 667 to 672 (Reference 3).

[0003]

However, in the above-mentioned conventional method, distortion at the block boundary in block coding is reduced to some extent, but it is still not enough.

An object of the present invention is to solve the above-mentioned problems,
It is an object of the present invention to provide a speech coding system with a better sound quality than the conventional one at a bit rate of 8 kb / s or less with a relatively small amount of calculation.

[0005]

A speech coding system according to the present invention has a linear prediction analysis unit for obtaining a spectrum parameter of an input speech signal divided into frames at regular intervals, and a sound source signal determined in the past. An audio having an adaptive codebook, a sound source codebook for vector-quantizing an excitation source of the input audio signal, and a gain codebook for quantizing respective gains of the adaptive codebook and the source codebook. In the encoding method, when searching the adaptive codebook, the excitation codebook, and the gain codebook for each subframe obtained by further subdividing the frame, the input speech signal and the spectrum parameter are used to search for the adaptive codebook, the excitation codebook, and the gain codebook. A perceptual weighted audio signal having the length of the subframe is obtained, and the current subframe A response signal obtained by giving the input filter to the synthesis filter using the spectrum parameter as an initial value is obtained by a predetermined length, and an overlap signal is obtained by weighting using the spectrum parameter. The adaptive codebook, the sound source codebook, and the gain codebook are searched in accordance with a signal obtained by connecting the overlap signal after an auditory sense weighted audio signal.

[0006]

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

First, the input audio signal x divided into subframes is converted to the unquantized LPC of the current subframe.
(Linear predictive coding) A weighting input signal _xw is created by an audibility weighting filter W using coefficients.

[0008] The transfer function W of the auditory weighting filter W
(Z) is represented by the following equation (1).

[0009]

Here, α _i is an unquantized LPC coefficient β, γ of the current subframe, a weighting coefficient, p
Is the LPC order.

[0011] As an initial value input audio signal x _w of the current sub-frame, LP unquantized of the next sub-frame
The zero input response of the synthesis filter S ′ using the C coefficient is obtained by the length of the overlap length L ₀ , and the weight is further weighted by the perceptual weighting filter W using the unquantized LPC coefficient of the next subframe. Thus, an overlap signal v is created. Here, when the current subframe is the last subframe of the frame, the nonquantized LPC coefficients of the current subframe are used instead of the nonquantized LPC coefficients of the next subframe.

The overlap signal described in Reference 3 is an input audio signal of the next subframe. However, the signals to be represented by the adaptive code vector, excitation code vector, and gain code vector of the current subframe are an input audio signal of the current subframe and an influence signal on the next subframe by the input audio signal of the current subframe. Therefore, it is better to adopt the influence signal on the next subframe by the input audio signal of the current subframe as the overlap signal, and the distortion at the block boundary of the block coding caused by encoding for each subframe Can be reduced more effectively.

[0013] The signal connected the overlap signal v after the weighted input signal x _w and x, referred to as extended weighted input signal.

Using the signal of the previous subframe as an initial value,
Seeking a zero input response of the synthesis filter S using the quantized LPC coefficients for the current subframe by the length of the subframe length L _s, the initial value obtained signals, quantized in the next sub-frame The quiescent response of the synthesis filter S ′ using the LPC coefficient is obtained by the length of the overlap length L ₀ , and further, the unquantized LP of the current subframe is obtained.
Weighting the subframe length portion with an auditory weighting filter W using C coefficients, and weighting the overlap length portion with an auditory weighting filter W ′ using unquantized LPC coefficients of the next subframe To obtain the weighted influence signal f and subtract it from the extended weighted input signal x. A signal obtained by subtracting the weighting influence signal f from the extended weighting input signal x is defined as y. Here, if the current subframe is the last subframe of the frame, instead of the nonquantized LPC coefficients of the next subframe, the nonquantized LP of the current subframe is used.
C coefficient, and the quantized LP of the next subframe
Instead of the C coefficient, the quantized LP of the current subframe
A C coefficient is used for each.

