JPH0668677B2 - Speech recognition method and apparatus using vector division quantization - Google Patents

Description

TECHNICAL FIELD The present invention relates to a speech recognition method and apparatus using vector division quantization.

2. Description of the Related Art FIG. 1 is a basic circuit diagram of a voice recognition device. In the figure, 1 is a microphone, 2 is an analysis unit, 3 is a changeover switch, 4 is a standard pattern unit, 5 is an input voice pattern unit, and 6 is distance calculation. Reference numeral 7 denotes a minimum value detecting portion, 8 denotes a recognition result portion, and the distance calculating portion 6 and the minimum value detecting portion 7 form a pattern matching portion. In FIG. 1, the voice coming from the microphone 1 is analyzed to extract a pattern for recognizing the features of the voice pattern. In the system for a specific speaker, the analysis result of each recognition target word of the speaker is registered as a standard pattern before recognition, and at the time of recognition, the standard pattern of each recognition image word is input. The parameters of the voice patterns are compared to select the closest recognition target word, that is, the word having the smallest distance. In the case of an unspecified speaker, a standard pattern that can absorb individual differences is used.

FIG. 2 is a diagram showing an example of an analysis method using a band pass filter group (BPF). In the figure, the voice "3" (/ san /) is transmitted through a 16 channel band pass filter group (all bands are
It is a time change figure of the spectrum pattern analyzed (BPF analysis) by 200-600HZ). One unit of the time axis is 18 ms, and if you take a cross section at a certain time, it becomes the spectrum at that time, and the actual recognition process is all digital processing, and the spectrum of a horizontal row at a certain time i Let the intensity value be the feature vector ai (= ai ₁ ai ₂ ai ₃ … ai ₈ … ai ₁₆ ),
The input voice pattern (here, the voice pattern "3") is A = a ₁ a ₂ ... Ai ... aI (I = 32).

Therefore, the voice pattern is expressed as follows.

A = a ₁ a ₂ ... Ai ... aI (1) ai is a quantity representing the feature of the voice at time i, and is generally a vector value, and A is the feature vector ai [i = 1 to 32.
(When I = 32)], and I corresponds to the length of the voice pattern A.

The vector ai is called a feature vector and is represented by ai = (ai ₁ , ai ₂ ... Aiq ... aiQ) (2). Q is the order of the vector and corresponds to the channel number 16 of the bandpass band filter group in the example of FIG.

Similarly, the standard pattern of the word n is Bn, and Bn = b ₁ nbn ₂ ... Bjn ... BJn ... (3). At this time, bjn is a feature vector of the standard pattern of the word n at time j, and has the same degree as the feature vector ai of the input pattern A. Jn represents the length of the standard pattern of word n, and n is a serial number indicating the word name,
Considering the recognition word set of N words and assuming Σ, Σ = {n | n = 1,2 ... N} (4) However, when it is not necessary to specify a specific word, the subscript n is omitted, and B = b ₁ b ₂ … bj… bJ …… (5) bj ＝ (bj ₁ , bj ₂ ,… bj ₈ … bjQ ) ... (6)

In the voice recognition process, the standard pattern Bn of all the words in the recognition word set for the input pattern A is subjected to pattern matching while being time-sighted, and the word n closest to the input pattern A is searched from the N words.

FIG. 3 shows a mapping model for time emmetropization.
In the above example, the standard pattern B of the word "3" is aligned with the time axis of the input pattern by the mapping function.
Usually, the mapping function is expressed by j = j (i) (7), which is called a distortion function.

If this distortion function is known, the time axis of the standard pattern B is converted by the equation (7) to obtain the time axis i of the input pattern A.
However, in practice, this distortion function is unknown, so that one pattern is artificially distorted so that it is most similar to the other, i.e., the distance is minimized and the optimal distortion is obtained. I try to define the function.

Figure 4 shows DP (Dynamic Prog) for implementing the above principle.
ramming; dynamic programming) A diagram for explaining an example of a matching method. Now, when the distortion function j (i) is considered as a function that distorts the time axis of the standard pattern B, this distortion function j (i) Therefore, the pattern B is converted into the following pattern B '.

B ′ = Bj ( ₁ ) bj ( ₂ ) ... bj (i) ... bj (I) ... (8) The distortion function is, for example, (a), j in consideration of the time distortion development of the actual voice pattern. (I) is (approximately) a monotonically increasing function, (b), j (i) is (approximately) a continuous function, (c), j (i) is a value in the vicinity of i, and so on. In addition, there are almost infinite distortion functions that satisfy these conditions. Among them, the distortion function j (i) is determined so that B ′ is most similar to the input pattern A, that is, the distance is the smallest. For this purpose, first, the time axis of the standard pattern B is set to the distortion pattern j (i) in the input pattern A
The pattern B'is obtained by mapping on the i-axis of the above. At this time, the distortion function j (i) that minimizes the distance between the pattern A and the pattern B'is the optimum distortion function. This input pattern A
And the mapping pattern B ′ is It is represented by. Here, ‖ ‖ indicates the distance between two vectors. Then, the problem of minimizing the distance in the above equation (9) is Is defined by Generally, D (A, B) is called the time-normalized distance or the distance between patterns, and d (i, j) is the vector ai and bj.
It is the distance between and and is usually called the vector distance.

In FIG. 5, the (i, j) plane shown in FIG. 4 is abstracted into a grid surface, and for each grid point, the inter-vector distance d (i, j) corresponding to its coordinates (i, j) is calculated. If the above equation (10) is considered on this plane, (1,
Starting from 1), we will search for the optimal path (path) to (I, J). In this case, the paths that move from the state of i-1 to the state of i-1 are limited to three as shown in the figure. Often. Note that the matching window is for preventing extreme time distortion, and the matching window satisfies the above three conditions (a) to (c) regarding the time-sequentialization. Now, for each i of i = 1, 2 ... I, optimal control is performed as to which state j should be moved to next, and the
Considering the case where the evaluation function of the formula is minimized, the initial condition is g (1,1) = d (1,1) …… (12) The recurrence formula is The distance between the patterns is D (A, B) = g (I, J) (14), and the calculation of the above equation (13) is performed using the grid points (i,
It will be done in the increasing direction of j). That is, g (i, j) is the minimum sum of distances from the (1,1) point to the (i, j) point.
1) g (i−1, j), g (i−1, j−1), g (i−1, j) already obtained for j, (j−1), and (j−2) in stage
-2), g (i, j) in the state j of the i-th stage is obtained.

FIG. 6 is a block diagram of a processor that executes the above-mentioned DP matching processing. In the figure, 11 is A memory, 12 is B memory,
13 is a d (i, j) calculator, 14 is a g (i, j) calculator, and 15 is a G
(J) memory, 16 is a control unit, and a (i, j) calculation unit 13 is ai
And the distance between vectors of bi is calculated, and the shortest distance g (i, j) to reach (i, j) is calculated by the g (i, j) calculation unit 14, and these are processed in parallel. g (i, j); j = G when calculating I to J
(J) The memory 15 contains g (i-1, j); j = 1 to J. Also, for min, the smaller _one of g ₁ and g ₂ is detected, and the smaller one is put into g.

When the DP matching method is used, at least I × J × N (where N is the number of registered words) if no matching window is provided, as is clear from item (1) of equation (13). Requires calculation of times.

In order to reduce the distance calculation amount by the DP method, a split method that takes an onomatopoeic unit has been proposed. In this split method, the distance calculation for each frame of the input speech is limited to a predetermined number (K). This is to reduce the amount of calculation of distance by making it possible to search only for the matrix in the case of DP matching, by storing only in the form of a matrix by going only to the onomatopoeia (chord book). Only the word standard pattern is subjected to vector quantization by this split method, and vector quantization is not applied to the input voice. Thus, in this split method,
A distance matrix between the analysis frame of the input voice and the pseudophony (vector) stored in advance is created. In this distance matrix, the horizontal axis is the frame number of the input voice,
The vertical axis is an onomatopoeic (vector) number, and DP matching between the standard pattern stored as a vector number series and the input voice is performed by referring to this distance matrix.

FIG. 7 is a block diagram showing an example of a recognition system based on the split method, in which 20 is an input unit, 21 is an analysis unit,
22 is a vector distance table, 23 is an onomatopoeia standard pattern (also called codebook), 24 is a word dictionary storage unit, and 25 is
The DP matching section 26 is a word identification section.

The input speech 20 is spectrally analyzed by the analysis unit 21, and the distance to the onomatopoeia standard pattern 23 is calculated for each frame to create the distance table 22. The input speech frame and the word dictionary 24 are matched by the DP matching 25, and the word having the minimum distance pattern is output by the word identifying unit 26 as a recognition result. With this split method, the number of calculation of the inter-vector distance becomes I × K, which is greatly reduced as compared with the conventional method (I × J × K) where vector quantization is not performed.

