JP3088163B2 - LSP coefficient quantization method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a compression coding method for an audio signal, and more particularly to a quantization method for a line spectrum pair.

[0002]

2. Description of the Related Art Parameters representing the characteristics of the vocal tract of a speech signal include a linear prediction coefficient and a reflection coefficient. Among them, a line spectrum pair (hereinafter referred to as L
SP coefficient) has advantages such as relatively little distortion with respect to quantization, good interpolation characteristics, and easy determination of stability. Therefore, it can be said that the LSP coefficient is an extremely effective parameter in the field related to the high-efficiency compression coding technology of audio. The LSP coefficient can be directly quantized (see Document 1),
It can be indirectly quantized by a PCM method (see Document 2).

Reference 1: Noboru Kanmura, Fumitada Itakura, "Speech Information Compression by Line Spectrum Pair (LSP) Speech Analysis / Synthesis", IEICE Transactions, Vol.
(1981/8) Reference 2: FKSoong & BHJuang, “Optimal Quantizati
on of LSP ParametersUsing Delayed Decision ", Proc.
ICASSP-90 (April 1990) The LSP coefficient is expressed as a vector having about 6 to 12 dimensions (elements) in the case of a normal audio signal.
The elements of each vector have the property of being arranged in ascending order with respect to frequency, and the values of the elements are not independent. Therefore, for the quantization, for example, a vector quantization method disclosed in Reference 3 is extremely effective.

Literature 3: Kohei Hoshi, Tetsuo Funada, "Recognition of Plosives by Two-Step Vector Quantization", IEICE Technical Report, SP90-6 (May 25, 1990)

[0005]

Generally, vector quantization of LSP coefficients is performed by storing a large number of vectors as a table, searching for a vector similar to the LSP coefficient of the input speech signal, and outputting the code. . Since the number of vectors prepared in advance is limited by the bit rate and the search time, quantization distortion may increase depending on the properties of the audio signal.

The present invention provides an LSP for adjacent speech frames.
Focusing on the fact that the coefficients have a correlation, the table vectors are supplemented.

That is, there is a correlation between LSP coefficients obtained from adjacent audio signals, and LSP coefficients obtained from adjacent audio frames are similar. Accordingly, it is an object of the present invention to provide a method for sufficiently utilizing the correlation to improve the quantization efficiency.

[0008]

SUMMARY OF THE INVENTION The present invention relates to a multi-stage quantization method for LSP coefficients (vectors).

In the present invention, the first-stage quantization table is composed of a fixed section storing a predetermined vector and a variable section storing a replaceable vector.

After the quantization of each frame is completed, the oldest vector in the variable section is replaced with a composite vector of the vectors determined in the current frame. That is, it corresponds to the output code of the current frame and the LSP of the current frame.
Replace with the vector corresponding to the coefficient.

[0011]

According to the present invention, vector quantization is performed instead of scalar quantization in order to effectively utilize the correlation between each element of the LSP coefficient. However, if the vector quantization is performed as it is, the size of the quantization table becomes very large. For example, first, the vector quantization is roughly performed with about 10 bits, and the quantization error is determined by using another quantization table. Vector quantization using a small quantization table is performed a plurality of times, for example, vector quantization is performed using a quantization table different from the previous two quantization tables. By repeating, the effect of vector quantization is used.

Also, the LS for the adjacent voice frame
In order to effectively use the correlation of the P coefficient, a part of the quantization table of the first stage is made variable, and the value of the LSP coefficient quantized in the previous frame is stored in the vector of the quantization table of the first stage. Used as

[0013]

FIG. 1 is a block diagram of a quantizer for LPC coefficients to which the present invention is applied. Here, it is assumed that two quantization tables of the first stage (first stage) and the last stage (second stage) are prepared.

The quantization table 11 in the first stage is composed of a variable portion 11a and a portion 11b composed of a predetermined fixed vector.

First, the optimal vector V ₁₀ selected in the first stage of the previous frame and the optimal vector V ₂₀ selected in the second stage are added by the adder 15, and the result is combined in the previous frame. The vector (quantized LSP coefficient) is obtained, and the oldest vector in the variable portion 11a of the quantization table 1 is replaced with the quantized LSP coefficient (composite vector) of the previous frame.

Next, the error evaluator 13 evaluates the error of all the vectors in the quantization table 11 with the LSP coefficient of the input speech, and selects the optimum vector V ₁ .
The first-stage optimal vector V ₁ and its sign (index) C ₁ are output.

Next, the vector V ₁ selected in the first stage is subtracted by the subtractor 16 from the LSP coefficient of the input voice to obtain a quantization error vector in the first stage.

The quantization tables 12 are all composed of predetermined fixed vectors.

In the error evaluator 14, the quantization table 12
For all vectors, performs error evaluation for the quantization error vector at the first stage, to select the best vector, and outputs the second-stage optimal vector V ₂ and the sign (index) C _2.

FIG. 2 shows an example of an inverse quantizer corresponding to FIG. The first-stage quantization table 21 is a variable part 21
a and a portion 21b composed of a predetermined fixed vector. The fixed part 21b corresponds to 11b in FIG.

First, the oldest vector in the variable portion 21a of the quantization table 21 is replaced with the composite vector (V ₁₀ + V ₂₀ ) of the previous frame.

Next, the vector V ₁ of the optimum first stage, indicated at C ₁ from the quantization table 21 is taken out.

The quantization table 22 corresponds to 12 in FIG. It consists of a predetermined fixed vector.

[0024] from the quantization table 22, the vector V ₂ optimum second stage, indicated at C ₂ is taken out.

The optimum vector V ₁ in the first stage and the optimum vector V _{2 in} the second stage are added by the adder 23 to form a quantized LSP coefficient (composite vector).

The following table shows the result of comparison between the conventional method (scalar quantization of LSP) and the effect of LSP quantization by the encoder of FIG. The vocal tract parameters used in the experiment were obtained by performing a 10th-order analysis with an analysis frame length of 40 ms. The conventional method sets the LSP coefficient of each order to 4, 4,
Non-linear scalar quantization is performed with 3, 3, 3, 3, 3, 3, 2, and 2 bits. According to the present embodiment, vector quantization is performed with 10 bits in the first stage and 12 bits in the second stage. It was done. The number of variable partial vectors in the first stage is six. Each of the quantizers is configured using learning data of about 3 minutes, and the test was performed using data for about 15 seconds that is not included in the learning data.

[0027]

[Table 1]

Here, the smaller the required number of bits, the higher the compression efficiency. Also, it is said that the smaller the spectral distortion is, the better the sound distortion becomes. The larger the SNR (signal-to-noise ratio), the better.

In the embodiment shown in FIG. 1, the vector of each stage and the sign thereof are determined for each quantization stage. However, as in the method described in Reference 2, a plurality of vectors are determined at each quantization stage. By searching for candidates of the final stage and comprehensively evaluating the quantization error after the completion of the final stage, the vector of each quantization stage and
The special number may be selected.

[0030]

According to the present invention, a variable section is provided in the first-stage quantization table, and the latest synthesized vector (quantized LSP
), The correlation of LSP coefficients between adjacent frames can be used effectively, and the quantization efficiency can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram of a quantizer to which the present invention is applied;

FIG. 2 is a block diagram of an inverse quantizer corresponding to FIG. 1;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Quantization table 11a Variable part 11b Fixed part 12 Quantization table 13 Error evaluator 14 Error evaluator 15 Adder 16 Subtractor 23 Adder

────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Yoshihiro Ariyama 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (56) References JP-A-2-296300 (JP, A) JP-A-3-288433 (JP, A) JP-A-4-171500 (JP, A) JP-A-4-363000 (JP, A) JP-A-63-117878 (JP, A) JP-A-2-186836 ( JP, A) Japanese Patent Publication No. 7-27400 (JP, B2) Japanese Patent Publication No. 7-120961 (JP, B2) 1990 IEICE Autumn National Convention Lecture Papers, Volume 1, "A-130 LSP Parameters Examination of efficient vector quantization method ”, p. 1-132, (published September 15, 1990) Proceedings of 1991 IEEE International Conference on Acoustics, Speech and Signal Processing, Vol. 1, K. Ozawa et al, "4 kb / s Improved CELP Coder with Efficient Vector Quantitation", p. 213- 216 (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 11/00-13/08 G10L 19/00-21/06 H03M 7/30 H04B 14/04 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. An image processing apparatus comprising at least two stages of quantization tables, wherein the first stage of vector quantization is performed by using L
Performs on the SP coefficient as a target, performs each vector quantization after the next stage on the quantization error up to the previous stage, evaluates the quantization error for each stage, or comprehensively evaluates the quantization error of all stages. In a quantization method of LSP coefficients for determining one vector corresponding to each stage of vector quantization and outputting respective codes corresponding to the vectors, the first stage of the quantization table is determined in advance. After the quantization of each frame is completed, the oldest vector of the variable unit is determined by the vector determined in the frame. A method for quantizing LSP coefficients, characterized in that: