KR20130080271A - A method and device for klt based domain switching split vector quantization - Google Patents

Abstract

PURPOSE: A Karhunen-Loeve transform (KLT) based domain switch split vector quantization method and an apparatus are provided to use a the KLT based SVQ apparatus and algorithm, thereby reducing an abnormal signal distortion part of the domains which is caused by the interaction between the KLT domain and an original domain. CONSTITUTION: A KLT domain converter (101) converts an inputted audio signal vector into a small sized sub-block. The KLT domain converter outputs a KLT vector signal by the KLT conversion of the converted audio signal vector. A first quantizing device (120) compares each of the values of the inputted audio signal vector with the codes of each code book and prints out the nearest code book. A second quantizing device (103) compares each of the values of the KLT vector with the codes of each code book and prints out the nearest code book. A final code book determiner (105) compares the code vector which is quantized in an original domain and the KLT domain of the audio signal, respectively, and selects the nearest code vector. [Reference numerals] (101) KLT domain converter; (104) Inverse KLT domain converter; (105) Final code book determiner

Description

MELT-based domain switch split vector quantization method and apparatus {A METHOD AND DEVICE FOR KLT BASED DOMAIN SWITCHING SPLIT VECTOR QUANTIZATION}

The present invention relates to a speech signal processing method. Specifically, the present invention relates to a method of selecting a codebook in a split vector quantization (SVQ) algorithm.

It is very important to efficiently quantize a linear spectral frequency (LSF) coefficient representing a short-term correlation of a speech signal for high quality speech encoding in a speech encoder. The optimal linear prediction coefficient value of a linear predictive coeffieient (LPC) filter is obtained by dividing an input signal into frames and minimizing the energy of prediction error for each frame.

LSF data converted in LPC increases coding efficiency, but there is a problem in that LSF data becomes more complicated in multiple dimensions. As the order of VQ (Vector Quantizatio) increases, the computation becomes complicated and the memory requirements increase. In order to solve this problem, Split Vector Quantization (SVQ) has been proposed.

When directly quantizing the coefficients of the LPC filter, the characteristics of the filter are very sensitive to the quantization error of the coefficients, and the stability of the LPC filter after coefficient quantization is not guaranteed. Therefore, LPC coefficients should be converted to other parameters with good quantization properties and quantized, and mainly converted to reflection coefficients or LSF. In particular, LSF values are closely related to the frequency characteristics of speech, so most recently developed standard speech compressors use the LSF quantization method.

The LSF quantization method uses interframe correlation of LSF coefficients for efficient quantization. That is, the LSF of the current frame is predicted and the prediction error is quantized from the LSF value information of the past frame without directly quantizing the LSF of the current frame. The LSF 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.

This method has a weak problem with channel error because it uses the reconstructed previous frame. If an error occurs in the channel, the frame that is modified due to this error can ruin all subsequent data. This is because the previous frame used for removal is a reconstructed frame. Therefore, it is not used in a real time system which is sensitive to channel. However, this algorithm is not a problem when used in systems that do not affect the channel. For example, text to speech (TTS), telephone answering devices (TADs), voice recorders, and emergency callback systems do not affect the channel. Since this system is not a real-time system, the channel error problem can be ignored.

In general vector quantization, it is impossible 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 of them has been developed. This is called split vector quantization (SVQ). For example, in the 10th-order vector quantization using 20 bits, the vector table has a size of 10 × 220. However, a split vector quantization method of dividing into two fifth-order subvectors and assigning 10 bits each to each other is provided. If used, the size of the vector table is only 5 × 210 × 2. Dividing into more subvectors reduces the size of the vector table, which saves memory and reduces the search time. However, it does not have enough correlation between the vector values.

Meanwhile, a lot of researches on LSF quantization have been conducted to improve the performance of speech coding by compressing LSF. Resolution-constrained quantization (RCQ), which uses a fixed bit rate, results in a cell size that depends on the probability density function (PDF) of input data. This results in outliers in distortion. The abnormal signal distortion is a factor that degrades the performance.

