JP3095758B2 - Code Vector Search Method for Vector Quantization - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a code vector search method of vector quantization for searching for an optimum code vector most similar to an input vector.

[Summary of the Invention]

The present invention provides a code vector search method of vector quantization for searching for an optimum code vector most similar to an input vector from among a plurality of types of code vectors. By calculating the probability of selection, and searching for the highest selection probability according to the code vector selected in the past at the time of vector quantization,
An object of the present invention is to provide a code vector search method of vector quantization that can reduce the amount of calculation for searching for an optimum code vector.

[Conventional technology]

As one method of quantization, for example, a waveform encoding of a speech signal, a waveform in an analysis / synthesis system, a spectrum envelope parameter, and the like are not quantized for each sample value, and a group of a plurality of sample values (input vector) is collectively obtained as There is a quantization method represented by one identification code (index), and this quantization method is generally called vector quantization (vector quantiz).
ation: VQ) or multidimensional quantization, block quantization.

For example, in the vector quantization of a waveform, a sampled waveform is cut out for a certain period of time and framed, and a waveform pattern for each frame, that is, a vector composed of a set of sample values for each frame is identified by one identification code. It is represented by For this reason, various waveform patterns (that is, code vector: codevector or template: templa
te) is stored, and an identification code is given to each code vector. A table showing the correspondence between these identification codes and code vectors is shown in a codebook (codebook).
I'm calling That is, in the vector quantization of the waveform, each code vector of the code book is compared with the input vector for each frame of the input waveform (pattern matching), and the code vector having the highest similarity to the input vector (optimum code) The waveform of the section (frame) is represented by the identification code corresponding to the vector. For this reason, the code book takes into account the bias of the distribution of the pattern of the input waveform, and when expressing various input voices using a limited number of code vectors, it is appropriate to minimize distortion as a whole. It is necessary to store code vectors. As a specific method, from a set of original waveform patterns to be represented by code vectors,
For example, it is preferable to calculate a centroid (centroid) in the pattern space based on a distortion scale (distance scale or similarity) and store it in a codebook. Therefore, the stored pattern (code vector) changes depending on how the distortion measure is selected.

Further, for example, in vector quantization of linear predictive coding (LPC) frequently used in a speech analysis / synthesis system, a speech deformation after sampling is extracted for each frame cut out for a certain time. 1 to p
LPC parameters, for example, the partial autocorrelation coefficient (PARCO
R coefficient), Line spectrum pair (LSP)
A set of parameters is set as a set of vectors, which are represented by one identification code. As a result, it is possible to realize extremely low bit encoding while maintaining the clarity of the speech, although the quality is lower than the quality of the waveform code speech.

The linear prediction analysis will be described. That is, generally, for an audio signal or a video signal, information compression is performed by encoding a difference signal (prediction residual) between a value predicted using correlation between adjacent samples and an actual sample value. Can be. The above-described linear prediction analysis is used for such a predictive coding prediction method. Here, the linear prediction analysis is a method of adjusting a prediction filter coefficient as an LPC parameter representing a linear prediction filter to minimize the sum of squares of the difference signal (prediction residual). According to the analysis of the speech waveform to which the technique of the linear prediction analysis is applied, the properties of the speech waveform and its spectrum can be efficiently and accurately expressed with a very small number of parameters (the above k parameters). Moreover, those parameters can be obtained by relatively simple calculations.

[Problems to be solved by the invention]

By the way, in the vector quantization as described above, when an optimum code vector most similar to an input vector is searched to obtain an identification code, a search is performed for all code vectors in the code book. for that reason,
The amount of computation required for code vector retrieval is increasing. For this reason, for example, when real-time processing is to be performed, the hardware becomes large-scale.

Therefore, the present invention has been proposed in view of the above-described circumstances, and provides a code vector search method of vector quantization that can reduce the amount of calculation required for searching for an optimum code vector in a code book. It is intended to do so.