Firstly, it searches an adaptive code vector that minimizes the error E _a in the following formula (2).

[0016]

However,

[0018]

## EQU1 ##

Here, sa _d is a synthesis filter S, S ′ and an auditory weighting filter W, W of an extended adaptive code vector a _d created by connecting L ₀ 0s after the adaptive code vector of the delay d. Perceived weighted synthesized signal,
g _a is the optimal gain of the perceptually weighted synthesized signal of the extended adaptive code vector.

The optimum gain g _a of perceptual weighting synthesis signal sa _d expansion adaptive code vector a _d is given by the following equation.

[0022]

This equation is substituted into equation (2) to obtain the following equation.

[0024]

However,

[0026]

## EQU1 ##

Next, for the selected adaptive code vector, a sound source code vector that minimizes the error E _e of the following equation (7) is searched.

[0029]

Here, se _i ^⊥ is a synthesis filter S, S ′ of an extended excitation code vector e _i created by connecting L ₀ zeros after the excitation code vector of index i.
A perceptual weighting filter is W, the perceptual weighting synthesis signal se _i by W ', already selected extended adaptive code vector a _d extended excitation code vector e which is generated by orthogonally with respect to the perceptual weighting synthesis signal sa _d of _{i is} an orthogonalized perceptual weighted composite signal, and g _e is the orthogonalized perceptual weighted composite signal se of the extended excitation code vector e _i.
i ^{最適 is} the optimal gain. g _e is given by the following equation (8).

[0031]

This equation is substituted into equation (7) to obtain the following equation.

[0033]

Here,

[0035]

Finally, a gain code vector that minimizes the error E _g in the following equation (12) is searched for the selected extended adaptive code vector a _d and the selected extended excitation code vector e _i .

[0037]

Here, (G1 _k , G2 _k ) is a gain code vector of index k.

As (G1 _k , G2 _k ), not the gain code vector itself but, for example, the power of the residual signal estimated from the quantized power of the weighted input signal and the LPC coefficient, and the power of the extended adaptive code vector Alternatively, a gain code vector converted by a matrix calculated using the power of the extended sound source code vector may be used.

[0040]

Next, the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention.

In the following description, when the current subframe is the last subframe of the frame, the unquantized LPC (linear predictive coding) coefficient of the next subframe is the quantized LPC coefficient of the current subframe. The LPC coefficient that is not present, and the quantized LPC coefficient of the next subframe means the quantized LPC coefficient of the current subframe.

In FIG. 1, an audio signal divided for each frame (for example, 40 ms) is input from an input terminal 1 and sent to a linear prediction analysis circuit 2 and a subframe division circuit 3.

The linear prediction analysis circuit 2 performs a linear prediction analysis of the input speech signal, and applies the obtained spectral parameters to the weighting filter 4, the synthesis filter 14 and the weighting filter 15 of the overlap signal generation circuit 5, and the influence signal calculation circuit. 6. Output to the adaptive codebook search circuit 7, sound source codebook search circuit 8, gain codebook search circuit 9, and spectrum parameter quantizer 17.

In the spectral parameter quantizer 17,
The LPC coefficient input from the linear prediction analysis circuit 2 is converted into a spectral parameter to be quantized (LPC
No conversion is performed when quantizing the coefficient itself. ) And quantize the spectral parameters (eg, convert LPC coefficients to LSPs (line spectrum pairs)
), And converts the quantized spectral parameters into LPC coefficients, and converts the converted LPC coefficients into an influence signal subtraction circuit 6, an adaptive codebook search circuit 7, a sound source codebook search circuit 8 , And to the gain codebook search circuit 9, and further sends the quantized spectral parameter index to the multiplexer 13.

The weighting filter 4 which has received the input audio signal divided into the subframe length (for example, 8 ms) from the subframe division circuit 3, converts the unquantized LPC coefficient of the current subframe inputted from the linear prediction analysis circuit 2. The input audio signal having the sub-frame length is weighted by the perceptual sense according to the equation (1), and is transmitted to the connection circuit 16.

