JPH0577080B2 - - 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. This method is based on the time series B ⁿ of the feature vector of the uttered word n to be recognized.
=b ⁿ ₁ ...b ⁿ _j ...b ⁿ _J is held as a standard pattern in advance, and the input utterance pattern A = a ₁
...a _i ...a _I (hereinafter referred to as input pattern) is compared with the stored standard pattern, and the word name of the most similar standard pattern is taken as the recognition result. At this time, as a method for determining the inter-pattern distance by associating the time axes i and j of the standard pattern and the input pattern, a DP matching method is used which performs nonlinear correspondence using dynamic programming. DP matching uses recurrence formula calculation such as equation (1) to calculate the distance between vectors a _i and b _j d ⁿ (i,
This solves an optimization problem that minimizes the sum of j).

g ⁿ (i, j) = d ⁿ (i, j) + min P = 0, 1, 2 [g ⁿ (i-1, j-p)] (1) Here, g ⁿ (i, j) is (1,1) to (i,
j) is the cumulative value of d ⁿ (i, j). For details on 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). It has been described.

The recognition algorithm based on this DP matching principle is currently the most commonly used, but the problem is that it requires a large amount of calculation. A method for reducing the amount of calculation is the clockwise DP method described in Japanese Patent Application Laid-Open No. 58-98796. This method is (1)
Input pattern time axis i by recursive formula calculation such as Eq.
This improves real-time performance by processing in synchronization with the Processing synchronized with the time axis is realized by calculating the cumulative distance g ⁿ (i, j) for all words n and their time j at time i of the input pattern. Furthermore, as a method aimed at further speeding up the process, Japanese Patent Application No. 62-61732 describes a method in which the idea of pruning was introduced into the Crotwise DP method to speed up the processing. This method will be briefly explained below.

In this method, when proceeding with recurrence formula calculation, recurrence formula calculation is performed for (n, j) that satisfies a certain condition, and recurrence formula calculation is performed for (n, j) that does not satisfy a certain condition. The number of recurrence formula calculation circuits is reduced by incorporating a pruning process of omitting them. As a condition for pruning, for example, g ⁿ at time i
The minimum value gmin of (i, j) plus the margin α (beam width factor) is the threshold value θ at i.
There is a method used as (i). In this case, θ
Leave only (n, j) such that (i)>g ⁿ (i, j) as the subject of recurrence formula calculation, and other (n, j)
is excluded from the recurrence formula calculation.

When the input pattern a _i is input, recurrence formula calculation is performed by the clockwise method while performing pruning. Figure 2 shows the algorithm. In Figure 2, the obtained g ⁿ (i, j) and the threshold value θ
(i), and if g ⁿ (i, j) > θ(i), then j
=j+1 and omit the recurrence formula calculation. As described above, the processing is performed from i=1 to I-1 in synchronization with the input of the input pattern. When the last feature vector a _I of the input pattern is input, i=I
In the matching section, the process of equation (2) is performed, and the distance D(A,
B ⁿ ) is obtained.

D (A, B ⁿ ) = g ⁿ (I, J) = g ⁿ (J) + d ⁿ (I, J)
(2) However, g ⁿ (J) = ming _o (I-1, J) g ⁿ (I-1, J-1) g ⁿ (I-1, J-2) Next, the obtained D( A, B ⁿ ) are compared sequentially,
The minimum value is determined, and n corresponding to this value is output as a recognition result.

(Problems to be Solved by the Invention) In recognition processing that involves pruning in the conventional method, pruning at a certain time i in a word (n,
It is also possible that n in j) is of one type. In this case, even if the recurrence formula calculation is performed after i+1, words other than n will not be recognized as a result. However, in the conventional speech recognition apparatus, in order to obtain a recognition result, it is necessary to calculate the inter-pattern distance D (A, B ⁿ ) from the standard pattern for the pattern from the start to the end time of the input pattern. However, as described above, the recognition result is uniquely determined at the point when n becomes one type due to pruning. Therefore, in terms of obtaining recognition results,
Calculations from i+1 to I are redundant processing, which causes a reduction in recognition speed.

An object of the present invention is to provide a speech recognition device that can eliminate the above-mentioned wasteful calculations and obtain recognition results more quickly.

(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 speech feature vector time series B ⁿ = b ⁿ ₁ ...b ⁿ _j ...b ⁿ _J as a standard pattern for each word n, and a threshold storage section that stores the threshold value for pruning. and the feature vector a _i of the input speech at time i are sequentially read, and at each time i, the cumulative distance g of the distance d ⁿ (i, j) between the feature a _i of the input speech and the standard pattern in the standard pattern storage section is calculated. ⁿ
(i, j) is calculated for the value of (n, j) that satisfies the condition determined by the threshold value of the threshold storage section, and at that time, the value (n, j) that satisfies the condition is calculated.
If the value of n in j) is one type, a matching unit outputs the word n as a recognition result, a cumulative distance storage unit stores the cumulative distance obtained by this matching unit, and the matching unit These are each part of the determination unit that outputs the word n that gives the minimum value of the cumulative distance g ⁿ (I, J) determined at time I as a recognition result.

(Operation) When the speech recognition device according to the present invention proceeds with recurrence formula calculation, if there is only one type of n of (n, j) that is not pruned, the speech recognition device according to the present invention performs word n without having an input end of the input pattern. The feature is that it outputs as a recognition result. The fact that n in (n, j) that is not pruned is now one type means that words other than n will not become recognition results even if recurrence formula calculations are performed any further. Therefore, when n becomes one type, by outputting n as a recognition result, subsequent recurrence formula calculations can be omitted, and recognition processing can be speeded up.

A speech recognition device with the above functions has a function to determine whether there is one type of n in (n, j) that is not pruned, and if there is one type of n, it outputs n as a result at that point. , if there are two or more types, it can be realized by continuing the recurrence formula processing at i+1.

(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.

The standard pattern storage unit 1 in FIG. 1 stores recognition target words n (1≦n≦N) uttered in advance
Each time series data of is a standard pattern B _o = b ⁿ ₁ … b ⁿ _j …
_It is assumed that the threshold value θ(i ⁾ for pruning at time i is stored in the threshold storage unit 2 in advance. The uttered input pattern A is analyzed in real time and the feature vector a _i
The data is sequentially input to the matching unit 3 as time-series data.

In the matching section 3, for each input a _i ,
Perform recurrence formula calculation with pruning for n, j and calculate g ⁿ
Find (i, j). For this process, a method is used in which the idea of pruning is introduced into the conventional clockwise DP method (see the above-mentioned document, Japanese Patent Application No. 61732/1983). The matching unit 3 reads the threshold value θ(i) from the threshold storage unit 2 and calculates g ⁿ (i, j)<θ(i)
Recurrence formula calculation is performed for (n, j) that satisfies . As the recursion formula calculation process involving pruning in the matching section 3, the conventional process shown in FIG. 2 can be used. Matching section 3
In the process in step 1, the cumulative distance g ⁿ (i, j) found at time i is held in the cumulative distance storage section 4 . These cumulative distances are read out from the matching unit 3 and used when calculating the recurrence formula of i+1.
Here, in the matching section 3, in addition to the above processing, it is determined whether or not n of (n, j) is one type.
If it is a type, n is output as the recognition result. The algorithm of the matching section 3 including this process is shown in FIG. The area within the double frame is a process for determining whether n is one type. n' is a flag indicating whether or not there is a j that is not branched for the word n, and if there is such a j, n' becomes the value of the word number n. Also, nn
is the number of n pieces of (n, j) that are not pruned in i. After the processing of i is completed, if the determination processing of nn<2 is satisfied, n' is output as the recognition result.

This algorithm does not prune (n,
If there are always two or more types of n in j), the recurrence formula calculation up to time I-1 and the processing of equation (2) at I are performed as in the conventional method, and input pattern A and all standard patterns Bn and Find the distance between patterns. In the determination unit 5, the input pattern A obtained in the matching unit 4 and all standard patterns Bn, n=1...N
Outputs the standard pattern that gives the minimum distance among the inter-pattern distances as a result.

(Effects of the Invention) As described above, according to the present invention, when there is only one type of n in (n, j) that is not pruned, the amount of calculation is reduced by immediately outputting n as a result. , it becomes possible to realize a speech recognition device with increased recognition processing speed.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention;
FIG. 2 is a flowchart showing the processing algorithm in the matching section of the conventional system, and FIG. 3 is a flowchart showing the processing algorithm in the matching section of FIG. 1... Standard pattern storage section, 2... Threshold storage section, 3... Matching section, 4... Cumulative distance storage section, 5... Judgment section.

Claims (1)

[Claims] 1 Speech feature vector time series B ⁿ = of each word n
b ⁿ ₁ …b ⁿ _j …b ⁿ _J is stored as a standard pattern in a standard pattern storage unit, a threshold storage unit stores a pruning threshold value, and a feature vector a _i of input speech at time i is stored. Sequential reading, at each time i, the cumulative distance g ⁿ (i, j) of the distance d n (i, j) between the input voice feature a _i and the standard pattern in the standard pattern storage unit is calculated as the cumulative distance g ⁿ (i, j) in the threshold storage unit. Find the value of (n, j) that satisfies the condition determined by the threshold value,
At that time, if there is only one value of n in (n, j) that satisfies the condition, there is a matching section that outputs the word n as a recognition result, and a cumulative section that stores the cumulative distance obtained by this matching section. Speech recognition characterized by having a distance storage unit and a determination unit that outputs a word n that gives the minimum value of the cumulative distance g ⁿ (I, J) found at time I by the matching unit as a recognition result. Device.