JPH01112298A - Voice recognition equipment - Google Patents

Abstract

PURPOSE: To efficiently execute free pruning correspondingly to a speaker or an environment change and to execute voice recognition processing at a high speed by learning a threshold by using a voice spoken in the past and information indicating the validity/invalidity of a recognized result. CONSTITUTION: A parameter for finding out an accumulated distance from past vocalization on an optimum path and finding out a suitable threshold from the found value is learned. The learning using the accumulated distance on the optimum path is executed only when the recognized result is correct, and if the recognized result is erroneous, the continuation of errors can be prevented by increasing the tree pruning threshold. Namely the validity/invalidity of the recognized result is inputted from a result checking part 8, and when the recognized result is inputted as an error, threshold parameters α, β set up at preset are read out from a threshold parameter storing part 3 and updated so that the threshold is increased. Consequently an optimum threshold for tree printing can be learned, a threshold suitable for a speaker or an environmental change can be set up, a recognition speed can be increased, and a recognition rate can be improved.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to an improvement in a speech recognition device that recognizes uttered speech at high speed.

(Prior Art) Speech recognition is an important technology for realizing excellent man-machine interfaces, and speech recognition devices are already being used in various fields. Most of the current devices are
A recognition method based on the pattern matching method is adopted.

In this method, uttered words to be recognized are stored in advance as standard patterns, and the input utterance pattern (hereinafter referred to as input pattern) is compared with the stored standard pattern. The recognition result is the word name of the standard pattern. At this time, as a method for associating the time axes of two patterns and finding the distance between the patterns, D
P matching method is used. Regarding Dr matching, see “Two-stage DP matching for efficient recognition of continuously uttered word sounds,” Nikkakei Electronics, 1983.
Pages 171 to 208 of the November 7 issue (hereinafter referred to as Reference 1)
) is described in detail. According to this literature,
7: The distance wED(A, B) between Tough A + 8 is defined as follows.

D (A, B) =, m, in[Σd(i, j
)]1 $1(i) 221 d(i, j) is the distance between vectors al and . The inter-pattern distance can be determined, for example, by the following recurrence formula calculation.

...(1) 2g (i, j) is a point in the ij plane spanned by 1+J (
1, 1) to (i, j), and is hereinafter referred to as the cumulative distance. D(A,B) calculates this recurrence formula as i=1...I,
It is obtained as g(I, J) obtained by going up to j=1...J.

Here, consider the i, j plane as shown in FIG. The above recurrence formula, as shown in Figure 2, (i-1, j), (i
-1, j-1), (i-1, j-2) to (i, j
) A three-tree path (a).

Allowing (b) and (c), from lattice point (1,1) to (I,
This is to find the path (hereinafter referred to as the optimal path) for (i, j) that provides the minimum -3= sum of the inter-vector distance @d(i, j) to J). The optimal path is determined by (a) and (b) when calculating equation (1).

Path information h (f
, j) for all (i, j),
After D is determined, the path retained from (I, J) is (
This can be obtained by backtracking all the way back to 1,1). For information on how to find the optimal path by backtracking, see "Applications of Niockel Dynamic Programming for Speech Recognition J.
bit, Vol, 15° No. 8, page 131 1
Details are given on page 42.

There have been many improvements to the recognition algorithm using DP matching described above, one of which is JP-A-58-9
There is a clockwise DP method described in No. 8796. This method calculates gn(i, j) using the time axis i of the input pattern as the outermost loop.
This method realizes matching processing that is synchronized with the input of feature vector ai and improves real-time performance. That is, in this method, at time i of the input pattern, all words n
and gfi (
i.

Find j).

In addition, a patent application filed in 1982-6 that sped up the processing by introducing the idea of pruning to the clockwise Dr method described above.
No. 1732 and Japanese Patent Application No. 62-219460. These methods will be briefly explained below.

The method of patent application No. 62-61732 is Clockwise D.
In the P method, for n, i, j whose cumulative distance gn(i, j) at time i is greater than or equal to a certain threshold θ(i), the recurrence formula calculation after time i+1 is omitted. be. This means that g″(i, j) is large (n.

i, j) are considered to be unlikely to be on the optimal path, and the recurrence formula calculation is omitted. As a result, the number of recurrence formula calculations to be performed is significantly reduced, and the recognition processing speed is increased. There are the following methods for setting θ(i).

(a) θ(i)=αi+β (b) θ(i)=gmin(i)+α (α, β are constants) (a) Assuming that the optimal cumulative distance increases, θ(i) is a linear monotonous increase of i. What is defined as a function, (b) is the cumulative distance g″(i, j) at each i.

J=i, ---Jn, n=1,-, θ is determined by giving a margin of α to the minimum value g m1n(i) of N. However, in this method, since the threshold parameters α and β for determining the threshold value are constant values, recognition errors and calculation amount due to inappropriate threshold value θ(i) may not be reduced.

The method of Japanese Patent Application No. 62-219460 is
This is a device that can deal with such problems in No. 1732, and has a function of learning threshold parameters from past utterances. The threshold parameters are learned using the following steps. First, after outputting the recognition results, matching is performed between the input pattern and the standard pattern that gave the recognition results, and an optimal path is determined by backtracking. Next, the cumulative distance gap+(i) on the optimal path, i=1. ..., 1 is calculated, and the threshold value θ(i) is g-p+':' for all i.
The threshold parameter α is set so that x) < θ(f).
Find ゛, β′. α and β used for the next recognition process are
Threshold parameter α for utterances X times in the past one collection
'(X), β'(x), x=1. ..., find it from X.

(Problems to be Solved by the Invention) In pruning in the conventional method, the parameters α and β are calculated using the same method regardless of whether the recognition result is correct or erroneous.

However, in the case of misrecognition, when determining the optimal path, the uttered word is matched with a different standard pattern, so the correct path with the correct word cannot be determined. Therefore, when misrecognition occurs, the threshold parameter α is inappropriate.

β may be learned, leading to further misrecognition.

An object of the present invention is to provide a speech recognition device that eliminates the above-mentioned problems and can always determine an appropriate pruning threshold θ.

(Means for Solving the Problems) The speech recognition device according to the present invention requires the following parts. In other words, the speech feature vector time series B'' of each word n =
b",...bai...b"Tm as a standard pattern, and a threshold parameter storage section that stores a threshold parameter that is a parameter for determining a pruning threshold. , the feature vector ai of the input voice at time i is read sequentially and the time series pattern A=al・
. . . ai . . . an input pattern storage section that holds them as ao, the characteristics ai of the input voice at each time i, and the standard pattern b'' of the standard pattern storage section.

Cumulative distance gn(s l 3 ) of "distance" (11j) to
a matching unit that calculates gn(r, J) a determination unit that outputs the word n that gives the minimum value as a recognition result; a result confirmation unit that determines whether the recognition result is correct; and a standard for the input pattern A and the recognition result in the input pattern storage unit when the result is correct. The optimum path calculation section reads the pattern Bn and calculates the optimum path, and if the result is correct, the optimum path calculation section calculates the result by -8.
= The threshold parameter is updated using the value of the cumulative distance on the optimal path determined, and if the result is incorrect, the threshold parameter in the threshold parameter storage is updated to increase the threshold value. These are the various parts of the threshold parameter determining section.

(Operation) The speech recognition device according to the present invention learns a threshold value using speech uttered in the past and information on whether the recognition results are correct or incorrect, and pruns in response to changes in the speaker or the environment. It is characterized by efficient and high-speed recognition processing.

As described above, pruning is performed using the threshold value θ(i) at each time i of the input pattern. Therefore, θ(i) is not less than the cumulative distance of the correct word on the optimal path, and
It is desirable to set it not too high. The feature of the present invention is to obtain cumulative distances on the optimal path for past utterances, and to learn parameters for obtaining an appropriate threshold value θ(i) from these values. Furthermore, learning using the cumulative distance on the optimal path is performed only when the recognition result is correct, and if the recognition result is an error, the pruning threshold is raised to prevent consecutive errors. Features. The operating principle will be explained below.

After the recognition process using the conventional method is performed and the results are output, the user inputs whether the results are correct or not in accordance with a prompt prompting the user to input whether the recognition results are correct or not. If the recognition result is correct, DP matching described in the above-mentioned document 1 is performed between the recognition result standard pattern B of word n and the retained basic pattern.In the matching, the recurrence formula The information h(z, j) of the path selected in the calculation and the cumulative distance g(i, j) are held for all (i, j), and the optimal path is obtained by backtracking.

The cumulative distance g-pt(t) on the optimal path is the optimal path b
-p+ (s) = J (1)...j(f)...
It is obtained as the cumulative distance g(i, j(i)) at j(I).

The cumulative distance g−p+ (1) on the optimal path at each i obtained in this way can be considered as the optimal value of the pruning threshold θ(i) at time i of the immediately preceding input voice. . Therefore, by correcting the current threshold parameters using this information, a more appropriate threshold can be set for the next recognition process. Furthermore, if the speaker or the environment changes, an inappropriate threshold θ may be set using the same parameters as before, but even in such cases, the above principle allows a more appropriate θ to be set for each utterance. can be set.

The above is the process when the recognition result is correct, but in the case of incorrect recognition, it is assumed that there is a strong possibility that the optimal path for the correct word has been pruned, and if the recognition result is higher than the current threshold θ. Threshold parameter α such that the threshold value is set
, β is updated.

(Example) Examples of the present invention will be described in detail below with reference to the drawings. FIG. 1 is a block diagram showing one embodiment of the present invention.

In the standard Bakun storage unit 2 in FIG. Threshold values α and β for determining the mowing threshold θ(i) are stored in advance. The uttered basic cover turn A is analyzed in real time and the feature vector ai
The data is sequentially input to the matching unit 4 as time-series data.

Furthermore, ai is simultaneously stored sequentially in the input pattern storage section 1 and held until the next input. Matching section 4
Now, for each input ai, a recurrence formula calculation is performed on n and j to obtain gn(i+j). For matching, a method (Japanese Patent Application No. 62-61732) is used in which the idea of pruning is introduced into the conventional clockwise DP method. As a method for determining the pruning threshold θ, here, a −th vehicle adjustment increasing function (θ(i)=αi + β) is used. After calculating the cumulative distance at i, the matching unit 4 reads α and β from the threshold parameter storage unit 3, calculates θ(i), and satisfies g'(i, j)<θ(i) (n , j). When al+1 is input, recurrence formula calculation is performed for (n, j) that satisfies the pruning criterion determined at time i. The matching section 4 performs the processing up to time I while performing pruning in this manner, and calculates inter-pattern distances between the base cover pattern A and all standard patterns Bn.

In determination group 5, the standard pattern that gives the minimum distance among the inter-pattern distances between the base cover pattern A obtained by the matching section 4 and all the standard patterns Bn, n=1...N is output as a result. Subsequently, the user inputs whether the recognition result is correct or not from the result confirmation section 8. The result confirmation unit 8 has a means for inputting whether the result is correct or not.
A device consisting of two keys can be used. In this way, when the result is correct or incorrect, the threshold parameter learning process is performed depending on whether the result is correct or incorrect.

First, a case where the result is correct will be explained. In this case, the conventional method (Japanese Patent Application No. 62-61732)
Similarly, the cumulative distance on the optimal path in matching the base cover pattern and the standard pattern that gave the recognition result is used. When the optimal path calculation unit 6 receives input from the user that the result is correct, it reads the input pattern A from the input pattern storage unit 1 and the standard pattern Bn that gave the result from the standard pattern storage unit 2, and reads the input pattern A from the input pattern storage unit 1 and the standard pattern Bn giving the result from the standard pattern storage unit 2, One-to-one matching is performed using the P matching method as shown in . During matching, the cumulative distance g at (i, j)
A path h(i, j) leading to (i, j) and (i, j) is maintained for all (i, j). The optimal path is obtained from (1, J) by backtracking the path information. The cumulative distance on the optimal path obtained in this way is g, pt(i), i = 1,...
・、■. Thereafter, the parameter determination unit 7 determines the threshold parameter α.

Perform learning of β. Otsu. , 1 is the cumulative distance on the optimal path, so the pruning threshold during matching must always be greater than or equal to this value. Parameter α.

The value of β can be obtained, for example, as a coefficient of the least squares approximation straight line of go. FIG. 3 shows θ(i) based on g*p+ and the determined α and β. In the figure, β is larger than the value determined as the coefficient of the least squares approximation straight line by an amount of margin Δβ.

The parameter determination unit 7 stores α obtained in this way.
, β have been uttered X times (X≧0) in the past are stored.

New threshold parameters are determined from these values and stored in the threshold parameter storage section 3. The threshold parameter is determined by taking the maximum value of α and β values X times. As a method for determining the threshold parameters α and β, in addition to a method of taking the maximum value of the past X times, a method of taking the average value of the past X times, etc. can be used.

Next, a case where the recognition result is erroneous recognition will be explained. If the user inputs an error in the recognition result, the currently set threshold parameters α and β are read from the threshold parameter storage unit 3 and set to the threshold value θ (
i) is updated so that it increases. For example, the updated value of α is
α,,=k・α,14. α. ,,=α−+a+T and B
Similarly, the updated value of β is also β. ,,−k・β. 4. β0. ,=
β. , +T, etc. can be used. In the above example in which a linear monotonically increasing function is used to obtain θ, θ1. ．． 〉θ. It is possible to set α and β to be +4.

In the above embodiments, the method of determining the threshold value θ was explained using a linear monotonically increasing function, but as another method,
Based on the minimum value at i (θ(i)−gmin
(i)+α) will be explained. In this case, the parameter α can be learned as follows. First, the minimum value g m1n (1) + i of the cumulative distance of each i used in the pruning process in the matching unit 4 during recognition.
=1 +...

(2) is stored in the parameter determination unit 7 at each time i. Thereafter, processing is performed in the same manner as in the above embodiment, and the optimal path calculation unit 6 calculates g-pl(t), i = 1,...
・,■ After determining g□
rr(1)=g. pl(i)-gmin(i),
Calculate s-1,..., ■ and get ga+. Find the maximum value of.

The parameter determination unit 7 stores g obtained in this way.
The maximum value of □ is stored for X times of past utterances (X≧0). α can be determined based on the average of these values or the maximum value.

(Effects of the Invention) In the high-speed speech recognition device according to the present invention described above, even if erroneous recognition occurs, it is possible to learn an optimal threshold value for pruning without leading to frequent erroneous recognition. Therefore, it is possible to set a threshold value that adapts to changes in the speaker, environment, etc. This solves the conventional problems caused by inappropriate threshold values, increases recognition speed, and improves recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram for explaining the state of matching in the conventional method, and FIG. 3 shows the processing performed by the parameter determining section in the embodiment of FIG. It is a figure for explaining. DESCRIPTION OF SYMBOLS 1... Input pattern storage part, 2... Standard pattern storage part, 3... Threshold parameter storage part, 4... Matching part, 5... Judgment part, 6... Optimal path calculation part, 7...parameter determination part, 8...result confirmation part.

Claims (1)

[Claims] Voice feature vector time series of each word n = b^n,
...b^n_j...b^n_j_m as a standard pattern; a threshold parameter storage section that stores a threshold parameter that is a parameter for determining a pruning threshold; Sequentially read feature vector a_i of input voice at time i, time series pattern A=a
_1...a_i...a_I and the input pattern storage section that stores the characteristics a_i of the input voice at each time i.
The cumulative distance g^n(i, j) of the distance d^n(i, j) between the standard pattern b^n_j in the standard pattern storage section and the standard pattern b^n_j in the standard pattern storage section is determined by the pruning condition determined by the parameters in the threshold parameter storage section. A matching unit that calculates the value of (n, j) that satisfies a determination unit that outputs, a result confirmation unit that determines whether the recognition result is correct, and an optimal path calculation unit that reads the input pattern A in the input pattern storage unit and the standard pattern Bn of the recognition result and calculates the optimal path if the result is correct. If the result is correct, the threshold parameter is updated using the value of the cumulative distance on the optimal path obtained by the optimal path calculation section, and if the result is incorrect, the threshold parameter is updated. A speech recognition device comprising: a threshold parameter determination unit that updates a threshold parameter in a parameter storage unit so as to increase the threshold value.