KR100389898B1 - Method for quantizing linear spectrum pair coefficient in coding voice - Google Patents

Abstract

PURPOSE: A method for quantizing a linear spectrum pair coefficient in coding voice is provided to select a code word minimizing a whole spectrum error by combining an error generated in an interpolation subframe with quantization. CONSTITUTION: A pth linear spectrum pair coefficient is extracted from a voice signal of an ith frame(201,202). A linear spectrum pair coefficient is extracted from the voice signal of the ith frame by subframes(207,208). Another linear spectrum pair coefficient used for synthesizing a voice of a previous frame is obtained. A third linear spectrum pair coefficient interpolated by subframes by linear interpolation is obtained. An index c minimizing a quantization error is obtained by using a codebook of linear spectrum pair coefficients.

Description

Quantization Method of Line Spectrum Pair Coefficients in Speech Coding

The present invention relates to a method of quantizing line spectrum pair coefficients in speech coding. More particularly, the present invention relates to an improved quantization method for selecting a codeword that minimizes an entire spectral error by combining an error occurring in an interpolation subframe with a quantization process. It is about a method.

In analytical synthesis and parametric speech coders having a low data rate of 8 kbps or less, speech is mainly represented by spectrum information and excitation information, and then quantized and transmitted.

This spectral information and excitation information are extracted for each block of voice called a frame, and for low data rate, the spectral information must be expressed within 24 bits / frame.

On the other hand, the speech decoder recovers the speech information from the transmitted bit stream and reproduces the synthesized sound. In this case, when the length of the subframe of speech is long, it is common to perform spectral interpolation to minimize the sudden change of the spectrum between frames.

However, since spectral interpolation is performed independently of quantization, the characteristic is irrelevant to quantization performance, and the longer the interval (interpolation subframe) synthesized by interpolation, the more the synthesized sound quality depends on interpolation characteristics rather than quantization performance. Therefore, there is a problem that information about interpolation must be transmitted as additional information in order to ensure high quality speech encoding of the speech encoder.

It is an object of the present invention to provide an improved method for integrating spectral quantization and interpolation processes instead of processing interpolation information as additional information, which has been devised to solve the above problems.

1 is a process diagram showing a conventional spectral quantization method.

2 is a process diagram showing a spectral quantization and interpolation method according to the present invention.

A method of quantizing line spectrum pair coefficients according to the present invention which achieves the above object is a method of vector quantizing line spectrum pair coefficients extracted from a speech signal, wherein the line spectrum pair coefficients of the order of p from the speech signal of the i-th frame are (Where k = 0 ,,,, p-1); Line spectrum pair count for each subframe from audio signal of i th frame (Where j = 0 ,,, s-1, and s are the number of subframes); Line spectrum pair coefficients used in speech synthesis of the entire frame Process of obtaining; Coefficient of line spectrum pair interpolated by subframe by linear interpolation Process of obtaining; And a process of obtaining an index c that minimizes the quantization error shown by the following equation using a codebook of line spectrum pair coefficients.

It is characterized by including. Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

Early research in the field of spectral quantization focused on what is the optimal spectral parameter in terms of improving quantization performance. Linear Predictive Coding (LPC) Coefficient, PARCOR (Paral CORrrelation) Coefficient, LSP (Line Specturm Pairs) Coefficient, and Impulse Response of Synthetic Filter are used as spectral parameters, and LSP coefficient is known to have the best performance. .

In addition, spectral interpolation has been studied for interpolation methods and interpolation intervals. Linear interpolation is mainly used as the interpolation method, and the narrower the interpolation interval, the less spectral distortion is found.

The 5.6kbps / s Vector Sum Excited Linear Prediction (VSELP) coder for the European mobile network Global Systems for Mobile Communication (GSM) improved linear interpolation from an interpolation codebook.

In addition, the 20㎳ M-LCELP (Multimode Learned CELP) proposed the interpolation spacing law in subframe by the method of split vector weighting. However, since the methods described so far are separated from the quantization process and interpolation process, even if the spectral parameters are quantized through the optimal quantizer, if the linearly interpolated spectral error is large, the spectral distortion of the entire frame increases.

Even if it is not optimally quantized but well interpolated, the spectral distortion of the entire frame can be relatively small.

As one of methods of considering spectral quantization and interpolation together, M-LCELP selects a plurality of candidates as spectral quantization values during spectral quantization. An interpolation quantizer process is performed for each selected plurality of candidates to determine a quantization value having a minimum spectral distortion as a final quantization result. However, this method has a disadvantage in that additional information is allocated and transmitted for interpolation quantization.

1 is a process diagram showing a quantization process of a conventional quantizer.

In step 101, the input voice A linear predictive analysis is performed to model the spectrum of. Through this, the voice is expressed by a linear prediction filter as shown in Equation 1.

Here, i is the analysis frame number, which means the current frame, Denotes the k-th order linear prediction coefficient (hereinafter referred to as LPC coefficient) in the i-th frame.

In step 102, the p-order LPC coefficient is converted into a line spectrum pair coefficient (hereinafter referred to as an LSP coefficient).

The conversion of the LSP coefficients from the LPC coefficients is a transform that can be inversely transformed. Since the quantization characteristics of the LSP coefficients are superior to those of the LPC coefficients, it is common to quantize the LSP coefficients instead of the LPC coefficients.

At this time, the LSP coefficient of the converted i th frame Let's do it.

In step 103, an LSP codebook is previously created from the training data. The mth codeword in the codebook is It is expressed as

In step 104, an optimal codeword is found using the outputs of steps 102 and 103. An error with any m-th codeword is expressed by Equation 2 below.

Here we assume weighted Euclidean distance as the distance measure. Is a weighting function.

If the number of codewords in step 103, that is, the codebook size is M, step 104 is to find a codeword index c that minimizes Equation 2 among the M codewords.

If this is expressed as an equation, Equation 3 below.

The index c is transmitted to the decoder and used for speech synthesis. Is allocated as in Equation 4 below.

If the analysis frame of speech is divided into s subframes, the value of Equation 4 becomes a parameter representing the spectrum of the last subframe, and for the other subframes, the quantized spectral parameter of the previous frame, that is, the previous frame. By interpolating with the LSP coefficient of the last subframe of, minimize the spectral change between subframes. The interpolated LSP coefficient at the j subframe of the current i th frame , the interpolation factor for the j subframe , The quantized LSP coefficient of the previous frame In this case, the interpolated LSP coefficient for each subframe can be obtained through linear interpolation according to Equation 5 below.

5b in Equation 5 denotes linear interpolation. It is expressed as.

The quantization methods of the models described so far do not reflect the interpolation characteristics of the subframes in the quantization process as shown in Equations 3 and 5. If the LSP coefficient of each subframe extracted In this case, the entire spectral error after interpolation is expressed as in Equation 6 below.

The total spectral error is expressed by Equation 7 below when m is selected instead of the codeword index c in step 104.

here, Assume that

therefore, Sufficient conditions for this are as shown in Equation (8).

Table 1 shows that the optimal codeword Seems not.

k code word 0ω _{o, k} code word1ω _{1, k} All frames Current frame Interpolation Frame Analysis Interpolation frame (select code word 0) Interpolation frame (select code word 1) 0 0.15 0.2 0.15 0.2 0.1 0.15 0.175 One 0.6 0.4 0.6 0.5 0.6 0.6 0.5

In this case, the LSP value of the current frame is quantized to 0.01, which is codeword 1, by Equation 2. The total error after quantized with codeword 1 and subjected to linear interpolation becomes 0.025625 by equation (6). However, if the quantization is codeword 0, the overall error is 0.015. Thus, Table 1 shows that optimal quantization does not mean optimal quantization and interpolation.

To solve this problem, the present invention uses Equation 6 as a quantization error equation.

2 is a process diagram illustrating a process of performing a spectral quantization / interpolation method (hereinafter referred to as JQI) according to the present invention.

Steps 201 to 203 are the same as steps 101 to 103 of FIG. 1.

In step 207, LPC analysis is performed for each subframe, and the LSP coefficient extracted in step 208 is used for an error equation of JQI. Step 206 performs the operation of the JQI. Accordingly, the quantization index c generated in operation 206 and input as operation 209 may be obtained as in Equation 9 below.

here, Denotes that the weighting function is a function of the analysis order k and the subframe j, unlike in Equation 3. Steps 209 and 210 are used in Equation 9 Corresponds to the process of obtaining.

In the process illustrated in FIG. 2, the quantization result of the example shown in Table 1 is codeword 1, and the total spectral error including quantization and interpolation is minimized.

Equation 9 is equal to Equation 3 by Equation 10 below.

Therefore, the JQI may be referred to as a generalized method including a conventional quantization method.

As described above, the spectral quantization and interpolation method according to the present invention has an effect of selecting a codeword that minimizes the overall spectral error by combining the error occurring in the interpolation subframe with the quantization process.

In addition, there is an advantage in that the weighting between the quantization subframe and the interpolation subframe can be different according to the selection of the weighting function.

Furthermore, there is an advantage that the speech coder using the conventional scheme can be improved without additional information.

Claims (1)

In the method of vector quantizing the line spectrum pair coefficient extracted from the speech signal, P-order line spectrum pair coefficient of order p from the audio signal of the i-th frame (Where k = 0 ,,,, p-1); Line spectrum pair count for each subframe from audio signal of i th frame (Where j = 0 ,,, s-1, and s are the number of subframes); Line spectrum pair coefficients used in speech synthesis of the entire frame Process of obtaining; Coefficient of line spectrum pair interpolated by subframe by linear interpolation Process of obtaining; And Using the codebook of the line spectrum pair coefficients, the process of finding the index c that minimizes the quantization error shown by A method of quantizing line spectrum pair coefficients, comprising.