The present invention aims at dividing a standard pattern and an input pattern vector by dividing a standard pattern and an input pattern vector in a voice recognition method and a device thereof by a split method for vector quantization of a feature vector, that is, a 16-dimensional vector consisting of 16 elements By dividing into an 8-dimensional vector consisting of 8 elements and an 8-dimensional vector consisting of the latter half of 8 elements, the amount of calculation required for pattern matching can be further reduced as compared with the split method, and the recognition speed can be improved. It was made for the purpose of achieving it.

In order to achieve the above-mentioned object, the present invention (1) divides an input voice pattern vector in a voice recognition method by a split method in which a feature vector is vector-quantized, and divides the input voice pattern vector with the divided input voice pattern vector. The distance between the pseudophony standard pattern is calculated, an inter-vector distance table is created for each, and DP matching between the standard pattern and the input voice pattern is performed in divided vector units, or (2) the feature vector is set as a vector. In a voice recognition device using the splitting method for quantizing, an inter-vector distance created by calculating a distance between a dividing unit that divides an input voice pattern vector and a pseudophony standard pattern for each divided vector of the divided input voice patterns. A word consisting of a table and a vector number sequence of segmented onomatopoeic standard patterns It is characterized by having a dictionary and a plurality of DP matching units that refer to the distance tables of the plurality of inter-vector distance tables and perform matching with the word dictionary in units of divided vectors. The configuration of the present invention will be described below based on an embodiment.

FIG. 8 is a block diagram for explaining one embodiment of the present invention, in which the number of vector divisions is set to 2. In the figure, 23a is one of the two pseudophonic standard patterns, 23b.
Is the other onomatopoeia standard pattern, 22a is the standard pattern 2
An inter-vector distance table corresponding to 3a, 22b an inter-vector distance table corresponding to the standard pattern 23b, and 24 a word dictionary storage unit composed of the vector number sequence of the onomatopoeic standard pattern of 23a and 23b divided into two. , 25a, 25b are DP matching parts which refer to the distance tables of 22a, 22b, respectively. The input voice 20 is spectrum-analyzed by the analysis unit 21, and each input frame vector is divided into two to calculate the distances from the standard patterns 23a and 23b.
The distance tables 22a and 22b are created respectively. The input voice frame and the word dictionary 24 are matched in units of the division vectors, and after addition, the DP matching units 25a and 25b perform matching. The word having the minimum distance pattern is output by the word identifying unit 26 as a recognition result.

Effect From the above description, according to the present invention, by performing vector division, the size of the onomatopoeia standard pattern can be reduced, and therefore the amount of calculation of the distance between vectors is further reduced as compared with the conventional split method. Therefore, the recognition speed can be improved. That is, if the number of onomatopoeia patterns by the split method is k, for example, if the number of vector divisions is 2, the onomatopoeia pattern is the best case. become. Also, since the number of dimensions of the divided vector becomes 1/2, the calculation amount of distance is become. For example, if k = 256, the calculation amount becomes 1/16. In this way, the amount of calculation of the inter-vector distance can be reduced.

[Brief description of drawings]

FIG. 1 is a basic configuration diagram of a voice recognition device, FIG. 2 is a diagram showing an example of voice analysis, FIG. 3 is a mapping model for temporal emmetropization, and FIG. Figure 5, No.
FIG. 6 is a grid-like plan view for time-sequentialization, FIG. 6 is a block diagram of a processor for performing DP matching processing, FIG. 7 is a block diagram for explaining an example of the split method, and FIG. FIG. 1 is a block diagram for explaining an embodiment of a voice recognition device according to the present invention. 20 …… input part, 21 …… analyzing part, 22,22a, 22b …… distance table between vectors, 23,23a, 23b …… pseudophonic standard pattern,
24 …… Word dictionary storage section, 25,25a, 25b …… DP matching section, 26 …… Word identification section.

Claims (2)

[Claims] 1. A speech recognition method using a split method for vector-quantizing a feature vector, dividing an input speech pattern vector, and calculating a distance between the divided input speech pattern vector and the divided pseudophony standard pattern. , A voice recognition method using vector division quantization, characterized in that a distance table between vectors is created and DP matching between a standard pattern and an input voice pattern is performed for each division vector. 2. A voice recognition device using a split method for vector-quantizing a feature vector, wherein a distance between a division unit that divides an input voice pattern vector and a pseudophony standard pattern is divided for each of the divided input voice pattern division vectors. An inter-vector distance table created by calculation, a word dictionary consisting of vector number sequences of divided onomatopoeia standard patterns, a distance table of the plurality of inter-vector distance tables is cited, and matching with the word dictionary is performed. A voice recognition device using vector division quantization, comprising: a plurality of DP matching units that perform division vector units.