JPH0574836B2 - - Google Patents

Description

[Detailed description of the invention] (Industrial application field) The present invention relates to a pattern recognition device.

(Prior art and its problems) A pattern recognition device calculates the similarity of a pattern expressed as a time series of feature vectors, such as voice, with a standard pattern (matching), and performs recognition between patterns. It is a device that performs For example, a speech recognition device registers the speech pattern to be recognized as a standard pattern in the device in advance, and during recognition operation,
This method is based on a method in which a comparison operation is performed between each standard pattern, the degree of agreement between the two is examined, the standard pattern that most closely matches is determined, and a recognition judgment is made. In this pattern matching method, it is important to use a measure that is stable against fluctuation phenomena specific to patterns, such as uneven expansion and contraction of the time axis caused by fluctuations in speaking rate in speech patterns. .
Regarding this, please refer to the Acoustical Society of Japan's speech research group materials.
It is described in detail in S73-22 (published in December 1971) "Comparison of various DP matching methods in speech recognition" (hereinafter referred to as document (1)). In addition, as a way to implement this method efficiently and with high precision, the DP matching method is effective, which uses dynamic programming (hereinafter abbreviated as DP method) to calculate the distance between patterns by normalizing fluctuations in the time axis. It is proposed that This matching method assumes that the pattern to be matched is a time series vector sampled at a constant period. For this reason, the storage capacity for storing each pattern is required to correspond to the length of each pattern. For example, in the case of a word speech recognition device, a capacity corresponding to the number of words is required, which requires an enormous storage capacity. Furthermore, since the calculation time increases in proportion to the pattern length, high-speed calculation is required.
On the other hand, in the vowel stationary part, the speech pattern is
Temporal changes in the frequency structure are small, and changes are large in transitional areas such as from vowels to consonants.
A device is known that utilizes this property to compress information on each pattern to reduce storage capacity (Reference (2) Japanese Patent Application Laid-open No. 137899/1983). Below, an overview of the pattern compression method of document (2) will be explained.

The audio pattern is expressed as a time series of feature vectors by sampling the output of an audio signal frequency analyzer, etc. at a constant cycle, and is expressed as a time series of feature vectors as follows: B=〓1,〓2,... _〓j , _〓J (1) . 〓 _j is the feature vector at time j. Pattern B in equation (1) is divided into K sections by K+1 breakpoints l ₁ to l _K+1 as shown in Figure 1a, and the central vectors in each section are 〓'1, _〓'2 ,
..., 〓' Let the sequence of _K be the compression pattern C. The central vector 〓' _k is the vector at the center of the k-th section (l _k , l _k+1 ) extracted as the representative vector of the section. That is, 〓′ _k = 〓 _(lk+lk+1)/2 .

The breakpoint sequence {l _k } is expressed as ‖〓′ k −〓 j ‖ when the magnitude of the vector error for 〓 _j where the representative vector 〓′ _k is included in the interval (l _k , l _k++1 ) is expressed as ‖〓′ _k −〓 _j ‖ , is determined optimally so that the vector error in each interval is minimized. That is, T= Min {l _k } _K+1 〓 ^k=1 _lk+1 〓 ^j=lk+1 ‖〓′ _k −〓 _j ‖ (2) By solving the minimization problem of equation (2), the optimal breakpoint is found. The sequence {l _k } is found. As a specific method for calculating equation (2), dynamic programming is used as follows.

Based on the initial value T (0, 0) = 0 (3), recurrence formula T (m, k) = Min m-δ≦l<m [T (l-1, k-1) + _n 〓 ^j=l ‖〓 _(l+n)/2 −〓 _j ‖〕 (4) The recurrence formula becomes m−δ≦l≦m−1 (5) m=l~J,k =1~K+1
By sequentially calculating the breakpoint sequence {l _k }
When dividing into K, between each division section (l _k , l _k+1 )
_The error d ₍ ^〓 ′ _{k ,} 〓 _j ) between the central vector 〓 _{′ k} uttered at The minimum breakpoint sequence {l _k } can be found as the optimal breakpoint sequence. When this optimal breakpoint sequence {l _k } is found, the median vector between each breakpoint 〓 _k = 〓 ( _lk+lk+1)/2 is found, and the optimally compressed compression pattern C=〓1, 〓2 , ..., 〓 _k , ..., 〓 _K is found. Each vector of the compression pattern C〓 _k
are taken as representative vectors within each section, and a pattern matching method is executed between them and an input pattern input later to perform recognition. In other words, in the compression pattern expressed by a series of representative vectors, the thinned out redundant vectors are interpolated with the extracted representative vectors, reproduced on the time axis before compression, and the DP matching method is applied. perform recognition processing.

