JPH06202700A - Speech encoding device - Google Patents

Abstract

PURPOSE:To provide the speech encoding device which obtains a reproduced speech of high quality irrelevantly to a bit rate. CONSTITUTION:This device is equipped with a 1st code book 11 which contains a singular value vector group and a 2nd code book 12 which contains a residue vector group, and the difference between the output of a multiplier 13 and a speech conversion vector obtained by the unitary conversion of a unitary converting circuit 14 is found and stored in an error evaluating circuit 16. The error evaluating circuit 16 searches for a combination of the singular value vector and residue vector corresponding to the least value among differences obtained as to all combinations of the sigular value vectors and residue vectors in the two code books.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coder, and more particularly to an adaptive predictive coder using singular value decomposition.

[0002]

2. Description of the Related Art Conventionally, a speech code using vector quantization, in which a parameter to be encoded of a speech signal is regarded as a vector, an optimum vector pattern is selected from a codebook prepared in advance and block coding is performed. There is a conversion method. In the adaptive predictive coding method, there are a linear prediction coefficient and a residual signal as main parameters to be coded. Since the residual signal is noisy and can be treated as a random number sequence signal, this residual signal is tens of samples. The blocks are collectively regarded as one vector, registered in the codebook, and subjected to block coding. FIG. 2 shows a speech coding apparatus adopting this method.
Shown in. The apparatus here utilizes that the voice generation mechanism can be approximated by a model including a sound source and vocal tract characteristics. That is, the sound signal Y uses the sound source X and the synthesis filter H,

[0003]

[Equation 1] Y = H · X

The fact that can be expressed as follows is utilized.

The operation of the speech coder shown in FIG. 2 will be described below. In the speech coding apparatus shown in FIG. 2, the input terminal 2
The input speech input to 1 is linearly analyzed by the analysis circuit 22 to obtain a linear prediction coefficient, and the synthesis filter 2 is calculated from the linear prediction coefficient.
Make up 3. Then, the subtracter 25 obtains the difference between the input speech and the signal obtained by passing one residual vector among the residual vectors registered in the codebook through the synthesis filter 23. The obtained difference is squared by the squarer 26 and input to the error evaluation circuit 27. The error evaluation circuit 27 stores the value,
Instruct the codebook to output the next residual vector. This operation is performed for all the residual signals registered in the codebook for one input signal. Finally,
The error evaluation circuit 27 finds the smallest value among the obtained values and outputs the index (number) of the noise signal corresponding to the value. The prediction coefficient previously obtained by the analysis circuit and this index are multiplexed by the multiplexing circuit 28,
It is sent to the transmission line. Note that the decoder also has the same codebook 24, and can synthesize an audio signal from the received prediction coefficient and index.

By the way, in the speech coding apparatus of FIG. 2, the amount of calculation for finding the optimum residual signal is very large. Therefore, as a method for solving this problem, a method using singular value decomposition has been considered. As mentioned above, the mechanism of voice generation is

[0007]

[Equation 2] Y = H · X

Can be expressed as The synthesis filter H is a lower triangular matrix and can be expressed by the following equation.

[0009]

[Equation 3]

Here, h (O), h (1) ... h (N-
1) is the impulse response of the synthesis filter. When the synthesis filter H is decomposed into singular values and diagonalized, Expression 3 is expressed as Expression 2.

[0011]

[Equation 4]

Here, d ₁ , d ₂ , ... d _N are called singular values. U and V are unitary matrices for diagonalization. When Formula 1 is rewritten using Formula 3, it is expressed as Formula 5.

[0013]

(5) U ^t Y = DV ^t X

Here, if U ^t Y = θ and VtX = η,

[0015]

[Equation 6] θ = Dη

[0016] It should be noted that θ is a signal (voice conversion vector) obtained by unitarily converting the voice signal. Since the singular value vector D is a diagonal matrix, each element can be expressed independently,

[0017]

(7) θ = dη

With this method, the convolution operation of passing the residual vector through the synthesis filter can be converted into the element-wise product-sum operation. (Generally, instead of the singular value d,
Linear predictive coefficients are used. )

[0019]

However, in the conventional speech coding apparatus, coding is performed by paying attention only to the quantization error of the residual vector, and the quantization of the singular value (generally a linear prediction coefficient) is performed. The conversion error is not taken into consideration. In other words, if Equation 7 is rewritten in consideration of the quantization error,
θ = (d + Δd) (η + Δη) Conventionally, the residual vector with the smallest Δη is selected from the codebook and Δd is ignored. When the number of quantization bits is large, Δd can be ignored because it is considered to be sufficiently small. However, when the number of quantization bits is small, that is, when the bit rate is low, Δd cannot be ignored and the reproduced voice is reproduced. There is a problem of deterioration. SUMMARY OF THE INVENTION An object of the present invention is to provide a speech coder that can obtain reproduced speech of high quality regardless of the bit rate.

[0020]

According to the present invention, a unitary conversion circuit for unitary-converting a voice signal to output a unitary-converted voice signal, and a first codebook for accumulating a singular value vector group, Outputting a second codebook for accumulating a residual vector group, a product of a specific singular value vector output from the first codebook and a specific residual vector output from the second codebook A multiplier, a subtractor for obtaining the difference between the unitary-converted speech signal and the product, and the specific singular value vector for the first and second codebooks each time the difference is input. And specifying the specific residual vector, storing the differences corresponding to all combinations of the singular value vector and the residual vector, and corresponding to the difference having the smallest absolute value. A speech coding apparatus characterized by comprising an error evaluation circuit that searches for a combination of a specific singular value vector and the specific residual vector and outputs an index of the singular value vector and the residual vector To be

[0021]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 2 shows a block diagram of one embodiment of the present invention in FIG. 1. The speech coding apparatus according to the present embodiment has a first codebook 11, a second codebook 12, a multiplier 13, a unitary conversion circuit 14, a subtractor 15, and an error evaluation circuit 16. The first codebook 11 stores singular value vectors optimally created by learning from a large number of voice samples in advance. Also, the second codebook stores residual vectors that have been optimally created by learning from a large number of voice samples in advance.

The input voice input to the input terminal 17 is input to the unitary conversion circuit 14 and the voice conversion vector θ.
Is converted into (n) (1 ≦ n ≦ N: n is an element number). On the other hand, the first codebook 11 and the second codebook 12
From (1) and (2), a specific singular value vector and a specific residual vector are respectively taken out, and the multiplier 13 obtains the synthesized speech conversion vector θk (n). That is, θk (n) = di
Perform (n) × ηj (n) (i and j are indices). Next, the subtractor 15 calculates the difference {θ (n) −θi between the speech conversion vector θ (n) and the synthesized speech conversion vector θi (n).
(N)} is calculated. Calculated difference {θ (n) −θi
(N)} is squared by the error evaluation circuit 16 and stored as an error ei. Note that ek = Σ {θ (n) −θi (n)}
It is ² . The error evaluation circuit 16 instructs the first codebook and the second codebook to output the next set when the combination of one set of singular value vector and residual vector is completed.

As described above, the voice conversion vector θ
(N) and all the combinations of the singular value vector and the residual vector (if the numbers of i and j are I and J, I × J sets), a synthesized speech conversion vector θk (n)
When the comparison with is completed, the error evaluation circuit 16 obtains the obtained I
A combination of the singular value vector and the residual vector corresponding to the smallest value among the × J error eks is searched for. Then, the search result (indexes i and j) is output to the output terminal 18.

In this embodiment, all combinations of the singular value vector and the residual vector stored in the two codebooks are calculated, but simple product-sum calculation, subtraction and square calculation are performed. The amount is small.

[0025]

As described above, according to the present invention,
In the applied predictive coding method using singular value decomposition, not only the residual vector but also the singular value vector is stored in the codebook, and the most suitable vector is selected from among them. It is possible to prevent deterioration of the quality of the audio signal generated by the rate.

[Brief description of drawings]

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

FIG. 2 is a block diagram of a conventional speech encoding device.

[Explanation of symbols]

 11, 12 Codebook 13 Multiplier 14 Unitary conversion circuit 15 Subtractor 16 Error evaluation circuit 17 Input terminal 18 Output terminal 21 Input terminal 22 Analysis circuit 23 Synthesis filter 24 Codebook 25 Subtractor 26 Squarer 27 Error evaluation circuit 28 Multiplexing circuit

Claims (1)

[Claims] 1. A unitary conversion circuit for unitarily converting an audio signal to output an audio conversion vector, a first codebook for accumulating a singular value vector group, and a second codebook for accumulating a residual vector group. , The specific singular value vector output from the first codebook and the second singular value vector
A multiplier for outputting the product of the specific residual vector output from the codebook of No. 1, a subtractor for obtaining the difference between the voice conversion vector and the product, and the first and The specific singular value vector and the specific residual vector are specified for the second codebook, and the differences corresponding to all combinations of the singular value vector and the residual vector are stored, and the absolute value A combination of the specific singular value vector corresponding to the small difference and the specific residual vector, and an error evaluation circuit that outputs the index of the singular value vector and the residual vector. Speech coding device.