JPH0449718B2 - - Google Patents

Description

[Detailed description of the invention]

INDUSTRIAL APPLICATION FIELD 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 The general configuration of a speech recognition device using pattern matching is as follows. A feature extraction means converts the input speech signal into a system of feature vectors by filter unbanking, frequency analysis LPU analysis, etc., and a feature vector sequence that is uttered in advance and extracted by the feature extraction means is used to recognize all words. A standard pattern storage means for registering as a standard pattern, an input pattern uttered for recognition and extracted by the feature extraction means, all standard patterns stored in the standard pattern storage means, and a feature vector. A pattern comparison means that calculates the similarity or distance as a series, and a determination means that determines and outputs the 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. Become. 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, the time axis of both It is necessary to expand and contract the pattern lengths of the two to make them the same and then compare them. At this time, since the utterance time length does not change 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 in terms of distance below). Devices that use dynamic programming are generally used to perform this kind of matching efficiently (hereinafter referred to as DP).
(referred to as matching). The DP matching method can be explained using a grid graph. Figure 1 is a grid graph, and the horizontal axis is the i coordinate corresponding to the input pattern T = a ₁ a ₂ ...a _I ,
The vertical axis represents the j coordinate corresponding to the standard pattern R ⁿ =b ⁿ ₁ , b ⁿ ₂ . . . b ⁿ _Jo . Matching the input pattern T and the standard pattern by non-linearly expanding and contracting the time axis means that on this lattice graph, path 1 indicating the correspondence between the feature vectors of both patterns is determined by some price standard, and this The purpose is to evaluate the distance between both patterns with respect to the route. When determining this route, limiting conditions are set in consideration of the nature of the voice. FIG. 2a shows an example of restrictive conditions for route selection. That is, in this example, the path to point (i, j) is from point (i-2, j-1) to point (i-1,
path 2 passing through point j), path 3 coming from point (i-1, j-1), or path 4 passing from point (i-1, j-1) to point (i, j-1). This means that only Kakashi can be taken. At this time, if the condition is that the starting ends of the input pattern and the standard pattern must correspond, the matching path is limited to the shaded area in FIG. This restriction is made to prevent extreme correspondences from occurring due to the fact that no matter how much the time axis expands or contracts, the same word cannot be expanded or contracted so drastically. If the distance between the vectors a _i and b ⁿ _j is d ⁿ (i, j), then the distance along the path between the input pattern T and the standard pattern R ⁿ is d ⁿ along that path.
It is defined as the weighted average of (i,j). A, b, c, d, and e on the route in FIG. 2 are the loads when the corresponding route is selected. In order for DP matching to be applicable, this load should be determined in such a way that no matter what path is selected on the lattice graph under the above-mentioned limiting conditions, the sum of the loads along that path will be constant. a=c=e=2, b=
If d=1, the sum of these loads is I+J ⁿ , a=b
If =c=1 and d=e=0.5, the sum of these loads will be J ⁿ and will be constant regardless of how the route is chosen.
Both of these are commonly used. Furthermore, by relating this load to j under the condition that the sum of the loads is constant, it is also 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 pattern T and the standard pattern R ⁿ is
Under the limiting conditions, the distance between the vectors d ⁿ (i,
j) is defined as the minimum value of the weighted average of j). 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. g = (i, j) = ming ⁿ (i-2, j-1) + ad ⁿ (i-
1, j) + bd ⁿ (i, j) g ⁿ (i-1, j-1) + cd ⁿ (i, j) g ⁿ (i-1, j-1) + cd ⁿ (i, j) g ⁿ ( i-1, j-2)+ed ⁿ (i, j-1)+dd ⁿ (i,
j)...(1) (Initial condition g ⁿ (1, 1) = d ⁿ (1, 1), D(T,
R) = g ⁿ (I, J ⁿ )/(sum of loads) where g ⁿ (i, j) is the load of the vector distance d ⁿ (i, j) from the starting point to the point (i, j) The minimum value of the sum, D(T, R ⁿ ) is the distance between the input pattern T and the standard pattern R ⁿ . ) Various other conditions can be considered for route selection. FIGS. 2b to 2j, etc. are other examples. In addition to this, various other modifications can be considered. The recurrence formula is rewritten to correspond to these route selection conditions. As mentioned above, as a function of the load with respect to j,
In order to give importance to a part of the matching results on the matching route, the weights on the route may be set as shown in FIG. 3, for example. In the case of a in the same figure, the sum on the path from the matching starting point to the ending point is I+ _Jo 〓 ^j=1 W ⁿ j,
In the case of b, _Jo 〓 ^j=1 W ⁿ j. That is, the weight sum along the path a depends on both the input pattern length and the standard pattern length, and b depends only on the standard pattern length. Here, according to the above explanation, _Jo 〓 ^j=1 W ⁿ j is constant regardless of how the route is selected. At this time, for example, the recurrence formula for the above equation (1) for calculating the cumulative distance for a is as follows in equation (1): a=1+1/2W ⁿ j, b=1+1/2 W ⁿ j, c=1 +W ⁿ j, d=1/2W ⁿ j and e=1+1/2W ⁿ j . The distance between the input pattern T and the standard pattern R ⁿ is becomes. n^ argmin ⁿ [D(T, R ⁿ )] is determined, and the word corresponding to the standard pattern R ⁿ is taken as the recognition result. The notation argmin ⁿ [f(x)] is f
It means x that minimizes (x). 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 W ⁿ j of the frame j that abuts the consonant part of the standard pattern. W ⁿ j
can be determined for each frame, so it is possible to finely assign weights that are most suitable for each standard pattern. If _Jo 〓 ^j=1 W ⁿ j is kept constant regardless of n, equation (2) can also be changed to D(T, R ⁿ )=g(I, J ⁿ ). 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 best matches the word as a whole, and for matching that emphasizes consonants, 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 recognition by performing matching for multiple types of weighting and comprehensively judging each matching result. Standard patterns and input patterns when paths are determined independently and recognized from the matching results along each path, or when a path is determined using a certain weighting method and the weighting method is varied along the determined path. In addition, a comprehensive evaluation method based on multiple matching results uses the weighted average of the results obtained by the various weighting methods as the final distance. Select several recognition result candidates using a method that uses the minimum distance as the recognition result or a certain weighting method, and then select each candidate using other weight dimensions. The distance between vectors d ⁿ (i,
j) only needs to be calculated once for each grid point,
Furthermore, 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 is an input terminal for an audio signal, and 6 is a feature extractor 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 inter-vector distance calculation unit, and in the i-th input frame, n=1,
2, . . . , N; j=1, 2, . . . J ⁿ , the intervector distance d ⁿ (i, j) is determined. d ⁿ (i,
j) can most easily be city distance. That is, a _i = (a _i1 , a _i2 , ..., a _in ), b ⁿ _j
= (b ⁿ _j1 , b ⁿ _j2 , ..., b ⁿ _jn ), it can be done as d ⁿ (i, j) = _n 〓 ^k=1 | a _ik − ^{b n} _jk |. 9 is an inter-vector distance storage unit which stores the inter-vector distances d ⁿ (i, j) calculated by the inter-vector distance calculation unit 8 as n=1, 2, . . . , N;
j, 1, 2, . . . J ⁿ is also memorized until it is no longer needed. That is, when selecting the route restriction conditions as shown in FIG. 3, the distance between vectors for two frames is stored. The inter-vector distance storage unit 9 is
It is composed of two storage areas, VDM1 and VDM2, and VDM1 is the distance between vectors of the current frame i,
VDM2 stores the distance between vectors of the previous frame, and when one input frame is updated, the contents of VDM1 are transferred to VDM2, and the new distance between vectors is stored in VDM1. 10 is a weighting coefficient storage unit, and in this embodiment, a case where weighting is performed as shown in FIG. 3a will be explained. It is assumed that there are K types of weights that are matched with one standard pattern, and that the k-th weight corresponding to the j-th frame of the standard pattern of the n-th word is W ⁿ _k j. The weighting coefficient storage unit 10 has n=1, 2..., N; k=1, 2,
..., K: j=1, 2, ..., weighting coefficients W ⁿ _k j for J ⁿ are stored. 11 to 13 are cumulative distance calculation units using K types of weighting coefficients.
The cumulative distance calculation unit k (k=1, 2, ..., K) calculates the weighting coefficient W ⁿ _k from the starting point (1, 1) to (i, j).
Weighted sum of distances between vectors for j g ⁿ _k (i, j)
is calculated for n=1, 2, ..., N. That is, under the constraint conditions shown in FIG. 3a, the following recurrence formula is calculated. FIG. 5 shows the detailed configuration of the cumulative distance calculating section k. 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 inter-vector distance storage section 9. The value of the recurrence formula necessary for the calculation is memorized until the calculation of the formula is no longer necessary.
ADM ₁ is the cumulative distance g(i,
j) (n=1, 2,..., N; j=1, 2,...
J ⁿ ), and ADM2 stores the previous frame i-
Cumulative distance g(i-1,j)(n=1,
2,...,N;j=1,2,... ^Jn ). ADM1 when one input frame is updated
The contents of are transferred to ADM2, and 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 inter-vector distance calculation unit 8 and the cumulative distance calculation units 11 to 13 calculate the inter-vector distance and cumulative distance using n=1, 2, ..., N; j=
1, 2, ...J ⁿ is performed every frame,
As soon as the input is completed, n=1, 2,...
For N, the final cumulative distance g ⁿ _k (I, J ⁿ ) by the K product class weighting method is stored in the cumulative distance storage unit 104.
It will be stored in ADM1. In Fig. 5, 105 is g ⁿ _k (I, J ⁿ ) obtained in this way.
This is the cumulative distance normalization unit that normalizes the . terminal 10
When the end of audio input is notified to terminal 7 and the total number of frames is notified to terminal 108, the contents of ADM1 are normalized, and the normalized result is output from terminal 106 to determination section 14 at the next stage. The normalized result is becomes. In FIG. 4, reference numeral 16 denotes a voice section detecting section, which detects the start and end points of the input voice, and a known method can be applied based on the power of the input voice, etc. Reference numeral 17 is a frame number counting section which counts every frame after the start of the voice section and finally obtains the length of the voice section. 14 is a determination unit that obtains the final recognition result from the normalized cumulative distance D ⁿ _k (I, J ⁿ ) obtained in the above manner. The following methods can be used for determination. () Distance D ⁿ with respect to the standard pattern R ⁿ of the input signal
of That is, let D ⁿ be the weighted average of the second weighting coefficient a _k for k of D _{n k} (I, J ^{n )} , and use the word corresponding to the standard pattern R ⁿ for n that minimizes D ⁿ as the recognition result. shall be. () Standard pattern R ^q (1) that gives the l-th minimum value from the minimum value for the normalized cumulative distance D ⁿ _ko (I, J ⁿ ) using the weighting coefficient W ⁿ _ko (j) for the standard pattern R ⁿ , R ^q(2) ,..., R ^ql is calculated, and for the obtained standard pattern R ^q(1) , R ^q(2) ,..., R ^ql ,
D ⁿ _k with or without D ⁿ _ko (I, J ⁿ )
The word corresponding to the standard pattern with the minimum weighted average as explained in () of (I, J ⁿ ) is taken as the recognition result. In the above embodiment, the matching paths for the weights W ⁿ _k j calculated for the standard pattern R ⁿ were obtained independently with respect to k.
For the standard pattern n, the weight W ⁿ _ko j
It is also possible to calculate the cumulative distance by other weights W ⁿ _k j along the matching path calculated for . At this time, the recurrence formula of equation (3) is changed as follows.

[Table] At this time, the configuration shown in FIG.
1 to calculate the formula (4), and add a signal line 18 that notifies the other cumulative distance calculation units 12 to 13 of the route obtained there, and the cumulative distance calculation units 12 to 13
can be changed to calculate equation (5). The determination process is the same as in the case of (). 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, The meaning of the notation is to execute B while A is true. Step 200 is a part for initializing before the recurrence formula calculation unit 103 calculates the recurrence formula. In step 201, the input pattern is matched with all standard patterns, and the cumulative distance between the input pattern and each standard pattern is obtained for each weighting coefficient. Step 202
is the sum of weighting coefficients along the route, which is a part that normalizes the cumulative distance, and is responsible for the processing performed by the cumulative distance normalization unit 105 of the embodiment. Step 203 is a process performed for each input frame, and all standard patterns n = 1, 2, ..., N
, the distance between vectors in all frames of the standard pattern and the cumulative distance for each type of weighting coefficient are determined. This is a process performed by the inter-vector distance calculation unit 8 and the cumulative distance calculation units 11 to 13 of the embodiment. FIG. 7 shows a case where the matching path is fixed to the one determined with respect to a certain weighting coefficient for each standard pattern, and the cumulative distance is calculated by varying 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. Note that in this embodiment, the cumulative distance normalization unit 105
However, if the weighting coefficients are determined so that _Jo 〓 ^j=1 W ⁿ j=constant, 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. However, by considering the overall matching result, more accurate recognition results can be obtained, and
By completing calculations for each input frame for all recognized words and weighting coefficients, real-time processing is possible.

[Brief explanation of the drawing]

Figure 1 is a diagram explaining DP matching, Figures 2 a to j are diagrams showing examples of constraint conditions for matching paths in DP matching, and Figure 3 is a diagram for explaining matching paths for performing DP matching with local emphasis. FIG. 4 is a block diagram showing the configuration of a pattern comparison device according to an embodiment of the present invention, FIG. 5 is a block diagram showing details of the cumulative distance calculating section in the same embodiment, and FIG. The figure is a diagram showing the operation in the same embodiment as a program.
The figure is a program diagram showing the operation in another embodiment. 6...Feature extraction unit, 7...Standard pattern storage unit, 8...Vector distance calculation unit, 9...Vector distance storage unit, 10...Weighting coefficient storage unit, 11
~13... Cumulative distance calculation unit, 14... Judgment unit, 1
03... Recurrence formula calculation unit, 104... Cumulative distance storage unit.

Claims (1)

[Claims] 1. An input signal is converted into a series of feature vectors a ₁ , a ₂ , . . .
a feature extraction means for converting into an input pattern T consisting of a ₁ ... a _I , and a series of feature vectors b ⁿ ₁ , b ⁿ ₂ ... b ⁿ _j ...
...b ⁿ Standard pattern R ⁿ consisting of _Jo (where n=1,
2, ..., N), and ^a plurality of types of weighting coefficients W ⁿ ₁ (1), W ⁿ ₁ (2), ..., W ⁿ ₁ ( J ⁿ ); W ⁿ ₂ (1),
W ⁿ ₂ (2), ..., W ⁿ ₂ (J ⁿ ); ...; W ⁿ _k (1), ..., W ⁿ _k
(J ⁿ ), and a distance between vectors a _i and b _j in a lattice graph with the frame of the input pattern T on the horizontal axis and the frame of the standard pattern R ⁿ on the vertical axis. inter-vector distance calculation means for calculating d ⁿ (i, j); inter-vector distance storage means for storing this inter-vector distance; A recurrence formula calculation means for calculating a distance, 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 characterized by: 2. The inter-vector distance calculation means calculates the inter-vector distance between the input pattern and the standard pattern for each frame of the input pattern and for each frame of all standard patterns, and the recurrence formula calculation means calculates the inter-vector distance between the input pattern and the standard pattern. Claim 1, characterized in that the distance is determined for each frame of the input pattern and for each frame of all standard patterns.
The pattern comparison device described in Section 1.