In the synthesis filter 14, the input speech signal of the current subframe input from the subframe division circuit 3 is set as an initial value, the excitation source is set to 0, and the quantization of the next subframe input from the linear prediction analysis circuit 2 is performed. LP not done
Using the C coefficient, a combined signal is created with an overlap length and sent to the weighting filter 15.

The weighting filter 15 weights the input signal from the synthesis filter 14 using the unquantized LPC coefficients of the next subframe input from the linear prediction analysis circuit 2 according to equation (1). And connection circuit 1
Send to 6. At this time, instead of using unquantized LPC coefficients, the spectral parameter quantizer 1
7 may be used.

In the connection circuit 16, the transmission signal of the weighting filter 15 is connected after the transmission signal from the weighting filter 4, and is transmitted to the influence signal subtraction circuit 6.

The influence signal subtraction circuit 6 calculates the influence signal from the previous subframe using the current subframe input from the spectrum parameter quantizer 17 and the quantized LPC coefficient of the next subframe. , Using the current subframe input from the linear prediction analysis circuit 2 and the non-quantized LPC coefficients of the next subframe to create a weighted influence signal,
, And sends the result to an adaptive codebook search circuit 7, a sound source codebook search circuit 8, and a gain codebook search circuit 9. When weighting, the quantized LPC coefficient sent from the spectrum parameter quantization circuit 17 may be used instead of using the unquantized LPC coefficient.

In the adaptive codebook search circuit 7, the signal input from the influence signal subtraction circuit 6, the unquantized LPC coefficients of the current subframe and the next subframe input from the linear prediction analysis circuit 2, and Using the quantized LPC coefficients of the current subframe and the next subframe input from the spectrum parameter quantizer 17 and the adaptive code vector input from the adaptive codebook 10,
The error E _a is calculated according to the equation (5), an adaptive code vector that minimizes the error E _a is searched, and the selected adaptive code vector is transmitted to the excitation codebook search circuit 8 and the gain codebook search circuit 9. Further, the delay d of the selected adaptive code vector is sent to the multiplexer 13.

In the excitation codebook search circuit 8, the signal input from the influence signal subtraction circuit 6, the unquantized LPC coefficients of the current subframe and the next subframe input from the linear prediction analysis circuit 2, and Quantized LPC coefficients of the current subframe and the next subframe input from spectrum parameter quantizer 17, selected adaptive code vector input from adaptive codebook search circuit 7, and input from excitation codebook 11 Using the obtained sound source code vector, an error E _e is calculated according to equations (9) to (11), a sound source code vector that minimizes the error E _e is searched, and a selected sound source code vector is searched for a gain codebook. The signal is sent to the circuit 9, and the index of the selected sound source code vector is sent to the multiplexer 13. Here, when calculating the error E _e, in order to reduce the calculation amount, the autocorrelation of the weighted synthesized signal se _i extended excitation code vector, Torankoso (M.Tranc
oso) and Atal (B. Atal), "Efficient Search Procedure for Selecting the Optimum Innovation in Stochastic Coders (Efficient Se)
arch Procedures for Select
ting the Optimum Innovati
on in Stochastic Coder
s) ", IEEE Tra Transactions on Sound, Speech and Signal Processing (IEEE Tra)
ns. Acoustic. , Speech, Signal
Processing), vol. Using the autocorrelation approximation method described in a paper (Reference 3) on pages 38, 385 to 396, the following equation (13) may be used.

[0053]

Here, hh is a weighted synthesis filter formed from a synthesis filter S using the quantized LPC coefficients of the current subframe and a weighting filter W using the non-quantized LPC coefficients of the current subframe. WS
Ee _i is the autocorrelation function of the excitation code vector at index i, im is
Impulse response length.

[0055] Further, when obtaining the cross-correlation between the weighted synthesis signal se _i and arbitrary vector v of the extended excitation code vector may be calculated as the following equation (14) in order to reduce the calculation amount.

