JPS63226695A - High-speed pattern matching system - Google Patents

Abstract

(57) [Summary] 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 pattern matching method, which is the main process of speech recognition for automatically recognizing speech uttered by humans.

(Prior art) Various technologies have been developed regarding pattern matching methods for speech recognition, but one of the most heavily used among them is the one published in ``Journal of the Acoustical Society of Japan, Vol. 42, No. 9 (Showa 61). There is a DP matching method as described in page 725 of "Published in September 2013" (hereinafter referred to as Document-1). This is considered to be extremely effective as a method for matching time axis distortion of audio. Furthermore, as an extension of the DP matching method to continuous word recognition, see page 727 of the above-mentioned document-1 or the specification of Japanese Patent Application No. 199098/1983! ! There is a clockwise DP method as described in . Although this method is described as a continuous word recognition method with syntactic control, it naturally also includes discrete word recognition as a special form. Here, for simplicity, the main part of the clockwise DP method will be explained in the form of f-discrete word recognition.

Assuming that word names are designated by numbers n, a word set (n l n = 1.2.-- N ) is set as a recognition target. Standard pattern for each word B" = gadfly, b, i'...>1n...v
Thinking about Jn. Here, j indicates time, and 111I'' means the characteristic of standard pattern B'' at time j. Similarly, the input voice button is A=甐1. Recommendation...Ship...Heavy industry.

Speech recognition is performed by finding the pattern distance D(A, B") between the input pattern A and the standard pattern "n", determining the minimum distance n, and using this as the recognition result.

In DP matching, the inter-pattern distance is calculated by, for example, the following dynamic measurement method calculation.

0 initial condition gn (1, 1) = dfl (1, 1) ...
・・・・・・・・・・・・・・・(1) 0 recurrence formula % formula % O distance between patterns D (A, B') = g" (I, J'),
・・・・・・・・・・・・・・・・・・・・・(3);c
koK d"(i, j)tiSpec1k mt Th4
"f) Distance d'(i, j)=ll-town"II. This is the integral form.

g"l+J) is called the optimal cumulative distance. Initially, this DP matching process was executed for each word, but in the clockwise DP method, it was improved to be executed for each word in parallel. That is, i, j, as shown in Fig.
At each time i of an input pattern in the space spanned by n, gfi(
The system calculates the optimal cumulative value of i, j), and then advances the time i and executes the process.

In actual calculations, it is not necessary to prepare a work area for the entire space in the figure, and for the i direction, time i and i
-], it is possible to proceed with the calculation of equation (2).

In this P matching, the problem is the amount of calculation required for the distance d(i, D. Normally book 1, To, n is 101(+
ms) would be a huge burden on normal hardware.

As a measure against this problem, No. 725 of the above-mentioned document-1
In a paper published on page 730 titled 1 Bump matching method in speech recognition -2,? It may be considered to apply vector quantization as described. In other words, the set of standard batan features (lf+
A set of code vectors ([h
), and each standard pattern B' has a number that specifies the code vector Ck closest to each 1b,' = k(n, j
) as a time series. Based on this, when executing DP matching, the distance D (kl) between the feature book i of the input pattern and each code vector is calculated and made into a table, and when calculating the recurrence formula, d'' (i, j) = D(k(n, j)) ---
−−−・・・・−−−−−−・Refer to (4) and (2
) calculate the formula. Although the number of distance calculations has been significantly reduced by using vector quantization, it is still estimated that the amount of calculations will be considerable.

(Problems to be Solved by the Invention) Approximately 256 code vectors are required in order not to cause a reduction in recognition 8 compared to the normal DP matching method. Even if one calculation of D(i, k) is completed in 40 μs, 1Qms is required for 256 calculations. In other words, most of the clock minutes (lQms) of i are used in calculating the distance between vectors, leaving no time for calculating the recurrence formula. For this reason, DP matching has conventionally been performed using high-speed dedicated hardware.

The purpose of the present invention is to improve the above-mentioned drawbacks of the clockwise DP method, which requires a large amount of calculation even when combined with vector quantization techniques, and to provide a pattern matching method for speech recognition that is high speed and inexpensive. do.

(Means to Solve the Problem) The pattern matching method of the present invention uses the clockwise DP method that employs the vector quantization method to match the input audio at the current clock based on the optimal cumulative value calculated in the past. The present invention is characterized in that it predicts a set of code vectors GEk for which distances to feature sets of a pattern need to be calculated, and eliminates the need to calculate distances to other code vectors.

(Operation/Principle) Originally, DP matching was performed from point (1, 1) to (I
, J''), the path that minimizes the sum of d'(Lj), that is, the cumulative value, is searched for.The optimal cumulative value g''(i, j) calculated in this process is (1 ,1) dn(i,j) on the optimal path from point to point (i,j)
It gives the cumulative value of Therefore, a large value of gn('+j) means that there is a low possibility that this point (i, j) is on the optimal route.

In the present invention, when % g''('+ J) is predicted to be large, d''(i, D=D(k(n, j)) in equation (4).
) is intended to reduce the amount of calculation by omitting the calculation.

