KR100446595B1 - Vector quantization method of line spectrum frequency using localization characteristics, especially searching optimum code book index using calculated distortion - Google Patents

Abstract

PURPOSE: A vector quantization method of a line spectrum frequency(LSF) using localization characteristics is provided, which adopts a spectral distortion measurement considering the localization characteristics of LSFs in order to provide further improved sound quality. CONSTITUTION: According to the method, a line spectrum frequency vector of P dimension is converted into a line spectrum frequency vector of (P-1) dimension, and minimum distortion and optimum code book index and code book index are initialized(300). A quantized line spectrum frequency vector of index k is searched in the code book using the code book index k(320). Localization restriction rate is determined(330), and distortion of the quantized line spectrum frequency of index k is measured. And minimum distortion is updated by replacing the minimum distortion with the distortion and replacing the optimum code book index with the index k(360). If the code book index k is larger than the size of the code book M, the optimum code book index is output(370), and otherwise, the code book index k is replaced with (k+1)(380).

Description

Vector Quantization Method of Line Spectrum Frequency Using Localization Characteristics

The present invention relates to a vector quantization method for quantizing line spectrum frequencies (hereinafter referred to as LSFs), and in particular, a method of measuring spectral distortion in consideration of localization characteristics of LSFs. The vector quantization (VQ) method adopted.

As digital voice communication technology develops, a technique for compressing voices more efficiently is required. For high compression of voice, a technique of modeling the generator is widely used. In voice communication, a voice call is made by transmitting parameters of the model. The voice generating organ is composed of a vocal track and a vocal cord, and vibration occurs in the gate, so that sound is generated while passing through the vocal cord. In general, the saints are represented by periodic impulse or random signals, and the gates are represented as prediction coefficients by a linear prediction technique. LSFs can be obtained based on linear predictive coding (LPC), which is widely used in voice coding because of their excellent quantization characteristics.

Quantization of LSFs spectral model parameters requires measurement of spectral distortion. Conventionally, the spectral distortion was measured by the following Euclidean method.

[Equation 1]

는 각각 P차수의 i번째 LSF의 원래값과 양자화된 값이다.Where ωi ^p and

Are the quantized and original values of the i th LSF of P order, respectively.

It is widely known that voice signals have important information on the peak or formant rather than the spectral valley portion on the spectral envelope. On this basis, it is well known that weighted Euclidean measurement schemes have been introduced, and that the sound quality is improved cognitively by this approach. The weighted Euclidean measurement method is as follows.

[Equation 2]

Where wi is the weight of the i th LSF.

When expressing the spectral envelope in voice coding, the reason for using LSF as compared to LPC, PARCOR (PARtial CORrelation) and LAR (Log Area Ratio) is because of the interlacing characteristics of LSF. That is, PSF LSFs have the following characteristics.

[Equation 3]

Recently, a new property has been discovered called localization of LSFs. This means that there is an interlacing characteristic between LSFs between P and P-1 orders. In other words,

[Equation 4]

However, in the existing Euclidean or weighted Euclidean method, there is a problem that the localization characteristics of the LSFs as described above are not considered.

The present invention was made to solve the above problems, and to provide a vector quantization method adopting the spectral distortion measurement method in consideration of the localization characteristics of LSFs when quantizing the LSFs to provide more improved sound quality do.

1 shows a process of processing a voice signal according to the present invention.

FIG. 2 is a block diagram illustrating the line spectrum frequency vector quantization process shown in FIG. 1.

3 is a flowchart illustrating a vector quantization process of line spectrum frequencies according to the present invention.

4A to 4C show experimental results for comparing the spectral envelope of the conventional method and the present invention.

The configuration of the present invention for achieving the above object is as follows.

를 찾는 코드북 탐색단계; 상기 P-1차원의 선스펙트럼 주파수 벡터와 상기 색인이 K인 양자화된 선스펙트럼 주파수 벡터 내의 모든 원소 ω_i ^P-1과 ω_i ^P(k) (여기에서 i의 값은 1에서 P까지의 값)간에 ,A method of vector quantizing a P-dimensional line spectrum frequency vector ω ^P = (ω ₁ ^P , ω ₂ ^P ,..., Ω _P ^P ) using a codebook includes a P-dimensional line spectrum frequency vector. Convert to a spectral frequency vector ω ^P-1 = (ω ₁ ^P-1 , ω ₂ ^P-1 ,… ω _P-1 ^P-1 ), with minimum distortion D _{w, 1} ^* , optimal codebook index k ^* and codebook Initialization step of initializing index K: Quantized line spectrum frequency vector having index k in the codebook using the codebook index k

Codebook searching step of finding; All elements ω _i ^P-1 and ω _i ^P (k) in the P-1 dimensional line spectrum frequency vector and the indexed K quantized line spectrum frequency vector, where i is a value from 1 to P )