[Means for solving the problem]

A code vector search method for vector quantization according to the present invention has been proposed to achieve the above object,
In a vector quantization code vector search method for searching for an optimum code vector most similar to an input vector composed of a plurality of pieces of information having correlation with each other from a plurality of types of code vectors prepared in advance, The probability of selection of each code vector corresponding to at least one code vector selected in the past in the past with respect to the code vector is obtained in advance based on the code vector sequence selected in the past, and vector quantization is performed. At the time
With reference to the selection probabilities of the respective code vectors according to at least one code vector selected in the past in the past, the search is performed in order from the higher one of the respective code vector selection probabilities.

[Operation] According to the present invention, when searching for an optimal code vector by vector quantization, the higher one of the code vector selection probabilities has a high probability of being the optimal code vector. Therefore, by searching the code vector selection probability in descending order, the optimum code vector can be quickly found.

[Example] Hereinafter, an example to which the present invention is applied will be described.

The code vector search method of the vector quantization according to the embodiment of the present invention is a vector quantization for searching for an optimum code vector most similar to an input vector from a plurality of types of code vectors. In the case where a correspondence table (for example, a table as shown in Table 1) of probabilities to be selected next to each code vector corresponding to the obtained code vector sequence is obtained and the current vector quantization is to be performed, According to the code vector sequence selected in the past, the search is preferentially performed in order from the higher-order (that is, the one with the highest probability) among the respective code vector selection probabilities in the correspondence table.

Here, in the present embodiment, vector quantization in linear prediction analysis (LPC) frequently used in a speech analysis / synthesis system will be described as an example.

In the linear prediction analysis used in the speech analysis / synthesis system, first, a speech waveform after being sampled at a sampling frequency fs is cut out for a fixed time (for example, 20 msec) and framed. In the vector quantization in this linear prediction analysis, the 1st to pth order extracted for each frame
LPC parameters, for example, the partial autocorrelation coefficient (PAR
A set of vectors (input vectors) is configured by setting the COR coefficient), line spectrum pair (LSP) parameters, and the like.

Here, as the LPC parameter, for example, a k parameter, which is a partial autocorrelation coefficient, represents a spectral envelope of a voice, and is obtained by performing a partial autocorrelation analysis on an input voice signal for each frame. At this time, a set of k parameters of adjacent frames has a correlation due to linguistic constraints or constraints due to human vocalization mechanisms. Since the set of k parameters has a correlation, a correlation also appears between code vectors selected by vector quantization. From this, for example, the optimal code vector that would be chosen in the frame that is currently about to be vector quantized is
The prediction can be made to some extent by looking at the code vector selected in the past, for example, the optimal code vector obtained by vector quantization of the frame immediately before (immediately before) the current frame.

Therefore, in this embodiment, in the vector quantization of the k parameter, a table of the current code vector selection probabilities according to the optimum code vector sequence searched and selected in the past is created in advance. That is, for example, a table of code vector selection probabilities as shown in Table 1 is created based on the frequency of the next selected code vector according to the most recently selected optimum code vector. As described above, as a method of calculating the frequency of the next selected code vector with respect to the most recently selected optimal code vector, the code vector of the code book is obtained at the time of training for creating the code book,
A method of obtaining the frequency of the next code vector with respect to the code vector selected immediately before, or using a code vector already created by the above training, vector quantization of an actual speech signal or the like for a predetermined time to obtain statistics Thus, there is a method of calculating the frequency of the next code vector with respect to the code vector selected immediately before.

In the present embodiment, when the current vector quantization is to be performed, the higher one of the code vector selection probabilities in Table 1, that is, the higher the probability is, according to the code vector sequence selected in the past. The search is performed in order from the one with the highest priority. In other words, when searching for an optimal code vector by vector quantization, the higher one of the code vector selection probabilities has a higher probability of being the optimal code vector. For this reason, if the search is performed in order from the one with the highest code vector selection probability, the optimum code vector can be found quickly (in a short time).

That is, in this Table 1, for example, the plurality of types of code vectors to eight prepared, the corresponding identification code for each of these code vectors and I ₀ ~I _7. In Table 1, the selection probability of each code vector in the current vector quantization with respect to the optimal code vector selected immediately before (in the past) is shown. Therefore, when the above-mentioned probability of the current code vector with respect to the most recently selected optimum code vector is high, there is a high possibility that the code vector is selected. Therefore, the order of the code vector searched by the current vector quantization is as follows.
The search is performed first from the code vector having the higher probability shown in Table 1.

