KR100322702B1 - Quantization and Decoding Method of Speech Signal Using Spectral Peak Pattern - Google Patents

Abstract

PURPOSE: A method of quantizing and decoding voice signal using a spectrum peak pattern is provided to reduce a calculation amount by 70 percentages while maintaining a performance of a split-linear spectrum pair vector quantization. CONSTITUTION: A quantization method comprises steps(100) of receiving a linear spectrum pair(LSP) vector, determining one of a plurality of linear predictive coding(LPC) spectrum peak patterns, and outputting a pattern index of the determined LPC spectrum peak pattern and the LSP vector. The quantization method comprises quantizing the output LSP vector using an M-LSP code book designed suitably for a property of a LPC spectrum peak pattern. After searching a code word close to the quantized LSP vector, an index of the code word is output(102). The quantization method comprises steps (104) of obtaining a difference between the quantized LSP vector and the most similar code word, quantizing the residual signal using an N-residual signal code book, finding a code work similar to the quantized residual signal, and outputting an index of the code word. There is sent to a decoder a bit stream that is formed using the spectrum peak pattern index generated in the step(100), the code book index generated in the step(102), and the code word index generated in the step(104,106).

Description

Quantization and Decoding Method of Speech Signal Using Spectral Peak Pattern

The present invention relates to a quantization and decoding method of a speech signal using a spectral peak pattern. In particular, when one line spectrum pair vector is input using a line spectrum pair vector quantization technique, a peak pattern of a corresponding linear predictive encoding spectrum is obtained. Determining one of the branch shapes, the line spectrum pair vector relates to a method of quantization using a corresponding codebook designed in advance for the characteristics of each spectral peak pattern and a suitable decoding method.

The present invention proposes an efficient novel spectral quantization scheme for low rate speech coders. In the field of digital speech coding, efforts have been made to express a speech signal at a lower transmission rate without compromising sound quality.

In particular, when an encoder based on a low rate linear predictive coding (LPC) of 4,8 k bps or less is intended to express a speech signal at a low bit rate without degrading syllables, information about a linear predictive coding (LPC) spectrum Efficient expression is essential.

The present invention proposes a new Line Spectrum Pairs (LSP) vector quantization technique, which will be referred to as Spectral Pattern Clssification (LSP Vector Quantization).

At present, it is difficult to allocate more than 32 bits per frame to the representation of linear predictive spectral information in the 2.4 ~ 4.8k bps low bit rate encoder environment developed by leading research groups in the United States, Europe and other countries. Generally, 20 to 27 bits are allocated. When using a scalar quantizer under such conditions, there is a problem that a lot of deterioration in sound quality.

On the other hand, the line spectral pair (LSP) coefficient is the most representative method of expressing linear predictive coding (LPC) spectrum, and the vector quantization method has been studied in various angles to reduce the transmission rate of the spectrum using the line spectrum pair coefficient.

Among them, distributed-line spectrum pair vector quantization (Split-LSP VQ) is represented with 25 bits per frame with spectral distortion less than 1 dB of line spectrum pair coefficient, but has a disadvantage of requiring a large amount of computation.

Therefore, the present invention solves the above-mentioned problems of the speech signal using the spectral peak pattern to reduce the calculation amount to 70% while maintaining the performance of the dispersion-line spectrum pair vector quantization (Split-LSP VQ). The purpose is to provide a quantization method.

In addition, another object of the present invention is to restore the residual signal and the line spectrum pair coefficients from the transmission bit stream encoded by the quantization method of the speech signal using the spectral peak pattern rate, and to output the vectored sum of the restored residual signal and line spectrum pair coefficients. The present invention provides a method of decoding a speech signal using a spectral peak pattern.

The quantization method of the speech signal using the spectral peak pattern of the present invention for achieving the above object

A first step of inputting a line spectrum pair vector to determine one linear predicted spectral peak pattern among the plurality of linear predicted encoded spectral peak patterns, and outputting a pattern index and a line spectrum pair vector in which the linear predicted encoded spectral peak pattern is determined;

A second step of quantizing the line spectrum pair vector output in the first step by using a pre-designed line spectrum pair codebook designed for the characteristics of the linear predictive spectral peak pattern, and outputting a codebook index most similar to the quantized line spectrum pair vector. ;

After calculating the difference (residual signal) between the quantized line spectrum pair vector and the corresponding codeword in the second step, the difference is quantized using the residual signal codebook, and the codebook index most similar to the quantized residual signal is obtained. Outputting a third step; And

A fourth step of generating a transmission bit string by inputting a spectral peak pattern index from the first step, a codebook index of a line spectrum pair vector from the second step, and a codebook index of the residual signal from the third step. It features.