A localization constraint determining step of allocating a predetermined value smaller than 1 and greater than 0 to the localization constraint α when the localization characteristic is satisfied, and assigning 1 to the localization constraint α if not satisfied; When the distortion of the quantized line spectrum frequency whose index is k is D _{w, l} ^k ,

는 색인이 k인 양자화된 선스펙트럼 주파수 벡터의 i번째 원소)에 의해 D_w,l ^k를 구하는 왜곡도 측정단계; 상기 색인이 k인 양자화된 선스펙트럼 주파수의 왜곡도 D_w,l ^k가 상기 최소 왜곡도 D_w,l ^*보다 작은 경우에는 상기 D_w,l ^*를 상기 D_w,l ^k로 치환하고, 상기 최적 코드북 색인 k^*를 상기 k로 치환하는 최소 왜곡도 갱신 단계, 상기 코드북 색인 k가 코드북의 크기 M보다 큰 경우에는 상기 최적 코드북 색인 k^*를 출력하고, 같거나 작은 경우에는 상기 코드북 색인 k를 k+1로 치환하고 상기 코드북 탐색단계를 반복하는 반복 여부 결정단계를 포함하는 것을 특징으로 한다.Where P is the number of dimensions of the line spectrum frequency, α is the localization constraint of the line spectrum frequency, w _i is the weight of the i-th line spectrum frequency, ω _i ^P is the i-th element of the P-dimensional line spectrum frequency,

Is a distortion measurement step of obtaining D _{w, l} ^k by the i th element of the quantized line spectrum frequency vector whose index is k; The case where the index is a distortion of the quantized line spectral frequency is also D _{w, l} ^k the minimum distortion is less than D _{w, l} ^* k is the substitution of the D _{w, l} ^* in the D _{w, l} ^k, If the optimum codebook index k ^* a minimum skewness updating step, the codebook index k which is substituted by the k is greater than the size M of the code book is equal to output the optimal codebook index k ^*, and or small, the codebook index k and repeating the codebook search step by substituting with k + 1.

The present invention uses the localization characteristics of LSFs to measure distortion in quantization of spectral envelope. That is, as the interlacing characteristic exists between the P-order LSFs of the original sound and the P-1 order LSFs, the interlacing characteristic is maintained between the quantized P-order LSFs and the P-1 order LSFs of the original sound. will be. In other words,

[Equation 5]

To apply this, the following distortion measurement method was introduced.

[Equation 6]

간에 수학식 5에 따른 국소화 특성 만족여부에 따라 다른 값을 갖는데, 국소화된 경우에는 1보다 작고 0보다 큰 소정의 값(0.8 정도의 값이 적절하다)을, 그렇지 않은 경우에는 1의 값을 갖는다.α is equal to ω _i ^P

It has a different value depending on whether localization characteristics are satisfied according to Equation 5, and if it is localized, it has a predetermined value smaller than 1 and larger than 0 (a value of about 0.8 is appropriate). .

This method allows vector quantization of LSFs. Vector quantization is a process of finding the VQ index k ^* of the LSF that minimizes Equation 6 among the LSFs candidates.

[Equation 7]

(k)는 P차 LSF코드북의 k번째 코드 원소의 i번째 값이고, M은 코드북 크기이다.here

(k) is the i th value of the k th code element of the P th order LSF codebook, and M is the codebook size.

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

1 shows a process of processing a voice signal according to the present invention.

First, the input voice is derived from the LPC coefficient (step 100), and the LSF is converted from the LPC (step 110). The optimal codebook index is retrieved using the vector quantizer using the localization characteristic by the transformed LSF as described above (step 120), and the retrieved index is transmitted through a transmission process.

FIG. 2 is a block diagram illustrating the line spectrum frequency vector quantization process shown in FIG. 1. The input LSF ω ^P (P-dimensional LSFs) is sequentially measured in a distortion degree with each quantized LSF in the LSF vector codebook 220 (200), and the index of the quantized LSF having the smallest distortion is found (210). Will pass to the stage.

3 is a flowchart illustrating a vector quantization process of line spectrum frequencies according to the present invention.

The vector quantization process using localization is a process of converting LSFs into vector codebook indexes. First, the minimum distortion degree, the optimal codebook index k ^* and the codebook index k are initialized as ∞, 1, and 1, respectively (step 300). In step 310, the LSF of the P-order of the original audio is converted into the L-SF of the P-1 order.

Next, the codebook index k is used to find a quantized LSF whose index is k in the codebook 220 (step 320).

