JPS5926960B2 - Time series pattern matching device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an improvement in a pattern matching device that is an important part of a pattern recognition system for speech and the like.

A system that automatically recognizes speech has applications in fields such as data input to computers and machine operation as a means of ergonomically superior pine-machine interface, and its practical application is desired from various quarters.

Pattern matching is a simple and relatively stable recognition method. In this method, standard phonetic patterns (hereinafter referred to as standard patterns) are registered in advance for each word in the language to be recognized, and the input pattern is compared (matched) with each standard pattern to find the best one. It is recognized by determining standard patterns that are very similar.
However, voice patterns contain a large amount of information, and registering ordinary analysis data as is requires a huge amount of memory, making the recognition device expensive. For this reason, instead of registering the characteristics of a word speech pattern over the entire time as a standard pattern, only a series of parameters that describe the phonemes that make up the word speech are registered as standard patterns, and the input speech is divided into phonemes. Attempts have been made to recognize a pattern by dividing it into units and then matching it with a standard pattern. However, with this method, it is extremely difficult to classify input speech into phoneme units, and it is easy to make incorrect classification operations, so a recognition system with sufficient performance has not been realized. Therefore, a practical method is to directly match the compressed standard pattern without segmenting the input speech. In this case, the time axes of the input pattern and the standard pattern are different (normally, the standard pattern is sampled at a relatively short constant period, and the input pattern is unequal length and sampled at a relatively long period). Therefore, it is necessary to perform matching by aligning the time axes of the two. For this reason, it is necessary to expand the phoneme sequence of the standard pattern to match the input pattern, but in this case, the expansion must not be extremely different from the standard duration of each phoneme. Another problem is how to efficiently perform matching. A method that uses dynamic programming is suitable for solving such problems, as proposed in the Japanese Patent Application No. 4911.
3815”.

However, since the cited example projects the input pattern onto the time axis of the standard pattern and compares it, (1) division is required in the dynamic programming recurrence formula, resulting in a large amount of calculation. In addition, (2) there was a drawback that the magnitude of the value calculated as the degree of similarity was affected by the time length of the standard pattern. The purpose of the present invention is to improve the disadvantages (1) and (2) of the above cited example by projecting a standard pattern onto the time axis of the input pattern and comparing it, contrary to the method cited in the cited example. The objective is to realize a speech recognition device with good performance.

Although the principles of the present invention can be effectively applied to a wide range of pattern recognition problems other than speech, the configuration of the present invention will be described in detail below using speech as an example. In general, audio patterns are patterns that have a time axis, so Qi-(・All.
a2l, ......ANi) is expressed as a time series.

Here, Qi is counted from the beginning of the voice and represents the characteristics of the first time point. As an example, a vector whose elements are the output levels of a speech analysis filter can be considered. Consider another time series with a similar configuration, and consider the problem of performing pattern matching with B as a standard pattern and A as an input pattern.

However, the time axis of B is compressed nonlinearly, and J is
It is assumed that it is considerably small compared to . wood*
In the apparatus of the present invention, each feature vector L)j of a standard pattern is stored in correspondence with the minimum duration Mj and maximum duration Nj allowed, and each feature vector b) of the standard pattern is
JJ of m・ and n・ where ・ is specified
The distance between each feature vector Qi of the input pattern and the feature vector 1b of the standard pattern associated with it is added for i to the total amount. An operation to minimize the j-field is performed by dynamic programming, and the obtained minimum value is taken as the degree of similarity S (A.B) between the input pattern A and the standard pattern B.

This means that, as shown in FIG. 1, J line segment sequences (point sequences) such as 11 to 14 are defined on the i-j plane, making j points correspond to l* i points.

A function that makes point i correspond to point j is Indicated by

Let us take as an example the case where distance 01j is used as a measure of similarity between 3. and b.

? The distance between 1 and [Bj is denoted by d(1.j). Thus, the similarity can be obtained by calculating J using the minimization problem under the constraint that the number of i's associated with I j's satisfies the condition (3).

The apparatus of the present invention is characterized in that the above-mentioned maximization problem is calculated using the following dynamic programming method.

That is, the degree of similarity is obtained by sequentially calculating in the range 1≦i≦, 1≦j≦J based on the recurrence formula.

A specific example of the recurrence formula (7) will be explained with reference to FIG. FIG. 2 is an enlarged view of the vicinity of point (1.j) on the ij plane.

Although the coordinate axes are omitted, it is assumed that the horizontal axis is the i-axis and the vertical axis is the j-axis. In the example shown in the figure, when the minimum length Mj is specified as 3 and the maximum length Nj is specified as 5, g(
1. j) is d(1-
Lj), (3≦l≦5) and g(1
-1. It shows how it is calculated based on j).
In this case, the values in brackets on the right side of equation (7) are l-3, l-4,
Three types of l-5 are generated. That is, the recurrence formula of the JIV (7) formula means that the minimum of the above three values of G3, G4, and G5 is selected and newly calculated as g(1.j).

The meaning of the recurrence formula in equation (7) is as above, so if j is now made to correspond to a stage in dynamic programming, g(1.j) is d(1,j) of the jth stage. and the j-1st
It is easily calculated if g(g), j-1) of the stage is known.

So starting from stage j-1, j=2, 3...
..., and finally reaches the stage j-J and calculates g(1.
In order to carry out the above principle, the apparatus according to the present invention can obtain 5:0 (1) Input pattern A=S1, 62, . . .
, input pattern buffer for holding part or all of the button, (2) standard pattern B=TOl, 1b2,...
..., 1bj...BJ and its attached Mj. standard pattern storage unit for holding nj,
(3) Distance calculation unit for calculating the distance d(1, j) of ΦITOj, (4) Distance register for temporarily storing the required number of d(1.j), (5) Recurrence formula (7) ), (6) recurrence formula calculation unit for calculating recurrence formula value g(1.j)
Its main part is an arithmetic register for temporarily storing the required number of .

FIG. 3 shows the functional connections of the above parts. In the figure, 41 is an input pattern buffer, 42 is a standard pattern storage section, 43 is a distance calculation section, 44 is a distance register, and 45 is a distance calculation section.
is a recurrence formula calculation unit, and 46 is an arithmetic register. Reference numeral 40 denotes a control unit, which has a function of controlling the operations of each of the above-mentioned units.

The recurrence formula calculation unit 45 calculates the required number of d(1.j) from 44,
46, read the required number of g(1.j-1), calculate recurrence formula (6) based on these, and obtain g(1.j
) is written to 46. The repetition for (g) and j) is 40
controlled by. As described above, the principle of performing pattern matching between standard pattern B and input pattern A by dynamic programming has been explained.

Each according to this principle?・ b ・ is optimally associated with ・ , and stable machining can be performed efficiently.

Moreover, the minimum and maximum duration constraints are also accurately realized. Furthermore, compared to Japanese Patent Application No. 113,815/1982, there is a great advantage in that the recurrence formula (7) does not include division. In addition, the value g(1.J) of the similarity S(A.B) has the feature that it is not affected by the standard pattern length J because it is an amount integrated in the j-axis direction. It enables recognition judgment by comparison. A basic embodiment of the invention is shown in FIG. This embodiment shows in particular detail the recurrence formula calculation section of the block diagram of FIG. Figure 4
0, 41, 42, 43, and 46 correspond to those in FIG. 3, respectively. From 450 to 455 is the recurrence formula calculation section 4 in Fig. 3.
This corresponds to 5.

451 is an adder circuit which adds the signal d from 44 and the signal S from 452.

452 is a first register which is set to O by the signal r from 450 and has the function of holding the output of 451 thereafter.

Signal d given sequentially from 44 by 451 and 452
The sum total is calculated and stored.

453 is an adder circuit which receives signals S1 from 452 and 46
The sum of the signals g1 is calculated and output as G2.

454 is a comparator circuit that compares G2 and G3 and allows the smaller value to pass.

Reference numeral 455 is a second register which, after being set to a sufficiently large value by the control signal r, holds the last one of the outputs of 454 which are sequentially applied.

G2 given sequentially from 453 by 454 and 455
The minimum value of is calculated and held at 457.

Reference numeral 450 denotes a recurrence formula calculation control section, which controls each section 451 to 455.

Furthermore, the above 41, 42, 43, 44, 46, and 450 are controlled by the control section 40. Prior to starting calculation of the recurrence formula, the control section 40 issues a control signal Rs to set the initial value of equation (8) in the calculation register.

After that, the parameter (1.j) of the recurrence formula is generated sequentially as signals 1 and j, and the recurrence? 7) have them calculate. An appropriate example of the order in which (1.j) is generated is to sequentially increment i in stage j, and when that stage is completed, increment j by 1 and move on to the next stage. Next, the operation of each part at an arbitrary point (1.j) will be explained.

When the signal j is applied to the standard pattern storage section 42, Bj is outputted as the signal b, and at the same time, Mj and Nj are outputted as m and n, respectively. Distance 1 Calculation unit 43 is C from 41
i (1=1 to I) are read out sequentially, and the distance between b and given as signal b is
The distance d(1.j) of J is calculated using equation (5) and written to the distance register 44. After that, the recurrence formula calculator changes i from 1 to J, and converts g(1,.j) to J
Works to calculate sequentially. The operation of calculating g(1.j) for a particular (1.j) is as follows. First, 450 issues a control signal r to reset 452 to 0 and 457 to a sufficiently large value. After that, 1=O, 1
, 2c......, 1'=i-1 is issued sequentially up to 5 (Nj-1), and the sum of each time point with respect to l is 4.
52 to calculate as S・(Mj-1)≦l≦(N
When l satisfies j-1), S1 is output by output command signal 0e1, and this is expressed as [ ] on the right side of equation (7).
This corresponds to the second term *41x1d(1-X.j). At this time, emit i〃i-1 and set g1 as g(1-1, .
j-1) is output from 46 and the sum with S1 is calculated as G2, which becomes one of the values in brackets [ ] on the right side of equation (7). Therefore, l changes from (Mj-1) to (Nrl)
All the values on the right side of equation (7) can be obtained while changing within the range of . 454 and 455 to find the minimum value of G2, g
The calculation result of equation (7) is obtained.

At this point, an output command signal 0e2 is issued to 455 to send the value of g to the calculation register of 46 and store it as g(1.j). As a result, the calculation of recurrence formula (7) becomes (1
．． j) has been completed. Next, i is increased by 1 and g(1+1,j) is calculated. The above calculations do not need to be performed for all (1.j). For example, it is impossible to calculate g(1.j) at a point where i<j, but the recurrence formula calculation is omitted at such a point. Therefore, the capacities of the distance register 44 and calculation register 46 may actually be small. Although the principle of the present invention has been explained above with reference to examples, these descriptions do not limit the scope of the present invention. In particular, in the above description, seasonal 1 and T
Although the similarity between Oj was evaluated using the distance d(t), j), this is just an example, and other measures may be used. When the quantity is so large that season 1 and 1bj are similar, such as inner product or correlation, the similarity becomes a maximization problem, contrary to equation (6),
The operation for selecting the minimum value in equation (7) and the examples can be replaced by the operation for selecting the maximum value, but such cases are also included in the scope of the present invention. Furthermore, as described above, the scope of application of the present invention is not limited to audio patterns, but is broadly applicable to patterns expressed as time series. In addition, in the above description, in order to maintain generality, we have described the case where Mj and Nj are specified for each j, but when Mj and Nj are each constant, there are special settings in the standard pattern storage section. You can fix the shape of the recurrence formula instead of memorizing it. For example, m・0j is always 1 and n・
When is always 2, equation (7) becomes a simple form like J.

If the standard pattern is compressed in the time direction at a constant rate, such a simple method is sufficient and effective in terms of simplifying the device, but in this way Mj. A case where nj is fixed is also included within the scope of the present invention. Also, the similarity value is not fixed at g(1.J), but g(1.J).
J) (where i is variable) is used as the degree of similarity.

[Brief explanation of the drawing]

FIG. 1 is a diagram explaining the principle of the present invention. In the figure, the horizontal axis shows the human pattern A-Cl, @2,...
・・・・・・・・・,61,・・・・・・・・・,QI is i
The vertical axis shows the standard pattern B-1b.
1,1b2,・・・・・・・・・,Bj,・・・・・・
..., Bj are arranged in correspondence with j. 11
, 12, . . . , 14, each 81 is associated with TOj. FIG. 2 is a diagram showing how the dynamic programming recurrence formula, which is the core of the present invention, is calculated. FIG. 3 is a diagram showing a basic embodiment of the present invention. Reference numeral 40 designates a control unit 41 as an input pattern buffer, 42 a standard pattern buffer, 43 a distance calculation unit, 44 a distance register, 45 a recurrence formula calculation unit, and 46 a calculation register.

Claims (1)

[Claims] 1 Input pattern A expressed as a time series of feature vector @_1 = @_1, @_2, ......, @_i, ......@
An input pattern buffer that holds _l in a format that can be referenced with respect to i, and a standard pattern B = ■_1, ■_2, ......, ■_j, expressed as a time series of feature vectors ■_j.
......, ■_J and a standard pattern storage unit for holding the minimum length m_j and maximum length n_j specified in advance for each ■_j in a format that can be referenced with respect to j, and @_i read from the input pattern buffer. A measure d indicating the degree of similarity between and ■_j read from the standard pattern storage unit
(i, j); and d(i, j) calculated by the distance calculation means with respect to specified j.
) in a format that can be referenced with respect to i, and the recurrence formula calculation result g(i
, j), and in each stage related to j, l is made variable between m_j and n_j, and d(i-l+1
, j) to d(i, j) has h(i-1,
d(i, j), which is the sum of j−1), is @_i and ■_j
are small enough to be similar, the minimum value of l is used as the new g(i
, j) in the calculation register, and calculates the degree of similarity between the input pattern A and the standard pattern B as the calculation result of the stage where j=J. Device.