In the method described above, the compressed pattern becomes a pattern that is most similar to the pattern before compression, so that pattern compression is possible without reducing the recognition rate.

However, in this pattern compression method, it is necessary to calculate the recurrence formula shown in equations (3) to (5) in order to obtain a compressed pattern. For example, both the input pattern and the standard pattern are compressed and recognized. For processing, a high-speed arithmetic circuit is required. Therefore, circuits that perform compression processing are not necessarily inexpensive.

(Object of the Invention) An object of the present invention is to provide a pattern recognition device that eliminates such conventional drawbacks, realizes pattern compression with an extremely small amount of calculations, and significantly reduces both storage capacity and amount of calculations. It is in.

(Structure of the Invention) According to the present invention, a sequence of feature vectors {b _j },
A pattern memory that holds a standard pattern expressed as (j=1, 2, ..., J), and a pattern change amount D _j of the vector sequence {b _j } stored in the pattern memory, (j=1 ,...J) over the entire pattern, and divide the vector sequence {b _j } so that the total variation D _J of the pattern variation D _j is divided into K equal parts by the number of divisions K determined from the pattern length J. Divide into K sections, determine the section point sequence {l _k } obtained by the above calculation, read the vector b' _k of the center position of each section determined by the {l _k } from the pattern memory, and calculate each section. As a representative vector of the compression pattern C=c ₁ ,
c ₂ , ... c _K ; a compression pattern memory that stores the compression pattern C; and a recognition processing unit that compares an arbitrary input pattern with the compression pattern. A pattern recognition device is obtained.

Further, according to the present invention, a pattern memory holding a standard pattern expressed as a sequence of feature vectors {b _j }, (j=1, 2, ..., J) and the vector sequence stored in the pattern memory are provided. The amount of pattern change D _j , (j=1, 2, ..., J) of {b _j } is calculated over the entire pattern, and the number of divisions K determined from the total amount of change D _J of the pattern change amount D _j is calculated. , divide the vector sequence {b _j } into K sections so as to divide the total pattern variation D _J into K equal parts, determine the segment point sequence {l _k } obtained by the above calculation, _k } of the center position of each section from the pattern memory and outputs a compressed pattern C=c ₁ , c _{2 ,} . . . c _K as a representative vector of each section; A pattern recognition device is obtained, comprising a compressed pattern memory that stores a pattern C, and a recognition processing section that compares an arbitrary input pattern with the compressed pattern.

(Detailed Description of Configuration) The present invention will be described in detail with reference to the drawings. The input pattern A is expressed by a series of feature vectors A=a ₁ , a ₂ , . . . , a _i , . . . , a _I (6) representing the frequency structure of the input. Further, the standard pattern B to be compared is expressed as B=b ₁ , b ₂ , . . . , b _j , . . . , b _J (7). In the present invention, the standard pattern B is divided into K+1 breakpoints l ₁ , l ₂ , . . . as shown in FIG.
Divided into K sections by ..., l _k , ..., l _K+1 , vectors b' ₁ , b' ₂ , ... in each section (l _k , l _k+1 )
..., b' _K is a compression pattern C. Each division (l _k ,
The vector b′ _k at the center of l _{k + 1}
The vector b _(lk+lk+1)/2 at the center position of is extracted as the representative vector b' _k of the section. The breakpoint sequence {l _k } is determined as follows. As shown in Figure 1b, the amount of pattern change D _j of standard pattern B
is expressed as inter-vector distance d _j =‖b _j -b _j-1 ‖, then the sum total of this inter-vector distance d _j over the pattern section is expressed as D _j = _j 〓 ^m=1 d _n (8). The breakpoint l ₁ ,...l _K+1 is the total variation D _J of the pattern variation D _j over the entire pattern.
The standard pattern B is determined by dividing the standard pattern B into K sections so that the standard pattern B is divided into K equal parts. In other words, if the amount of change divided into equal parts is ΔD, then the amount of change D _j of the vector sequence {b _j } from the beginning of the vector sequence {b _j } is the amount of change divided into equal parts, as shown in FIG. 1b. The point beyond ΔD is l ₂ ,
Determine as l ₃ ,..., l _k ,..., l _k+1 . The breakpoint sequence {l _k } is expressed as {l _k }=l ₁ , l ₂ , ..., l _k , ..., l _k+1 (9). the median vector between this breakpoints
If b′ _k is expressed as c _k again, the compression pattern C is expressed as C=c ₁ , c ₂ , . . . , c _k , . . . , c _k (10). By setting K to a value smaller than J, compressed pattern C can be stored in a memory with a smaller capacity than standard pattern B. The above is the principle of the present invention.