In addition, the general vector quantizer stores the codebook. Many studies are trying to make an optimal codebook during the training process, but in the real world, the source mismatch occurs because the learning and the test are different. . Due to such source mismatch, distortion is increased and a problem of degrading voice coding occurs.

The background technology of the present invention is described in Republic of Korea Patent Publication No. 10-0446630 (August 23, 2004).

An object of the present invention is to provide a method for selecting a codebook closer to an input speech signal and a vector in order to reduce performance degradation caused by abnormal signal distortion and source miss match during LSF quantization.

According to the spirit of the present invention for achieving the above object, the quantization apparatus, KLT domain conversion unit for converting the input voice signal vector to a small sub-block KLT transform to output a KLT vector signal, the input A first quantizer which compares the respective values of the speech signal vector with the codes of the respective codebooks and finds the closest codebook, and compares the respective values of the KLT vector with the codes of the respective codebooks to find the closest codebook. A second coder for outputting and a final codebook determiner for selecting the nearest code vector by comparing the quantized code vector in the original domain of the speech signal and the quantized code vector in the KLT domain with the source vector, respectively.

In one embodiment, the quantization device is a split vector quantization (SVQ) device.

Preferably, the apparatus further includes an inverse KLT domain transform unit for inverse KLT transforming and outputting the KLT vector output from the second quantizer, and the final codebook comparator includes an output from the first quantizer and the inverse KLT domain transform unit. Compare code vectors with source vectors.

In one embodiment, the input speech signal is a line spectral frequency (LSF) type signal, and the input speech signal is a difference value between the input current frame and the predicted current frame data.

^k를 찾고, 상기 제2 양자화기는 상기 KLT 변환부를 통해 변환된 X^k1 내지 X^kn의 값을 갖는 KLT 벡터 X^k의 각각의 값에 대해 각 코드북의 k¹ 내지 kⁿ 번째 코드들과 비교하여 가장 가까운 코드북

^k를 찾는다.More preferably, the first quantizer is the closest codebook in comparison to the k ¹ to k ^nth codes of each codebook for each value of the input speech signal vector x ^k having values of x ^k1 to x ^kn .

Finding ^k , the second quantizer is most compared to the k ¹ to k ⁿ th codes of each codebook for each value of the KLT vector X ^k having values of X ^k1 to X ^kn transformed through the KLT transform unit Nearest codebook

Find ^k

The KLT-based SVQ apparatus and algorithm according to the present invention reduces the abnormal signal distortion part of each other due to the interaction between the KLT domain and the original domain. By using two domains, the codevectors are widely located, reducing source mismatches.

When the SD (Spectral Distortion) is measured by comparing the performance of the KLT-based SVQ according to the present invention with the conventional SVQ, the KLT-based SVQ according to the present invention has a gain of 1 bit or more compared with the existing SVQ.

1 is a block diagram of a quantization apparatus using a KLT domain according to the present invention,
FIG. 2 is a flowchart of a quantization method implemented through the quantization device according to FIG. 1.

Specific details of other embodiments are included in the detailed description and the drawings.

Advantages, features, and methods of achieving them will be apparent with reference to the embodiments described below in conjunction with the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Is provided to fully convey the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

For efficient speech coding, the human vocal tract filter is modeled by Linear Predictive Coding (LPC) coefficients. Linear Spectral Frequency (LSF) has been proposed for better coding of LPC coefficients. Each frame of speech can be generalized as an all-pole filter, H (Z) = 1 / A (Z). Where A (Z) is an inverse filter composed of LPCs, and its value is represented using a Z-transform.

In Equation 1, P is the LPC order and a is the LPC coefficient. To define the LSF, an inverse filter can be represented by two polynomials as shown in Equation 2 below.

In Equation 2, P (z) is an even term of the inverse filter, and Q (z) represents an odd term. The LSF coefficient used for coding means the root of P (z) and Q (z). LSF obtained by Equation 2 is sorted in ascending order. Its properties help to make efficient LSF calculations, thus increasing compression efficiency.

In case of using 10th order LPC, 10th order LSF data is generated. As vector quantization increases as the dimension increases, vector quantization greatly increases the amount of computation. Therefore, there is a split vector quantization (SVQ) that properly divides and quantizes vectors to solve this problem. This method reduces computation, but split loss occurs because the inter-dimensional correlation is not partially used. Thus, SVQ is less than VQ.

1 is a configuration diagram of a quantization apparatus using a KLT domain according to the present invention, and FIG. 2 is a flowchart of a quantization method implemented through the quantization apparatus according to FIG. 1.

As a kind of encoding paradigm, there is a quantization method based on Karhunen-Loeve Transform (KLT). KLT is one of methods for removing correlation between vectors.

^k를 찾게 된다(S210). 제2 양자화기(103)는 KLT를 통해 변환된 X^k1 내지 X^kn의 값을 갖는 KLT 벡터 X^k의 각각의 값에 대해 각 코드북의 k¹ 내지 kⁿ 번째 코드들과 비교하여 가장 가까운 코드북

^k를 찾게 된다(S220).1 and 2, the SVQ quantization apparatus 100 according to the present invention includes a KLT domain transform unit 101, a first quantizer 102, a second quantizer 103, and an inverse KLT domain transform unit ( 104) and a final codebook determiner 105. When the input signal x ^k is input, the KLT domain converter 101 converts the input signal x ^k into a sub block having a small size, thereby generating the X ^k from the KLT vector signal (S200). A first quantizer (102) closest codebook compared to the respective codebooks of k ¹ to k ^n-th codes for each value of the signal vector x ^k which has a value of an input x x ^k1 to ^kn

^k is found (S210). A second quantizer 103 is the closest codebook compared to the respective codebooks of k ¹ to k ⁿ th code for each value KLT of vector X ^k has a value of X X ^k1 to ^kn converted by a KLT

^k is found (S220).

The first and second quantizers 102 and 103 vary in performance depending on bit allocation. To achieve optimal performance, bit allocation is determined by the covariance matrix and dimensions. The bit allocation equation is the same as Equation 3.

In Equation 3, c and c _i are bits of the entire bit and the i-th subblock, respectively, and k and k _i represent the dimension and the dimension of the i-th subblock. G _k is a constant that depends on the dimension, and C _Yi is a covariance matrix of the i th subblock in the KLT domain. Equation 3 is used for bit allocation of the second quantizer in the KLT domain, and the first quantizer is used by replacing C _Yi with C _Xi in Equation 3. Where C _Xi is the covariance matrix of the i th subblock in the original domain.

Each subblock is trained using the Linde-Buzo-Gray (LBG) algorithm. If many classes of codebooks are included in the codebook database, the SNR efficiency of the vector quantizer for speech signals will be further improved.

Then, the inverse KLT domain converter 104 inverse KLT transforms the quantized speech signal output from the second quantizer 103 again.

The KLT domain converter 101 and the inverse KLT domain converter 104 are configured in units of frames.

The final codebook determiner 105 selects the closest code vector by comparing the source code with the selected code vector in the two domains output from the first quantizer 102 and the inverse KLT domain transform unit 104 (S230). ).

That is, the SVQ according to the present invention selects a code vector having a small distortion in each domain by using a quantizer designed in the original domain and a quantizer designed in the KLT domain. Finally, the closest codevector is selected by comparing the source code with the selected codevector in each domain.

When the distortion values of the code vector and the source vector of the KLT domain are relatively smaller than the distortion values of the code vector and the source vector of the original domain, the distortion is reduced by selecting the code vector of the KLT domain. On the contrary, when the distortion values of the code vector and the source vector of the original domain are relatively smaller than the distortion values of the code vector and the source vector of the KLT domain, the distortion is reduced by selecting the code vector of the original domain.

This interaction reduces the abnormal signal distortions of each other. By using two domains, the codevectors are widely located, reducing source mismatches.

When the SD (Spectral Distortion) is measured by comparing the performance of the KLT-based SVQ according to the present invention with the existing SVQ, it can be seen that the KLT-based SVQ according to the present invention obtains a gain of 1 bit or more compared with the existing SVQ.

Although the present invention has been described with reference to the embodiments illustrated in the accompanying drawings, it is merely an example, and a person having ordinary knowledge in the art related to the present invention may have various modifications or equivalents thereto. I will understand that it is possible.

