KR20010084468A - High speed search method for LSP quantizer of vocoder - Google Patents

Abstract

PURPOSE: A high speed searching method for an LSP quantization apparatus of a sound coder is provided to reduce a searching calculating amount by reducing a searching target code vector range. CONSTITUTION: An optimal arranging location with respect to each code book is determined(S10). An arranging order of the predetermined number of cod vector in each code book is substituted by a new code book arranged according to an element value of a predetermined reference row(S20). An element value of a reference row of the arranged code book is compared with previous and next elements values of a corresponding row of target vector to determine a searching range(S30). An error calculating amount of which is much is obtained to determine an optimal code vector(S40).

Description

High speed search method for LSP quantizer of vocoder

SUMMARY OF THE INVENTION The present invention has to be carried out in a speech coding technique, which is a core technology representing multimedia and mobile communication system performance (sound quality). The present invention is to encode and transmit speech information that efficiently encodes a continuous speech signal in a digital form. A compression problem that reduces data volume without deterioration. Searches a high computational codebook when split vector quantization of line spectral pairs coefficients of a speech coder is used to compress a speech signal at a low data rate among multimedia services. The present invention relates to a fast search algorithm that can reduce the size of the codebook to be searched by using the ordering property of the line spectral frequency (LSF) coefficients to reduce the overall calculation without degrading the SD performance.

Currently, the speech coding used at a transmission rate of 16kbps or less has an analysis / synthesis structure, and reduces the amount of transmission data in a bandwidth-limited transmission environment by efficiently extracting and quantizing parameters representing voice signals instead of transmitting the voice directly. It can have a structure.

It is very important to efficiently quantize linear predictive coding (LPC) coefficients representing the short-term correlation of speech signals for high-quality speech encoding in low-rate speech encoders. The optimal linear prediction coefficient value of the LPC filter is obtained by dividing the input speech signal into frame units and minimizing the energy of the prediction error for each frame. The LPC filter is generally a tenth order pole filter, where many bits are allocated for quantization of the ten linear prediction coefficients used.

When directly quantizing the number of the LPC filter, there is a problem that the characteristics of the filter is very sensitive to the quantization error of the coefficient and the stability of the LPC filter after the coefficient quantization is not guaranteed.

Therefore, LPC coefficients should be converted to other transmission parameters with good quantization properties and quantized, and mainly converted to reflection coefficients or LSP. In particular, the LSP value is closely related to the frequency characteristic of speech, so most recently developed standard speech compressors use the LSP quantization method. Also, vector quantization is frequently used for more efficient quantization.

In vector quantization, it is not possible to quantize an entire vector at once because the vector table is too large and takes a long time to search. To solve this problem, a method of dividing an entire vector into several subvectors and independently quantizing each one has been developed. This is called split vector quantization.

Predictive SVQ, a method of adding a predictor to SVQ, uses inter-frame correlation of LSP coefficients for efficient quantization. That is, the LSP of the current frame is predicted and the prediction error is quantized from the LSP value information of the past frame without directly quantizing the LSP of the current frame. The LSP value is closely related to the frequency characteristic of the speech signal, and thus can be predicted in time and a fairly large prediction gain can be obtained.

In addition, a method for dividing vector quantization into several steps, a selective vector quantization method for selectively quantizing using two tables, and a linked split vector quantization method for selecting a table to use based on the boundary value of each subvector have been developed.

The conventional method described above, when quantizing LSP coefficients using VQ, most quantizers have a huge LSP codebook. In addition, there is a problem that a large amount of computation is consumed in the process of searching the optimal code vector of the codebook.

SUMMARY OF THE INVENTION The present invention has been made to solve the problems of the prior art, and an object of the present invention is to provide an efficient codebook search method that can reduce the amount of search computation by reducing the range of code vector to be searched by using the order property of LSP coefficient values. have.

As a technical idea for achieving the object of the present invention, in the LSP coefficient quantizer used in the low-rate speech coder, the process of determining the optimal alignment position for each codebook using a lot of speech data, and each codebook Replacing the sorting order of the predetermined number of codevectors with a new codebook arranged in descending order according to the element values of the predetermined reference column, and the element values of the reference column of the sorted codebook and the element values before and after the corresponding column of the target vector. According to a predetermined operation flow diagram, a process of determining a search range by forward and reverse comparison, and a process of determining an optimal code vector by obtaining an error criterion having a large amount of calculation only within the determined search range are presented.

1 is a block diagram of an encoding stage of a G.723.1 speech encoder.

2 is a block diagram of the structure of SVQ, which is a general LSP quantization scheme.

3 is a flow chart according to the present invention.

4 is a flowchart illustrating a fast code vector search method.

5 is a block diagram of a forward and reverse comparison scheme when selecting a code vector, FIG. 5A shows a forward comparison, and FIG. 5B shows a reverse comparison.

<Explanation of symbols for the main parts of the drawings>

1: framer 2: high pass filter

3: LP coefficient analyzer 4: LSP quantizer

5: LSP Decoder 6: LSP Interpolator

7: formant recognition weight unit 8: pitch calculation unit

9: Harmonic Noise Synthesizer 10: Impulse Response Calculation Unit

11: zero input response 12: pitch predictor

13: MP-MLQ / ACELP 14: pitch decoder

15: the decoder here 16: memory update unit

18: LSP decoder 20: LSP vector P

21: SVQ 22:

23: M subvectors 24: codebook target vector

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

1 is a block diagram illustrating an encoding stage of a G.723.1 speech encoder.

As shown in FIG. 1, when the continuous speech sample signal y [n] is applied to the framer 1, the framer 1 has the 240 input signal signal per frame and has two subframes. Convert to a signal. In addition, the converted signal is further divided into two to generate a signal having 240 samples per frame and four subframes.

In addition, the high-frequency pass filter (2) receiving the output of the framer 1 removes the DC component, and then transfers a signal to the LP coefficient analyzer (3), the LP coefficient analyzer (3) is LP (Linear) predictive) Perform the operation of coefficients. At this time, four LP coefficient combinations are generated for each frame, and these LP coefficient combinations are used to construct a short-term perceptual weighting filter.

In addition, the LSP quantizer 4 receives the output signal of the LP coefficient analyzer 3, converts the LP coefficients (Linear predictive coefficient) into LSP coefficients, and the LSP encoder 5 converts the LSP quantizer ( The output of 4) is applied to encode the LSP coefficients.

In addition, the LSP interpolator 6 receives the output signal of the LSP encoder 5 and performs linear interpolation using the LSP vector decoded in the current frame and the LSP vector decoded in the previous frame.

On the other hand, the formant recognition weighting unit 7 receives the output signals of the high pass filter 2 and the LP coefficient analysis unit 3 and performs formant recognition weighting filtering every frame. In this case, the output signal of the LP coefficient analyzer 3 means an LP coefficient on which quantization has not been performed.

In addition, the pitch calculation unit 8 performs two pitch calculations for every frame, one operation for the 0th and the first frame among the four subframes, and the other operation for the second and the third subframes. It is done.

In this case, one thing to mention is that the operation of each block from now on is not performed on the signal of the entire frame, but is a subframe based operation.

The harmonic noise forming unit 9 is a part for improving the quality of the encoded speech signal, and the impulse response calculating unit 10 forms a formant recognition weighting filtering signal W (z) for the closed loop analysis. , LSP interpolator (6) output signal ( ) And the harmonic noise shaping unit 9 provide an impulse response to the combined filtering signal of the output signal P (z), and the zero input response unit 11 provides a zero input response of the combined filtering signal and pitch The prediction unit 12 is an adaptive codebook excitation source, the MP-MLQ unit 13 is a fixed codebook excitation source for 6.3 kbps, and the ACELP 13 is a fixed codebook excitation source for 5.3 kbps.

In addition, the pitch decoder 14 is a part for decoding the adaptive codebook excitation source, and the excitation decoder 15 is a part for decoding the fixed codebook excitation source.

In addition, the memory update unit 16 is a final process of encoding the subframe, and the formant recognition weight filtering signal W (z) and the LSP interpolator 6 output signal ( ) And the harmonic noise forming section 9 are the parts for updating the output signal _Pi (z) to the memory.

2 is a block diagram of the structure of SVQ, which is a general LSP quantization scheme.

As shown in Figure 2, LSP vector P (20), SVQ (21), It consists of 22.

This configuration is the structure of a generic SVQ used to quantize LSF vectors. The LSF vector P, the target vector, satisfies the ordering property of the following equation.

In Equation 2, the error criterion E _{l, m} is p and Where p _m is the target vector for the mth codebook search. _{l, m} corresponds to the lth codevector of the codebook for the mth subvector. At this time, the optimal codevector value for each subvector is selected to minimize the next error criterion E _{l, m} , and finally the selected codebook index l is transmitted through the channel.

In Equation 2, the LSP codebook vector is Is divided into M subvectors, and each subvector is composed of L _m code vectors. Codebook sizes L ₀ , L ₁ , ..., L _M-1 for _M subvectors sometimes use the same bit, but sometimes additional bits are allocated for a particular subvector to improve sound quality performance. W _m is a weighting matrix for the m th subvector and is obtained from p, an unquantized LSF vector.

