JPH02238499A - Vector quantizing system - Google Patents

Abstract

PURPOSE:To secure the reduction of a quantizing distortion by executing a fuzzy vector quantization by using selectively a code vector in accordance with an input vector. CONSTITUTION:First of all, a quantizing distortion at the time when an input vector 106 is brought to vector quantization by a vector 405 being the nearest to the vector is calculated 406. Subsequently, a vector is added successively in order being near the input vector 106 from in candidate vectors, brought to fuzzy vector quantization 408, and a quantizing distortion is calculated. If the quantizing distortion increases, the added candidate vector is not used, this procedure is applied to the remaining candidate vectors, as well, and it is repeated until the vector to be used reaches a prescribed number of pieces or until the quantizing distortion is decreased and eliminated. A code vector (a representative vector in a code book) is used selectively for such a fuzzy vector quantization. In such a way, the quantizing distortion can always be set to a regular vector quantization or below.

Description

[Detailed description of the invention] [Industrial application field]

The present invention relates to a high-efficiency audio encoding device, and particularly to an audio encoding method suitable for obtaining high-quality reproduced audio at a high information compression rate. [Prior Art] Conventionally, various methods have been proposed as high-efficiency speech encoding methods. For example, ``Digital Information Compression'' by Kazuo Nakata.
(Kosaido Sanpo Publishing, Electronic Science Series 100),
Various methods are explained in an easy-to-understand manner, and a large number of methods related to waveform encoding methods and information source encoding methods (parameter encoding methods) are shown. [Problem to be solved by the invention 1 Among these methods, although the waveform encoding method has good sound quality, it is difficult to increase information compression efficiency, and the parameter encoding method has high information compression efficiency, but On the other hand, even if the amount of information is better, there is a limit to the sound quality, and sufficient quality cannot be obtained.In particular, information compression (nearly 10 kbps), which is between the bands that both are good at, is in the valley. On the other hand, as a hybrid method that combines the advantages of both methods, a multi-pulse method (for example, B, S
, Atal et al.
El of L P Excitation for
producing natural-sound
ngspeech at low bit rat
es” Proc. I CA S S P82,
S-5.10, (1982), etc.), TOR method (A
．． Ichikawa et al. , “A spe
echocoding method using th
ined-out residual”Proc.
ICASSO 85, 25.7 (1985))
etc. have been proposed in recent years, and various studies have been conducted. The situation is inadequate both in terms of sound quality and the cost required for processing. In general, various high-efficiency encoding methods focus on the fact that the presence of voice information is unevenly distributed, and achieve this by thickening the allocation of codes to the parts where information exists. We are actively promoting a method that focuses on the bias of information as a combination of multiple parameters, and assigns a thicker code to the part where audio information exists for a set of parameter combinations (called a vector). (for example, S. Roucos et al.
, ”S egmentquantization
for very-low-rate speech
coding"Proc.ICASSF' 82,
p, 1563 (1982)) has been attracting attention. In the vector quantization method, a table that associates vectors with codes is called a codebook, but in order to achieve high-quality speech encoding, it is necessary to create a high-quality codebook in advance. Creating a codebook requires the use of an extremely large amount of speech data, and there are also problems such as how large the size of the codebook should be. Regarding the codebook problem, a fuzzy vector quantization method has also been proposed in which the input speech and all vectors in the codebook are interpolated using a class function for each vector (for example, H.P. Tseng et al. ．．
, ”Fuzzy vector quantiza
tionapplied to hidden Mar
kov modelingICASSP 87.4
(1987)), a large amount of class function (membership function) information, which is information on the similarity with each vector, is required, so although it is expected that the sound quality will improve compared to the quality of the codebook, the It is not used as a technology. It is currently being considered for use in preprocessing such as speech recognition. Furthermore, in order to reduce the amount of bi-information, it has been proposed to apply the k-nearest neighbor law (KNN method), which compares the input speech with all vectors in the codebook and uses only k closest vectors (for example,
Nakamura et. There are limits to how much can be reduced. To solve this problem, we can eliminate the need for sorting processing and reduce the amount of information to be transmitted by calculating the neighborhood vectors for each vector in the codebook in advance and recording this information in the codebook. The inventors of the present invention have already proposed a method that can reduce the
3-240972). In any of the above fuzzy vector quantizations, the average quantization distortion can be reduced compared to normal vector quantization, but if you examine each case in detail, the quantization distortion may actually increase. . An object of the present invention is to provide a method that guarantees reduction of quantization distortion in fuzzy vector quantization.