(Embodiment) FIG. 2 shows a block diagram of a first embodiment of the present invention which operates based on the above principle. Further, a detailed diagram of the compression processing section 40 is shown in FIG. An audio signal input from a microphone or the like 10 is converted by the audio analysis unit 20 into a time-series pattern of feature vectors expressed by equations (6) and (7) representing a frequency structure. The pattern is stored in the audio pattern memory 30. Here, when creating a standard pattern from input audio, standard pattern B is stored in the audio pattern memory 30. The compression processing unit 40 reads the vector values b _j and b _j-1 from the audio pattern memory 30 and calculates the distance calculation unit 4 .
1, the vector distance d _j =‖b _j −b _j-1
Calculate D _j . By changing j from 1 to J in this process, the pattern change amount D _j , (j=1, . . . J) over the entire pattern is calculated and stored in the pattern change amount memory 43. On the other hand, the control section 70
Now, the compression ratio is determined based on the standard pattern length J.

Next, in the compression processing section shown in FIG. 3, the number K of compression vectors is determined by the compression rate determining circuit 71 using the constant C _J. By determining the number K of compressed vectors, the section change amount determination circuit 72 calculates the equally divided amount of change ΔD by dividing the total amount of change D _J over the entire pattern into K equal parts. Using this equally divided amount of change ΔD as a reference, k is sequentially varied from 1 to K, and kΔD is calculated by the multiplication circuit 73 and output to the comparison unit 47 of the compression processing unit. The comparison unit 47 sequentially changes J from 1 to compare the overall pattern change amount D _j and kΔD, and determines a break point by setting the point where D _j exceeds kΔD as j=l _k . By determining the breakpoints, k is incremented by 1 and processing is performed until k=K, thereby determining the breakpoint sequence {l _k } _{k=1 to K.} From this breakpoint sequence, the center vector between the breakpoints is the representative vector for the standard putter.
b _k and stored in the compressed pattern memory 50 as {C _k }
Store as . Each vector C _k of the compression pattern
are respectively taken as representative vectors within the interval, and a well-known pattern matching method is executed by the recognition processing unit 60 between them and the input pattern A that will be input later.

In the method of determining compression vectors in the first embodiment, the number of compression vectors is determined based on the pattern length J of the pattern to be compressed. In this method, the number K of compressed vectors is uniformly determined by the pattern length J, regardless of the total amount of change in the pattern. Therefore, the compression effect on so-called flat patterns, where the total amount of change in the pattern is small, is small. The principle of the second embodiment, which is improved in this respect, will be explained with reference to FIG.

In FIG. 4, a and b illustrate the pattern compression method according to the first embodiment, and c and d illustrate the method according to the second embodiment. In the first invention, the number of break points K is determined by dividing the pattern length J by a constant C _J that determines the compression ratio, as K=J/C _J. Therefore, the number of break points K is uniformly determined from the pattern length J. In FIG. 4a, the case of K=6 is shown. On the other hand, FIG. 4c shows a compression method for pattern _B2 having the same length _J as pattern B1 in FIG. . Figure 4 b and d are pattern B ₁ and pattern
It shows the total amount of change D _1J and D _2J of B ₂ , and the compression ratio C _J
are the same, the equally divided changes ΔD ₁ and ΔD ₂ are
Depends on the amount of pattern variation. In this case pattern B ₂
Since pattern B ₁ has a larger amount of variation than pattern B 1, ΔD ₂ >
ΔD becomes ₁ . From this, in the compression method according to the first embodiment, the degree of approximation between the compressed pattern and the original pattern is largely influenced by the amount of variation within the pattern. The second embodiment realizes a compression method that is independent of the amount of variation within the pattern. Now, the fourth
In c and d of the figure, the amount of change ΔD divided into equal parts is given as a constant value C _D , and the number of compressed vectors K is K=
When determined as D _2J /C _D , the amount of change ΔD that is divided into equal parts is
is unrelated to the pattern length J. That is, if the change amount ΔD (=D _J /K) to be divided into equal parts is set to a constant value C _D , and the break point sequence {l _k } is determined by the method of the first invention, it will depend on the pattern change amount. A compression pattern with a constant degree of approximation can be obtained without any problems.

The second embodiment will be described below. FIG. 5 is a detailed diagram of the compression processing section 40 and the control section 70 according to the present invention. The other parts have the same functions as those in FIG. 2 according to the first invention. Standard pattern B is currently stored in the audio pattern memory 30. Similarly to the first embodiment, the pattern change amount D _j of equation (8) is calculated and stored in the pattern change amount memory 43. In the control unit 70,
The number K of compressed vectors is determined by the compressed vector number determination circuit 74 by dividing the total amount of pattern change D _J by _{CD, using a constant C D that} determines the amount of change to be equally divided. By determining this number of compressed vectors K, the subsequent processing is performed in exactly the same way as in the first embodiment,
The break point sequence {l _k } _{k=1 to K} and the compressed pattern {C _k } are stored in the compressed pattern memory 50, and the recognition processing unit 60
The pattern matching method is executed.

(Effects of the Invention) In this embodiment, it is possible to compress a standard pattern using a representative vector with an extremely small amount of calculation, and compared to the conventional method, it is possible to significantly reduce both the amount of memory and the amount of calculation. The error caused by the above compression is small and does not reduce the recognition rate.
Note that by using a similar configuration for feature vector sequences other than voice patterns, it is possible to significantly reduce the amount of memory and the amount of calculation without similarly reducing the recognition rate.

[Brief explanation of drawings]

Fig. 1 is a principle explanatory diagram explaining the principle of compression processing in the present invention, Fig. 2 is a block diagram showing a first embodiment of the invention, Fig. 3 is a configuration diagram of a compression processing section and a control section, and Fig. 4 The figure is a diagram for explaining the principle of the second embodiment of the present invention, and FIG. 5 is a diagram showing an example of the configuration of a compression processing section and a control section in the second embodiment of the present invention. In the figure, 10...input terminal, 20...speech analysis section, 50...sound pattern memory, 40...compression processing section, 50...compression pattern memory, 60...
Recognition processing unit, 70...Control unit, 41...Distance calculation unit, 42...Accumulation unit, 43...Pattern change amount memory, 44...Comparison unit, 71...Compression rate determination circuit,
72...section change amount determining circuit, 73...multiplying circuit, 74...compression vector number determining circuit.

Claims (1)

[Claims] 1. Sequence of feature vectors {b _j }, (j=1, 2,...
..., J), and the pattern change amount D _j of the vector sequence {b _j } stored in the pattern memory, (j=
1,...J) over the entire pattern, and divide the vector sequence {b _j } so that the total variation D _J of the pattern variation D _j is divided into K equal parts by the number of divisions K determined from the pattern length J. Divide into K sections, determine the segment point sequence _{ l _k } obtained by the above calculation, and
a compression processing unit that reads a vector b′ _k at the center position of each section determined by the above from the pattern memory and outputs a compressed pattern C=c ₁ , c ₂ , . . . c _K as a representative vector of each section; 1. A pattern recognition device comprising: a compressed pattern memory that stores C; and a recognition processing section that compares an arbitrary input pattern with the compressed pattern. 2 Sequence of feature vectors {b _j }, (j=1, 2,...
..., J) and the pattern change amount D _j of the vector sequence {b _j } stored in the pattern memory, (j=1,
_{2 , .} Divide the vector sequence {b _j } into K sections and obtain the segment point sequence {l _k } obtained by the above operation.
is determined, and the vector b' _k at the center position of each section determined by {l _k } is read out from the pattern memory and the compressed pattern C=
A compression processing unit that outputs c ₁ , c ₂ , ...c _K , a compression pattern memory that stores the compression pattern C, and a recognition processing unit that compares an arbitrary input pattern with the compression pattern. A pattern recognition device characterized by: