JPH0527120B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field The present invention relates to a speech pattern comparison method in a speech recognition device.

PRIOR ART In recent years, speech recognition devices have been put into practical use to realize man-machine dialogue, but an important part of speech recognition is a section that matches feature patterns registered in a dictionary with feature patterns of input speech. . Normally, there are two problems in matching this speech feature pattern: the first is that the length of the speech varies each time it occurs, and the second is that the formants differ depending on the speaker, resulting in frequency fluctuations. It is a certain thing. A pattern matching method using dynamic programming (DP) is known for the first variation absorption method. this
The DP matching method expands or contracts the time length of a pattern to maximize the similarity between two patterns to be compared in order to absorb fluctuations in the time direction. Since the method calculates similarity as an index, the amount of calculation is large, and a huge amount of DP calculation is required to absorb frequency fluctuations. Note that a method for absorbing these two frequency fluctuations has not yet been established.

Purpose The present invention has been made in view of the above-mentioned circumstances, and provides a voice pattern comparison method that can absorb both time fluctuations and frequency fluctuations and match patterns with high accuracy with a small amount of calculation. This is what we are trying to provide.

Configuration In order to achieve the above object, the present invention provides a speech pattern comparison method for comparing speech patterns consisting of a time series of feature amounts obtained by frequency analysis of one speech, in which the results of frequency analysis are binarized. A voice pattern formed by finding a feature distribution and thinning the feature distribution in the frequency axis direction is used as at least one of the voice patterns to be compared, and the time of both voice patterns is calculated using dynamic programming. In a voice pattern comparison method that performs comparison by aligning in the axial direction, or compares voice patterns consisting of time series of feature values obtained by frequency analysis of two voices, the band power obtained as a result of frequency analysis. Detecting the peak of , and using the speech pattern showing the peak in time series as at least one of the speech patterns to be compared, and performing the comparison by aligning both speech patterns in the time axis direction using dynamic programming. It is characterized by The following will explain based on examples.

First, a normal DP matching method will be explained with reference to FIGS. 1 and 2.

First, let's consider comparing the pattern in Figure A and the pattern in Figure B in Figure 1. In Figure 1, Patterns 1, 2, etc., in which audio patterns are sampled at regular intervals along the time axis, are shown. Ori,
Each of these patterns 1, 2, . . . is called a frame. This DP method first uses the first frame in figure a and the frame in figure b.
The difference between the two waveforms is determined by correlating the first frame in the figure, and the shaded area in FIG. 2 is determined. Similarly, the first frame in figure a and the second frame in figure b, the first frame in figure a and the third frame in figure b...the second frame in figure a, the first frame in figure b, and the second frame in figure a. 2 frame and the second frame in figure b, etc., and the frames, that is, the correspondence on the time axis, are established so that the difference in waveform is minimized. Therefore, it is effective for patterns with little frequency fluctuation, such as when a and b are voices generated by the same person, but even though the waveforms are similar, such as the waveforms of the broken line and solid line in Figure 2, , if there is a difference in frequency, these cannot be considered as the same waveform. This phenomenon occurs when the speakers of voices a and b are different, and this is affected by the formant differences between individuals.

The present invention has been made to solve the drawbacks of the above-mentioned DP matching method, and its operating principle will be explained with reference to FIG. 3. First, the pattern sampled as described above is sampled in the frequency axis direction and the time axis direction, and in order from the lowest frequency, i = 1, 2, ... I, and in the time axis direction, j = 1, 2, ... 3,...J, and the two patterns are represented by A(i,j) and B(i,j). The next pattern to be registered in the dictionary is the filter group i=
1, 2, . . . I, set a threshold value, binarize, and register (A(i, j), j _A =1, . . . J _A ). on the other hand,
The recognized speech is similarly binarized and then thinned to become B(i,j), j _B =1, 2, . . . J _B . here
The problem is the correspondence between j _A and j _B , and this correspondence is shown in Figure 3. In Figure 3, when A(i, j _A ) is represented on the i-j _A plane, it becomes as shown in diagram a, and B(i,
When j _B ) is expressed on the i-j _B plane, it becomes as shown in diagram b. However, among the binarized values of 0 and 1, the 1 part is indicated by diagonal lines. _{At this time} , A( _i ,
The similarity r(j _A , j _B ) between j _A ) and B(i, j _B ) is defined by the following formula, r(j _A , j _B )= _I 〓 ⁱ⁼¹ A(i, j _A )・B(i, _jB )...(1) When the cumulative total of similarities from _jA = 1, _jB = 1 to _jA , _{jB is expressed as R(jA, jB ) ,} R( j _A , j _B ) = r ( j _A , j _B ) + maxR ( j _A , j _B −1) R ( j _A −1, j _B −1) R ( j _A −1, j _B ) ……( 2) Determine (j _A , j _B ) such that (where
max indicates that the maximum value within { } is to be used). Note that although equation (1) above takes the product,
This may be a theoretical calculation, or it may be possible to extract "1" by changing i of B(i, j _B ) and perform calculations only on that part. Also, the result of equation (2) is the number of frames I + J
It is also possible to normalize with It is also assumed that the starting end and ending end of each pattern are associated with each other.

FIG. 4 is a block diagram showing an embodiment of the present invention constructed according to the above operating principle.
is a microphone, 2 is a filter group, 3 is a voice section detection section, 4 is a binarization section, 5 is a switch, 6 is a dictionary section,
7 is a thinning section, 8 is a similarity calculation section, 9 is a j _A , j _B changing section, 10 is a similarity detection section, 11 is an R calculation section, 1
2 is a stepping part that advances j _A or j _B by one step, 13
14 is a maximum calculation unit for R, and 14 is a recognition result output unit. According to the present invention, since the pattern a has a width in the frequency axis direction and the width of the pattern b is narrow, the frequency varies depending on the speaker. Therefore, even if the pattern b fluctuates in the frequency axis direction, the fluctuation can be absorbed as long as it does not exceed the pattern width of a.

FIG. 5 is a diagram showing another embodiment of the present invention. In this embodiment, a peak detection section 15 is provided before the dictionary section 6, and the peak detection section 15 detects the frequency of the characteristic pattern of the voice. Detect the peak of , and register that pattern in the dictionary section. On the other hand, when a pattern to be matched is input, it is set to 0,
It is binarized to 1 (the part that becomes 1 at this time is called the feature distribution part), and the similarity between this and the dictionary pattern is maximized over time using dynamic programming according to equation (2) above. This is a method that performs expansion and contraction for comparison.

Effects As is clear from the above description, according to the present invention, it is possible to provide a highly accurate voice pattern comparison method that can absorb both time fluctuations and frequency fluctuations with a small amount of calculation.

[Brief explanation of the drawing]

Figures 1 and 2 are diagrams for explaining the DP matching method, Figure 3 is a diagram for explaining the principle of the present invention, and Figures 4 and 5 are examples of the present invention, respectively. FIG. 2 is a block diagram for explaining. 1...Microphone, 2...Filter group, 3...Speech section detection section, 4: Binarization section, 5...Switch,
6... Dictionary section, 7... Thinning section, 8... Similarity calculation section, 9... j _A , j _B changing section, 10... Maximum similarity calculation section, 11... R calculation section, 12... ...j _A (j _B ) Step unit, 13...R maximum calculation unit, 14...Result output unit, 15...Peak detection unit.

Claims (1)

[Claims] 1. In a speech pattern comparison method for comparing speech patterns consisting of a time series of feature quantities obtained by frequency analysis of speech, the result of frequency analysis is binarized to obtain a feature quantity distribution part, The audio pattern formed by thinning the feature distribution part in the frequency axis direction is used as at least one of the audio patterns to be compared, and both audio patterns are matched in the time axis direction using dynamic programming and compared. A voice pattern comparison method characterized by performing the following. 2. In a speech pattern comparison method that compares speech patterns consisting of a time series of feature quantities obtained by frequency analysis of speech, a peak of band power obtained as a result of frequency analysis is detected and the peak is shown in a time series. A voice pattern comparison method characterized in that a voice pattern is used as at least one of the voice patterns to be compared, and the comparison is performed by aligning both voice patterns in the time axis direction by dynamic programming.