JPH0577079B2 - - Google Patents

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 employ a recognition method based on a pattern matching method. 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 the two patterns and finding the distance between the patterns, a DP matching method is used that performs nonlinear matching using dynamic programming. Regarding DP matching, see "Two-stage DP matching for efficiently recognizing continuously uttered word sounds," Nikkei Electronics, November 7, 1983 issue, pages 171 to 208 (hereinafter referred to as Document 1).
is described in detail. According to this document, the distance D (A, B) between patterns A and B is defined as follows.

D (A, B) = ^min _j=j= (i) [ _i=1 〓 ^I d (i, j)] d (i, j) is the distance between vectors a _i and b _i .
The inter-pattern distance D can be determined, for example, by the following recurrence formula calculation.

g(i,j)=d(i,j)+ming(i
-1, j) g (i-1, j-1) g (i-1, j-2) (a) (b) (c) ...Equation (1) g (i, j) is determined by i, j This is the minimum cumulative value of the inter-vector distance d from point (1, 1) to (i, j) in the ij plane, and will be referred to as cumulative distance hereinafter. D(A, B) is obtained as g(I, J) obtained by performing this recurrence formula calculation up to i=1...I, j=1...J.

Here, consider the i,j plane as shown in FIG. As shown in Figure 2, the above recurrence formula is (i-
1, j), (i-1, j-1), (i-1, j-2)
Allowing three paths (a), (b), and (c) from lattice point (1, 1) to (i, j), vector distance d(i, j ) to find the path (hereinafter referred to as the optimal path) for (i, j,) that provides the minimum sum. The optimal path is calculated using equation (1).
Path information h(i, j) indicating which path was selected among (a), (b), and (c) is retained for all (i, j), and after D is determined, (I , J) by backtracking back to (1, 1). For information on how to find the optimal path using backtracking, see ``Application of dynamic programming in speech recognition'', bit, Vol. 15. No. 8.
This is explained in detail on pages 131 to 142.

There have been many improvements to the recognition algorithm using DP matching described above, one of which is the clockwise DP method described in Japanese Patent Application Laid-Open No. 58-98796. In this method, the time axis i of the input pattern is set as the outermost loop and g ⁿ (i, j)
This is a method that realizes matching processing that is synchronized with the input of feature vector a _i by calculating , and improves real-time performance. That is, in this method, at time i of the input pattern, g ⁿ (i,
Find j).

In addition, Japanese Patent Application No. 62-61732 and No. 62-219460 disclose methods in which the above-mentioned Crotwise DP method is speeded up by introducing the idea of pruning. These methods will be briefly explained below.

The method of patent application No. 62-61732 is to
In the DP method, the cumulative distance g ⁿ (i, j) at time i
For n, i, and j whose values are greater than or equal to a certain threshold value θ(i), the calculation of the recurrence formula after time i+1 is omitted. This means that g ⁿ (i, j) is large (n, i,
j) is considered to be unlikely to be on the optimal path, and calculation of the recurrence formula is omitted. As a result, the number of recurrence formula calculations to be performed is significantly reduced, and the recognition processing speed is increased. The following methods are available for setting θ(i).

(a) θ(i)=αi+β (b) θ(i)=g min(i)+α (α, β are constants) (a) Assuming that the optimal cumulative distance increases, θ(i) is linearly monotonous for i. What is defined as an increasing function, (b) is the cumulative distance g ⁿ (i, j) at each i, J=i,...Jn, n=
θ is determined by giving a margin of α to the minimum value g min(i) of 1, . . . , N. However, in this method, since the threshold parameters α and β for determining the threshold value were constant values, there were cases in which recognition errors and calculation amount due to inappropriate threshold value θ(i) were not reduced.

The method of patent application No. 1983-219460 is
This is a device that can deal with such problems in No. 61732, and has a function to learn 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. Then, the cumulative distance on the optimal path
g _ppt (i), i = 1, ..., 1, and set the threshold parameters α', β so that the threshold θ(i) satisfies g _ppt (i) < θ(i) for all i. Find ′. α and β used in the next recognition process are threshold parameters α′(x), β′(x), x=
Obtain from 1,...,X.

