JPS604506B2 - Pattern similarity calculation method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a calculation method for comparing two patterns represented by time series of feature vectors and calculating the degree of similarity between these patterns.

2. Description of the Related Art Recently, technologies such as voice synthesis and recognition have made remarkable progress, and attempts are being made to add voice functions to various electronic devices.

In particular, in such technology in the audio field, when recognizing an input audio signal, a pattern comparison is performed between a feature vector pattern formed from the input audio and a standard pattern registered in advance. The input signal is recognized based on the comparison result. Comparison between patterns of this type has conventionally been performed using a time axis expansion/contraction matching method using dynamic programming. However, in this type of time-axis expansion/contraction matching, optimization is performed at each point in time, and the matching is performed with time-axis expansion/contraction that ultimately maximizes pattern similarity, which requires a significant amount of calculation. Become. Furthermore, in matching between the same categories, there is a disadvantage that time axis expansion/contraction that is significantly different from linear expansion/contraction cannot be considered. On the other hand, simply associating feature vectors with maximum similarity often results in incorrect time expansion and contraction. The present invention was developed in view of the above-mentioned problems in conventional pattern similarity calculations, and its outline is to reduce the amount of calculation by simply associating vectors that maximize the similarity at each point in time. At the same time, by comparing with the correspondence obtained by linear expansion/contraction, it is possible to prevent the time axis expansion/contraction from becoming too large. Next, the present invention will be explained in detail using the drawings.

As an example of comparing two patterns, pattern A consisting of the time series number A,,......An (the bold letters hereafter represent vectors) of feature vectors, and It is assumed that the similarity is calculated by selecting the correspondence between the pattern B and the pattern B consisting of the series B, , &...Bm.

FIG. 1 is a diagram for explaining a method of expanding and contracting the time axis.
In the same figure, the number n of time series of pattern A and pattern B
The number m of time series is not necessarily the same. Therefore, although it is not possible to make the time series of the patterns correspond in a proportional relationship, first, the correspondence is made so that a nearly proportional relationship is obtained as shown by the black circles in the figure. When the feature vector BJ has already been correlated according to the calculation method that will be described later, three points in the vicinity of the i-th point of the pattern B, for example, j, j+. is selected, and the feature vector Bi,
Bi ten. , Bj+2, the vector B with the highest degree of similarity
Select k.

On the other hand, when the vector of pattern B in which the feature vectors Ai+, are made to correspond proportionally is BI, then the above Ai+, and B,
The similarity between the feature vector Ai and the vector Bk is compared with that between the feature vector Ai and the vector Bk. If the similarity between the vectors Ai and B is greater by comparing them, the feature vector AM is associated with the vector 8, and the comparison operation is then proceeded to the next point in time. Furthermore, the feature vector AM
In a state where the similarity between the vector BI and the vector BI is large, the feature vector Ai+,
It is also possible to associate the vector with the highest degree of similarity with Ai+, and proceed to the next point in time. The above-described similarity comparison operation is performed for n time-series feature vectors A, to An to be associated with pattern B. FIG. 2 is a block diagram showing an apparatus for executing the calculation method of the present invention.

M, is a memory that stores the standard pattern A;
, are read out in response to instructions given from the central control unit C, and the feature vector of the standard pattern at a certain point in time is given to the next stage memory M2. Similarly, M3 and M4 are memories provided for input pattern B, respectively.
The pattern is written into the memory M3, and the extracted feature vector is written into the memory M4. The contents of the memories M2 and M4 in which the feature vectors are written are given to the similarity calculation unit L, and the similarity between each feature vector of the standard pattern and the input pattern is calculated based on the above-mentioned comparison principle. be done. The calculation results are transferred to and held in memory M5.
Already, in the process of linear calculation near the i-th time point in the input pattern B, the similarity between the feature vector AM and the vectors Bi, Bi1, Bi12 (hereinafter mode Q) is stored in the memory M5.
For nonlinear calculations near the first time point, the similarity between the feature vector Ai and the vector B1, or the feature vector A and the vector BH,B,,B,1 (hereinafter referred to as mode 8) is ) is calculated and stored. X.
is a maximum value detection unit that detects the maximum value among the similarities held in the memory location. When the calculation is executed in the above mode o, the maximum value and the frame number in the time series in which the feature vector is located are written into the holding unit M6 via the gate G. On the other hand, when calculation in mode 8 is executed, the maximum value and its frame number are written into the holding unit M7. The modes written in holding unit M6 and holding unit M7 are and mode 8.
The calculation results are compared in the comparison section X2 based on a command given from the control section C, and the larger value is selected and recorded in the addition memory M8. On the other hand, as a result of the comparison operation performed by the comparison unit X2, the time-series frame number in the input pattern of the feature vector selected as having the maximum similarity is transferred to the control unit C. The control unit C refers to the transferred frame number when calculating the mode Q at the next time,
The memory M is used to calculate the similarity similar to the above calculation operation.
Issue a command to send the feature vector to 4. The above calculation operation is executed sequentially for each frame number and each calculation result is added to the memory M.
8, and from the contents of the memory port, among the referenced standard patterns, for example, the standard pattern with the highest degree of similarity is used to recognize the input pattern. As described above, according to the present invention, by simply making a correspondence with a vector that has the maximum degree of similarity at each point in time to reduce calculations compared to the conventional method, and at the same time, by comparing it with a correspondence obtained by linear expansion and contraction. , it is possible to prevent large scale expansion and contraction of the time axis, and more efficient pattern similarity can be obtained with a small amount of calculation.

[Brief explanation of drawings]

FIG. 1 is a principle explanatory diagram for explaining the calculation method according to the present invention, and FIG. 2 is a block diagram of a device that implements the principle of the present invention. A,, An: Feature vector of pattern A, B,
, &...Bm Feature vector of pattern B, C: control unit, L: similarity calculation unit, X,: maximum value detection unit, X2: comparison unit, X8: addition memory, M. , M2, M3, M4, M5
, M6, M7: memory. Figure 1 Figure 2

Claims (1)

[Claims] 1 Time series of feature vectors A_1, A_2, ...A_n
Pattern A consisting of vectors B_1, B_2...B_
In the calculation method for calculating the similarity with pattern B consisting of m, the similarity is calculated by sequentially reading out the memory that stores the feature vector of pattern A, the memory that stores the vector of pattern B, and the contents of each memory. A calculation unit that calculates a feature vector A_i_+_1 of the next time point (i+_1) for a certain time point i, a vector B_k with the highest degree of similarity among the feature vectors near the j-th time point in pattern B, and pattern A and the pattern. The feature vector A_i_+
A calculation unit that calculates the vector B_1 of the pattern B associated with _1, a comparison unit that compares the degrees of similarity of the vector B_k and vector B_1, and a calculation unit that calculates the degree of similarity at the next point in time based on the comparison result. and a control unit that sets a feature vector, and when the degree of similarity of vector B_1 to the vector A_i_+_1 is greater than vector B_k, vector B_1 or pattern B is provided.
The feature vector A_i_+_1 is associated with the feature vector A_i_+_1 that has the highest similarity to the vector A_i_+_1 near the first point in time. If B_k has a greater similarity to the feature vector A_i_+_1 than the vector B_1, the vector B_k is assigned as a feature. A pattern similarity calculation method characterized by calculating the similarity between patterns by making a correspondence with the vector A_i_+_1.