JP2683734B2 - Audio coding method - Google Patents

Description

The present invention relates to a method for coding a speech waveform with a very small amount of information, and particularly, under a coding delay of about 30 milliseconds, 1 This relates to an effective coding method when the number of bits per sample point is about 1 bit or less. "Prior Art" As a means for compressing and encoding a voice signal sequence,
There is a method of extracting spectrum envelope information by linear prediction analysis on a frame-by-frame basis, and transmitting the code by combining the code of the parameter and the code corresponding to the prediction residual. In particular, when the amount of information is 1 bit or less per one sampling point (sampling point), it is known that the waveform distortion can be reduced by using the vector quantization method for both codes. Among them, as an example of a method of encoding waveform information, as described in a speech encoding method (Japanese Patent Application No. 61-181770), a linear prediction residual signal is used in the frequency domain and a weighted scale of the spectrum envelope. There is a method of vector quantization. Specifically, as shown in FIG. 4, a prediction residual signal of the input voice is obtained by the inverse filter 11, and the residual signal is subjected to the discrete cosine transform (DC
T) unit 12 converts the frequency domain DCT coefficient, rearrangement sub-vector division unit 13 rearranges these coefficients to divide into a plurality of sub-vectors, and weighted vector quantization unit 14 adaptively for each frame. Using the obtained weighted distance, the one closest to the codebook vector is selected for each subvector to obtain the waveform code output. The input speech is subjected to linear prediction analysis by the linear prediction analysis unit 15, and the analysis output is quantized by the vector / scalar quantization unit 16 to obtain a spectrum envelope coding output. In addition, a multi-pulse drive type encoding method, a time domain code drive type encoding method and the like are also known as encoding methods corresponding to the waveform information of the linear prediction residual. As a method of encoding the spectrum envelope information, a method of converting a linear prediction coefficient into an LSP parameter and performing vector-scalar quantization (Japanese Patent Application No. 57-204849) is known. In addition, various quantization methods using parameters such as PARCOR coefficient and logarithmic cross-section ratio are known and put into practical use. As described above, there are many known methods for separating the spectrum envelope information and the waveform information, obtaining the respective codes, and transmitting them as a set. However, in either encoding,
First, determine the sign of the spectrum envelope information from the input speech,
It is common to determine the sign of the waveform information based on the sign. In this method, there is no guarantee that the waveform distortion of encoding can be minimized when the information of the spectrum envelope and the information of the waveform are paired. This is because there are many combinations of both types of information that can further reduce waveform distortion. An object of the present invention is to provide a speech coding method for reducing the coding distortion of speech signal sequence information with a constant coding delay and a constant amount of information. "Means for Solving Problems" This invention attempts to quantize waveform information for each of a plurality of quantization candidates of spectral envelope information, and to obtain the spectral envelope information and the waveform information that can minimize the coding distortion. The feature is that the waveform distortion is significantly improved by searching for the combination. Particularly, the search is performed by quantizing the spectrum envelope information by vector-scalar quantization and gradually changing the scalar part. This point is different from the conventional method in which the waveform information is quantized with only one quantized value that approximates the spectral envelope information of the input voice. [Embodiment] FIG. 1 shows an improved coding method. Here, it is assumed that a voice waveform sampled at 8 kHz is input and 256 samples are processed as one frame. The spectrum prediction information is extracted by the linear prediction analysis unit 15 as a linear prediction coefficient. The coefficient is the line spectrum pair (LSP)
The data is converted into parameters and quantized by the vector / scalar quantization unit 16 (Japanese Patent Application No. 57-204849). On the other hand, the waveform information is a voice coding method (Japanese Patent Application No. 61-181770).
No.), vector quantization is performed on the linear prediction residual signal obtained by the inverse filter 11 in the frequency domain using the weighted scale of the spectrum envelope. Specifically, the prediction residual signal is converted to a DCT coefficient in the frequency domain by the discrete cosine transform (DCT) unit 12, and the coefficient is rearranged to be rearranged by the subvector division unit 13 to be divided into a plurality of subvectors. Using the weighted distance adaptively obtained for each frame by the weighted vector quantization unit 14, the one closest to the codebook vector is selected for each subvector. A coding distortion is improved by a combination of codes of both pieces of information with a combination of the coding of the spectrum envelope information and the coding of the waveform information described above as a basic system. This basic system is the fourth
It is the same as the encoding shown in the figure. Specifically, a large number of candidates for the quantization points of the spectrum envelope information are prepared, and the code of the waveform information is obtained for each of the candidates by the same procedure as in the case of the basic system. The local decoding unit 21 locally decodes the waveform code in the same procedure as the decoder. The resulting waveform distortion is measured by the waveform distortion measuring unit 22, and the code set with the smallest distortion is selected by the selector.
It is selected at 23 and used as a transmission code. As a method of preparing a code candidate for the spectrum envelope information, all the quantization points can be candidates. However,
When Ebit is the amount of information assigned to the spectrum envelope information, 2 ^E candidates are created, so when E reaches about 20,
The number of candidates becomes unrealistic. Therefore, in the present invention, the optimum code is locally obtained by gradually changing the code of the scalar part of the vector / scalar quantization. An example of this method (local optimization search) is shown in FIG. As the initial value of the central value of the search for the quantization point, the one obtained by the same quantization method as that of the basic system is used. That is, the LSP parameter of the input speech and the quantization point that minimizes the squared distance are selected from the vector code codebook, and the quantization error is scalar-quantized for each parameter element. Next, by changing only one element of the parameter to another scalar quantization point, a plurality of vector quantization point candidates are created near the center value (step S ₁ ). At this time, in the case where each element of the parameter is changed to only the quantization point adjacent to the scalar quantization point, the number of candidates is the same as the number K of elements of the parameter, that is, the number of dimensions of the vector. When changing to all the quantization points in each element, the number of candidates is N given by the following equation. N = Σ (2 ^{b (i)} −1) where b (i) is the number of scalar quantization bits assigned to the i-th element. I tried to quantize the waveform information for these candidates, measured the quantized waveform distortion of each candidate (step S ₂ ), and asked if there was a candidate that was better than the waveform distortion of the central value. examined (step S _3), if there is a candidate to be improved returns to step S ₁ the vector quantization point of a spectral envelope that became most waveform distortion smaller as a new center value (step S _4). When the waveform distortion does not become smaller than that of the center value even if any one element of the parameter is changed, the repetition is stopped, and this center value is used as the quantized value of the spectrum envelope information and paired with the waveform information at that time. Transmission information. In this iterative optimization, the number of scalar code elements,
In other words, the linear prediction analysis stops at about half the number of iterations. This local optimization search may be applied to some vector codes, or the vector code may be determined in advance and limited to that code. In the case of the former, the minimum number of waveform distortion obtained in the process of search, in the case of vector code that has extremely large distortion, by stopping the search of the scalar code, the number of searches can be reduced without degrading the performance. Can be reduced. Apart from this, when the shape of the spectrum envelope after quantization is extremely different from the shape of the spectrum envelope of the original sound, the search can be stopped without trying to quantize the waveform. As a scale for evaluating the difference in the shape of the spectrum envelope at this time, a scale that emphasizes the peak portion of the spectrum envelope is effective. Further, the waveform distortion can be easily evaluated not only by the squared distance but also by using a distance scale with auditory correction. "Effects of the Invention" As described above, the coding method of the present invention searches for a combination of the information of the spectral envelope and the information of the waveform that can further reduce the waveform distortion. Therefore, the waveform distortion can be significantly reduced without increasing the transmission information at all. FIG. 3 compares the SNR for three types of information with the conventional encoding (that is, the basic system). Is a conventional encoding method using a quantized value of one spectrum envelope, and is the basis of. Is a case of encoding by a method of first determining a vector code and searching for a scalar code, and is a case of performing a search of a scalar code for each vector. Even for about 2d
In the case of B, it can be seen that it is further improved.
The spectrum envelope information has 8 bits in the vector part,
16 bits are allocated to the scalar part, and the average is about 100 times in the case, and the average is about 1000 times in the case. The above SNR improving effect is not limited to the coding of the embodiment, and various spectral envelope parameter quantization methods (PARCOR coefficient, logarithmic cross-sectional area ratio, etc.) and various residual waveform coding methods (multi-pulse drive type, time It can be applied to the combined coding of the area code driven type, etc., and the same improvement effect can be expected. Further, the amount of calculation required for the encoding method of the present invention is several ten times or more that of the basic system, but it has the following practical advantages. ○ Even if the basic system is simplified, the deterioration of the final coding performance can be reduced. ○ Coding search operations, which occupy most of the operations, are easy to perform parallel processing in units of the basic system coding.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a functional block diagram showing the configuration of an embodiment of the present invention, and FIG. 2 is a flow chart showing a local optimization search method for quantization points of spectrum envelope information in this embodiment. FIG. 3 is a diagram in which the coding performance of the embodiment is represented by SNR and is compared with the performance of the basic system, and FIG. 4 is a functional block diagram showing the configuration of the conventional coding method.

Claims (1)

(57) [Claims] A voice signal is divided into analysis sections for each of a plurality of samples, and a vector scalar-quantized code of spectrum envelope information indicating the spectrum envelope shape of the voice signal in the analysis section, and
Using the vector-scalar quantized spectrum envelope information, in the method of encoding with the vector quantized code the waveform information indicating the linear prediction residual obtained from the speech signal in the analysis section, the vector of the spectrum envelope information・ The code of the scalar part in the scalar quantized code is changed one element at a time, the waveform information is quantized each time it is changed, and the distortion of the coded speech at that time is calculated, and the distortion is minimized. A speech coding method, wherein a quantized value of spectrum envelope information is searched for, and a code of the spectrum envelope information and a code of waveform information corresponding to the spectrum envelope information are used as output codes. 2. The search, when the change for each element of the scalar portion is completed, determines whether the minimum distortion up to that point is smaller than the coded speech distortion before changing the code of the scalar portion, and if it is smaller, The code consisting of the changed scalar code and the vector part in the vector / scalar quantization code is regarded as the vector / scalar quantization code, and the code of the scalar part is changed again by one element to search for the minimum distortion. When the determination is not small, the vector / scalar quantization code and the code of the waveform information corresponding to the vector / scalar quantization code are used as the output code.
Speech coding method described in the paragraph.