JPS6175982A - Pattern recognizer - Google Patents

Abstract

PURPOSE:To prevent a decrease of misrecognition due to mixture of similar patterns without averaging a standard pattern, by applying a detailed discrimination method among the results of recognition and the 2nd and lower candidate groups. CONSTITUTION:The feature vector trains of each pattern are stored to a standard pattern memory part 4 and an input pattern memory part 5. Then an input pattern A and a standard pattern B<1> are sent successively to a matching part 6 in 1-P times per l for calculation of distance. Meanwhile (k) units of candidates of recognition are selected by a selection part 7 based on the difference between DP matching distances D (A, B<1>). Then the information on an optimum DP path is stored. When the selection is over with all of P units of standard patterns, a discrimination feature part between two patterns B<x> and B<y> of the lower k-th and (k-1)-th places among the candidates of recognition is sent to a weighting distance calculator 9 from a weight coefficient memory part 8 according to the prescribed order. At the same time, the DP optimum path information on a pattern B<x> and the pattern A is sent to the calculator 9 from the selection part 7 respectively. Thus a weighting distance is calculated.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to an improvement in a system for recognizing patterns such as speech and characters.

(Prior Art and its Problems) In recent years, a number of practical pattern recognition systems, such as phonetic and handwritten characters, have been released and are being used in industry. Most of these recognition systems operate based on the patterning method. That is, data such as the user's voice and handwritten characters are input into the system when the system is first used, and the pattern is registered as a standard pattern.

Thereafter, when a pattern to be actually recognized (referred to as an input pattern) is input, a comparison with the standard pattern group is performed, and the distance between the input pattern and each standard pattern is calculated. By comparing these distances m4, this input pattern l is selected as the closest standard pattern. As a result, the input pattern is determined to be identical to this standard pattern.

The pattern matching method described above has a simple principle and has been proven to have a relatively high recognition rate. However, even in the pattern matching method described above, similar word sounds such as "Wakayama" and "Okayama", similar monosyllabic sounds such as "ki" and "gi", "riya" and "miya", etc. The current situation is that recognition is difficult and there are many errors. In addition, in character recognition, there is a similar problem with similar characters such as "7", "ke", "ri", etc. As a countermeasure, as shown in the specification of Japanese Patent Application No. 1988-88909, A method has been put into practice in which similar parts of similar standard patterns are averaged in advance. However, this method averages a part of the standard pattern, so the pattern may vary considerably! The problem of declining rates remained.

(Object of the Invention) The present invention has been made by focusing on these conventional problems. In other words, among the recognition results and the second and lower candidate groups, a detailed classification method as described below is adopted (this is because similar patterns exist, rather than averaging the standard patterns). The present invention provides a pattern recognition system that can improve the reduction in recognition rate.

(Structure of the Invention) According to the present invention, there is provided an analysis unit that analyzes input data and creates feature patterns 7, a standard pattern storage unit that stores patterns supplied from the analysis unit as standard patterns, and a standard pattern storage unit that stores patterns supplied from the analysis unit as standard patterns; an input pattern storage unit that stores an input pattern to be input;
a matching unit that performs a comparison process between an input pattern and a standard pattern; a selection unit that extracts a top recognition complement group based on the comparison result; and a weighted distance that calculates a distance by weighting the discriminative features between the standard patterns. 3 in total, and the weighted distance l! According to the present invention, in a pattern recognition system, a plurality of recognition candidates are first extracted, and then detailed identification is performed based on weighted distances. By doing so, more reliable recognition results can be obtained.

(Principle of the Invention) Here, the principle of non-invention will be explained using recognition of word sound patterns as an example. However, the object of the present invention is not limited to voice recognition.

When recognizing the input pattern "OKAYAMAJ," the standard pattern fcol'-OKAYAMAJ, l'-WAKA is used.
Assume that similar patterns such as Y-A, MAJ, and [OKUYAMAJ are included. In the case of C, the distance between the input pattern "OKAYAMAJ" and the three standard patterns is
Although it is smaller than the distance from other standard patterns, if the distance from the correct recognition result "OK A Y A M A J" is the minimum (limit f (i.e., there is a possibility of misrecognition). This is because l KA l and I
Despite the fact that there is a clear difference between KUI, +01 and IWAI, the similarity of the features in the remaining parts is affected by KA and IKUI.

+(This is because the discriminative features such as Jl and IWAl are masked.

Therefore, instead of determining the recognition result as the one with the smallest distance tafJs, we leave some of the ones with small distances as recognition candidates, and perform detailed identification as described below.
In the end, one recognition result is obtained. Now, the recognition candidates for the input pattern [OKUYAMAJ are, in descending order of distance, 1'-OKUYAMAJ, 2nd place 1'-WAKAY
AMAJ, 3rd place: “Suppose it is OKAYAMAJ.

First, the 2nd and 3rd place will be identified, and then the first place will be identified from that result to obtain the recognition result. Discriminative features between these standard patterns in advance, ie different phonemes 101 and IWAI.

Find the positions of IUI and IAI in advance.

Then, weighting is applied to the discriminative feature, distance calculation is performed again with the input pattern, and discrimination T is performed using the emphasized distance of the discriminative feature. In other words, the distances of the masked discriminative features IWAI and 101 of the standard pattern from the input pattern 101 are emphasized and identified by the similar portion lKAYAMA1 of the pattern, and a new recognition result is obtained. Here, between the distance between 1WAl and 101 and the distance between 101 and 101, the distance between +0+ and 101 is smaller, so as a result of emphasis, the overall weighted distance is equal to "OKAYAMAJ" of the input pattern and "l" of the standard pattern. ``OKAYAMAJ's group is smaller, correct ranking, 2nd place 1'''OKAYAMAJ 3rd place 1''
-WAKAYAMAJ can be obtained.

Next, 1st place 1'-OKUYAMAJ and new 2nd place "Q-K"
AYAMAJ
In the same way, the first correct recognition result is “OK A
, Y AM A J is obtained.

In this example, the number of recognition candidates is three, but detailed identification can be performed using the same method even when there are two or four or more recognition candidates.

Here, the weighting coefficient for emphasizing the discriminative feature N5 is calculated and stored in advance before recognition is performed. Discrimination fF between this standard pattern? An example of a method for determining a weighting coefficient that emphasizes the second part will be given below. Generally, the distance between speech feature vectors is small when both belong to the same phoneme φζ, and large when they belong to different phonemes. Therefore, the part where the distance between the two patterns is larger than the other can be regarded as a different phoneme, that is, a discriminative feature. Now, standard pattern vector sequence B = b, , b, position, b, ,...
, b- and B r =b, f , b, f, . . . Bl,
, . . . b-, and perform DP matching between the two. 1st
The figure shows the matching. Distance d (b, , b, (1)) of the solid line part on the DP optimal bus
If it is larger than the others, then each corresponding part of the bass, L, can be treated as a different phoneme, that is, a discriminative feature. Then, weighting coefficients W and f(i) can be given to emphasize that part.

(Example) Hereinafter, an example of the present invention will be explained with reference to the drawings. FIG. 2 is a block diagram of a speech recognition device in one embodiment of the present invention.

In the figure, 1 is a microphone, and 2 is an n-dimensional feature parameter vector sequence (al +22 , am ) for input audio.

. . . An analysis unit that sequentially converts into axl, and is composed of a filter bank, a Fourier transformer, etc.

3 is a switch that switches to A1m1+ when creating a standard pattern, and switches to B side when recognizing an input pattern. 4.5 is
The n-dimensional feature parameter vector sequence created by the analysis unit 2 is stored in the standard pattern storage unit and the input pattern storage unit, respectively, as a standard pattern B = b1. b, ,...,b,...,B=b,
, b, ,...+ b b *Input pattern A= a
It is stored as l, afi, .--a□.

