JPH09120300A - Vector quantization device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide the vector quantization device which can perform the vector quantization of a speech signal with good efficiency. SOLUTION: A means 21 takes a linear predictive analysis of an input speech signal S2 at different analytic window positions in the same frame to find a 1st LSP(linear spectrum pair) coefficient xi and a 2nd LSP coefficient yi , where the order (i) is a complex value; and a quantizer 22 performs the vector quantization of the 1st LSP coefficient xi , an inverse quantizer 23 performs the inverse quantization of the output quantized value CODE1 of the quantizer 22, and a means 24 finds the difference vector Δi consisting of the difference between the inverse quantized value xqi and the 2nd LSP coefficient yi of the same order. A means 26 decides plural properties of the signal S2 to obtain pieces mO-mK of mode information corresponding to the properties, a means 50 selects quantizers 51-53 corresponding to the mO-mK, and the selected quantizers 51-53 quantize Δi , so that the quantized value COD2 in transmitted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for encoding a voice by compressing the information of the voice signal, and more particularly to an LSP (Line Spectrum Pai) which represents the spectrum envelope information of the voice.
r) The present invention relates to a vector quantizer which extracts a coefficient (line spectrum pair coefficient) and vector-quantizes an LSP coefficient.

In recent years, this type of device has been widely used not only in wired telephones but also in digital mobile communications such as mobile telephones and car telephones, and is capable of efficiently compressing voice information while maintaining voice quality. Is required.

[0003]

2. Description of the Related Art Conventionally, in a speech coder, an LPC coefficient (linear prediction coefficient) representing frequency spectrum envelope information of a speech signal is transmitted, which is equivalent to the LPC coefficient and is excellent in quantization characteristic and interpolation characteristic. A method of converting to LSP coefficients and then quantizing is widely used.

In the voice coding system adopted in recent years for digital mobile communication such as mobile phones and car phones,
To further improve the quantization efficiency, vector quantization is used to quantize the LSP coefficient.

FIG. 10 shows a block diagram of a conventional vector quantizer and its description will be given. In FIG.
Reference numeral 11 is a codebook, 12 is a selection switch, 13 is an error minimizing unit, and 14 is a subtractor.

The input signal indicated by S1 is an LSP coefficient vector consisting of 10-dimensional LSP coefficients. Also, the codebook 1
1 stores L pieces of 10-dimensional code vectors, which is the same as the LSP coefficient vector S1, and indexes L to 1 are assigned to the L pieces.

The subtractor 14 obtains an error between the two by subtracting one code vector sequentially selected by the switch 12 from the LSP coefficient vector S1. The error minimization unit 13 controls the switch 12 so that the code vector having the smallest error is selected from the code book 11.

When vector quantization is performed in such a configuration, a vector having the smallest error (for example, Euclidean distance) from the LSP coefficient vector S1 is selected from the L code vectors, and the selected code is selected. By transmitting the vector indexes 1 to L, the transmission information is compressed.

[0009]

The above-described conventional voice coding method by the vector quantizer is widely used regardless of whether it is wired or wireless.

In wireless communication, effective use of frequency is required, and therefore reduction of transmission rate is essential. Therefore, in order to enhance the effect of information compression by vector quantization, the coding frame length (the number of input voice samples coded in one coding process), which is a unit of the coding process, tends to be long. .

However, if the length of the coded frame is increased, the ability to follow the time change of the voice may be deteriorated. Therefore, it is often the case that vector quantization is performed in units of subframes obtained by further dividing the frame. However, there is a contradiction that the information to be transmitted increases. This has been a big problem in the quantization of linear prediction coefficients, which is essential in speech coding.

To solve this problem, various methods have been proposed so far. For example, there are techniques such as matrix quantization, predictive vector quantization, and difference vector quantization.

Matrix quantization is an extension of vector quantization, and is a method of quantizing a plurality of vectors by treating them as one matrix. Although the degree of information compression is high, the distance calculation between matrices is performed in a matrix in a codebook. Since it has to be repeated for several minutes, there is a drawback that the amount of calculation becomes huge.

Prediction vector quantization and difference vector quantization are techniques for realizing information compression by utilizing the fact that there is a correlation between LSP coefficients between adjacent frames. However, when these methods are used for radio or the like, a problem of transmission error occurs.

In wireless, transmission errors due to the effects of fading are unavoidable. Since prediction vector quantization and difference vector quantization use past LSP coefficients, if a transmission error occurs and an error occurs in the LSP coefficient, there is a big problem that the influence continues to the subsequent frame.

In the speech coding apparatus, the error of the LSP coefficient is fatal, and if the speech is reproduced with the wrong LSP coefficient, there is a high possibility that a signal completely different from the input speech will be reproduced.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a vector quantizer capable of performing efficient vector quantization of a voice signal.

[0018]

FIG. 1 shows the principle of the present invention. The vector quantizer shown in this figure analyzes the input speech signal S2, extracts a linear prediction coefficient, and quantizes and transmits the linear prediction coefficient. The feature of the present invention is that the input speech signal S2 is transmitted in the same frame. The first LSP in which the order i takes multiple values by performing linear prediction analysis at different analysis window positions in
Analyzing means 21 for obtaining the coefficient x _i and the second LSP coefficient y _i
A vector quantizer 22 for vector quantizing the first LSP coefficient x _i ; an inverse quantizer 23 for inverse quantizing the output quantized value CODE1 of the vector quantizer 22; and an inverse quantizer 23 Output dequantized value x _qi and second LSP coefficient y _i
The difference detection means 24 for obtaining the difference vector Δ _i consisting of the difference between the same orders with and the plurality of properties of the input audio signal S2 are determined, and the mode information m according to the determined properties.
The determining means 26 for obtaining 0 to mK, and the first to Kth difference vector quantizers 5 corresponding to a plurality of properties of the input speech signal S2.
1 to 53 and the mode information m output from the determination means 26
First to Kth difference vector quantizers 5 corresponding to 0 to mK
1-53, and a selecting means 50 for supplying the difference vector Δ _i to the selected first-Kth difference vector quantizers 51-53. That is, the difference vector CODE2 quantized by the vector quantizers 51 to 53 is transmitted.

[0019]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 2 is a block diagram of a vector quantization apparatus according to an exemplary embodiment of the present invention.

In FIG. 2, reference numeral 21 denotes an LPC analysis unit, 22
Is a vector quantizer, 23 is an inverse quantizer, 24, 27,
28 is a subtractor, 25 is a vector division unit, 26 is a mode determination unit, 29 and 30 are error minimization units, 31, 32, 33,
34, 35, 36, 37, 38, 39, 40 are selection switches, B0, B1, B2, B3, C0, C1, C2, C
3 is a codebook.

In the vector quantizer comprising such components, as shown in FIG. 3, the input voice signal S2
Multiple times (two times in this example) within one frame in
Consider the case of vector-quantizing the LSP coefficient obtained by performing the PC analysis. However, what is LPC analysis?
This is a method of estimating the frequency characteristics of a voice signal by using the close correlation between voice samples. In other words, it is a method of estimating the coefficient of the filter that approximates the characteristics of the vocal tract in the voice generation model from the input voice.

Speech quality is improved by performing such LPC analysis a plurality of times. However, when vector quantization is performed a plurality of times within a frame as described in the prior art, compared to the case where LPC analysis is performed once. The amount of information to be transmitted increases by the number of times.

The LPC analysis unit 21 shown in FIG. 2 shows a case where the LPC analysis is performed twice, and the LPC analysis at the first and second times is performed.
As shown in FIG. 3, the LSP coefficients obtained by the PC analysis are LSP1 _i , (i = 1, ..., 10), LSP, respectively.
_Let 2 _i , (i = 1, ..., 10).

The LSP1 _i thus obtained is vector-quantized by the vector quantizer 22 to determine the index CODE1 to be transmitted. Next, C
Using the ODE1 as an input, the inverse quantizer 23 uses the LSP1
dequantized value LSP1q _i of _i, a (i = 1, ..., 10 ) determined. The inverse quantizer 23 used here is the same as the inverse quantizer used on the receiving side (decoder side).

The subtractor 24 causes LSP1q _i and LSP2
_Find the difference Δ _i between _i . _{Here, Δ i = LSP1q i -LSP2 i} , (i = 1, ..., 10) is (1). The subtraction method may be Δ _i = LSP2 _i −LSP1 q _i , (i = 1, ..., 10) (2). In the following, the case where the formula (1) is used will be described.

The existence range of the value of each order of the LSP coefficient is always between 0 and π. Where π is the Nyquist frequency. Therefore, the LSP coefficient exists between 0 and 1 when normalized by π.

On the other hand, it is known that the range of existence of the values of the respective orders of the difference Δ _i obtained by the equation (1) is narrower and is approximately between −0.25 and 0.25. That is, the range of the distribution of Δ _i is almost half of the existing range of the original LSP coefficient (here, LSP2 _i ).

The distribution of Δ _i also differs depending on the nature of the voice. Therefore, it is more efficient to quantize Δ _i by paying attention to the nature of speech than to simply quantize LSP2 _i .

Here, the mode information representing the property of the audio signal S2 of the frame currently being encoded is used. The voice signal S2 is divided into several modes such as voiced part, unvoiced part, unvoiced to voiced part, and voiced to unvoiced part.

The distribution of Δ _i also differs depending on the mode. Therefore, Δ _i for each voice property, that is, for each mode
Quantization can be performed more efficiently than conventional vector quantization by preparing an optimal codebook for. However, the codebook needs to be learned in advance.

As described above, according to the present invention, the input voice signal S2 is divided into a plurality of modes by utilizing the nature of voice, and the difference between the LSP coefficients is efficiently vector-quantized by using the codebook prepared for each mode. It has a structure to be converted. However, the above logic holds even when the LSP1 _i and LSP2 _i are interchanged.

Here, LSP (LSP1q_iAnd LSP2
_i) Difference Δ_iAnd the difference Δ _iDepending on the mode
Different vector quantizers (each codebook B0 to B3 and C0
Explain the advantage of quantization using C3)
You.

4 and 5 show histograms of LSP differences Δ _i . The LSP used here is obtained by LPC analysis of data of 400 single sentences (about 30 minutes) obtained by A / D conversion at a sampling frequency of 8 KHz. The analysis order was 10th.

[0034] FIG. 4 is a delta _i voiced portion of speech, and FIG. 5 is a delta _i of the part sound is changed to unvoiced portion from the voiced portion.
In FIG. 4 and FIG. 5, 1, 2, 7, and 8 are provided for ease of explanation.
The following is an example.

As is apparent from FIGS. 4 and 5, the average value of Δ _i is almost 0 and is localized in the vicinity of the average value. This is because the temporal change of the LSP coefficient is small.

Also, Δ_iDistribution depends on the degree of
It is clear that the lower the difference, the smaller the variance.
In addition, high-order differences have large variance. Also, depending on the mode
Also Δ _iIt can be seen that the distribution forms of are different.

[0037] Therefore, delta _i (1 to 10 order) low-order side (e.g. 1-4 order) of delta _L and higher side (for example, 5 to 10 order) of delta vectors split section into two vectors of _H 25 Split with
If each vector Δ _L , Δ _H is quantized using a quantizer that is optimal for the nature of the speech of the input frame, high quantization efficiency can be obtained. However, the number of divisions of Δ _i may be two or more.

That is, any one of the codebooks B0 to B3 is selected by the switch 31 according to the mode information modes 0 to 3 output from the mode determination unit 26, and the code vector of the selected codebook (for example, B0) is selected. The switches 32 are sequentially selected, and the selected one code vector is subtracted by the subtractor 27.
The error between the two is obtained by subtracting from Δ _L in Δ, and the error minimization unit 29 controls the switch 32 so that the code vector having the smallest error is selected from the code book B0. There is. Further, the codebooks C0 to C3 are also controlled in the same manner, and the description thereof will be omitted.

Next, the operation of the vector quantizer having such a configuration will be described. The input audio signal S2 is processed in units of frames including a fixed number of samples. That is,
The voice signal S2 in the frame is input to the LPC analysis unit 21, and the LPC analysis unit 21 performs the LPC analysis twice at different analysis window positions to obtain two sets of LSP coefficient LSP.
1 _i , LSP2 _i , (i = 1, ..., 10) are obtained.

Here, regarding the calculation of the LSP coefficient, it is also possible to first obtain the LPC coefficient (α coefficient) or the PARCOR coefficient and then convert the LPC coefficient into the LSP coefficient. on the other hand,
In the mode determination unit 26, the characteristics of the audio signal S2 are analyzed to determine the mode, and the mode information modes 0 to 3 are output. As a method for discriminating modes, any method already proposed can be used.

Here, as described above, the voice is converted to the unvoiced part,
It is classified into four modes: voiced to unvoiced change part, unvoiced to voiced change part, and voiced part change mode.
Assign numbers 1, 2, and 3. Therefore, the mode information has four values of modes 0 to 3.