Accordingly, the true scope of the present invention should be determined by the technical idea of the appended claims.

100: SVQ quantization apparatus 101: KLT domain converter
102: first quantizer 103: second quantizer
104: inverse KLT domain conversion unit 105: final codebook determination unit

Claims (16)

A KLT domain conversion unit for converting the input voice signal vector into a small sub-block and performing KLT conversion to output a KLT vector signal;
A first quantizer which compares respective values of the input speech signal vector with codes of respective codebooks to find and output the closest codebook;
A second quantizer which compares respective values of the KLT vector with codes of respective codebooks to find and output the closest codebook; And
And a final codebook determiner for selecting the nearest code vector by comparing the quantized code vector in the original domain of the speech signal and the quantized code vector in the KLT domain with a source vector, respectively. The method of claim 1,
Wherein the first quantizer or the second quantizer sets bit allocations according to covariance matrices and dimensions using the following equation:

Where c and c _i are bits of the entire bit and the i subblock, k and k _i represent the dimensions of the dimension and the i subblock, G _k is a constant that depends on the dimension, and C _Yi is the KLT domain. Or the covariance matrix of the i-th subblock in the original domain. The method of claim 1,
The quantization device is
A quantization device, characterized in that the split vector quantization device. The method of claim 1,
The quantization device is
And an inverse KLT domain transform unit for inverse KLT transforming and outputting the KLT vector output from the second quantizer. The method of claim 4, wherein
The final codebook comparator compares the code vectors output from the first quantizer and the inverse KLT domain transform unit with a source vector. The method of claim 5, wherein
And the input voice signal is a signal of a linear spectral frequency type. The method according to claim 6,
The input speech signal is a difference between the input current frame and the predicted current frame data. The method of claim 5, wherein
The first quantizer is the closest codebook compared to the k ¹ to k ^nth codes of each codebook for each value of the input speech signal vector x ^k having values of x ^k1 to x ^kn .

Finding ^k , the second quantizer is most compared to the k ¹ to k ⁿ th codes of each codebook for each value of the KLT vector X ^k having values of X ^k1 to X ^kn transformed through the KLT transform unit Nearest codebook

A quantization device characterized by finding ^k . Converting the input speech signal vector into a sub-block having a small size and performing KLT conversion to output a KLT vector signal;
A first quantization step of finding and outputting the closest codebook by comparing respective values of the input speech signal vector with codes of respective codebooks;
A second quantization step of finding and outputting the closest codebook by comparing respective values of the KLT vector with codes of respective codebooks; And
And a final codebook determination step of comparing the quantized code vector in the original domain of the speech signal and the quantized code vector in the KLT domain with the source vector to select the nearest code vector. The method of claim 9,
The first or second quantizer is a quantization method, characterized in that for setting the bit allocation according to the covariance matrix and the dimension using the following equation:

Where c and c _i are bits of the entire bit and the i subblock, k and k _i represent the dimensions of the dimension and the i subblock, G _k is a constant that depends on the dimension, and C _Yi is the KLT domain. Or the covariance matrix of the i-th subblock in the original domain. The method of claim 9,
The quantization method
A quantization method characterized by using a split vector quantization algorithm. The method of claim 9,
The quantization method
And an inverse KLT domain transformation step of inverse KLT transforming and outputting the KLT vector generated in the second quantization step. 13. The method of claim 12,
The final codebook comparison step is a quantization method, characterized in that for comparing the source code and the code vector output in the first quantization step and the inverse KLT domain transform step. The method of claim 13,
And the input voice signal is a signal of a linear spectral frequency type. 15. The method of claim 14,
The input voice signal is a difference between the input current frame and the predicted current frame data. The method of claim 13,
The first quantizer is the closest codebook compared to the k ¹ to k ^nth codes of each codebook for each value of the input speech signal vector x ^k having values of x ^k1 to x ^kn .

Finding ^k , the second quantizer is most compared to the k ¹ to k ⁿ th codes of each codebook for each value of the KLT vector X ^k having values of X ^k1 to X ^kn transformed through the KLT transform unit Nearest codebook

^A method of quantization, characterized by finding ^k .

Images