6 uses the dynamic programming method described in Japanese Patent Application No. 49-2418 to perform DP matching to normalize fluctuations in the time direction, and creates a standard pattern BJ.
== b 1J, b, l,...B:,...b
j and input pattern A = al l 82 r”'a +
This is a matching part that calculates the DP matching distance with +”'a 1. However, I = + 7
(RI is a mapping function used for time axis normalization of input pattern A and standard pattern B'. 7 is a mapping function used for normalizing the time axis of input pattern A and standard pattern B'. The pattern is extracted as a recognition candidate, and its rank, pattern name, and each distance d(al,
b, , (1)), mapping function I = 44(i), etc. 8 is between similar standard patterns Bx, B
This is a weighting coefficient storage part in which a weighting coefficient W■) for emphasizing the discrimination feature S of "" is stored. 9 is a weighted distance calculation unit which selects a pattern BX, which is one of the recognition candidates, from the selection unit 7. Each distance 1'lId (a Ir b7x
(+ρ, and mapping function j = j Calculate the weighted distance to the person. Figure 3 is a diagram showing the state of DP matching between A and BX. Multiply the weighting factor W7(Y) to emphasize it.10 is 2
This is an identification unit that compares the magnitude of the two weighted distances and excludes the larger one from recognition candidates. The selection unit 7, the finding distance calculation unit 9, and the identification unit 10 will be explained in detail in FIG. 4.5.6, which will be described later.

Next (about the operation of the device configured as above)
Explain below. After the standard pattern storage unit 4 and the input pattern storage unit 5 have finished storing the % signature vector sequence of each pattern, the input pattern person and the standard pattern 7B are stored for A=1 to P times, 111
The signal is sent to the zero-order matcher 6, where distance and music calculations are performed. During that time, each DP matching distance D (A, B
), the selection unit 7 successively selects recognition candidates, and the optimum DP bus information is stored. After all the standard 6772 patterns have been sorted, detailed discrimination is performed between the two patterns BX and BY in the lower ranks among the recognition candidates according to the ranking. Discrimination feature between the two patterns w7, 7. is sent from the weighting coefficient storage unit 8, and information on the DP optimal bus of the pattern Bx and input pattern A is sent from the selection unit 7 to the weighted distance calculation unit 9, and weighted distance Dw (A, B) = pattern Hyf This is done in the same way. Then, two weighted distances 4Dw (A, B
)D(A, 113'') is sent to the identification unit 10, where the recognition candidate with the larger landing distance is removed.

Next, the same operation is performed between the remaining recognition candidate pattern and the k-2nd pattern, and one more recognition candidate is removed. The above operation is repeated -4 times, k-1 recognition candidates are removed, and finally, the recognition candidate with the smallest actual finding distance is output as the recognition result.

FIG. 4 is a block diagram showing a specific example of the selection section 7. 71 is a comparator, 72 is a recognition candidate table, and k D
The P matching distance, its ranking, and information on the DP optimal bus are stored. 73 is a selection control section which controls each section as follows. From the matching section 6 to the DP matching distance D(A
, B) is sent to the comparator 71 (the recognition candidate table 72).
) is sent to the comparator 71 and compared in size. And D(A
, B)<D(A, Bk), the comparison circuit ζ compares the DP matching distance D(A, Bk") of the 11th place. Repeating the above, A(A
,B)>D(A,B''), the matching unit 6 receives the information of the DP optimum bus of the standard pattern 7B, and the m+1st data 4 in the g knowledge candidate table and the standard Swap the data of pattern B.

As described above, from among the DP matching distance data sent sequentially, pieces of data are stored in the recognition candidate table in descending order.

FIG. 5) is a block diagram showing a specific example of the tribute distance Pith calculation unit 9. 91 and 92 are memories for storing the distance on the DP optimum bus sent from the recognition candidate table 72; 93 is a memory for storing the marketability coefficient W sent from the weighting coefficient storage unit 8; xy( 41. 94, 95, and 96 are the memories, respectively.
They are multipliers, adders, and accumulators. Reference numeral 97 is an approach distance calculation control section that controls each section as follows, and emits a signal according to the time chart shown in FIG. First, the control unit 97 sends a clear signal C! to the accumulator 96.
and reset the accumulator to 0. Next, the mapping function jx(+) is read out from the memory 92 in accordance with the signal i issued by the control section 97. Next, the weighting coefficient Wxy is read out from the memory 93 and input to the multiplier 94.

10,000, the distance is calculated from the memory 91 according to the signal i, and added using the accumulator 96 and the adder 95, and the category name of this recognition candidate pattern is defined as the weighted distance D (A, Bx). is output with

FIG. 7 is a block diagram showing a specific example of the identification section 10. 101, 102, 103, and 104 are registers, 105 is a comparator, and 106 is a switch circuit. 10
7 is an identification control section which controls each section as follows. The register 1011 receives the weighted distance D (A, B), W from the accumulator 96 of the weighted distance calculation section 9.The register 103ζ receives the category name of the recognition candidate pattern B from the weighted distance calculation control section 97. recorded and remembered. Thereafter, along a similar path, the weighted distance @D (A, By) is stored in the register 102I, and the category name of the recognition candidate pattern BY is stored in the register 104#. Next, the comparator 105 has two weighted distances @D (A,
B ), D (lA, By) are sent, and W
W magnitude comparison is performed.

Based on the magnitude result, the switch circuit 106 is switched, and the category name of the recognition candidate pattern with the smaller collision distance is output as the recognition result. This description is not intended to limit the scope of the present invention. Special
The determination unit 7 selects recognition candidate abstraction methods in descending order of DP matching distance. ,B)-mmunD(A,B
) I < T.

TB, where TD satisfies the above equation. However, TD may be used as a threshold value of the DP distance, or may be a mixture of the above 20,000 equations, and may be TBX that meets the above equation among those with a small DP matching distance. Further, in the identification unit IO, identification was performed by comparing the magnitudes of two distances in order from the lowest recognition candidate and excluding the one with the larger distance, but the method is not limited to this method, for example,
It is also possible to perform two comparisons in rows 7 and 1' in order from the top. It is clear that the principles of the present invention can be effectively applied to patterns other than speech.

For example, when applied to character recognition, it is effective in distinguishing between extremely similar characters such as "ma" and "ma".

(Effects of the Invention) According to the present invention, in the pattern recognition system, there are similar standard patterns, and the difference in their discriminative features is caused by the variation in other common parts (the resulting difference). Even if the object is hidden, the detailed recognition described above can be performed to obtain a correct recognition result, and the recognition rate can be improved.

[Brief explanation of the drawing]

1 and 3 are diagrams for explaining DP matching, FIG. 2 is a block diagram showing an embodiment of the present invention, and FIG.
．． Figures 5 and 6 (respectively, selection units) are block diagrams showing configuration examples of the weighted distance calculation unit and the identification unit, and Figure 7 is a time chart for explaining the operation of the weighted distance calculation units. In the figure, 1...Mike, 2...Analysis section-13.
...Switch circuit, 4.--Standard pattern storage section, 5..°.Input pattern storage section, 6.....
・Matsuchinobu section, 7911101 selection section, 8...
- Weighting coefficient storage unit, 9... Weighted distance calculation unit, 10... Identification unit, 71... Comparator,
72... Recognition candidate table, 73...
Selection control unit, 91, 92, 93...--Memory, 94...-Multiplier, 95...-Adder,
96... Accumulator, 97... Weighted distance calculation control unit, 101, 102, 103
, 104...Register, 105...Comparator, 106...Switch circuit, Fig. 1

Claims (1)

[Claims] an analysis section that analyzes input data and creates a characteristic pattern; a standard pattern storage section that is connected to the analysis section at least when registering a standard pattern and stores the standard pattern supplied from the analysis section; and a standard pattern storage section that is connected to the analysis section at least during recognition. an input pattern storage unit that stores the input pattern supplied from the analysis unit; a matching unit that performs a comparison process between the input pattern and a standard pattern; and a selection unit that extracts a group of upper recognition candidate candidates based on the comparison result; A pattern recognition device comprising: a weighted distance calculation section that calculates a distance by weighting a discrimination feature between standard patterns; and an identification section that selects the recognition candidate group based on the weighted distance.