(Problems to be Solved by the Invention) In pruning in the conventional method, the parameters α and β are learned in the same way 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, inappropriate threshold parameters α,
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, there is a standard pattern storage section that stores the feature vector time series B ⁿ =b ⁿ ₁ ...b ⁿ _j ...b ⁿ _jo of the speech of each word n as a standard pattern, and a parameter that determines the pruning threshold. a threshold parameter storage unit that stores threshold parameters; an input pattern storage unit that sequentially reads the feature vector a _i of input speech at time i and holds it as a time series pattern A=a ₁ …a _i …a _I ; Characteristics of input audio at time i
The cumulative distance g ⁿ (i, j) of the distance d ⁿ (i, j) between a _i 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. The matching unit calculates the value of (n, j) that satisfies
a determination unit that outputs, as a recognition result, a word n that gives the minimum value of the cumulative distance g ⁿ (i, j) found in
a result confirmation unit that indicates whether the recognition result is correct; 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; In this case, 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 in the threshold parameter storage section. These are the parts of the threshold parameter determining unit that update the threshold value to make it higher.

(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,
It is desirable to set θ(i) so that it is not less than the cumulative distance of the correct word on the optimal path and 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 conventional recognition process 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, the standard pattern B ⁿ of the word n of the recognition result and the retained input pattern are determined as described in the above-mentioned document 1.
Perform DP matching. In matching, the information h(i, j) of the path selected in the recurrence formula calculation is
and cumulative distance g(i,j) are maintained for all (i,j), and the optimal path is obtained by backtracking. The cumulative distance g _ppt (i) on the optimal path is obtained as the cumulative distance g (i, j(i)) on the optimal path h _ppt (i) = j(1)...j(i)...j(I) .

The cumulative distance g _ppt (i) on the optimal path for each i obtained in this way is the time i of the immediately preceding input voice.
It can be considered as the optimal value of the pruning threshold θ(i) in . Therefore, using this information,
By correcting the current threshold parameters, 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 can be used to set a more appropriate threshold value for each utterance. It becomes possible to set θ.

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.
The threshold parameters α and β are updated so that a threshold higher than the current threshold θ is set.

(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 pattern storage unit 2 in FIG. 1, each time sequence data of the recognition target word n uttered in advance is stored as a standard pattern B _o ,
The threshold parameter storage unit 3 stores in advance thresholds α and β for determining the pruning threshold θ(i). The uttered input pattern A is analyzed in real time and is sequentially input to the matching unit 4 as time-series data of the feature vector a _i . Also, a _i is simultaneously stored sequentially in the input pattern storage section 1 and held until the next input. The matching unit 4 performs a recurrence formula calculation on n and j for each input a _i to obtain g ⁿ (i, j). For matching, the conventional method of black-wize DP is used.
A method that introduced the idea of pruning into the law (patent application 1983-
61732). To find the pruning threshold θ, here we use a linear monotonically increasing function (θ(i)=
αi+β) is used. Matching section 4
After calculating the cumulative distance at i, reads α and β in the threshold parameter storage unit 3 and calculates θ(i),
Find (n, j) that satisfies g ⁿ (i, j) < θ(i). When a _i+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 input pattern A and all standard patterns Bn.

In the determination unit 5, the input pattern A obtained in the matching unit 4 and all standard patterns Bn, n=1
. . . A standard pattern giving the minimum distance among the inter-pattern distances with N is output as a result. Subsequently, the user checks whether the recognition result is correct or not in the result confirmation section 8.
Input from The result confirmation unit 8 has a means for inputting whether the result is correct or not, and for example, a device including two keys corresponding to the correct or incorrect results 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 (patent application 1986-
61732), the cumulative distance on the optimal path in matching the input pattern and the standard pattern that gave the recognition result is used. Optimal path calculation section 6
When receiving input from the user that the result is correct, 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 standard pattern Bn that gave the result from the standard pattern storage unit 2. One-to-one matching is performed using the DP matching method. During matching, keep the cumulative distance g(i,j) at (i,j) and the path h(i,j) leading to (i,j) for all (i,j). . The optimal path is obtained from (I, J) by backtracking the path information h.
The cumulative distance on the optimal path obtained in this way is
Let g _ppt (i), i=1,...,I. After that, in the parameter determining section 7, the threshold parameters α and β are set.
Learn about. Since g _ppt is the cumulative distance on the optimal path, the pruning threshold during matching must always be equal to or greater than this value. Parameters α, β
The value of can be obtained, for example, as a coefficient of the least squares approximation straight line of g _ppt . FIG. 3 shows θ(i) based on g _ppt and the determined α and β. β in the figure 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 and β determined in this manner for X times of past utterances (X≧0). New threshold parameters are determined from these values and stored in the threshold parameter storage section 3. To find the threshold parameter,
The method of taking the maximum value of α and β is used. 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 the threshold value θ(i) is increased using them. Update as follows. The updated value of α is, for example, α _oew = k・α _pld ,
α _oew = α _pld + T Similarly, the updated value of B is β _oew =
Methods such as k·β _pld and β _oew =β _pld +T can be used. In the above example where a linear monotonically increasing function is used to find θ, by giving k a value of 1 or more or a positive value T, θ _oew >
It is possible to set α and β that result in θ _pld .

In the above embodiments, the method of determining the threshold value θ was explained using a linear monotonically increasing function. However, another method is to use the minimum value at i (θ(i)=g min(i )+α) will be explained. In this case, the parameter α can be learned as follows. First, the minimum value g min(i), i=1,..., I of the cumulative distance of each i used in the pruning process in the matching unit 4 during recognition is stored in the parameter determining unit 7 at each time i. put. Thereafter, the process is performed in the same manner as in the above embodiment, and after the optimal path calculating section 6 calculates g _ppt (i), i=1,..., I, the parameter determining section 7 calculates g _diff (i)
=g _ppt (i)−g min(i), i=1,...I is calculated and g _diff
Find the maximum value of. The parameter determination unit 7 stores the maximum value of g _diff determined in this manner for X times of past utterances (X≧0). α can be determined based on the average or maximum value of these values.

(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 one embodiment of the present invention;
FIG. 2 is a diagram for explaining the state of matching in the conventional method, and FIG. 3 is a diagram for explaining the processing performed by the parameter determining section in the embodiment of FIG. 1. 1... Input pattern storage section, 2... Standard pattern storage section, 3... Threshold parameter storage section, 4
...Matching section, 5... Judgment section, 6... Optimum path calculation section, 7... Parameter determination section, 8... Result confirmation section.

Claims (1)

[Claims] 1 Speech feature vector time series B ⁿ = each word n
b ⁿ _i …b ⁿ _j …b ⁿ _jo as a standard pattern; a threshold parameter storage unit that stores threshold parameters that are parameters for determining the pruning threshold; Sequentially read feature vector a _i of input voice of i, time series pattern A=
an input pattern storage section to hold as a ₁ ...a _i ...a _I ;
^{At each time i, the distance d n} _{( i} ^,
a matching unit that calculates the cumulative distance g ⁿ (i, j) of j) for the value of (n, j) that satisfies the pruning condition determined by the parameters of the threshold parameter storage unit; a determination unit that outputs as a recognition result the word n that gives the minimum value of the cumulative distance g ⁿ (I, J) found at time I; a result confirmation unit that determines whether the recognition result is correct; and if the result is correct. an optimal path calculation section that reads the input pattern A in the input pattern storage section and the standard pattern Bn of the recognition result and calculates the optimal path; and if the result is correct, the optimal path calculated by the optimal path calculation section a threshold parameter that updates a threshold parameter using the cumulative distance value of and, if the result is erroneous, updates the threshold parameter in the threshold parameter storage unit to increase the threshold value; A speech recognition device comprising: a determining section.