[Means to solve the problem]

In order to achieve the above object, fuzzy vector quantization is performed by selectively using representative vectors in the codebook (hereinafter referred to as code vectors) according to the input vector. This code vector selection consists of a means for selecting a code vector candidate to be used, a means for evaluating the relationship between the candidate vector and the input vector, and a means for determining a vector to be used based on the evaluation result. . [Operation] The operation of typical procedures of the present invention will be explained. When the voice to be transmitted is input, feature vectors are extracted in the analyzer and compared with code vectors in the codebook in order, and the closest vector is selected. When the present invention is applied to conventional fuzzy vector quantization, a predetermined number of code vectors are selected as candidate vectors, for example, in order of proximity to the input vector. Furthermore, when applying a method in which neighboring vectors are registered in advance for each code vector (Japanese Patent Application No. 63-240972), the neighboring vectors become candidate vectors, but among them, they are ordered in order of proximity to the input vector. I'll keep it. When using quantization distortion as an evaluation criterion for selecting a vector, first, the quantization distortion when an input vector is vector quantized using the vector closest to it (so-called normal vector quantization) is calculated. Next, vectors are sequentially added from among the candidate vectors in order of their proximity to the input vector, and fuzzy vector quantization is performed to calculate quantization distortion. At this time, if the quantization distortion decreases, the added candidate vector is officially used. If the quantization distortion increases,
The added candidate vectors are not used. This procedure is applied to the remaining candidate vectors and repeated until a predetermined number of vectors is used or the quantization distortion no longer decreases. In this way, by selectively using code vectors for fuzzy vector quantization,
Quantization distortion can always be kept below normal vector quantization. (Embodiment 1 An embodiment of the present invention will be described below with reference to the drawings. Fig. 1 is a block diagram for explaining an embodiment of the present invention. The communication path in the opposite direction is omitted because it would complicate the diagram. The audio from the buffer memory 103 is input to the buffer memory 103 of the configuration. This memory is provided to adjust the time of the following processing and to prevent interruption of input audio.
4, pitch information 107, spectrum information 10
6. Level information 105 is obtained. Spectrum information 1
06 is a fuzzy vector quantization unit 10 to which the present invention is applied
8 to obtain a vector code 109 and a class function (membership function) 110.6 The vector code 109, class function 110, pitch information 107, and level information 105 are sent to the receiving unit 113 via a transmitting unit 111 and a transmission line 112. It will be done. On the receiving side, the vector code 109' received by the receiving section
, class function 110'. Pitch information 107'' level information 105t is added to fuzzy vector dequantization unit 114, spectrum information 115 is restored, and pitch information 1
07' is added to the synthesis unit 116 together with the level information 105'. The synthesizer 116 decodes the audio waveform, passes through a two-sided buffer memory 117 for output, converts it into an analog signal by a digital/analog (D/A) converter 118, and reproduces it as an output audio 119. Each part will be explained in detail below. FIG. 2 is a diagram for explaining the analysis section 104. In this embodiment, the analysis section uses a power spectral envelope (PSE) analysis method. The PSE analysis method is described in detail in the paper "Power Spectral Envelope (PSE) Speech Analysis and Synthesis System" by Nakajima et al., Journal of the Acoustical Society of Japan, Vol. 44, No. 11 (1986-11). Here we will give an overview. In FIG. 2, a pitch extraction unit 201 extracts pitch information (pitch frequency or pitch period) of input speech. As a pitch extraction method, a known method such as a correlation method or an AMDF method may be used. The waveform cutting unit 203 cuts out a waveform section for analyzing spectrum information from the input voice, and cuts out a section of about 20 to 60 ms. It is often a fixed length section, but it depends on the pitch period,
The length may be made variable to about three times that length. The extracted waveform is sent to the Fourier transform unit 204 and transformed into a Fourier series. At this time, after multiplying the extracted waveform by a commonly used window function such as a Hamming window, zero data is embedded before and after the waveform to create 2048 points of data, and by using fast Fourier transform (FFT), high speed and frequency resolution are achieved. Highly accurate data can be obtained. The Fourier coefficient expressed in absolute value becomes the frequency component of the cut-out waveform, that is, the spectrum. When the cut-out waveform has a periodic structure, the spectrum has a line spectrum structure due to pitch harmonics. The pitch resampling unit 205 extracts only the harmonic components (line spectrum components) of the pitch frequency from the spectrum information obtained by FFT. The data extracted in this manner will be considered below in association with the period π during cosine series expansion, which will be described later. The power spectrum conversion unit 206 squares each component of the spectrum and converts it into a power spectrum. Furthermore, the logarithmization unit 2
07 logarithms each component to obtain a logarithmic power spectrum. The level normalization unit 208 absorbs level fluctuations based on the loudness of the input audio, but the next cosine conversion unit 209
, they may be extracted all at once. The cosine transform unit 209 uses data obtained by resampling the logarithmic power spectrum and approximately expresses it using a cosine series of finite terms. The number of terms m is usually set to about 25. The power spectrum envelope is expressed as follows. Y=A. + A, cosλ+A, cos2λ+-+
Amcos mλ (1) The coefficient A is the difference between the resampled power spectrum data and Y according to equation (1).
It is calculated so that the multiplicative error is minimized. The 0th term 80 of the coefficient represents the level of the input, so the level information 105
As, A1,...,A. is output as spectrum information 106. Next, a fuzzy vector quantizer having a vector selection function according to the present invention will be explained using FIG. In FIG. 3, a codebook 401 stores the values of elements of codevectors and their codes. When the spectrum information (input vector) 1o6 is input to the distance calculation unit 402, each code vector is read out from the codebook 401, the distance from the input vector 106 is calculated, and a distance value 403 is output. Here, the distance measure is a Euclidean distance weighted to each element of the vector, but it goes without saying that other suitable measures may be used. Furthermore, it is also possible to limit the range of code vectors targeted for distance calculation by using the pitch information 107 or the like. Candidate vector selection section 404 selects code vector candidates to be subjected to vector evaluation, which will be described below. Here, with reference to the distance value 403, a predetermined number (C) of vectors with the smallest distance are selected and output as a code 405 of candidate vectors sorted in descending order of distance. In addition to the above criteria for selecting candidate vectors, the distance value may be smaller than a predetermined threshold, or the number of vectors may be less than or equal to a predetermined number.
In addition, the distance value may be less than or equal to a predetermined threshold value. In addition, when targeting all code vectors in the codebook,
This candidate vector selection section is not necessary. The vector selection unit 406 calculates and evaluates quantization distortion for the candidate vector using the following procedure. Minimum value d of distance value 403 to input vector, Ii. 1 is the quantization distortion when the input vector is vector quantized (so-called normal vector quantization) using the nearest neighbor vector, so this is first used as the criterion for evaluation. Next, candidate vectors other than the nearest neighbor vector are combined with the nearest neighbor vector one by one, fuzzy vector quantization is performed, and quantization distortion is calculated. Fuzzy vector quantization is described in detail in the literature by Nakamura et al., "Normalization of spectrograms using fuzzy vector quantization" (Journal of the Acoustical Society of Japan, Vol. 45, No. 2 (1989)) and the literature cited therein. Therefore, I will provide an overview here. In fuzzy vector quantization, an input vector is expressed by degrees of belonging to a plurality of code vectors. The degree of membership is quantified using a class function (membership function). An example of how to find the class function is shown in the following equation. Now, when we target C code vectors (Ve,..., Vc), input vector Xi and code vector V
．． Let the distance between the two points be dtk. If the input vector does not match any code vector, the class function u+h for each code vector is determined by the following equation. uIk: Here, p is a parameter called fuzziness, usually 1
．． The value should be around 5. If the input vector matches any of the code vectors, the value of the class function for that code vector is set to 1, and the others are set to O. Next, reproduce the vector from the class function (inverse quantization operation)
I will explain about it. The reproduction vector Xk is expressed as a linear combination of code vectors. Xk = error (distance) between input vector Xk and reproduction vector x, l is quantization distortion due to fuzzy vector quantization. Fuzzy vector quantization is performed using the nearest neighbor vector and the remaining candidate vectors one by one, and the quantization distortion of each is determined. The minimum value of these quantization distortions is d mln
In the following cases, select the candidate vector that gives this minimum value. Also, the minimum value at this time is rewritten as dllIn. Next, the remaining candidate vectors are added one by one to the nearest neighbor vector and the currently selected vector, and the same procedure is repeated until there are no more candidate vectors. The above is a case in which all candidate vectors that reduce quantization distortion are selected. In addition, the process may be terminated when the number of selected vectors reaches a predetermined number. Furthermore, in the above method, all remaining candidate vectors are evaluated each time in order to select one vector. As a simplified method, vectors may be added one by one in order of decreasing distance to the input vector, and if the quantization distortion at that time is smaller than the quantization distortion before addition, that vector may be selected. The above is an example using quantization distortion as an evaluation criterion. Besides this, by having a local decoder? A vector can also be selected using an error scale in the same dimension as the force signal as the evaluation criterion. In addition to this, it is also possible to select vectors based on positional relationships in vector space. FIG. 4 is a conceptual diagram for explaining this. For simplicity, the number of dimensions of the vector is assumed to be two. In the figure, Xk is an input vector and a V-avoiding nearest neighbor vector. If Vl is the vector to be evaluated, ■, and vt
The reproduction vector (x
k') comes on the straight line connecting Ve and ■1. Therefore, the condition for the distance difference between Xk' and X to be smaller than the distance between V and It is a certain thing. From another perspective, this means that when the distance between Xh and Ve and the distance between Xh and V1 are known, it can be determined by the size of the angle formed by the three vectors V, Xk, and vI. . Specifically, vector inner product calculation may be performed. The fuzzy vector quantization unit 408 refers to the vector code 407 that is the output of the vector selection unit 406, and calculates
The input vector is fuzzy vector quantized using the selected vector 1. Specifically, the class function is calculated based on the above-mentioned equation (2). The output is the selected vector code 109 and the class function 110. When the number of selected vectors is variable, information on that number is also output. Furthermore, since the sum of the class functions is 1 due to their nature, it is only necessary to output the number of vectors, which is one less than the number of vectors. Further, when the number of actually selected vectors is less than a predetermined number (fixed), the series function value for the remaining number may be set to O. In the above, a case has been described in which the vector selection function of the present invention is applied to conventional fuzzy vector quantization. In contrast, we will briefly explain the case where it is applied to fuzzy vector quantization of the type in which neighboring vectors are registered in advance for each code vector (Japanese Patent Application No. 63-240972), which has already been proposed by the inventors of the present invention. Explain. In this case, the candidate vectors are the nearest neighbor vector and its neighboring vectors registered in advance. Therefore, the function of candidate vector selection section 404 is greatly simplified. Further, among the outputs of the fuzzy vector quantization unit 408 at the final stage, the vector code is only the nearest neighbor vector code. Candidate vectors that are not selected can be determined by setting the class function value to O. Next, the decoding side (receiving side) will be explained. FIG. 5 is a diagram for explaining the fuzzy vector inverse quantization section 114. When the vector code 109' is received, the corresponding code vector V is retrieved from the codebook 701.
l is read out. This and the received class function util
Using lQ', the vector reproducing section 702 reproduces the vector according to the above-mentioned equation (3). It goes without saying that the codebook 701 on the receiving side has the same content as the codebook 401 on the transmitting side. Reproduction vector Xh' = (A1' + 2' l..., A.'
) is sent to the synthesis section 116 as spectrum information 115. Next, the combining section 116 will be explained using FIG. 6. In the same figure, the logarithmic power spectrum reproducing section 801
Transmitted level information A. ″105′ and each element A%HA2′ of the reproduction vector (spectral information 115)
... Logarithmic power spectrum Y' using IAI1' 8
02 is obtained according to the following equation. Y'=A,'+A. ″cosλ+A2′cos2λ
+=-+Am' cos mλ (4) The regenerated logarithmic power spectrum Y' 802 is transformed (1/2) log -'' in an inverse logarithmic transform section 803 to obtain a zero-phase spectrum 804, and then the inverse Fourier transform section 805.The inverse Fourier transform unit 805 performs an inverse fast Fourier transform (IFFT).
), a speech segment 806 is obtained. The speech segments 806 are added together in a waveform synthesis section 807 while being sequentially shifted by pitch intervals according to the pitch information 107', and are output as reproduced speech 808. The effects of the present invention in Examples are shown in FIGS. 7 and 8. This is a plot of the relationship between quantization distortion and codebook size, showing cases where the vector selection function of the present invention is applied and cases where it is not applied. FIG. 7 shows an example applied to normal fuzzy vector quantization, and FIG. 8 shows an example applied to fuzzy vector quantization using a codebook in which neighboring vectors are registered in advance. (Effects of the Invention) According to the present invention, quantization distortion can always be made smaller than conventional fuzzy vector quantization, so high quality audio can be transmitted with the same amount of information.Furthermore, with the same quality, the amount of information can be reduced. Furthermore, by applying fuzzy vector quantization that uses a codebook in which neighborhood vectors are registered in advance, it is possible to further reduce the amount of information. is used as an example, but it goes without saying that it can be used for anything that has information with a similar structure.

[Brief explanation of drawings]

FIG. 1 is a block diagram explaining the system configuration of an embodiment of the present invention, FIG. 2 is a diagram explaining the analysis section, FIG. 3 is a diagram explaining the fuzzy vector quantization section, and FIG. 4 is a diagram explaining the vector quantization section. Diagram 5 explaining vector selection based on positional relationships
FIG. 6 is a diagram for explaining the fuzzy vector inverse quantization section, FIG. 6 is a diagram for explaining the combining section, and FIGS. 7 and 8 are diagrams for explaining the effects of the embodiment. Explanation of symbols 101...Input audio, 103, 117...Buffer memory, 104...Analysis section, 106, 115...Spectrum information, 107, 107'...Pitch information, 10
8...Fuzzy vector quantization section, 109, 109'
...vector index, 110,110'...class function, 114...fuzzy vector inverse quantization unit, 11
6...Synthesizer, 120...Output audio.

Claims (1)

[Scope of Claims] 1. Means for analyzing the characteristics of an input signal as a set (vector) of a plurality of parameters; a codebook in which a code is assigned to a vector) and the code and the representative vector are stored in association with each other; a vector representing the characteristics of the input signal obtained by the analysis means and the representative vector of each cluster in the codebook; means for comparing and determining to which cluster a vector of an input signal belongs; means for selecting a representative vector from a plurality of representative vectors in the codebook based on a predetermined evaluation criterion; and the input vector A vector quantization method comprising: a representative vector of a cluster to which the vector belongs; and means for quantizing an input vector using information on the selected representative vector. 2. The target of the evaluation is a representative vector whose distance to the input vector is less than a certain value, or whose similarity with the input vector is more than a certain value. One-term vector quantization method. 3. The object of the evaluation is a predetermined number of representative vectors predetermined in descending order of distance to the input vector or in descending order of similarity to the input vector. Vector quantization method for the first term in the range. 4. The targets of the evaluation are representative vectors whose distance to the input vector is less than a certain value, or whose similarity with the input vector is more than a certain value, and whose number is less than a predetermined number. A vector quantization method according to claim 1, characterized in that: 5. Scope of claims characterized in that the objects of the evaluation are a representative vector of a cluster to which the input vector is determined to belong, and a plurality of neighboring vectors registered in advance to the representative vector. Vector quantization method for the first term. 6. The vector quantization method according to claim 1, wherein the evaluation criterion is a quantization error of the input vector. 7. The vector quantization method according to claim 1, wherein the evaluation criterion is a positional relationship between the input vector and the representative vector of the evaluation target in a feature space. 8. Vector quantization according to claim 1, wherein the evaluation criterion is an error between the input signal and the output signal obtained as a result of performing local decoding. method. 9. The selection criteria in the representative vector selection means are:
2. The vector quantization method according to claim 1, wherein the representative vector to be evaluated is sequentially evaluated, and if the evaluation value is improved, the representative vector is selected. 10. The selection criterion in the representative vector selection means is to simultaneously evaluate a plurality of representative vectors to be evaluated;
The vector quantization method according to claim 1, characterized in that a representative vector whose evaluation value improves the most among those evaluation values is selected. 11. Vector quantization according to claim 1, characterized in that the number of representative vectors selected by the representative vector selection means is the number of all representative vectors with improvement in the evaluation value. method. 12. A patent characterized in that the number of representative vectors selected by the representative vector selection means is within a predetermined number and is the number of representative vectors that improve the evaluation value. A vector quantization method according to claim 1.