In order to achieve the above object, a method of decoding a speech signal using the spectral peak pattern of the present invention,

Extracting spectral peak information, line spectrum pair vector quantization information, and residual signal quantization information from the transmission bit stream, and loading the residual signal line spectrum pair codebook and the residual signal codebook from the spectral peak information;

A second step of restoring the residual signal from the residual signal quantization information extracted in the first step;

A third step of recovering line spectrum pair coefficients from the line spectrum pair vector quantization information extracted in the first step: and

And a fourth step of outputting the vector sum of the residual signal reconstructed in the second step and the line spectrum pair coefficient reconstructed in the third step.

In the quantization method of an audio signal using the spectral peak pattern, in the case of the tenth order line spectrum pair spectrum, it is preferable that the peak pattern of the linear predictive encoding spectrum of the first step is eight. In addition, the speech spectral peak pattern of the first step is determined from the slope of the linear predictive encoding spectrum of speech, and the slope of the linear predictive encoding spectrum is determined from the line spectrum pair coefficients.

Hereinafter, with reference to the accompanying drawings will be described in detail with respect to the present invention.

1 is a flowchart showing quantization of line spectrum pair coefficients in the quantization method of an audio signal using a spectral peak pattern according to the present invention.

A line spectrum pair (LSP) vector is input to determine one linear predicted spectral peak pattern among a plurality of LPC spectral peak patterns, and the pattern index 11 and the line spectrum having the linear predicted spectral peak pattern determined. Output the pair vector 13 (100).

The line spectral pair (LSP) vector having the spectral peak pattern determined is pre-made through learning, that is, in step 100 using an M-line spectral pair codebook pre-designed for the characteristics of the LPC spectral peak pattern. Quantize the output line spectrum pair vector.

In addition, the codeword most similar to the quantized line spectrum pair vector 17 is found, and the index of the codeword is output. That is, codebook index 1 (15) is output (102).

In step 102, the difference between the quantized line spectrum pair vector 17 and the codeword most similar to the line spectrum pair vector 17 is obtained (hereinafter, referred to as a residual signal; Residual). Quantize using, find the codeword most similar to the quantized residual signal, and output the index of the codeword. That is, the codebook index 2 (19) is output.

A bit string is formed by using the spectral peak pattern index 11 generated in step 100, the codebook index 1 15 generated in step 102, and the codeword index 2 19 generated in step 104. After making 106, this bit string is sent to the decoder.

2 illustrates a method for obtaining a line spectrum pair vector in a method of decoding a speech signal using a spectral peak pattern according to the present invention. The process of obtaining a line spectrum pair vector from a bit string quantized and transmitted in FIG. This is a flow chart.

In step 200, the spectral peak information from the transmission bit stream and the line spectrum pair vector quantization information 21 corresponding to step 102 of FIG. 1 and the residual signal quantization information 23 corresponding to step 104 of FIG. ). Also, from the spectral peak information, a line spectrum pair codebook and a residual signal codebook corresponding to this information are obtained.

Residual signal 25 from the residual signal quantization information extracted in step 200 Reconstructs and outputs (202) the line spectrum pair coefficients (27) from the line spectrum pair vector quantization information. Restore the output. (204)

In step 206, the two signals 25 and 27 reconstructed in steps 202 and 204 are summed and made as final outputs. That is, the final output 29 Becomes

3 is a flowchart illustrating a method of classifying spectral peak patterns based on line spectrum pair coefficients. In an embodiment of the present invention, the relative positions of negative stack spectral peaks are well represented in the line spectrum pair coefficients. As shown in FIG. 3, the peak pattern was classified into eight types.

In step 300, the slope of the voice spectrum is examined.

When the slope of the spectrum changes from '+' to '-' there is a peak in the spectrum.

The equation for obtaining the slope s _i of the speech spectrum from the line spectrum pair coefficient w _i is as follows.

Where | A (e ^iw ) | ^-2 represents negative spectrum and can be obtained through linear predictive encoding analysis.

Also, the sign (.) Function

Is defined as

In expression (1), | A (e ^iw ) | ² is the linear predictive coding coefficient {α ₀ , α ₁ ,. Using α _p } is expressed as follows.

Thus, | A (e ^iw ) | ² is represented by the following formula (3).

Differentiate equation (4) with respect to w

| A (e ^iw ) | ^-4 takes the inverse of the result of Equation (4) and takes the inverse, so it is always positive for any w and therefore does not affect the sign (.) Of Equation (1).

Therefore, equation (1) can be expressed as follows.

In the present invention, the gradient sign vector is {0, w ₁ ,. w _p , Ⅱ} is a value obtained by determining the slope of {s ₀ , s ₁ ,. s _{p + 1} }.

4 is a graph showing the correlation between the tenth order linear predictive encoding spectrum and the line spectrum pair coefficient in the quantization of a speech signal using the spectral peak pattern according to the present invention, and the tenth order linear predictive coding spectrum and the line spectrum pair coefficient This example shows the correlation between.

The slope of the spectrum is represented by '+' and '-', and the peak of the spectrum is between the line spectrum pair coefficients (w ₁ , w ₂ ), (w ₆ , w ₇ ), (w ₈ , w ₉ ), respectively. do.

In step 302 to step 324 of FIG. 3, the spectrum is classified into eight types using the position of the peak information obtained from the spectrum slope information.

In the case of the example of FIG. 4, in step 302, it becomes 'yes', and in step 304 to 308, it becomes 'no', and in step 320, it becomes 'no', so it is classified as pattern 4.

Table 1 shows the probability distributions for the eight patterns. From these results, it can be seen that the probability of occurrence of the eight patterns is almost similar.

Table 1. Occurrence probability for each pattern

The following voice data was used to evaluate the performance of the present invention.

Two men and two women recorded 25 sentences and recorded a total of 100 sentences. A line spectrum pair vector was extracted from the recorded data. The performance of the linear predictive encoding-line spectrum pair vector quantization (SPC-LSP VQ) of the present invention was evaluated using 11694 line spectrum pair vectors extracted from 69 sentences.

For comparison, variance-line spectrum pair vector quantization (Split-LSP VQ) was designed from the same conditions, and the following Spectral Distortion Measure was used as a comparative scale.

SD is defined as follows.

Where A (e ^iw ), (e ^iw ) denotes a linear predictive coded polynomial before and after quantization. And the unit of SD is dB.

SD integrates the quantization error for all vectors in the spectral domain. The smaller the value, the smaller the quantization error.

First, the performance was evaluated using 11694 vectors used for vector quantization learning. This result is shown in Figure 5, where the X-axis is the number of bits per frame and the Y-axis is the average SD value obtained by equation (7).

Here, the linear predictive encoding-line spectrum pair vector quantization (SPC-LSP VQ) method has an SD smaller by 0.2 dB at a quantization rate of 11 to 21 bits / frame compared to the dispersion-line spectrum pair vector quantization method (Split-LSP VQ). It can be seen.

5 is a graph evaluating the performance of the conventional quantization method and the quantization method according to the present invention using data used for learning in one embodiment of the present invention, and FIG. 6 is not used for learning in one embodiment of the present invention. It is a graph evaluating the performance of the conventional quantization method and the quantization method according to the present invention using the data.

FIG. 6 shows the performance evaluation of 2163 vectors not used for learning, although the linear predictive and linear-spectrum pair vector quantization (SPC-LSP VQ) methods, although not as well as in FIG. There is no worse result than this scatter-line spectrum pair vector quantization (Split-LSP VQ).

The computations of the linear predictive code and the line-spectrum pair vector quantization (SPC-LSP VQ) method are compared with the scatter-line spectrum pair vector quantization method (Split-LSP VQ). The amount of computation in the vector quantizer depends on the calculation of the distortion measure during the search process.

The scatter-line spectral pair vector quantization (Split-LSP VQ) method and the linear predictive code and the line-spectrum pair vector quantization (SPC-LSP VQ) method use the Euclidean distance, so the comparison of the calculations for the calculation Will depend.

The time to calculate the first Euclidean distance is one plus two times. That is, three operations are required, and the time required to calculate the p-order line spectrum pair vector requires 2p times multiplication and 4p-1 operations.

Thus, the difference in computation amount depends on the number of search steps and the order of the line spectrum pair vectors used for the Euclidean distance calculation.

Table 2 shows the number of searches and the amount of computation for n-bit variance-line spectrum pair vector quantization. In the case of splitting one line spectrum pair vector into two, that is, using two codebooks, the size of the two codebooks is 2 Becomes the total search count is 2 × 2 Becomes

The order of the line spectrum pair vector used for the Euclidean distance calculation is 5. Also, if n is odd, each of the two codebooks Is assigned, the number of searches is The order of the line spectrum pair vector is 5th order.

Table 2. Search Counts and Computations for n-bit Distributed-Line Spectrum Pair Vector Quantization

Table 2. Comparative Example of Split-LSP VQ and SPC-LSP VQ

In the case of n-bit SPC-LSP VQ, the spectral pattern is classified into eight types and expressed in three bits, and the search frequency is quantized by the two-level codebook. Becomes And the order of the line spectrum pair vector is 10, Equation (6) should be calculated for the input vector.

The calculation requires 20 cosine (SOS) and sine function (SIN) calculations, 42 times, and 39 times for one element of the line spectrum pair vector, and 200 times, the cosine and sine function calculations, In total, 891 operations are required.

Assuming a cosine function calculation of 20 operations, 4891 operations are required. In addition, since the comparison of FIG. 3 is to be performed to classify the spectral peak into one of eight types, an additional operation is required.