(k) (여기에서 i의 값은 1에서 P까지의 값)간에Then, all elements ω _i ^P-1 in the P-1-dimensional line spectrum frequency vector transformed in step 310 and the quantized line spectrum frequency vector whose index is found in step 320 are k and

(k) (where i is a value from 1 to P)

If the localization characteristic is satisfied, a predetermined value smaller than 1 and greater than 0 is assigned to the localization constraint α, and if it is not satisfied, 1 is assigned to the localization constraint α (step 330). That is, in the conventional vector quantization process, since the localization characteristics of the LSFs are not taken into consideration, the value of the localization drug rate α in Equation 6 becomes '1'. However, in the present invention, since the localization characteristics of the LSFs are taken into consideration, the values of the localizing agent ratio α are different according to whether localization characteristics of the LSFs are satisfied in Equation 6. That is, in the present invention, when the distortion is measured, the localization constraint is set to 1 when not satisfying the localization characteristic as shown in Equation 5, and to a predetermined value smaller than 1 and greater than 0, for example, 0.8 when the localization characteristic is satisfied. Done.

Then, the distortion degree of the quantized line spectrum frequency whose index is k

Obtained by (step 340). That is, the distortion of the quantized LSF whose index is k is measured to be relatively lower than the case where the localization characteristic does not satisfy the localization characteristic.

In step 350, the distortion degree D _{w, l} ^k measured in step 340 is compared with the minimum distortion degree D _{w, l} ^* initialized in step 300. The in step 350 the result of comparison, the measured distortion (D _{w, l} ^k) is the minimum distortion (D _{w, l} ^*) is smaller than is the measurement of the minimum distortion (D _{w, l} ^*) in step 360 Substitute the distortion degree D _{w, l} ^k and then replace the optimal codebook index k ^* with the current index k. As a result of the comparison in step 350, if the measured distortion degree D _{w, l} ^k is greater than or equal to the minimum distortion degree D _{w, l} ^* , the process immediately proceeds to step 370.

If the codebook index k is less than or equal to the size M of the codebook, replace the codebook index k with k + l and repeat the process from step 320 to step 360.If the codebook index k is larger than M, the optimal codebook index k ^* is substituted. It is delivered to the decoder (steps 370-380).

4A to 4C show experimental results for comparing the spectral envelope of the conventional method and the present invention. 4A shows the original speech waveform, FIG. 4B shows the comparison of the spectral envelope by the conventional quantization method, and FIG. 4C shows the comparison of the spectral envelope by the quantization method of the present invention. Comparing the drawings, it can be seen that the spectrum of the conventional method is not well matched in the first formant F1 and the second formant F2 known to have important information of speech. On the contrary, it can be seen that when the method of the present invention is used, the matching of the spectrum is improved compared to the existing method.

According to the present invention, when quantizing the LSFs, the distortion of the quantized line spectrum frequency whose index is k is calculated by dividing it into a case where the localization characteristic of the LSFs is not satisfied and a case where it is not satisfied. By searching the codebook index, the matching of the spectrum can be improved as compared to conventional methods to provide better sound quality.

Claims (1)

In the method of vector quantizing a line spectral frequency vector ω ^P = (ω ₁ ^P , ω ₂ ^P ,..., Ω _P ^P ) in a P-dimensional manner, Convert the P-dimensional line spectrum frequency vector into a P-1-dimensional line spectrum frequency vector ω ^P-1 = (ω ₁ ^P-1 , ω ₂ ^P-1 , ..., ω _P-1 ^P-1 ), An initialization step of initializing the minimum distortion degree D _{w, l} ^* , the optimal codebook index k ^* and the codebook index K; Quantized line spectrum frequency vector with index k in the codebook using the codebook index k

Codebook searching step of finding; All elements ω _i ^{P-1 in the} spectralized line spectrum frequency vector of the P-1 dimension and the quantized line spectrum frequency vector whose index is K;

(Where i is a value from 1 to P),

A localization constraint determining step of allocating a predetermined value smaller than 1 and greater than 0 to the localization constraint α when the localization characteristic is satisfied, and assigning 1 to the localization constraint α if not satisfied; When the distortion of the quantized line spectrum frequency whose index is k is D _{w, l} ^k ,

Where P is the number of dimensions of the line spectrum frequency, α is the localization constraint of the line spectrum frequency, W _i is the weight of the i-th line spectrum frequency, ω _i ^P is the i-th element of the P-dimensional line spectrum frequency,

Is a distortion measurement step of obtaining D _{w, l} ^k by the i th element of the quantized line spectrum frequency vector whose index is k; The case where the index is a distortion of the quantized line spectral frequency is also D _{w, l} ^k the minimum distortion is less than D _{w, l} ^* k is the substitution of the D _{w, l} ^* in the D _{w, l} ^k, A minimum distortion updating step of replacing an optimal codebook index k ^* with the k; If the codebook index k is larger than the size M of the codebook, the optimal codebook index k ^* is output. If the codebook index k is the same or smaller, the codebook index k is replaced with k + l, and the codebook search step is repeated. Vector quantization method of the line spectrum frequency using a localization characteristic comprising a.

Images