(Example) FIG. 2 is a configuration example of a discrete word recognition lid implementing the principle of the present invention. The audio waveform input from the microphone 10 is frequency-analyzed by the analyzer 20, converted into a feature vector, and input into a microprocessor 30 as a time series. A code vector Ck is stored in the code book 40, and a standard pattern B' for each word n is stored in the standard pattern storage section 50.
n, j) is stored as a time series. D memory 41
is the code vector (distance D between Ck and input vector car 1)
(k) is temporarily stored. The g memory 60 is (
2) It serves as a working memory for calculation of the recurrence formula, and gn
(ilj) and gn(i-Lj) are stored for required n and j. In addition, these 40.4
The 1°50.60 memory may be configured as an area on the main memory of the microprocessor 30.

When the first feature vector 4 of the input pattern is input, the function of the microprocessor 30 causes the gmemo I760
The following initial setting process is performed for the part g''(i-1, j).

gn (1, 1) = D (k (n, 1)) That is, = k (n, l) is read out for each word from the standard pattern memory 50, and the corresponding code vector is read from the code book 40. Calculate the distance to the feature vector set with reference and set it as the initial value g"(1, 1). Note that j
Setting a sufficiently large value ■ for the 'xc1 part is shown in Figure 6(a) of the specification of Japanese Patent Application No. 199098/1983.
The same is true for .

Generally, the process shown in FIG. 3 is executed at time i.

First, when a feature vector ship is input, all contents D(k) of the D table are reset to ■. Then for each word n, j=1.2. ...The following processing is performed on Jfl. From the g memory 60, the past optimal cumulative value g
n(i-11j)+gn(i-1,j-1)+g"
(i-Lj-2) and determine their minimum value g.

This value is compared with the threshold value θ(i) (block 100) to test whether (i, D at this point is on the optimal route. Since the optimal cumulative value increases with time i, , which is predetermined as a monotonically increasing function. When g>θ(i), this point (i, j
) is determined to have no trap on the optimal route, and the recurrence formula and distance calculation are omitted, and g″′(i.

For j), set ■. When g≦θ(i),
This point (i, j) is predicted to be on the optimal path, and the following recurrence formula calculation is performed.

First, =k(n,j) is read from the standard pattern memory 50, and based on this signal, D in the D table 41 is
See (k). If this is the case, D (assuming that the value of 10 has not been calculated yet, calculate the distance between the feature vector book 1 and the code vector Ch read from the codebook) and set it as D, and set it as D(k). Put it on the D table. If D(k) is not ω, it is assumed that the distance between the code vector Ck and feature vector book 1 has been calculated, and D=D(k).

This series of processing predicts that the distance between the code vector Ck and the feature vector set specified by % k = k (n, j) is required, assuming that n, i, and j at this point are on the optimal path. This method performs distance calculations and avoids duplicate distance calculations. Thus, in block 110 g
+D−+g”(it j), and a recurrence formula calculation equivalent to (2), (41) was performed. This new optimal cumulative value is g memo IJ 6
0K written. By repeating this process for the required j and n, the recurrence formula calculation at time i is completed. In block 120, the area g''(i, J) and the area g''(i-Lj) in the g memory are switched, and the optimal cumulative value calculated at the current time i is used as past data, and the next The process moves to time i+1.

In this way, the processing up to the time adjustment is performed, and when the input speech is completed and i = I + ], the inter-pattern distance D (A, B') is obtained. These are compared and the recognition result is determined and output as the smallest word ==''.

Although the principle of the present invention has been described above based on examples, these do not limit the scope of the present invention. Especially the third
Various modifications can be considered to the determination process of block 100 in the figure. Regarding the method of determining the threshold value θ(i), in addition to the method of manually defining it in advance, there is also a method of defining the threshold value θ(i)
Possible modifications include setting the value to the minimum value of j). Also, calculate the distance between it and ■k by j+n
It is also possible to do this in advance outside the loop. These modifications are within the scope of the present invention.

In addition, in the above explanation, formula (2) was used as the basic recurrence formula, but [Nikkei Electronics, 1983,
The principles of the present invention are also applicable to various deformed recurrence formulas as shown in Table IJK on page 184 of the May 7 issue.

Furthermore, the present invention can be used for continuous word recognition similar to the clockwise DP method described in Japanese Patent Application No. 56-199098.

(Effects of the Invention) The number of distance calculations can be reduced by limiting the code vectors that require distance calculations at the current time using the past optimal cumulative values.

[Brief explanation of the drawing]

FIG. 1 is a diagram explaining the principle of the present invention, FIG. 2 is a block diagram showing an embodiment of the invention, and FIG. 3 is a flow chart explaining its operation. 10...Microphone, 20...Analysis section, 30...Microprocessor, 40...
...Code book, 4] ...D table, 5
0...Standard pattern memory, 60......g
memory.

Claims (1)

[Claims] A codebook consisting of a set of code vectors ■_k representing the characteristics of speech, and a number k (n, j) (j = 1, 2...J^n) that specifies the code vector corresponding to each word n.
means for storing a standard pattern B^n expressed as a time series, means for temporarily retaining the characteristic ■ of the input speech pattern, means for calculating the distance D(k) between ■ and each code vector ■__k; means for calculating the optimal cumulative value g^n(i, j) of the distance D(k(n, j)) for each word by dynamic programming; A high-speed method characterized by limiting the set of code vectors ■_k whose distance to ■ must be calculated at the current clock based on the optimal cumulative value, and calculating the distance D(k) only for these. pattern matching method.