Table 3 shows the calculation of Split-LSP VQ and SPC-LSP VQ when 26 bits and 27 bits are allocated per frame.

As can be seen from Table 3, the calculation of SPC-LSP VQ is about 70% more than that of Split-LSP VQ. Therefore, when SPC-LSP VQ quantizes a line spectrum pair vector with the same bits / frame as compared to Split-LSP VQ, it can be seen that the SPC-LSP VQ shows excellent performance with a small amount of calculation.

As shown in the above experimental results, the spectral pattern classification-line spectrum pair vector quantization (SPC-LSP VQ) method shows superior performance compared to the conventional scatter-line spectrum pair vector quantization (Spit-LSP VQ) method. We maintain about 70% of the variance-line spectrum pair vector quantization. In addition, the spectral distortion between the line spectrum pair vector before quantization and the line spectrum pair vector after quantization by the spectral pattern classification-line spectrum pair vector quantization (SPC-LSP VQ) method is about 1 dB when quantized at 27 bits per frame. Was evaluated.

As described above, in the quantization method of an audio signal using a spectral peak pattern according to the present invention, when a line spectrum pair vector is input, a peak pattern of a linear predictive encoding spectrum corresponding thereto is determined, and the determined line spectrum pair vector is a spectrum. By quantizing using the corresponding codebook designed to suit the characteristics of the peak pattern, it has the effect of greatly reducing the computation amount while maintaining the performance of the conventional distributed-line spectrum pair vector quantization.

In addition, it is possible to effectively reduce the transmission rate of a speech encoder using a linear predictive encoding spectrum without degrading the sound quality, and by determining the line spectrum pairs using a spectral peak pattern, the memory of the vector quantization codebook becomes small, and the line spectrum quantization or decoding In this case, the time is shortened.

In addition, the quantization error may be reduced by implementing the difference between the quantized line spectrum pair vector and the corresponding codeword, that is, the multi-stage quantization through the residual signal information.

1 is a flowchart illustrating quantization of line spectrum pair coefficients in a quantization method of an audio signal using a spectral peak pattern according to the present invention.

2 is a flowchart showing a method of obtaining a line spectrum pair vector in a method of decoding an audio signal using a spectral peak pattern according to the present invention.

3 is a flowchart showing a method of classifying spectral peak patterns based on line spectrum pair coefficients.

4 is a graph showing the correlation between the tenth order linear predictive encoding spectrum and the line spectrum pair coefficient in quantization of a speech signal using the spectral peak pattern according to the present invention.

5 is a graph evaluating the performance of the conventional quantization method and the quantization method according to the present invention using data used for learning in one embodiment of the present invention.

6 is a graph evaluating the performance of the conventional quantization method and the quantization method according to the present invention using data not used for learning in one embodiment of the present invention.

Claims (5)

In the quantization method of an audio signal using a spectral peak pattern, A first step of inputting a line spectrum pair vector to determine one linear predicted spectral peak pattern among the plurality of linear predicted encoded spectral peak patterns, and outputting a pattern index and a line spectrum pair vector in which the linear predicted encoded spectral peak pattern is determined; A linear spectrum encoding vector for quantizing the line spectrum pair vector output in the first step using a pre-designed line spectrum pair codebook designed for the characteristics of the linearly predicted stacked trim peak pattern, and outputting a codebook index most similar to the quantized line spectrum pair vector. Step 2; After calculating the difference (residual signal) between the quantized line spectrum pair vector and the corresponding codeword in the second step, the difference is quantized using the residual signal codebook, and the codebook index most similar to the quantized residual signal is obtained. Outputting a third step; And A fourth step of generating a transmission bit string by inputting a spectral peak pattern index from the first step, a codebook index of a line spectrum pair vector from the second step, and a codebook index of the residual signal from the third step. A quantization method of a speech signal using a spectral peak pattern. In the audio signal decoding method using the spectral peak pattern, Extracting spectral peak information, line spectrum pair vector quantization information, and residual signal quantization information from the transmission bit stream, and loading the residual signal line spectrum pair codebook and the residual signal codebook from the spectral peak information; A second step of restoring the residual signal from the residual signal quantization information extracted in the first step; A third step of recovering line spectrum pair coefficients from the line spectrum pair vector quantization information extracted in the first step; And And a fourth step of vector summing and outputting the residual signal reconstructed in the second step and the line spectrum pair coefficient reconstructed in the third step. The quantization method of an audio signal according to claim 1, wherein in the case of the tenth order line spectrum pair spectrum, the peak pattern of the linear predictive encoding spectrum of the first step is eight. The quantization method of a speech signal according to claim 1, wherein the speech spectral peak pattern of the first step is determined from a slope of a linear predictive encoding spectrum of speech. The method of claim 1, wherein the slope of the linear predictive encoding spectrum of the first step is determined from a line spectrum pair coefficient.