JPH0291700A - Quantizer and inverse quantizer - Google Patents

Abstract

(57) [Abstract] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a quantizer/inverse quantizer that compresses the amount of information using a codebook consisting of a finite N codewords.

[Prior Art] As a method for compressing input vectors such as parameters and graphic patterns representing the frequency spectrum characteristics of speech into a small amount of information, a codebook consisting of a finite N number of codewords is prepared in advance, and a codeword is selected from among them. There is a vector quantization method that selects the code word closest to an input vector and represents the input vector only by its code number. The vector quantization method can very effectively compress the amount of information for input vectors in which there is a correlation between each component in the vector, such as the above-mentioned audio spectrum feature parameters and graphic information. A typical example of a quantizer/inverse quantizer using the conventional vector quantization method is given in Document 1 “5peech Coding Ba5ed Upon”
Vector Quantizalon A, Buzo, A, 11. Gray, J.R., R.M., G.
ray, and J, D, Markel IEEE Tr.
answers on ASSP vol, 28
．． No. 5 (1980) P, 562-574, Fig. 2 and Fig. 3. In Document 1, the input speech signal is divided into a phonological vector representing frequency spectrum characteristics, amplitude information, pitch period, voiced sound/
The information is expressed as unvoiced sound information, the phoneme vector is compressed using vector quantization, and the other information is compressed using one-dimensional vector quantization, that is, scalar quantization.

3 and 4 show the configuration of a quantizer/inverse quantizer that performs quantization/inverse quantization of phonetic vectors (hereinafter referred to as input vectors) and amplitude information described in Document 1.

In the figure! is an input vector input terminal, 2 is an amplitude information input terminal, 3 and 23 are codebooks consisting of N code words, 6 is a vector quantization means, IO is an amplitude information quantization means, and 11 is a code number transmission path. , 12 is a quantized amplitude information transmission line, 26 is an inverse vector quantization means, 28 is an amplitude information inverse quantization means, 29 is an output vector output terminal, and 30 is an amplitude information output terminal.

Next, the operation of the above conventional example will be explained. First, input vector
is input to the vector quantization means 6 in the quantizer via the input vector input terminal 1. The vector quantization means 6 inputs the input vector X and each of the N code words (Ci
l i = 1, -, N) distance (d(X, (:1)l
i = 1. ---, N), and output the code number of the optimal vector code word that gives the minimum distance in the code to the inverse vector quantization means 26 in the inverse quantizer via the code number transmission path 11. do. Here, the distance function d defines an evaluation measure of vector quantization distortion, and absolute value distance, Euclidean distance, etc. are used. Then, the inverse vector quantization means 26 outputs the code word 9 specified by the code number input from the codebook 23 via the output vector output terminal 29. At this time, the minimum distance is the quantization distortion of this quantizer/inverse quantizer with respect to the current input vector, and the smaller this value is, the better the quantizer/inverse quantizer is. On the other hand, the amplitude information is inputted via the amplitude information input terminal 2 to the amplitude information quantization means IG in the quantizer. The amplitude information quantization means IO quantizes the input amplitude information and outputs the result to the amplitude information dequantization means 28 in the dequantizer via the quantized amplitude information transmission line 12. The amplitude information dequantization means 28 dequantizes the amplitude information and outputs the result via the amplitude information output terminal 30.

[Problem M to be Solved by the Invention] A characteristic required of a quantizer/inverse quantizer is to minimize the average quantization distortion for various input sequences. Now, let us consider a case where the input sequence is a phonetic vector of speech or graphic pattern information. There are times when these input vectors change rapidly over time and times when they are relatively stable. The quantizer is not an optimal quantizer/inverse quantizer that provides the minimum quantization distortion.

A method to achieve an even smaller average quantization distortion than conventional quantizers/quantizers is to predict the current input vector using past quantization results, and then use the predicted value of the input vector There are quantizers and inverse quantizers that use the prediction error vector quantization method to quantize the error for the prediction error vector. For relatively stable input vector sequences, conventional quantizers and
The quantization distortion obtained by the inverse quantizer remains almost constant and does not decrease any further, whereas in the case of the quantizer/inverse quantizer using the prediction error vector quantization method, Quantization distortion gradually decreases over time. From this, it can be said that a quantizer/inverse quantizer using the prediction error vector quantization method is a better quantizer/inverse quantizer for an input vector sequence with little change. However, quantizers and inverse quantizers that use this prediction error vector quantization method have considerably larger quantization distortion than conventional quantizers and inverse quantizers for input vector sequences that change rapidly. The disadvantage is that the average quantization distortion for various input vectors often ends up being large.

