JPS59173883A - Pattern comparator - Google Patents

Abstract

PURPOSE:To obtain the result of recognition with high accuracy by having matching among plural types of weighting and giving the overall judgment to each matching. CONSTITUTION:The sound signal is supplied to a feature extracting part 6 and converted into a series of feature vector. The inter-vector calculation and the cumulative distance calculation are carried out every frame through an inter-vector distance calculating part 8 and cumulative distance calculating parts 11-13. The final cumulative distance obtained by a weighting method is stored as soon as the input is over, and the contents of said distance are normalized and delivered to a deciding part 14. The part 14 uses the word corresponding to a standard pattern with which the weighting average is minimum as the result of recognition.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a pattern comparison device, and particularly to a pattern comparison device applicable to speech recognition.

Conventional configuration and its problems A general configuration of a speech recognition device using pattern matching is as follows.

A feature extraction means converts an input speech signal into a thread of feature vectors using a filter bank, frequency analysis LPC analysis, etc., and recognizes a series of feature vectors that are uttered in advance and extracted by the feature extraction means to create a standard pattern for all words. standard pattern storage means to register as a standard pattern storage means, input cover patterns uttered to be recognized and extracted by the feature extraction means, all standard patterns stored in the standard pattern storage means and feature vectors as a series. It consists of a pattern comparison means for calculating similarity or distance, and a determination means for determining and outputting a word corresponding to the standard pattern with the highest degree of similarity (smallest distance) as a recognition result as a result of the pattern comparison.

At this time, even if the same speaker utters the same word, the duration of the utterance differs each time, so when comparing the standard pattern and the input pattern using the pattern comparison means,
It is necessary to expand and contract the time axes of both, align the pattern lengths of both, and compare. At this time, since the change in utterance time length does not occur uniformly in each part of the uttered word, it is necessary to expand and contract each part non-uniformly. The best results have been obtained when the expansion/contraction is performed in such a way that the similarity between the two patterns to be compared is maximized (the distance is minimized; this will be explained below in terms of distance). A device that uses dynamic programming to perform such matching efficiently is common (hereinafter, this matching will be referred to as DP matching).

The DP matching method can be explained using a grid graph. Figure 1 is a grid graph, and the horizontal axis is the input cover turn T.
The i coordinate corresponding to = a1 a2...aI, the vertical axis is the standard pattern R=bb...b? n
It represents the j coordinate corresponding to 2. Matching the input cover turn T and the standard pattern by non-linearly expanding and contracting the time axis means that on this lattice graph, path 1 that indicates the correspondence between the feature vectors of both patterns is determined by some ridge standard, and the The purpose is to evaluate the distance between both patterns.

When determining this route, limiting conditions are set in consideration of the nature of the voice. The second figure) is an example of a restriction condition for route selection. That is, in this example, the route to point (111) is route 2 passing from point (i-2, 1-1) to point (l-1, ]).
or path 3 coming from point (i-1,)-1) or point (1-
1, j-2) to point (i, j-1). At this time, if the condition is that the starting and ending ends of the input turn and the standard cut turn must correspond, the matching path will be limited to the shaded area in FIG. This restriction is
This is to prevent extreme correspondences from occurring due to the fact that no matter how much the time axis expands or contracts, it is unlikely that the same word will expand or contract so drastically.

Let the distance between the vectors for a and b be d''(i, i)1
], then the distance along the path `` between the input cover turn T and the no turn of the standard pattern H is d'' (i,
i) is defined as the weighted average of a on the route in Figure 2
, b, c, d, and e are the loads when the corresponding route is selected. In order for DP matching to be applicable, the load should be determined in such a way that no matter what path is selected on the grid graph under the above-mentioned limiting conditions, the sum of the loads along that path will be constant. a
=c=e-rei b=d=1, then the sum of this load is I+
J, a=b=c=1.

If d=e−〇, 6, the sum of these loads becomes ①, which is constant regardless of how the route is selected. Both of these are commonly used. Furthermore, by making this load a function of j under the condition that the sum of the loads is constant, it is possible to perform operations such as increasing the load on a portion of the route that is to be matched with greater emphasis.

The distance between the input cover turn T and the standard pattern Rn is defined as the minimum value of the weighted average of the inter-vector distances dn(i, i) under the limiting conditions.

That is, by solving the following recurrence formula, the minimum value of the weighted average and the path that provides the minimum value can be determined.

Margin below In addition to congestion, various other conditions can be considered for route selection.

FIGS. 2(b) to 2(1) etc. are other examples. In addition to this, various other modifications can be considered. According to these route selection conditions, the recurrence formula is rewritten into a corresponding one.

As described above, in order to place emphasis on a part of the matching results on the matching path by setting the load as a function regarding j, the weights on the path may be set as shown in FIG. 3, for example. The same figure (a
), the matching starting point or sum depends on both the input pattern length and the standard pattern length in (a), and depends only on the standard pattern length in (■).・It is constant regardless of how it is spread.

At this time, for example, the recurrence formula for the above formula (1) for calculating the cumulative distance for (a) is n=1 in formula (1).
+-! -Wn(i), b=1+LWn(S), c=
1 +W"(j)2 d=-!-w"(1), e = 1 + -! −
W''(j)2 The distance between the input cover turn T and the standard pattern Rn is as follows.

n=argmin(D(T,R''))] and take the word corresponding to the standard pattern Rn as the recognition result.a
The notation rqmin [f (X) ] is f (X
) means X that minimizes. At this time, if it is desired to perform matching that emphasizes the consonant part of the word, it is sufficient to increase the weight Wn(j) of the frame j that abuts the consonant part of the standard pattern. Since W”(j) can be determined for each frame, if we use each standard butter, we can obtain the formula (2
) can also be set as D(T, R'')=g(1, ■'').

As described above, weighted DP matching has superior features compared to normal DP matching, which evaluates all frames equally. However, it has the following problems.

That is, for example, the matching path when emphasis is placed on consonant parts and the matching path when all frames are evaluated equally are generally different, and it is expected that the recognition results will be different in both cases. Evaluating all frames equally means finding the distance that matches the word as a whole best, while matching that emphasizes the consonant part means finding the distance that matches best locally. This means that the standard pattern that is closest in distance as a whole and the standard pattern that is locally closest are generally different. Therefore, simply introducing a weighting method does not necessarily lead to an improvement in the recognition rate.

OBJECTS OF THE INVENTION It is an object of the present invention to provide a pattern comparison device that can solve the above drawbacks and obtain highly accurate recognition results.

Structure of the Invention The pattern comparison device of the present invention is configured to perform matching for multiple types of weighting and perform recognition by comprehensively judging each matching result. are determined independently and recognized from the matching results along each route, or a route for a particular weighting method is determined, and along this determined route, the standard pattern and input pattern are calculated when the weighting method is variously changed. Can be recognized from distance,
In addition, the method of comprehensive evaluation from multiple matching results is also
A weighted average of the results obtained by the various weighting methods mentioned above is used as the final distance, and the recognition result is the lowest /J% of that distance. It is possible to select candidate results and use other weighting methods to obtain the final recognition results for each of the obtained candidates. By calculating the cumulative distance obtained, the inter-vector distance "('11)" only needs to be calculated once for each grid point, and the recognition result can be obtained at the same time as the input is completed.

DESCRIPTION OF EMBODIMENTS FIG. 4 shows a first embodiment of the present invention. In the figure, 5
6 is an input terminal for an audio signal, and 6 is a feature extraction unit that converts the input audio signal into a series of feature vectors. Reference numeral 7 denotes a standard pattern storage unit, which stores in advance a series of feature vectors for each recognized word obtained by the feature extraction unit 6 as a standard pattern prior to recognition. 8 is an intervector distance calculation unit, and n=1.2 in the first frame of input.
゜・・・・・・, N; j=1.2.・・・・・・Jn
, the inter-vector distance dn(i,j) is determined. ”('+3) can most easily be the urban distance. In other words, a, -(a, , a, , -
---・-, a, ), b", -111121m
) (b71. b72...."" b7m) and Rutki, d" (l l 1) = f l a 1 k
b 7 kl = 1. Reference numeral 9 denotes an inter-vector distance storage unit which stores the inter-vector distance “(l) calculated by the inter-vector distance calculation unit 8.
IJ) with n-=1.2. --------, N; j
=1.2.・”・■0 is also memorized until it is no longer needed.

That is, when selecting the path restriction conditions as shown in FIG. 3, the distance between vectors for two frames is stored.

The inter-vector distance storage unit 9 is composed of two storage areas, V, DMl and VDM2, where VDMl stores the inter-vector distance of the current frame l, V DM2 stores the inter-vector distance of the previous frame V-4, and stores the inter-vector distance of the input frame When is updated by one, the contents of vDMl are moved to VDM2, and a new inter-vector distance is stored in VDMl.

10 is a weighting coefficient storage unit, which in this embodiment is shown in FIG.
, ) will be explained below. 1
The type of weighting to be matched with the two standard patterns is set to type , and the weight corresponding to the first frame of the standard number turn of the nth word is set to W'' (D. The weight coefficient storage unit 10 stores n =1.2.....

N = 1.2. -----, K; j = 1. ．． 2
．． -=., J" is stored. This is a cumulative distance calculating section using weighting coefficients for each of the types 11 to 13. Cumulative distance calculating section k (k=1. 2.・・・・・・
.

K) is the weighting coefficient W from the starting point (1, 1) to (1+j)
The weighted sum of vector distances qH(i
+j) with n=1. Thank you... Calculate N. That is, under the constraint conditions shown in FIG. 3(a), the following recurrence formula is calculated.

Margin below − ■ shield Man child 1♂ FIG. 6 shows the detailed configuration of the cumulative distance calculation section.

The recurrence formula calculation unit 103 is a part that calculates equation (3). 101 and 102 are terminals into which the contents of the inter-vector distance storage section 9 are input; 100 is a terminal into which the contents of the weighting coefficient storage section 10 are input; and 104 is a cumulative distance storage section which is used to input the contents of the vector distance storage section 9. ADMl stores the value of the recurrence formula necessary for calculation until it is no longer necessary to calculate the recurrence formula, and ADMl calculates the cumulative distance q (i
, i) (n=1.2.--..., N:, j=1.
2. ...-, I"), and AD'M
2 is the cumulative distance q ('-'
+]) ("='+ 2+ """+ N; j=1.2.
・・・・・・ xn> is memorized. When one input frame is updated, the contents of ADMl are moved to ADM2,
The newly calculated cumulative distance is stored in ADM1.

The recurrence formula calculation unit 103 calculates the recurrence formula of equation (3) from the cumulative distances stored in ADM1 and ADM2, and the inter-vector distances stored in VDM1 and VDM2.

As described above, the vector distance meter thin part 8. The calculations of inter-vector distances and cumulative distances in the cumulative distance calculation units 11 to 13 are based on n=1.2. −−−−−−, N: = 1.2°・・
... is performed every frame for In, and at the same time as the input is completed, n=1.2.・・・・・・、
For N, the final cumulative distance using K types of weighting methods,
qSho (I, 1”) KaM product distance storage unit 1o 4
It will be stored in oA DM 1. In Figure 5, 1o6 is the qtsuji obtained in this way (","
) is the cumulative distance normalization unit. When terminal 10 is notified that the audio input has ended and terminal 10B is notified that all frames have been deleted, the contents of 8DM1 are normalized and the terminal 1
The normalized result is output from 06 to the next stage determination unit 14. The normalized result is shown in Fig. 4. Reference numeral 16 denotes a voice section detection unit that detects the start and end points of the input voice, and a known method can be applied from the power i of the input voice. Reference numeral 17 denotes a frame number counting section which counts every frame after the start of the voice section and finally obtains the voice section length I.

14 is the normalized cumulative distance pH<
This is a determination unit that obtains the final recognition result from The following methods can be used for determination.

1) The standard pattern for n that minimizes Dn, where the distance of the input signal to the standard pattern Rn is set to pn, that is, the weighted average of the second weighting coefficient α for k of D■ (I, J'') Let the word corresponding to Rn be the recognition result.

11) Weighting coefficient “kl j) for standard pattern Hn”
For the normalized cumulative distance DHn (I, J"), the standard pattern that gives the fourth minimum value from the minimum value
Rq(1), R(1(2),...R
q(A) and the obtained standard patterns Rq(1), R
q(2), , , , , , For Bq(7), D
D: (I
, r''), the word corresponding to the standard pattern with the minimum weighted average as explained in 1) is taken as the recognition result.

In the above embodiment, the matching path for the weight Wk(i) calculated for the standard pattern H is k
However, for the standard pattern n, the cumulative distance by other weights Wh(+) is calculated along the matching path calculated for the weights WH,(j). You can also do that. At this time, the recurrence formula of equation (3) is changed as follows.

Margin below ○ O■ Mi°− ? 111---------------1111---
−−−−)■ 10++ At this time, the configuration shown in FIG.
4), and a signal line 18 is added to notify the other cumulative distance calculation units 12 to 13 of the route obtained there, so that the cumulative distance calculation units 12 to 13 calculate equation (5). You can change it to . The determination process is the same as in case 1).

FIG. 6 shows the above operation expressed by a program, in which cumulative distances are calculated independently for each type of weighting coefficient. You can also follow this program when implementing it with software.

In addition, in the program description, DOWHILE! ■] The meaning of the notation NDDO is to execute B while A is true.

Step 200 is a part for initializing the recurrence formula calculation section 103 prior to the recurrence formula calculation section. Step 20
Reference numeral 1 denotes a part that performs matching between the incoming cover turn and all standard patterns, and the cumulative distance between the incoming cover turn and each standard pattern is obtained for each weighting coefficient.

Step 202 normalizes the cumulative distance using the sum of weighting coefficients along the route, and corresponds to the process performed by the cumulative distance normalization unit 106 of the embodiment.

Step 203 is a process performed for each input frame.
All standard patterns n=1.2. ..., N, the cumulative distance for each type of inter-vector distance 2 weighting coefficients in all frames of the standard pattern is determined. Inter-vector distance calculation unit 8゜cumulative distance storage unit 1'1 of the above embodiment
This is the process performed in steps 1 to 13.

FIG. 7 shows a case where the matching path is fixed to the one determined for a certain weighting coefficient for each standard pattern, and the cumulative distance is calculated by varying only the weighting coefficient. In this example, steps numbered the same as in FIG. 6 represent similar functions as in FIG. The only difference is the content of step 203, which has already been explained.

If the coefficients are determined, there is no need for this normalization.

Effects of the Invention The pattern comparison device of the present invention is configured to introduce various weighting coefficients into the matching path and comprehensively judge the results. In addition, by considering the overall matching result, more accurate recognition results can be obtained, and by completing calculations for each input frame for all recognition word 2 weight coefficients, p1 real time can be obtained. It is now possible to process this.

[Brief explanation of drawings]

Figure 1 is a diagram explaining DP matching, Figure 2 (a) -
(i) is a diagram showing an example of constraint conditions for matching paths in DP matching, FIG. 3 is a diagram showing an example of weighting on matching paths for performing DP matching with local emphasis, and FIG. 4 is a diagram showing the present invention. FIG. 5 is a block diagram showing the configuration of a pattern comparison device according to an embodiment of the present invention; FIG.
This figure is a diagram showing the operation in the same embodiment as a program, and FIG. 7 is a diagram showing the operation in another embodiment as a program. 6...Feature extraction unit, 7...Standard pattern storage unit, 8...Vector distance calculation unit, 9.
... Inter-vector distance storage unit, 1Q... Weighting coefficient storage unit, 11-13... Cumulative distance calculation unit, 14... Judgment unit, 103... . . . Recurrence formula calculation unit, 104 . . . Cumulative distance storage unit. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 1
figure

Claims (2)

[Claims] (1) The input signal is a series of feature vectors a1. a2.・・・
・Feature extraction means for converting into an input pattern T consisting of al...al, a series of feature vectors, bg...
・Standard pattern storage means for storing a standard pattern FF' consisting of bn...b■o (where n=1, courtesy..., N), and a plurality of types of standard pattern Hn associated with the standard pattern Hn. Weighting factor W (1) 2 (2), ...
Lie (J”): w; (1) + WS (2) , ,,,,
,,, W”, (J”) ; ,,,,,, :
Wk(1). . . . , in a weighting coefficient storage means for storing Wζ(J”) and a lattice graph in which the horizontal axis is the frame of the input cover pattern T and the vertical axis is the frame of the standard pattern H, ai and The distance between the vectors d (11
1) intervector distance calculation means for calculating the intervector distance, intervector distance storage means for storing the intervector distance, and recurrence for calculating a plurality of types of cumulative distances for the input cover turn and the standard pattern using the plurality of types of weighting coefficients. A formula calculation means, a cumulative distance storage means for storing a plurality of types of cumulative distances obtained by the recurrence formula calculation means, and a determination means for obtaining a final recognition result from the plurality of cumulative distances. A pattern comparison device. (2) The inter-vector distance calculation means calculates the inter-vector distance between the input cover pattern and the standard pattern for each frame of the input cover pattern and for each frame of all standard patterns, and the recurrence formula calculation means calculates the distance between the vectors of the input cover pattern and the standard pattern. 2. The pattern comparison device according to claim 1, wherein the cumulative distance is determined for each frame of an input pattern and for each frame of all standard patterns.