The LSP1 _i is quantized by the quantizer 1 to obtain the code CODE1 which is the quantization result. As the quantizer 1, any previously proposed quantization method can be used.

Next, CODE1 is input to the inverse quantizer 23.
Is LSP1_iQuantized value of LSP1q_iIs sought
You. Then, in the subtractor 24, LSP1q_iAnd LSP2
_iDifference Δ_iIs calculated by the following equation. Δ_i= LSP1q_i-LSP2_i(I = 1, ..., 10) (3) Δ_iIs input to the vector division unit 25, and the low-order side (for example,
1 to 4) _LAnd the higher side (for example 5-10
) Consisting of_HTwo vectors are generated.

Next, Δ _L and Δ _H are vector-quantized, respectively. The code books B0 to B3 and C0 to C used at that time are quantized.
3 is selected according to the mode information modes 0 to 3. Here, codebook B0, codebook B1, codebook B2, codebook B3
Is a codebook for Δ _L , and modes 0, 1, 2, and
3 is supported.

Codebook C0, codebook C1, codebook C
2, codebook C3 is a codebook for Δ _H , and corresponds to modes 0, 1, 2, and 3, respectively. It is assumed that each codebook has been learned in advance for each mode.

Vector quantization of Δ _L and Δ _H is performed in mode3.
The case of performing vector quantization when the input voice signal S2 is a voiced portion will be described as an example. First, m
The codebook B3 is selected for Δ _L by the selection operation of the switch 31 in accordance with the information of the node 3. Codebook B3
The error between all the code vectors stored in the matrix and Δ _L is calculated by the subtractor 27, and the code vector with the smallest error is selected by the selection operation of the switch 35 under the control of the error minimization unit 29. Then, the index of this selected code vector is transmitted.

However, the Euclidean distance can be used as a measure of the error. Further, in order to reduce the processing amount, it is possible to evaluate the error with some code vectors in the codebook by performing some preliminary selection.

The procedure for Δ _H is the same as that for Δ _L except that codebook C3 is selected, and a description thereof will be omitted. The description is omitted because the same is applied except that the codebook used is different when another mode is selected.

Since the decoding side (reception side) has exactly the same codebook, it is easy to use L from the received index.
The quantized value of SP can be reproduced. In this embodiment, as many codebooks as the number of modes are required, but Δ _L , Δ
_Since the bias of the distribution of _H is used, the quantization can be efficiently realized with a smaller size codebook (with less transmission information) than the conventional vector quantization in which the LSP coefficient is quantized as it is.

Further, PARC is used instead of LSP coefficient difference.
You may make it quantize the difference of OR coefficient (It may be called a K parameter or a reflection coefficient.). In this case, the point that the difference between PARCOR coefficients is quantized is different from the above-described embodiment, and the other parts are the same, so the description will be omitted.

Here, in order to verify the performance of the quantizer of the present invention (the vector quantizer 22 shown in FIG. 2), the quantization error (conversion error) by the quantizer of the present invention and the quantum of the present invention will be described. 24 bits already proposed instead of the rectifier 22
The quantization error when the LSP2 _i is quantized by using both of the two-division / two-stage vector quantizer (hereinafter referred to as a reference quantizer) is compared.

It is known that the reference quantizer has characteristics equivalent to 30-bit scalar quantization and has high quantization efficiency. Further, as a measure of the performance of the quantizer, the LSP coefficient before quantization and the LSP after quantization are used.
The LPC cepstrum distance (hereinafter referred to as CD) with the SP coefficient is used. CD is a distance measure in the frequency domain and is defined by equation (4).

[0053]

(Equation 1)

Here, C _x (i) and C _y (i) are LPC cepstrum coefficients obtained from the LSP coefficient before quantization and the LSP coefficient after quantization, respectively. Also, p = 30. The smaller the CD value, the smaller the quantization error due to quantization.

40 was not used for learning the codebook for evaluation.
A single sentence (about 5 minutes) was used. The sampling frequency is 8
KHz. FIG. 6 shows C in the voiced part of the voice signal S2.
FIG. 7 shows the CD characteristics in the unvoiced part, FIG. 8 shows the CD characteristics in the part where voiced changes to unvoiced, and FIG. 9 shows the CD characteristics in the part where voiced changes to unvoiced.

In any of the modes shown in FIGS. 6 to 9, the quantizer of the present invention having a smaller number of bits than the reference quantizer exceeds the performance of the reference quantizer. Particularly, in the voiced mode shown in FIG. 6, the quantizer of the present invention having only 14 bits achieves the same performance as that of the reference quantizer of 24 bits.

Further, in the mode of changing from voiced to unvoiced shown in FIG. 8 and in the mode of changing from unvoiced to voiced shown in FIG. 9, it is found that a codebook size larger by 2 bits than the mode of the voiced part is required. It was

It is considered that this is because the variance of the LSP difference in the voiced mode is greater than the variance of the LSP difference in the voiced mode, and the variance of the LSP difference in the unvoiced to voiced mode is larger.

As described above, according to the vector quantizing device of this embodiment, it is possible to perform vector quantizing much more efficiently than the conventional vector quantizing device.

[0060]

As described above, according to the vector quantizing device of the present invention, the LSP coefficient is efficiently vector-quantized by utilizing the property of the voice, so that the efficient vector quantizing of the voice signal is performed. There is an effect that can be done.

[Brief description of the drawings]

FIG. 1 is a principle diagram of the present invention.

FIG. 2 is a block diagram of a vector quantization device according to an exemplary embodiment of the present invention.

FIG. 3 is a diagram for explaining the relationship between LPC analysis and frames.

FIG. 4 is a diagram showing a histogram of LSP differences in voiced voice.

FIG. 5: LS in the part where the voiced part changes to the unvoiced part
It is a figure which shows the histogram of the difference of P.

FIG. 6 is a diagram showing a CD characteristic in a voiced part.

FIG. 7 is a diagram showing a CD characteristic in an unvoiced part.

FIG. 8 is a diagram showing CD characteristics in a portion where voiced voice changes to unvoiced voice.

FIG. 9 is a diagram showing CD characteristics in a portion where voicelessness changes to voiced sound.

FIG. 10 is a block configuration diagram of a vector quantization device according to a conventional example.

[Explanation of symbols]

21 analysis means 22 vector quantizer 23 inverse quantizer 24 difference detection means 26 judgment means 50 selection means 51-53 1st-Kth difference vector quantizer S2 input speech signal x _i first LSP coefficient y _i second LSP coefficient CODE1 vector quantized value of first LSP coefficient x _i x _qi vector quantized value inverse quantized value of CODE1 Δ _i difference vector m0 to mK mode information CODE2 quantized value of difference vector Δ _i

Claims (5)

[Claims] 1. A vector quantizer which analyzes an input speech signal to extract a linear prediction coefficient, quantizes and transmits the coefficient, and performs linear prediction analysis on the input speech signal at different analysis window positions within the same frame. By analyzing the first LSP coefficient x _i and the second LSP coefficient y _i , the vector quantizer performing vector quantization of the first LSP coefficient x _i , and the vector quantizer An inverse quantizer that performs inverse quantization of the output quantized value, an output inverse quantized value of the inverse quantizer, and the second LSP coefficient y
difference detecting means for obtaining a difference vector composed of differences of the same degree with _i , judging means for judging a plurality of properties of the input voice signal, and obtaining mode information according to the judged properties, and the input voice First to Kth difference vector quantizers corresponding to a plurality of properties of the signal, and the first corresponding to the mode information output from the determining means.
To Kth difference vector quantizers, and selecting means for supplying the difference vectors to the selected first to Kth difference vector quantizers, A vector quantization device, which transmits the difference vector quantized by a difference vector quantizer. 2. A vector quantizer, wherein the order i is a tenth order of 1, ..., 10. 3. The mode information corresponds to a voiced portion of the input audio signal, a voiceless portion thereof, a voiceless to voiced portion thereof, and a voiced to voiceless portion of the input audio signal. The vector quantization device according to claim 1 or 2, characterized in that 4. A vector division means for dividing the difference vector into a difference vector of a low order group and a difference vector of a high order group;
As Kth difference vector quantizers, first to Kth difference vector quantizers for low-order groups that quantize the difference vectors of the low-order group and high-order groups for quantizing difference vectors of the high-order group A first to a Kth difference vector quantizer, wherein the selecting means selects the first to the Kth difference vector quantizers for the low-order group and the high-order group according to the mode information. The vector quantization device according to claim 1, wherein the selected difference vector quantizer for the low-order group and the high-order group transmits a difference vector quantized. 5. The vector quantizer according to claim 1, wherein PARCOR coefficients are used instead of LSP coefficients.