In addition, when considering the case where auxiliary information such as amplitude information is simultaneously compressed in addition to the input vector as in the conventional example, there are times when the auxiliary information changes little and there is no need to transmit it. It is considered effective to use the amount of information used for quantization of the input vector in addition. However, in conventional quantizers and inverse quantizers, input vectors and auxiliary information are quantized independently, so optimal quantization according to such inputs cannot be performed.

The present invention aims to solve the above-mentioned problems and to obtain a quantizer/inverse quantizer with smaller average quantization distortion than conventional methods.

[Means for Solving the Problems] A quantizer/inverse quantizer according to the present invention includes a codebook composed of M vector codewords, N−M prediction error vector codewords, and a vector codeword. Determine which of the quantization distortions obtained by the quantization means/inverse vector quantization means, the prediction error vector quantization means/inverse prediction error vector quantization means, or the vector quantization means and the prediction error vector quantization means is smaller. A comparison means for making a judgment is provided.

In addition, when compressing the information amount of one input vector and auxiliary information at the same time, in addition to the above method, the predicted value of auxiliary information obtained using the quantization results of past auxiliary information and the current auxiliary information are a discriminating means for calculating the distortion between and comparing it with a predetermined threshold; and when the discriminating means determines that the distortion is smaller, the result of multiplying the prediction error vector code word by an arbitrary gain coefficient and the quantizer; prediction error vector quantification means for calculating an optimal prediction error vector code word and gain coefficient that minimizes quantization distortion between the prediction error vectors calculated within;
Furthermore, when the comparison means selects the optimum prediction error vector code word, the apparatus further includes auxiliary information quantization means for quantizing the gain coefficient instead of the auxiliary information and outputting it to the inverse quantizer.

[Operation] The quantizer according to the present invention compares the quantization distortion when using the optimal vector code word with the quantization distortion when using the optimal prediction error vector code word using the comparison means, thereby determining the current input value. For vectors, it is selected whether to perform conventional vector quantization or prediction error vector quantization.

Further, the inverse quantizer selects which inverse quantization should be performed based on the input code number from the vector code.

[Embodiment] An embodiment of the present invention will be described below with reference to FIGS. 1 and 2, in which the same parts as in FIGS. 3 and 4 are denoted by the same reference numerals. FIG. 1 is a block diagram of a quantizer, and FIG. 2 is a block diagram of a dequantizer. In the figure, 4 and 24 are positive integers M
Codebooks A, 5, and 25 consisting of vector codewords are N
- Codebook B consisting of M prediction error vector codewords, 7
is a prediction error vector quantization means, 8 is a comparison means, 9 is a discrimination means, 13 is an amplitude information quantization result, 14 is a discrimination result,
15, 17 and 21 are code numbers, 16 and 18 are quantization distortions, 19 is a gain coefficient, 20 is a comparison result, 22 is a control means, and 27 is an inverse prediction error vector quantization means.

First, the input vector is input to the vector quantization means 6 and the prediction error vector quantization means 7 via the input vector input terminal l, and the amplitude information is input to the discrimination means 9 and the amplitude information quantization means via the amplitude information input terminal 2. 10 is input.

The determining means 9 uses the quantization result 13 of past amplitude information to obtain a predicted value of current amplitude information, and obtains a prediction error between this predicted value and the input amplitude information. This prediction error is compared with a predetermined threshold value, and the determination result 14 is output to the prediction error vector quantization means 7 and the amplitude information quantization means 10. On the other hand, the vector quantization means 6 processes M vector code words H:1 in the codebook A4.
Calculate the distance (d(X,C:1)lt=1, .-, M) between the input vector x and the optimal vector that gives the minimum distance. The code number 15 of the code word and the minimum distance at that time are quantized distortion 1
6 and is output to the comparison means 8. Further, the prediction error vector quantization means 7 predicts the current input vector using the quantization result of the past input vector represented by the transmitted code number 21 and the amplitude information quantization result 13 up to the current processing time. A value Xp is obtained, and a prediction error vector Xe between this predicted value Xp and the surface input force vector X is calculated. If the prediction error of the amplitude information is equal to or greater than the threshold value as the discrimination result 14 input from the discrimination means 9, then the prediction error vector codewords (Gili
= u+ t , -, N) and the prediction error vector Xe (d(Xe, C1) li = M+ 1 ,
-, N), and outputs the code number 17 of the optimal prediction error vector code word that gives the minimum distance among them and the minimum distance at that time to the comparison means 8 as the quantization distortion 18.

If the prediction error of the amplitude information is less than the threshold value as the discrimination result 14 input from the discrimination means 9, based on the first equation and the second equation shown below, N−M pieces in the codebook BS are determined. prediction error vector code word (Cil i=M+ 1, -
, N) and the prediction error vector ×e with gain (dg(Xe, C1)li −M+ 1 , − ,
N) and gain coefficient (g(Xe, C1)Ii=M+ 1,
-, N), and outputs the code number 17 of the optimal prediction error vector code word that gives the minimum distance and the minimum distance at that time to the comparison means 8 as quantization distortion 18, and outputs the code word A gain coefficient 19 for the amplitude information is outputted to the amplitude information quantization means IO. Incidentally, when the Euclidean distance is used as the distance function d, the equation 2-1 becomes a very simple equation, and the solution can be easily obtained.

dg (Xe, (:i) = d (Xe, g *
(:i) ---The second formula comparing means 8 compares the two quantization distortions 16 and 18 inputted from the vector quantizing means 6 and the prediction error vector quantizing means 7, and compares the comparison result 20. is outputted to the amplitude information amount quantization means 10, and furthermore, the code number 21 with small quantization distortion is outputted to the prediction error vector quantization means 7 and the code number transmission line 11. The amplitude information quantization means IO determines that the prediction error of the amplitude information is less than the threshold value as the determination result 14 input from the determination means 9;
If the comparison result 20 inputted from the comparison means 8 selects the optimal prediction error vector code word, the gain coefficient 19 inputted from the prediction error vector quantization means 7 is quantized, and the quantized The information is output to the determination means 9 and the quantized amplitude information transmission line 12 as the amplitude information quantization result 13 together with additional information indicating that the information is a gain coefficient. In cases other than the above, the amplitude information quantization means IO quantizes the amplitude information, and outputs the amplitude information quantization result 13 together with additional information indicating that the quantized information is amplitude information to the determination means 9 and the quantization It is output to the amplitude information transmission line 12.

When the quantized information input via the quantized amplitude information transmission line 12 is amplitude information, the amplitude information dequantization means 28 in the dequantizer dequantizes it and sends it to the amplitude information output terminal 30. Output to. When the input quantization information is a gain coefficient, a predicted value of the current amplitude information is obtained from past quantized amplitude information and outputted to the amplitude information output terminal 30, and the dequantized gain coefficient is inversely predicted. It is output to the error vector quantization means 27. When the code number input via the code number transmission line 11 specifies a code word in the codebook A24, the control means 22 outputs the code number to the inverse vector quantization means 26, and outputs the code number to the inverse vector quantization means 26. When specifying a code word within, the code number is output to the inverse prediction error vector quantization means 27.

The inverse vector quantization means 26 reads the vector code word in the codebook A 24 specified by the input code number, and outputs it to the output vector output terminal 29. The inverse prediction error vector quantization means 27 reads out the prediction error vector code word in the codebook B25 specified by the input code number, and uses the quantization result of the past input vector to calculate the predicted value of the current input vector. seek. When no gain coefficient is input from the amplitude information inverse quantization means 28, the result of adding the predicted value of the code word input vector to the prediction error vector is output to the output vector output terminal 29. When a gain coefficient is input from the amplitude information inverse quantization means 28, the result obtained by adding the predicted value of the input vector to the value obtained by multiplying the predicted error vector code word by the gain coefficient is outputted to the output vector output terminal 29.

In the above embodiment, an example was shown in which the gain coefficient is quantized and output to the inverse quantizer when it is determined that quantization of amplitude information is unnecessary. However, when there is no auxiliary information of the input vector such as amplitude information , in the above example 2.9, 10, 12, 13,
The configuration is such that 14, 19, 20, 28, and 30 are removed. Further, a configuration in which the gain coefficient is always quantized and used is also possible. In addition to the conventional vector quantization means and prediction error vector quantization means, it is equipped with another quantization means such as prediction error vector quantization means with different prediction conditions, and also includes a prediction error vector quantization means with different prediction conditions. By configuring the codeword to be included in the codebook, the average quantization f can be further improved.

[Effects of the Invention] As described above, according to the present invention, the selection of the conventional vector quantization means and the predicted vector quantization means is automatically performed according to the input vector, and A quantizer/inverse quantizer with small average quantization distortion can be realized. In addition, if the characteristics of the input vector sequence are known to some extent, an optimal quantizer/inverse quantizer can be realized by optimally selecting the ratio of the number of vector code words M to the number of prediction error vector code words N-M. It is configured so that it can be done.

When quantizing auxiliary information at the same time, according to the present invention, when quantization of the auxiliary information is deemed unnecessary, it is possible to use the surplus information for quantizing the input vector, and the quantization of the entire quantizer properties can be further improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a quantizer in an embodiment of the present invention, FIG. 2 is a block diagram showing an inverse quantizer in the embodiment, and FIG. 3 is a block diagram showing a conventional quantizer. FIG. 4 is a block diagram showing a conventional inverse quantizer. In the figure, 1 is an input vector input terminal, 2 is an amplitude information input terminal, 3 and 23 are codebooks, and 4 and 24 are codebook A.
, 5 and 25 are codebook B, 6 is vector quantization means, 7
is a prediction error vector quantization means, 8 is a comparison means, 9 is a discrimination means, IO is an amplitude information quantization means, 11 is a code number transmission path, 12 is a quantized amplitude information transmission path, 13 is an amplitude information quantization result, 14 is the discrimination result, 15.17 and 21 are code numbers, 16 and 18 are quantization distortions, 19 is gain coefficient, 20
is a comparison result, 22 is a control means, 26 is an inverse vector quantization means, 27 is an inverse prediction error vector quantization means, 28 is an amplitude information inverse quantization means, 29 is an output vector output terminal, 3
0 is an amplitude information output terminal. Agent Masuo Oiwa Figure 2 Figure 1 Figure 3 Figure 4

Claims (1)

[Claims] 1. In a quantizer/inverse quantizer that performs information compression using a codebook consisting of a finite N codewords, the quantizer has M positive integers smaller than N. , a codebook consisting of N-M prediction error vector codewords, and an optimal vector codeword closest to the current input vector to be quantized from among the M vector codewords. a vector quantization means for calculating the quantization distortion of the selected input vector; and a vector quantization means for calculating the quantization distortion of the selected input vector; prediction error vector quantization means for selecting an optimal prediction error vector codeword closest to the prediction error vector from error vector codewords and calculating its quantization distortion; and the selected optimal vector codeword and the optimal prediction error. The inverse quantizer has a comparison means for selecting a codeword giving a smaller quantization distortion among the vector codewords and outputting its code number, and the inverse quantizer uses the same codebook as in the quantizer and the quantizer. an inverse vector quantization means for outputting a vector code word when the code number input from the quantizer specifies one code word among the M vector code words; and a code number input from the quantizer. inverse prediction error vector quantization means for outputting a result of adding the prediction error vector codeword to the predicted value of the input vector when specifies one codeword among the N-M prediction error vector codewords; A quantizer/inverse quantizer comprising: a control means for switching between the vector quantization means and the inverse prediction error vector quantization means based on a previous code number. 2. In a quantizer/inverse quantizer that simultaneously compresses the amount of information of an input vector and auxiliary information related to the input vector, the first quantizer has the same configuration as in claim 1, and also quantizes auxiliary information. A predicted value of the current auxiliary information is calculated using the auxiliary information quantization means and past auxiliary information quantization results, and a prediction error of the predicted value for the input current auxiliary information and a predetermined threshold value are calculated. It has a discrimination means for comparison,
The prediction error vector quantization means in the quantizer calculates an optimal prediction error vector code word when the determination means determines that the prediction error is equal to or greater than the threshold value, and the determination means If the prediction error is determined to be smaller than the threshold value, the quantization distortion between the result of multiplying the prediction error vector code word by an arbitrary gain coefficient and the prediction error vector obtained in the quantizer is minimized. The auxiliary information quantization means determines that the prediction error is smaller than the threshold, and the comparison means in the quantizer determines the optimal prediction error vector code word and gain coefficient. When an error vector code word is selected, the auxiliary information is not quantized, but the gain coefficient is quantized and output to the inverse quantizer, and the inverse quantizer has the same configuration as in claim 1, and in addition, Performs inverse quantization of the auxiliary information when the auxiliary information is input,
When a gain coefficient is input instead of the auxiliary information, the auxiliary information inverse quantization means dequantizes the gain coefficient and outputs it to the inverse prediction error vector quantization means to calculate and output the predicted value of the auxiliary information. The inverse prediction error vector quantizer in the inverse quantizer is configured such that the input code number specifies one code word among the N-M prediction error vector code words, and the gain coefficient is input. quantizer/inverse quantizer, characterized in that the quantizer/inverse quantizer is configured to output a result obtained by adding a value multiplied by a gain coefficient of the specified prediction error vector code word to the predicted value of the input vector when