In order to apply the fast search method in the present invention, conversion of an existing codebook is required. This is a process of replacing codebooks arranged in descending order with existing codebooks based on experimentally determined specific columns in order to easily determine the range of code vectors to search.

Selection of a specific column is performed separately for each codebook, and experimentally finds a column having a minimum average search range. The average search range is the order value of the elements of the nth column of the codebook arranged around each nth column and the values of the positions of n + 1 and n-1 of the target vector using the target vector. The average number that satisfies.

As shown in Equation 3 below, the element value of the n-th column of the target vector must be less than the element value of the n-th column of the codebook, and the element value of the n + l-th column of the target vector must be greater than the element value of the nth column of the codebook.

When the optimally determined specific column for each codebook is called the reference column N ₀ , N ₁ , ..., N _M of each codebook and the 10th-order LSF vector is assumed as the target vector, the following equations 4 and 5 are used to search. The search range of the codebook is determined by comparing the element values of the reference column of the codebook to be compared with the element values before and after the reference column in the target vector and excluding code vectors that violate the ordering quality.

Here, as shown in Equation 4, comparing element values of the Nth column of a certain code vector with elements of the (N-1) th column of the target vector to determine whether the order property is satisfied is referred to as forward comparison. As in Equation 5, comparing the element values of the Nth column of the code vector with the element values of the (N + 1) th column of the target vector is called backward comparison.

3 is a flow chart according to the present invention.

As shown in FIG. 3, the present invention provides a process for determining an optimally optimal alignment position for each codebook by using a large amount of speech data in the LSP coefficient quantizer used in a low-speed speech coder (S10), and each codebook. Replacing the sort order of the predetermined number of code vectors with a new codebook arranged in descending order according to the element values of the predetermined reference column (S20), and before and after the corresponding column of the reference value of the sorted codebook and the corresponding vector of the target vector. Determining a search range by forward and reverse comparison of element values according to a predetermined operation flowchart (S30), and determining an optimal code vector by obtaining an error criterion having a large amount of computation only within the determined search range (S40). .

4 is a flowchart illustrating a fast code vector search method.

As shown in FIG. 4, f1, f2, b1, and b2 represent code vector element values and target vector element values used for forward and reverse comparison, respectively. Here, since each codebook is sorted in descending order, only a starting point satisfying the ordering property is obtained in the forward comparison, and the rest of the codebooks satisfy the ordering property of the forward order.

In the reverse comparison, a process of obtaining a specific start point and an end point of a code vector set satisfying an ordering property is illustrated in an operation flowchart as shown in FIG. 5.

5 is a block diagram of a forward and reverse comparison scheme when selecting a code vector, FIG. 5A shows a forward comparison, and FIG. 5B shows a reverse comparison.

As shown in FIG. 5A, the process of determining a codebook search start point includes calculating a LSP vector p (S100), initializing a variable i to 0 (S110), p _{n + 1} and Comparing and determining the size of _{i + 64, n} (S120), the size of p _{n + 1} in the process (S120) _{If i + 64, n,} the process of increasing the variable i by 64 (S130), and the size of p _{n + 1 When i + 64, n} is greater than (S140) storing the variable i, initializing the variable j to the value of the variable i stored variable (S150), and the size of p _{n + 1} Comparing and determining the size of _{j + 16, n} (S160) and the size of p _{n + 1} in the process (S160) _{In the case where j + 16, n} is less than _16, the process of increasing the variable j by 16 (S170) and the size of p _{n + 1 If i + 16, n} is greater than (S180) for storing the variable j, initializing the variable k to the value of the stored variable j (S190), and the size of p _{n + 1} Comparing and determining the size of _{k + 4, n} (S200), the size of p _{n + 1} in the process (S200) If less than _{k + 4, n} increases the variable k by 4 (S210) and the size of p _{n + 1} If greater than _{k + 4, n} (S220), storing the variable k, initializing the variable m to the value of the stored variable k (S230), and the size of p _{n + 1} Comparing and determining the size of _{m + 1, n} (S240) and the size of p _{n + 1} in the process (S240) _{If m + 1, n,} the process of increasing the variable m by 1 (S250) and the size of p _{n + 1 If m + 1, n} is greater than the step of storing the variable m + 1 (S260), and the step of setting the m + 1 calculated through the above process (S270).

As shown in FIG. 5B, the process of setting the codebook search end point may include calculating the LSP vector p (S300), initializing the variable i with L _m (S310), p _n−1 and Comparing and determining the size of _{i-64, n} (S320) and the size of p _n-1 in the process (S320) _{In the} case of larger than _{i-64, n,} the process of reducing the variable i by 64 (S330) and the size of p _{n-1 If i-64, n} is smaller than the step of storing the variable i (S340), initializing the variable j to the value of the stored variable (S350), and the size of p _n-1 Comparing and determining the size of _{j-16, n} (S360) and the size of p _n-1 in the process (S360) If the size larger than _{j-16, n,} the process of reducing the variable j by 16 (S370), and the size of p _n-1 If smaller than _{j-16, n} (S380) for storing the variable j, initializing the variable k to the value of the stored variable (S390), and the size of p _n-1 Comparing and determining the size of _{k-4, n} (S400) and the size of p _n-1 in the process (S400) If greater than _{k-4, n,} the process of reducing the variable k by 4 (S410), and the size of p _n-1 If less than _{k-4, n} (S420) for storing the variable k, initializing the variable m to the value of the stored variable (S430), and the size of p _n-1 Comparing and determining the size of _{m-1, n} (S440) and the size of p _n-1 in the process (S440) _{In the} case of larger than _{m-1, n,} the process of reducing the variable m by 1 (S450) and the size of p _{n-1 When it} is smaller than _{m-1, n} , the process includes storing the variable m-1 (S460), and setting the m-1 calculated through the above processes as an end point (S470).

The codebook search range is determined by the start point and the end point obtained by the above processes of the flowchart.

Although the present invention has been described with reference to exemplary embodiments, these descriptions should not be interpreted in a limiting sense. It will be apparent to those skilled in the art that the present invention may be variously modified, combined, or implemented in various ways with reference to the detailed description of the invention. Accordingly, the following claims should be considered to cover all such modifications and embodiments.

As can be seen from the above description, the present invention provides a process for determining an optimally optimal alignment position for each codebook by using a large amount of speech data in an LSP coefficient quantizer used in a low-rate speech coder. Replacing the order of the predetermined number of codevectors with a new codebook arranged in descending order according to the element values of the predetermined reference column, and preserving the element values of the reference column of the sorted codebook and the element values before and after the corresponding column of the target vector. Determining a search range by forward and reverse comparison according to an operation flowchart, and determining an optimal code vector by obtaining an error criterion having a large amount of computation only within the determined search range, and encoding a codebook for a target vector to be searched. Sort codevectors in descending order based on the size of element values in specific columns in the vector Line up. Then, since the optimal codevector values that minimize distortion in the sorted codebook are almost similar to the target vectors, these values can be assumed to have an ordering property. Under this assumption, the element values of a particular column sorted in descending order are compared with the values of adjacent columns in the target vector. Therefore, only a large amount of computational distortion is calculated for code vectors that satisfy the ordering quality. Otherwise, the computation of distortion values is omitted. When searching each codebook, the search target can be reduced by using the order property of the LSP vector, and the overall computation amount can be reduced by reducing the size of the searched codebook by using the order property of the line spectral frequency coefficient. There is an effect to reduce.

Claims (6)

In the high-speed search method using the ordering of the LSF coefficients to reduce the amount of computation required during codebook search without attenuating the spectral distortion (SD) performance in the split VQ quantizer by the split VQ method used in the low-speed speech coder, A first step of rearranging the codebooks according to element values of a specific column in order to easily determine the range of searched code vectors; And a second process of determining a search range using a sequence property between a given target vector and aligned code vectors and obtaining an optimal code vector. The method of claim 1, wherein the first process comprises: determining an experimentally optimal alignment position N _m for each codebook using a plurality of voice data; A method of fast searching for an LSP codebook of a speech coder, comprising: replacing a sorting order of L _m codevectors in each codebook with a new codebook arranged in descending order according to element values of a predetermined reference string. The method of claim 2, wherein the determining of the alignment position N _m is performed by selecting a specific column for each codebook, and experimentally searching for a column having a minimum average search range. Fast navigation method. The method of claim 1, wherein the second process is a process of determining search ranges by forward and backward comparison of element values of a reference column of the codebook arranged in the first process and element values before and after the corresponding column of the target vector according to a predetermined operation flowchart. And determining an optimal code vector by obtaining an air reference E _{l, m} having a large amount of computation only within the determined search range. The method according to claim 4, wherein in the searching range determining process, an element value of the n th column of the codebook arranged around each n th column using a target vector for the sorted codebook, and n + l and nl of the target vector. A fast search method for an LSP codebook of a speech coder, characterized in that an element value of a position is an average number satisfying ordering quality. The method of claim 4, wherein E _{l, m} follows the following equation. P _m : target vector for m-th codebook search : the lth codevector of the codebook for the mth subvector W _m : Weighting matrix for mth subvector