At this time, for example, it is possible not to search for a code vector having a low probability. Whether to perform the search based on the probability in this manner can be as shown in Table 2, for example. In the second table, for each of the optimal code vectors selected immediately before shown in the first table, the code vectors are arranged in descending order of the probability of selection of each code vector in the current vector quantization.

In the second table, for example, the code vector searched by the current vector quantization is set to the fourth highest probability,
By making the remaining code vectors with low probabilities non-searchable, it becomes possible to find the optimum code vector in a short time without searching all the code vectors in the codebook. In addition, for example, when there are 1024 types of code vectors to be searched, a significant reduction in the amount of computation can be expected by searching for, for example, up to the 100th with the highest probability.

That is, in the present embodiment, it is possible to reduce the amount of calculation for searching for the optimum code vector, and also to shorten the processing time. Therefore, for example, even when real-time processing is to be performed, it is possible to prevent an increase in hardware configuration.

In the above-described embodiment, the optimal code vector selected immediately before is used for obtaining the selection probability. However, the present invention is not limited to this, and based on the optimal code vector selected in two or more past frames. Alternatively, the selection probability of the currently selected code vector may be determined.

Here, FIG. 1 shows a specific example of a configuration (that is, a configuration of encoding and decoding) of a partial autocorrelation type speech analysis / synthesis system of an input speech signal to which the embodiment of the present invention is applicable.

In FIG. 1, the analog audio signal waveform supplied via the input terminal 101 of the audio analysis system (encoding) is
An analog / digital (A / D) converter performs sampling at a predetermined sampling frequency, and this waveform data is added to adders 103 _{1 to} 103 _p , 108 _{1 to} 108 _p , and multipliers 104 _{1 to} 104
_p , 106 _{1 to} 106 _p , delay device 107 _{1 to} 107 _p , correlation calculator 105 ₁ to 10
In 5 _p it is transmitted to the configured lattice prediction filter circuit 115. Waveform data supplied to the lattice prediction filter circuit 115, first, the sent to the correlation calculator 105 _1, also sent to the delay unit 107 _1, after being delayed a predetermined time in the delay unit 107 ₁ is sent to the correlation calculator 105 _1. From the correlation calculator 105 _1, so that the above k parameter (k ₁₎ is obtained by the correlation calculation between the waveform data and the delayed data. This k parameter (k ₁ ) is calculated by the multiplier 10
4 ₁ , 106 ₁ The multiplier 104 the waveform data is _1, the aforementioned multiplier 106 ₁ is the delayed data is supplied, each multiplier 104 _1, 106 ₁ respectively supplied data to the k parameter (k ₁₎ Is multiplied. Multiplier 10 above
The output from the 4 ₁ differential data is obtained by being transmitted to the adder 108 ₁ being subtracted from the delayed data. Further, the output from the multiplier 106 ₁ is differential data obtained by the subtraction is transmitted to the adder 103 ₁ from the waveform data. Hereinafter, adders 103 _{2 to} 103 _p , 108 _{2 to} 108 _p , and multiplier 104 ₂
~ 104 _p , 106 ₂ ~ 106 _p , delay unit 107 ₂ ~ 107 _p , correlation calculator 105
_A similar filtering process is performed in _{2 to} 105 _p .

By such filtering processing is performed, the adder 103 the prediction residual waveform data, which is the difference _p input signal and the prediction value signal is obtained, the prediction residual waveform data, transmitted to the correlation detection circuit 109 Is done. The correlation detection circuit 109 obtains amplitude information A and pitch information T from the predicted residual waveform data.

Further, each k parameter (k _{1 to} k _p ) from each of the correlation calculators 105 _{1 to} 105 _p is transmitted to the encoder 110, and the encoder 110 performs, for example, the above-described vector of the embodiment of the present invention. It is output after the quantization process of each k parameter is performed using the quantization code vector search method. The amplitude information A and the pitch information T are also output via the encoder 110. These outputs are encoded outputs.

Next, as a speech synthesis system (decoding), the encoder 11
The output from 0 is sent to the decoder 120,
Processing opposite to 0 is performed. The pitch information T via the decoder 120 is sent to the pulse generator 121, and the pulse generator 121 outputs a pulse signal based on the pitch information T. This pulse signal is transmitted to a changeover switch 123 that is switched according to the property of the signal, that is, from voice sound to unvoiced sound. Also, random noise from the random noise generator 122 is sent to the changeover switch 123. The output from the changeover switch 123 is sent to the multiplier 124. Further, the amplitude information A is also sent to the multiplier 124, and data approximate to the predicted residual waveform data can be obtained from the multiplier 124.

The estimated residual waveform approximation data is added to adders 125 _{1 to} 125 _p , 127
_{1 to} 127 _p , multipliers 126 _{1 to} 126 _p , 128 _{1 to} 128 _p , delay unit 129 ₁ to
It is supplied to the constructed lattice prediction filter 135 in 129 _p. Each of the decoded k parameters is sent to each of the multipliers 126 _{1 to} 126 _p and 128 _{1 to} 128 _p . In the lattice-type prediction filter circuit 135, the prediction residual waveform approximation data is filtered through a path opposite to that of the lattice-type prediction filter 115. Here, the delay units 129 _{1 to} 129 _p and the Adders 127 _{1 to} 127 _p , Multipliers 128 ₁ to
Data sequentially treated with 128 _p has a the adder 125 of _the outputs of after passing through the respective adders 125 ₁ to 125 _p.

Data approximating the waveform data obtained by such a filtering process is converted into an analog synthesized voice waveform by a digital / analog (D / A) converter 130 and output from an output terminal 141.

In addition, the present invention is not limited to the above-described embodiment. For example, the present invention is also applicable to a case where waveform data of an audio signal waveform is framed in the time axis direction and vector-quantized. For example, the waveform after being converted is cut out for a predetermined time that does not decrease the correlation and is framed (or the frame is further divided into blocks), and the waveform pattern for each frame (or for each block), that is, for each frame ( An input vector is composed of a set of sample values for each block. At this time, the present invention can be applied when searching for an optimal code vector in each frame (block). Also in this case, the same effect as described above, that is, the amount of calculation for code vector search can be reduced, and even when real-time processing is performed, an increase in the size of the device configuration can be prevented.

Further, for example, an audio signal or the like is framed (or block-formed), and waveform data for each frame (for each block) is subjected to, for example, fast Fourier transform (FFT), discrete cosine transform (DCT), Karhunen-Loeve transform ( The present invention is also applicable to vector quantization of an input vector composed of FFT coefficients, DCT coefficients, and the like obtained by performing orthogonal transformation processing such as KL transformation. In this case, the same effect can be obtained.

In addition, the present invention is a signal other than the audio signal or the video signal, if the signal is composed of mutually correlated information,
It is applicable to the vector quantization.

[Effects of the Invention] According to the code vector search method for vector quantization of the present invention, among a plurality of types of code vectors prepared in advance, an input vector constituted by a plurality of pieces of information having a correlation with each other is most frequently used. In a code vector search method of vector quantization for searching for a similar optimum code vector, for each code vector, a probability that each code vector corresponding to at least one previous code vector selected in the past is selected. Based on the code vector sequence selected in the past, it is obtained in advance, and at the time of vector quantization, referring to the selection probability of each of the code vectors according to at least one previous code vector selected in the past, By searching the above code vector selection probabilities in order from the highest one, the optimal code vector This makes it possible to reduce the amount of subtraction for file search and shorten the processing time.

Therefore, for example, even when the real-time processing is to be performed, the hardware does not increase in scale.

[Brief description of the drawings]

FIG. 1 is a block circuit diagram showing a configuration of a partial autocorrelation type speech analysis / synthesis system according to a specific example to which the embodiment of the present invention is applied.

Claims (1)

(57) [Claims] 1. A vector quantization code vector search for searching an optimum code vector most similar to an input vector composed of a plurality of pieces of information having correlation with each other from a plurality of types of code vectors prepared in advance. Determining, for each code vector, a probability of selecting each code vector in accordance with at least one previous previously selected code vector based on a previously selected code vector sequence. At the time of vector quantization, reference is made to the selection probabilities of the respective code vectors according to at least one code vector selected in the past in the past, and the search is performed in order from the higher one of the respective code vector selection probabilities. A code vector search method for vector quantization, characterized in that: