JPH0361955B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (a) Technical Field of the Invention The present invention relates to a speech recognition method, and more particularly to an improvement in a method for matching patterns expressed in a time series of vectors representing features of speech.

(b) Prior art and problems In general, in order to recognize speech by pattern matching, standard patterns of speech are registered in memory as a word dictionary for each word, and input speech patterns and standard patterns are registered in memory. It is matched.

In this case, it is desirable that the input speech pattern and standard pattern be normalized on the time axis with respect to the pronunciation of the word as much as possible.

However, when generating one word, the duration length of the word sound and the length of the phoneme in the word sound are
Generally, it varies depending on the person pronouncing it and the situation when pronouncing it.

In this way, dynamic programming (DP matching) and discriminant function method are representative methods for pattern matching of two patterns of different lengths represented by a time series of vectors representing speech characteristics. Are known.

However, although dynamic programming is effective for normalizing nonlinear expansion and contraction of the time axis, it is difficult to deal with stochastic variations in parameters.

Further, although the discriminant function method is effective for stochastic variation of parameters, it has a drawback in that there is a problem in normalization in the time axis direction.

(c) Object of the Invention In order to eliminate the above-mentioned drawbacks, the object of the present invention is to provide a pattern matching device that combines the dynamic programming method and the discriminant function method and has the advantages of both methods.

(d) Configuration of the Invention The configuration of the present invention includes a means for matching an input time series pattern with a standard time series pattern by dynamic programming, and a means for matching an input time series pattern with a standard time series pattern based on the matching result of the matching means. A conversion means for converting the series to the same length as the standard time series pattern is provided, and weighting is performed to weight central elements of parameters between the new input time series pattern converted by the conversion means and the standard time series pattern. This method performs pattern matching by calculating distance using .

(e) Embodiment of the invention FIG. 1 is a block diagram illustrating an embodiment of the invention.

Divide the duration of the audio into segments of equal time units, and calculate the feature vector of the audio in all of the segments as vector a→
₁ , a→ ₂ , a→ ₃ , ..., a→ _I , and for the standard time series pattern, the vector b→ ₁ , b→ ₂ , b→ ₃ , ...,
When b → _J , memory 1 stores the input time series pattern A={a→ _i } _i=1~I , and memory 3 stores the standard time series pattern B={b→ _j } _j=1~J . Stored.

In the DP matching section 2, memories 1 and 3
The input time series pattern A and the standard time series pattern B read out are matched by dynamic programming.

For details on matching using dynamic programming, see "Journal of the Acoustical Society of Japan," vol. 27, No. 9, pp. 483-490, September 1971.
Generally speaking, the duration length of each of the two voice patterns A and B is divided into equal time units into times 1 to I and times 1 to J, and each is divided vertically on a plane. The axes and horizontal axes are scaled, and the feature vector of the voice at each time is determined.

That is, find vectors a→ ₁ , a→ ₂ , a→ ₃ , ..., a→ _I and vectors b→ ₁ , b→ ₂ , b→ ₃ , ..., b→ _J and calculate 2
The optimal route 10 is obtained so that the distance between the two voice patterns A and B is minimized, and the optimal route 1 at that time is
The matching distance is obtained as the cumulative distance along 0.

That is, first, let the distance g ₁ , ₁ = |a→ _i −b→ _i |, and g _i,j
is the corresponding vector of speech patterns A and B at points i and j at the minimum point from the three points g _i-1,j , g _i-1,j-1 , g _i,j-1 shown in Figure 2. It is obtained by adding the distance between |a→ _i −b→ _i |. By finding g _{i,j while successively increasing i, j} in this way, g _I,J is finally found as the optimal matching distance between speech patterns A and B, and this calculation process At each grid point of the matrix, by storing in P _i,j which of the previous three points the optimal path passed through, the optimal matching path can also be found at the same time.

Voice patterns A and B obtained in matrix 4
In order to obtain a new matching pattern C between input time-series patterns A using the optimal matching distance g _{I,J between} From the point (I, J) of matrix 4, j=J, J-1,...
..., 3, 2, 1 while tracing back the optimal path, and by setting the vector a _i corresponding to j, and when there are multiple corresponding vectors a _i , c _j is the average of them, A matching pattern C is obtained.

Here, M is a variable for averaging when there are multiple vectors a _i corresponding to j. i=I,j
=J, M=0 becomes the starting point c→ _J =a→ _I . optimal route
As shown in Figure 2, P _i,j is from the point (i-1, j),
+, if the path to point (i, j) is taken, then point (i
-1, j-1) to point (i, j), 0, from point (i, j-1) to point (i, j)
If P _i,j is +, then M=1 at the previous grid point. Also, P _i,j
When is 0 or -, M=0.

If we go back one step from the starting point (I, J), then P _i,j
If is in the 0 direction, i=i-1, j=j
−1, M=0, so c→ _J-1 = a→ _I-1 , −
If it is in the direction of , c→ _J-1 = a→ _I , and if it is in the + direction, i=i-1, M=M+1
Therefore, c→ _J = (a→ _I + a→ _I-1 )/2.

Furthermore, if P _i,j goes back one point in the + direction,
Since i=i-1 and M=M+1, c→ _J = {2
(a→ _I +a→ _I-1 /2+a→ _I-2 }/3=(a→ _I +a→
_I-2 )/3
This shows that three a→ _i are averaged.

In the converting unit 5, the matching pattern C={c→ _j } _{j=1 to J} obtained as above is stored in the memory 6. The new matching pattern C is converted into the standard time series pattern B in the calculating unit 7. _In order to weight _the distance calculation between
A distance D is calculated from the memory 8 that stores _J using vectors w→ ₁ , w→ ₂ , . . . , w→ _J representing weights, and the final matching distance is output.

For example, the distance D is However, N is the number of dimensions of the voice pattern vector.

As an example of the weight W, first, from a number of utterance patterns related to a certain word, an average pattern is determined and set as B, and then each utterance pattern is set as A, and C is calculated using the method of the present invention. demand. From this, the weight W _jk is determined.

However, C ^l _jk is an element of the voice pattern C obtained by the l-th utterance, and L is the number of times the voice is uttered to obtain the weight.

By performing weighting as described above and calculating the distance, matching using dynamic programming that can deal with stochastic variations in parameters can be performed. That is, by increasing the weight of the central element of the parameter,
Distance calculations are performed by reducing the weight of relatively unimportant elements.

(f) Effects of the Invention As explained above, since the present invention can change the weight of distance calculation, it is possible to change the weight of the distance calculation. Since we can provide a parameter matching method that combines the advantages of the discriminant function method, which is effective against stochastic variations in And the effect is great.

[Brief explanation of drawings]

FIG. 1 is a block diagram illustrating an embodiment of the present invention, FIG. 2 is a diagram illustrating selection of a matching path, and FIG. 3 is a flowchart for determining a new matching pattern. 1, 3, 6, and 8 are memories, 2 is a DP matching section, 5 is a conversion section, and 7 is an arithmetic section.

Claims (1)

[Claims] 1. Means for matching an input time series pattern with a standard time series pattern by dynamic programming, and based on the matching result of the matching means, matching the input time series pattern with the standard time series pattern. A conversion means for converting the length of the time series, and weighting that weights central elements of parameters is used between the new input time series pattern converted by the conversion means and the standard time series pattern. A pattern matching device that performs pattern matching by calculating distance.