[0056]

Where H is a weighting synthesis filter WS
Where H ^T represents the transposed matrix of H.

[0058] extended adaptive weighted combination of the code vector signal sa _d and also when obtaining the cross-correlation with any vector v, similarly, may be calculated by the following equation (15).

[0059]

In the gain codebook search circuit 8, the signal input from the influence signal subtraction circuit 6, the unquantized LPC coefficients of the current subframe and the next subframe input from the linear prediction analysis circuit 2, and Quantized LPC coefficients of the current subframe and the next subframe input from spectrum parameter quantizer 17, selected adaptive code vector input from adaptive codebook search circuit 7, and excitation codebook search circuit 8 Using the selected sound source code vector input from the above and the gain code vector input from the gain codebook 12,
A gain code vector that minimizes the error E _g is searched according to 2), the selected gain code vector is sent to the gain code book search circuit 9, and the index of the selected gain code vector is input to the multiplexer 13.
Send to

In the present embodiment, the unquantized LPC coefficients weighted by the auditory sense in the adaptive codebook search circuit 7, the sound source codebook search circuit 8, and the gain codebook search circuit 9 are used. The quantized LPC coefficient sent from the quantizer 17 may be used.

In this embodiment, the adaptive codebook search circuit 7, the excitation codebook search circuit 8, and the gain codebook search circuit 9 have the same overlap length. A long may be used.

[0063]

As described above, according to the method of the present invention, when searching the adaptive codebook, the sound source codebook and the gain codebook, the spectral parameter determined by the input speech signal and the result of the linear prediction analysis thereof. To obtain an auditory weighting signal having the length of the subframe length, and further, using the auditory weighting signal and the spectral parameter, obtain an overlap signal having a predetermined length, and after the auditory weighting signal, The adaptive codebook, the sound source codebook, and the gain codebook are searched for using the signal to which the overlap signal is connected.

As a result, the speech signal represented by the adaptive code vector, excitation code vector, and gain code vector of the current subframe is composed of the input speech signal of the current subframe and the next subframe based on the input speech signal of the current subframe. In this case, the input signal of the next subframe is used as an overlap signal by using the input signal of the current subframe on the next subframe as the overlap signal. Compared with the scheme (the scheme described in Reference 3), distortion at the block boundary of block coding caused by encoding for each subframe can be reduced more effectively.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Input terminal 2 Linear prediction analysis circuit 3 Subframe division circuit 4 Weighting filter 5 Overlap signal creation circuit 6 Influence signal subtraction circuit 7 Adaptive codebook search circuit 8 Sound source codebook search circuit 9 Gain codebook search circuit 10 Adaptive codebook 11 Sound source codebook 12 Gain codebook 13 Multiplexer 14 Synthesis filter 15 Weighting filter 16 Connection circuit 17 Spectrum parameter quantizer

Claims (2)

(57) [Claims] 1. A linear prediction analysis unit for obtaining a spectrum parameter of an input speech signal divided into frames at regular intervals , and a sub-frame obtained by further subdividing the frame.
From the input audio signal and the spectral parameters
An audibility weighting unit for obtaining an audibility weighted audio signal;
An adaptive codebook having a sound source signal determined in the past, a sound source weighting signal as a target signal, a sound source codebook unit for vector-quantizing the sound source signal of the input sound signal, the adaptive codebook and the sound source code. in the speech coding method and a gain codebook for quantizing the respective gains of the book, the audibility
The weighted signal is used as an initial value for the spectral parameter.
Overlap by a predetermined length using
Overlap signal generator for calculating a signal, the O
The burlap signal is converted to the auditory weighting signal of the subframe length.
No. of connected to the rear, longer target signal than the subframe length
And calculating the adaptive codebook, the excitation codebook, and the gain codebook for the target signal . 2. The search for the adaptive codebook, the sound source codebook, and the gain codebook, wherein lengths of the overlap signals connected after the perceptual weighting audio signal are set to be different from each other. 2. The speech coding method according to 1.