JPS60111300A - Continuous word voice recogntion method - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION (5) Technical Field of the Invention The present invention relates to a continuous word speech recognition method, and particularly to a method for recognizing continuous word speech from both ends using a level building method.

This invention relates to a continuous word speech weaving method that makes it possible to reduce the total matching distance tti.

0 Prior Art and Problems Recognizing consecutive words uttered continuously is extremely difficult compared to recognizing isolated words, but in recent years research in this area has become active.・Programming) Methods for continuous word speech recognition using matching methods have been established.

Furthermore, the same results as the two-stage DP matching method are obtained,
Level building with a small amount of calculation
(Building) method has been proposed. In the level building method (hereinafter referred to as the LB method), as is well known, matching scanning is performed for each nine-digit number (level), and in the second level, the input end is fixed.

In the first and second and subsequent levels, the word with the minimum matching distance and its cumulative distance are determined for each input frame end point within a certain range, and the start end is free, and each frame within a limited range is Among the ending word strings, the word string with the minimum cumulative distance is determined, and the results are accumulated for each level.

By the way, in general, in the matching space between the standard pattern and the input pattern, a limit is placed on expansion and contraction for one hour, and a slope limit is provided. FIG. 1 shows a diagram conceptually representing the calculation aspect of the conventional LB method. for example,
When the slope limit is set to 1/2 to 2, expansion and contraction in one hour is allowed from 1/2 to 2 times. For simplification of explanation.

It is assumed that all referenced standard patterns have the same length. In this case, the part indicated by the toll line in Figure 1 is
This is an area that requires calculation. Namely.

From the starting point of the input pattern, the range of slope 1/2 and slope 2 is limited, while 9. From the end point of the input pattern corresponding to each level, the slope is limited to 1/2 to 2 in the opposite direction. If the amount of calculation can be further reduced under the same slope restriction as the To case, it is thought that it will be possible to speed up continuous word speech recognition.

0 Object and Structure of the Invention The present invention aims to improve the above points, and in a continuous word speech recognition method using the LB method, matching is also performed in the reverse direction from the end of the input pattern, and the matching distance is The purpose is to reduce the overall amount of calculation. Therefore, the continuous word speech recognition method of the present invention is equipped with a dictionary storing standard patterns of word sounds, and represents acoustic features obtained by analyzing unknown input speech in which a plurality of words are uttered consecutively. In a continuous word speech recognition method that performs recognition by matching an input cover turn with the above-mentioned standard pattern using a level-building method, recognition candidate words are recognized using a level-building method from the beginning of the input cover turn under a predetermined slope restriction. By calculating the digit-by-digit cumulative similarity between the standard pattern of
Find a standard pattern sequence with a large degree of similarity, and use the word sequence corresponding to the obtained standard pattern sequence as the recognition result "C".
It is characterized by output. Hereinafter, a description will be given according to an example while referring to the drawings.

Embodiment of the invention Fig. 2 is an explanatory diagram of matching distance calculation according to an embodiment of the invention, Fig. 3 is a system diagram of an embodiment of the invention, and Figs. FIGS. 6 and 7, which are explanatory diagrams of matching distance calculation according to another embodiment, are diagrams for explaining the slope restriction related to the present invention.

In the case of the example shown in FIG. 2, the maximum number of digits (recognized number) allowed for the input of continuous word speech recognition is now RmαX, Ilmax''=6.And.

For example, calculations are performed in the same manner as in the conventional LB method up to the (Rmax -1) digit. Incidentally, the slope limit is from 1/2 slope to
It is assumed that the slope is 2 and the lengths of the referenced standard patterns are all equal. here.

The total amount of matching distance is the same as before.

The range of calculation is from the end point of the input pattern to the frame t' of the input pattern corresponding to the '(Rmax 1 )th digit, taking into account the slope restriction of 2 in the opposite direction. That is, distance calculation is performed for the range of the person shown in FIG.

Parameters used in the following explanation are determined as follows.

I: Input cover turn length.

Dn(z): Minimum cumulative distance of R word strings ending in t frames of input.

Nn('): Last word name of the word string corresponding to DR(ri).

BR(t)2NR(starting point position -1 (back pointer).

D' (=): Minimum accumulated distance in the backward direction when one reference standard pattern terminates in the input t frame.

N'(su): Word name corresponding to D'(i) Ro: Number of recognized words - (R): Frame of R word string of recognition result.

As above <-, as in the conventional LII method, °C91<R
R in the range of <Rmax-1 and 1 <: i <T,
1 - The minimum cumulative distance DR(t) of the word string is determined and stored, and the word name N, (i) 1 Back pointer BR (
') should also be noted.

On the other hand, calculate one digit in the reverse time direction from the end of the input, that is, calculate the range B in Figure 2.

Although it is similar to the calculation of the first Lebel, it is performed according to the reverse time series, and the backward minimum cumulative distance D'(j) and the simple IflJN'(return.(Rmax -1)) up to the eclipse at that time are The reason why the range of the !/= product distance calculation in the positive direction is set to the range A in FIG. As is clear from comparing the conventional example shown in FIG. 1 with the case of the example shown in FIG.

From forward and backward LB method operations.

After all, Drt (I) is DR(1) = m, i (DR-1(Q +D' (
i + 1) ).

The number of recognized words R6 is.

It is given by Ro=αr,m,in(DR(I)). Here, αrgrrLin() means calculating the variable iR that gives the minimum value in [].

What are the recognition results?

g (Rol) = ar, qmtn (DR1(S+D'(i+1)) is the initial value, 2×R(R6-1
within the range of.

-(R-1)=B R(g(R)) It can be obtained from studying. Thus, the recognition result.

Nf (g(])), Nt(g(2)), −, Nn
, -1(t (R8-1) ), N'(!(Ro-1
)) is determined.

Next, a system according to an embodiment of the present invention will be described with reference to FIG.

The audio signal input from the microphone 1 is acoustically analyzed in the parameter extraction unit 2, and is extracted 9, for example, a moment M corresponding to the first formant frequency, a moment M corresponding to the second formant frequency, a low frequency power, Characteristics of the input audio, such as high frequency power, are extracted. In the single@registration mode, the input pattern consisting of the time-series information of the line pattern is sent to the word dictionary 4 via the switching section 3, and is registered as a standard pattern corresponding to each word.

In the recognition mode, the input cover turn may contain a plurality of consecutive words, and the input cover turn is sent to the collation determination section 5 in order to recognize unknown input speech. The input cover turn is 1 line and 6 (-stored.@1
Stage = ttt sound 1 (9 is for the J-th level, similar to the conventional LB method, the input end is fixed, the termination point is free, and the input pattern stored in the input buffer 6 and each read out to the dictionary buffer 7) Performs matching with standard patterns corresponding to words.

The calculation range at this time is determined based on the slope limit given in advance by the slope limiter 8 (the slope limiter 8).

The iK stage calculation unit 10 is for the second level and subsequent levels.

Perform the same calculations as before, with free start and end points.

On the other hand, the backward calculation unit 11 performs 81 calculations in the reverse time direction from the end of the input pattern, similar to the calculation by the first stage calculation unit 9. The same applies to the slope restriction given by the slope restriction section 8. After all consecutive input cover turns have been input, the calculations by the first stage calculation section 9 and the second stage calculation section 10 and the calculations by the backward calculation section 11 are performed as follows.
Can be processed in parallel. The final determination unit 12 determines, from the calculation result of the step subtraction unit 10 and the division result of the backward total n unit 11, that the sum of the minimum cumulative distances obtained by combining these is:
Select the smallest one. The recognition result output unit J3 outputs the determination result by the final determination unit J2.

Output as a recognized word string.

In the above embodiment, the number of digits in the reverse time direction is set to 1, but of course it may be 2 or more digits. Figure 4 shows
Figure 2 shows the corresponding creation JI mouse when the number of digits in the reverse time direction is 2, and Figure 5 shows the total when the number of digits in the reverse time direction is 3. FIG. 1 is a diagram showing the amount of 1′1.

As can be seen from FIGS. 4 and 5, in these cases as well, the calculation cost is reduced in the same way as in the case of FIG.

By the way, the slope restriction in the above embodiment was determined based on the fact that the lengths of the referenced i-prefecture patterns are equal; , it is not possible to easily limit the slope of the LB method. In order to deal with this problem, for example, it is recommended to do the following.

In advance, refer to (compress and expand the voting pattern to keep the length consistent.9) By doing this, even if a slope limit of 1/2 to 2 is given as described above, Therefore, variations between standard patterns are avoided.

Alternatively, it may be handled as follows. For example, from the maximum length Jmar and the 7i% minimum length Jmin of the reference standard pattern, as shown in FIG. Please keep this in mind. If the slope is limited as shown in FIG. 7 from the maximum values Amα, τ and the minimum value 1min, any path that is inappropriate as a static path for matching will be excluded from the calculation.

According to the present invention, as described in detail, continuous word sounds are muttered (1
In the process, by performing matching using the IJB method from both ends of the input cover turn, it is possible to reduce the calculation time. Example 1-1. In the example shown in FIG. 2, compared to the conventional example shown in FIG.

[Brief explanation of drawings]

Fig. 1 is a diagram conceptually representing the total l'F' by the conventional LB method, Fig. 2 is an explanatory diagram of calculating the mating distance according to an embodiment of the present invention, and Fig. 3 is a diagram showing the calculation of the mating distance according to an embodiment of the present invention. Kazumi 1/1fl
Example I system block diagram, FIGS. 4 and 5 are explanatory diagrams of matching distance calculation according to another embodiment of the present invention,
FIGS. 6 and 7 are diagrams for explaining the incorrect slope restriction of the present invention. In the figure, 2 is a parameter extraction unit, and 4 is a word dictionary. 5 represents a comparison determination section. ! Office NT Applicant Fujitsu Limited

Claims (3)

[Claims] (1) Equipped with a dictionary in which standard patterns of Sokekei speech are stored, input cover patterns representing acoustic features obtained by analyzing unknown input speech in which multiple words are uttered in succession, and the above-mentioned standard patterns. In a continuous word speech recognition method in which recognition is performed by comparing and recognizing words using the level building method C2, the standard pattern of the recognition candidate word and the corresponding input word are recognized using the level building method starting from the starting point with a predetermined slope restriction for the above-mentioned cover pattern. By not only requesting digit-by-digit similarity with cover turns, but also matching in the reverse time direction from the end using the level-building method. Continuous word speech B characterized in that a standard pattern sequence having the maximum similarity with the cough entry cover pattern is selected, and a first word sequence corresponding to the obtained cough standard pattern sequence is output as a recognized result. Weaving method. (2) The slope limit is determined to a predetermined value after compressing and expanding the standard pattern in advance to unify the length. Continuous word speech recognition method. (3) The slope limit is determined by determining the maximum and minimum lengths of the input partial patterns corresponding to one standard pattern from the maximum and minimum lengths of the foreboard pattern, The continuous word speech recognition method according to claim 1, characterized in that the maximum value and the minimum value are determined based on two factors.