JPS6265090A - Pattern comparator - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a pattern comparison device that identifies an input pattern by comparing an input pattern with a plurality of types of standard patterns represented by a series of feature vectors. In particular, the present invention relates to a pattern comparison device applicable to recognition of consecutively uttered word sounds.

Conventional Technology If voice, which is the most natural means of generating information for humans, could be used as an input means for a human-machine system, the effect would be enormous. In this case, it is desirable for the speech recognition device to be able to recognize continuously uttered speech as a condition for recognizing more natural speech.

A pattern comparison device using the so-called two-stage DP method, which uses two stages of dynamic programming (hereinafter referred to as DP), has already been put into practical use as a pattern comparison device that is effective in recognizing continuously uttered word sounds. .

However, this two-stage DP method involves many duplicate calculations and requires a huge amount of calculation.Therefore, a pattern comparison device that can perform real-time processing using this method is required to have extremely high-speed processing, and the device is also complex and expensive. It has no choice but to become a thing.

In order to eliminate the above drawbacks, the ○(N)DP method is used as a pattern comparison device that requires significantly less calculation.

Pattern comparison devices using the CWDP method, ACDP method, etc. have been proposed. However, in order to implement these methods in hardware, a large capacity memory that can be accessed at high speed is required.

The present invention provides a pattern ratio axis device that eliminates these drawbacks, and is particularly directed to patent application No. 58-48600.
The A CD P (Aug
Mented Continuous DP Match
ing) for the purpose of applying the law. This method can be applied to both cases where the number of words is unknown and the number of words is known, and automatic control is also possible.

Before explaining the present invention, the ACDP method will be explained next. Hereinafter, for the sake of simplicity, a case where the number of words is unknown will be explained. Although the pattern comparison device of the present invention can be used to recognize various input patterns,
In the following, continuous word speech that is continuously uttered will be explained as an example.

The sounds uttered by humans change from time to time depending on the person, and are non-linearly expanded and contracted in time with respect to a standard pattern that serves as a reference. In order to compare this non-linearly expanded and contracted input pattern with the standard pattern and perform input speech recognition, we need to non-linearly associate each feature vector of the input pattern with the standard pattern, and determine which standard pattern the input pattern corresponds to. It is necessary to calculate which is most similar to However, even though this input voice expands and contracts nonlinearly, it does not become abnormally long or short.

Focusing on the physical characteristics of such input patterns, when comparing input patterns and standard patterns, rather than comparing all possibilities without limit, the validity determined by the physical properties of the input patterns Try to make comparisons within a range of possibilities.

The input audio signal is processed by a pattern comparison device through frequency analysis, LPC analysis, PARCOR analysis, correlation analysis, etc.
It is converted into a series of several sets of numerical values (feature vectors), and each vector is compared between the series of feature vectors of this input pattern and the series of feature vectors of the standard pattern to be compared.

This comparison value for each vector, that is, the cumulative distance, which is the sum of distances between vectors, is used as a measure of pattern similarity. When calculating this cumulative distance, comparing each vector for all combinations would result in an enormous amount of total n, making it impossible to put it into practical use as a pattern comparison device.

Considering a plane with the input pattern on one axis and the standard pattern on the other axis (hereinafter referred to as the ij plane), the combination of each vector of the input pattern and the standard pattern is a plane on the ij plane. can be shown by each grid point (hereinafter simply referred to as a point). Therefore, calculating the distance between vectors for all combinations means calculating the distance between vectors at each point, and calculating the cumulative distance means calculating the distance between the feature vector of the input pattern and the corresponding standard pattern. The distance between the vectors and the feature vectors is sequentially calculated and summed. The correspondence between the feature vectors of the input pattern and the standard pattern, that is, the sequence of points selected in the process of calculating this cumulative distance, is called a path. Limiting the range of comparison in consideration of the physical properties of the input patterns described above means setting constraints on route selection. Here, terms and symbols used in the following description will be explained.

A: Input pattern (A-al, a2..., ao,...
....

al) al: Feature vector I of the i-th frame of the input pattern A: Number of frames of the input pattern Rn: nm-th quasi-butter 7 (Hn=bnj, bn2
．． ,,,,bnj.

. . . , bnyn) bnj = Feature vector 1 of the [th ] frame of the nth standard pattern: Number of frames of the nth standard pattern n: Word name (word number) 1≦n≦N N: Total number of standard patterns dn ( i, i): Inter-vector distance Dn (s:t) between the feature vector bnj of the j-th frame of the n-th standard pattern and the feature vector a of the i-th frame of the input pattern Dn(s:t): Partial distance of the input pattern (turns a3 to
Minimum cumulative distance between at and standard pattern Rn (hereinafter referred to as partial cumulative distance) D(i): The minimum cumulative distance between the first to i-th frames and the combination pattern of the optimal combination of each standard pattern. Distance between patterns (hereinafter referred to as end cumulative distance) N(i): Constitutes the combined pattern when determining the optimal combination of each standard pattern for the input pattern from the first to the i-th frame. Number indicating the last standard pattern (hereinafter referred to as the last word name)B(
i): Frame number indicating the frame one before the starting point frame of N(i) (hereinafter referred to as pack pointer) Dn(i, j): Partial pattern am+1 of the input pattern
The minimum value for m of the minimum cumulative distance between ~a and the partial patterns bn1 to bnj of the standard pattern Rn (hereinafter referred to as intermediate cumulative distance) B''(i, i): Dn(i+i) ヲfcf Input back pattern cD starting point position (hereinafter referred to as intermediate back pointer) However, the path for DP matching is Dn (s:t).

D(i) is a path whose basic axis is the axis on the input pattern side, and Dn(i, i) is a path whose basic axis is the axis on the standard pattern side.

Continuous word speech recognition using DP matching is efficiently performed by calculating the following recurrence formula. That is, B(i)=BiN(i)=n", where m", n' is m that satisfies formula (1), and n is i
If you calculate sequentially for = 1 ~ engineering, you get N(I), N(B(I
)) , N(B(B(I)) ) ,..., N(B(
J...B(I)...))) The input words are found in the reverse order.

FIG. 2 is a diagram explaining the physical meaning of the calculation of equation (1). b is an example of the constraint condition of the path when calculating Dn(m:i). That is, the route leading to point ('+j) is point (i-2, 1-1), (i-1, j-1).

('-' r j-2), and when the cumulative distance is calculated, the distance between vectors at each point is weighted according to the selected route, as represented by the number shown on the route. If weights are given in this way, no matter what route is selected under the route constraint conditions shown in the figure, the sum of the weights along that route will be equal to the number of frames of the input pattern and will be constant.

a shows how partial patterns from the n+1 frame to the i-th frame of the input pattern are matched with the n-th standard pattern Hn. Straight line 1 has a slope of %, straight line 2 has a slope of 2, and the path of DP matching is limited to the area sandwiched by both straight lines under the constraint b. 3 is an example of a selected matching path.

In order to calculate equation (1), Dn(m+1 : i
) must be calculated for each of i=1 to I over a range of m related to i and 10. That is, in the example of FIG. 2a, the inter-vector distance and cumulative distance must be calculated for all grid points within the triangular area surrounded by straight line 1.2 every time i changes.

The ACDP method substantially reduces the amount of calculation by approximately calculating Dn(m+1:i) by introducing the idea of word spotting.

FIG. 3 is a diagram explaining the concept of the ACDP method.

FIG. 5A is a diagram showing an example of constraint conditions for a route. That is, the point (
'+1) is the point (i-2, j-1), (1-1
, j-1) and (i-1, j-2), and the number shown on the route represents this weight. The essential difference from the one shown in Figure 2 is that if the weights are given in this way, no matter what route is chosen under the route constraint conditions shown in the figure, the weight along that route will be The point is that the sum is equal to the number of frames of the standard pattern and remains constant.

To summarize the idea of the ACDP method, (1) Dn(i+Jn) +Bn according to the constraint condition of the path whose basic axis is the axis on the standard pattern side as shown in Figure 3a.
Calculate (i, Jn).

j2) m+1-n=(i, In), 1) n(
t, ■”) to Dn(m-r+1:i) ~Dn(m+
approximately estimate r+1:i). That is, Dn(m+1:1)=Dn(i,In)fist-Jn. Here, r is the range of estimation of D''(m+1:i).

From (1) and (2), we were able to approximately estimate Dn(m+1:i) with the starting point at any point in B"(111)-"1m+1≦Bn(ijn)+r and the ending point at l. become.

When this method is applied to continuous word speech recognition, the result is as follows.

(1) Initial conditions Ryu) = Q, 13 Yu) = 0 Dn (-1, 1) = Dn (o, j) = oo, for
n = = 1 - N, j = 1 ~ J'' (2) Execute (3) ~ ('7) for i = 1 ~ I) For n = 1 ~ N (
4) Execute (6) (4) Initial value Dn (l, 1) - dn
(i, 1) (5) For i = 3 ~ Jn (6) For r' = o ~ r m + 1 = Bn (i, Jn) (D (m') + rf'(m' + 1 : 1) 1 (7 4)
−07=1°−0,. . 70. . -14,N,3(8)
Judgment process (determining the word string from B(i), N(i) in the reverse order as described above) As is clear from the above process, according to the ACDP method, dn(l++ Since the calculation of step (41, (to) including the calculation of ) only needs to be performed once for each i and n, the amount of calculation is significantly reduced compared to a normal two-stage DP.

However, if the distance between the vectors dn(i++) and the intermediate cumulative distance Dn(111) are changed to n=1 to N for every 7 V -mu i,
.

In order to obtain j=1 to Jn, fairly high-speed access is required as a memory for storing standard patterns. For example, set the frame period to 5m5ec
, number of standard patterns N=300. If the average number of frames of the standard pattern is 1 = 30, the average number of grid points to be calculated per frame of the input pattern is N 1 = 9000.
If the number of dimensions of the feature vector is 15, the number of accesses to the standard pattern storage memory per frame of the input pattern is 9000×15=135000.

In order to do this within 5 mBeC, an access time of 37 n5ec or less is required. In addition, the amount of memory required to store the standard pattern is 1sskB if one element of the feature vector is expressed in one byte, which means that 135kB of high-speed memory that can be accessed in 37 n5ec or less is required. This makes the device very expensive.

Problems to be Solved by the Invention It is an object of the present invention to significantly reduce the number of memories that require high-speed access in a continuous turn comparison device based on the ACDP method.

Means for Solving the Problem The pattern comparison device of the present invention converts the input signal into feature vectors a 1 , a 2 r...+ a 1 r...,
a feature extraction means for converting into a series of feature vectors bn1. bn2゜..., bn3. ...,bn■
A standard pattern storage means for storing a standard pattern Hn consisting of n (where n=1 to N), and a lattice graph having an input frame i as a horizontal axis and a standard pattern frame j as a vertical axis. When matching with, for each of the feature vectors bnj of each frame j of the standard pattern, DP matching with the width of W from j=1 to
(2) A DP matching means that performs the matching calculation over the entire range of the input pattern, such that the matching calculation is performed in the order of 0, and when the calculation of one area is completed, the calculation of the next adjacent area similar to the above is performed in the same manner. It is equipped with

The input signal is transformed into a feature vector a1, a2,..., a
,,... .

Convert to a series of aI, and create a series of feature vectors bn1゜bn
3. ..., bnj, . , when matching with the standard pattern Rn, for each of the feature vectors bnj of each frame j of the standard pattern, i In the axial direction, perform DP matching with the width W in the order of j = 1 to ■0,
When the calculation for one area is completed, the matching calculation is performed for the entire range of the input pattern, such that calculations for the next adjacent area similar to the above are performed in the same manner.

Embodiment FIG. 4 is a diagram showing the principle of the present invention. If the limiting conditions for the route are chosen as shown in FIG. 3, and 8.9 is a straight line parallel to the j-axis and separated from each other by W (≦In/2) frames, then the cumulative distance along the straight line 8 is Dn(i.

)), it is possible to calculate the intermediate cumulative distance Dn(i, i) of the area sandwiched by both straight lines (shaded area), and furthermore, the intersection of the straight line 8 and the straight line t, )n can be calculated. 10
．． If the intersection of straight line 9 and straight line j=J'' is 11, then 1
The terminal cumulative distance D(i
) can be calculated if the terminal cumulative distance before straight line 8 is determined. The intermediate calculations can be performed horizontally from bottom to top as shown by arrow 12. In this way, when the frame period is T, the calculation for the shaded area A is performed as described above during the time WT, and the calculation for the shaded area B is performed as described above during the next WT. In addition, it is possible to proceed with the calculations described above sequentially for each time WJ.

In this way, the jth frame of the standard pattern becomes W
It can be constant while matching with the input pattern of the frame, so
The access time required for the standard pattern memory can be slowed down.

That is, the number of grid points in each shaded area is WJ, and calculations for each grid point must be performed during TW (T
is the frame period), then the calculation time allowed for one frame of the standard pattern is TW/IN, which is equal to the time to read the data for one frame of the standard pattern, and the calculation time allowed for one frame of the standard pattern is TW/IN. If we apply the numerical value of W
= J/2, so T (1/2)/ TN-=T/
2N = 5 msec/2X300: 8.3 μsec, and if the order of the feature vector is set to 16 as before, the access time allowed to read out the elements of one vector of the standard pattern is 8.3 psec/1.
5=553 nsec, which means that the access time is approximately 16 times longer than the value required in the conventional example. However, this time, high speed is required of the memory for storing the input pattern (the first frame of the standard pattern must be read NW times during the time TW/J. For example, TW/T/NW=
T/NT=5msec/300X30=0.561ts
It is necessary to read out 16 feature vectors during ec. Therefore, the access time allowed to read one element of the vector is 37 n5ec. The required capacity is, for the above example, an average of 15 W = 15
X30=450 bytes is enough, leading to a significant cost reduction.

The above is the principle of the present invention, but W can be fixed or variable. When W is fixed, basically, when the minimum value of work 0 is ■□, W≦J, /2
The easiest way is to do this, but it costs 0! It is not very effective when very short items are included. However, if memory is used, W>Jll.

Even in the case of /'2, calculations can be performed using the above concept. In other words, Do(i, In) itself can be calculated using the width of the input frame of the jade badge, but from now on, D
In order to find n(i), it is necessary to find the meridian cumulative distance to i-J, 7/2. Therefore, W)T,
/2, calculate Dn(1) at i and part of W-IJ2Fvv (t, In).

It is sufficient to memorize Bn(i, ■''). Also, in order to compensate for the above-mentioned drawback, it is also possible to make W variable according to Jn.

In this embodiment, a case will be described in which W is made variable depending on In.

FIG. 6 is a detailed diagram illustrating the invention in detail.

Processing when the input pattern is the first frame will now be described. Each of the diagonal lines A, B, etc. will be called a block. At this time, if Dn(i-IJ''/2)) has not been determined, the above calculation is performed for the grid points indicated by ○ in the same figure, and Dn(i-[:In/2)),
Dn(i −(In/2) +1 ), ・, Dn(i
).・is i of the block before this block
=1.2. ... The intermediate cumulative distance Dn('++) at these grid points in Jn becomes the initial value of the intermediate cumulative distance Dn('++) in this block. When Dn(i-CI''/2)') has been determined, the calculation for the standard turn corresponding to n is skipped.

By doing the above, the width of the block for each standard pattern Rn is calculated as Wn=(In/2)+1 for every n. To determine whether Dn(i-(In/2)) has been found, first
This can be determined from the fact that if initialization t and te are set in cx:□, then when it has been determined, it should be a finite value other than 0 or (1). Or, when Dn(i) is calculated, Gn(
i) - 〇, a flag Gn(1) is set to Gn(i) = 1 when it cannot be determined, and the initial value is set to G''(i
)=1 for i=1~I, and Dn
If we set G''(i) - 0 each time (i) is determined, it can be determined whether Dn(i) has been determined by looking at G''(i).

FIG. 1 shows a feature extractor according to an embodiment of the present invention based on the above principle, which converts an input audio signal into a series A of feature vectors a. Reference numeral 102 is a word standard memory unit, in which N words as recognition school meals are each stored in a standard pattern Rn.
= b nl,...Br),,...

It is registered in advance in the form of a feature vector as bnIn (1≦n≦N). Reference numeral 103 denotes an inter-vector distance calculation unit, which calculates the distance dn (mr + ) between the feature vector ai of the first frame of the input pattern and the feature vector b of the n-th standard word ζ turn Hn, and calculates the distance dn (mr + ) between the Memorize until it disappears. In this example, the vector distance between the frame immediately before the frame for which the partial cumulative distance is being calculated and the frame is stored until the intermediate cumulative distance of the relevant frame is calculated. dn(m,)) is, for example, ~
It can be defined as the city distance of and bnj. That is, let the dimension of the vector be L, and the radius = (aml, aXn2..., a
mL), bnj=(bo, 1°bnj2. . . . , ,
(mr + )= Σ D
mk-bnjk'=1. In addition, in this embodiment, the direction of calculation is the uniaxial direction, and for the vector b of the j-th frame of the n-th standard pattern Hn, an input pattern in the range of 1-(Tn/2)≦m≦i is calculated. Find dn(m,j) for the above.

Reference numeral 104 is an intermediate cumulative distance calculation unit, and the cl”(m
, j), the intermediate cumulative distance D''(m, j).

An intermediate back pointer Bn (m, j) is determined. Reference numeral 113 denotes a terminal cumulative distance calculation unit, which includes D (e), B (e), N
- is determined in the same range as d0(m++). When the matching path constraint conditions shown in Figure 3a are adopted,
Dn (m, j) and B'' (m, j) are determined by the intermediate cumulative distance calculation unit 104 as follows.

When j = 1, Dn (m, 1) dn (m, j) B'' (m, 1) = m ) ≧ 3, where p (e), B (e), N) is the terminal cumulative distance calculation Part 11
3, it can be found as follows.

That is, calculate p=Bn(m, 5) m'=p-r+++p+r for m', and use the recurrence formula to find B as)=m''N-nl.However, m'', n'' are m that satisfies formula (2)
'.

It is n.

112 determines the execution range of the calculation of the distance between vectors and the calculation of the cumulative distance for each n, and D(
1-(Jn/2)) has not been determined, d'' (m, j) and Dn (m, j) are calculated within the range of m, respectively, by the inter-vector distance calculation unit 103.The interim cumulative distance calculation unit 1
04. An instruction is given to the terminal cumulative distance calculation unit 113.

The end cumulative distance D- obtained as above is stored in the end cumulative distance storage section 105, the bank pointer B (e) is stored in the back pointer storage section 106, and the last word number N) is stored in the end word storage section 107. is memorized.

In addition, Dn (m + i) + Bn (m, j) (however, j =
12 + ”' + 1 ”; n=1.2 + ・
...+N) is temporarily stored in the partial cumulative distance calculation unit 104 until it is no longer needed. In this embodiment, the intermediate cumulative distances of the frames immediately before and two frames before the frame for which the intermediate cumulative distances are being calculated are stored until the intermediate cumulative distances are calculated.

Furthermore, the terminal cumulative distance stored in the terminal cumulative distance storage unit 105) is a necessary initial value of equation (1), and D
) can be memorized until Dn(m+1.1) is determined.

Reference numeral 108 denotes a voice section detecting section, which determines a voice section from the magnitude of the input signal and the like. Voice section detection unit 1
08 detects that audio input has started, the frame number counting section 109 starts counting each frame. The above processing was for the i-th frame, and the count value of the frame number counting unit 109 is set to i.

Therefore, the same processing as described above is performed every time the frame advances by one. The frame number counting unit 109 starts counting when a voice section is detected, and is reset when the voice section ends. Last word storage unit 107. Therefore, in the pack pointer storage unit 106, N(i), B(i)
is i=1.2. ..., 3 will be remembered about ■.

The segmentation section 110 is a pack pointer storage section 1.
06 to issue an instruction to read a predetermined pack pointer. That is, the segment/yeon section 110
issues a value i to the pack pointer storage unit 106,
The pack pointer B (
1) is read out.

Segmentee 7. Upon receiving the value B (i) from the pack pointer storage section 106 , the transfer section 110 issues the same value to the bank pointer storage section 106 . Therefore, when the C1 voice section detection section 108 detects the end of the voice input, the final value I of the frame number counting section 109 is supplied to the segmentation section 110, and the segmentation section 110 first stores the value in the bank pointer. Issued to Section 10 et al. Thereafter, according to the operation described above, the pack pointer storage unit 1
From 06, B (I), B (B (I), B (B (
An output of B(I))) , -=, o will be obtained from time to time. These values are the frame at the end of the second-to-last word, the frame at the end of the third-to-last word,
This is the frame at the end of the fourth word, and since N(i) is a word that ends in the l frame, if this value is given as is to the last word storage unit 107, the last Recognition results are obtained in reverse order starting from the word.

Note that in order to obtain recognition results in the original order, this order conversion is performed in the pointer storage unit 106.
The process may be performed on the output of the last word storage section 107 or on the output of the last word storage section 107.

In the above embodiments, for convenience of explanation, the point (m
, j), the inter-vector distance, intermediate cumulative distance, intermediate pack pointer, etc. are dn("+ 1) + s(
mIIL"("+]) etc., but these storage areas do not need to be prepared over the entire range of i, i, n.
Dn("+l)+" at the lattice point of the first stage
(mI ] ) is calculated using the i-1th stage for the route shown in FIG. 3, as is clear from the recurrence formula above.

For dn(m+] ') (However, In "l
-102 to l) only need to be stored. Also, if we calculate these values for j = 1 to Jn for each n, then for "(m++)" and "Bn(l++)", we will have to calculate for one block regarding n, so when There is no need to remember it.

Based on this idea, a more detailed explanation will be given with reference to FIG. Of course, this can also be followed when implementing it using software. In the same figure, the notation NDDO means executing Y while X is satisfied.
The notation O means to execute Y until the condition X is satisfied, and the notation ENDIF means to execute Y until the condition X is satisfied.
Otherwise, each means to execute Z.

In addition, the notation regarding intermediate cumulative distance, intermediate pack pointer, distance between vectors, etc. will be changed as follows. That is, the third
When using the route shown in the figure, the calculation of the intermediate cumulative distance and intermediate pack pointer of the 1st stage is based on the intermediate cumulative distance and intermediate pack pointer of the i-1st stage and 1-2nd stage, and the intermediate pack pointer of the i-1st stage and the j-1st stage. It is sufficient to store the inter-vector distance of the first input pattern of the block, and the intermediate cumulative distance and intermediate puncture pointer in the frame of the first input pattern of the block can be calculated using the intermediate cumulative distance of one frame and two frames before, the intermediate pack pointer, and one frame. It is sufficient that the previous inter-vector distance is stored. In the end, the required memory is "(il)".

For Nn(i,+), 1=-1~k. 3 steps in the range of +1, for dn(i, i) i=0 to 1o+1
(However, -1 means one frame before the last frame of the immediately preceding block, and 0 means the last frame of the immediately preceding block.), one frame before the last frame of each block. and the final frame is Dn('I)
If dn(i, i) in the final frame of LE3n(lIIL each)07 is memorized, it is possible to calculate the recurrence formula, and we will perform calculations for one block for each word n. For example, "('l]) for 3 stages of 1=-1 to J0+1.

Since Bn('l]) can be commonly used for all n by setting 1=-1 to J-+1,
These are D(io, k), B(io, k) (however, 1o=-1~Ifn+1, mo~2), i
= 0 ~ engineering. 2 steps of +1 (7) dn(i, +) is 1=o
-IIn+1, it can be commonly used for all n, so it can be used as d(io,
k) (However, 10=0~7 +1.=o~2
), and the memory for storing the initial value of the next block is Dn(1+l)+Bn('
+I), d''('+
+) respectively Dn(' l ]) lBn(' l I
Ldn(]) (where i=o ~1゜j=1~■0, where i=o is the last frame, and i=1 is the frame immediately before the i-most frame). Significant memory savings can be achieved.

In FIG. 6, steps 201 and 202 are initialization steps. Step 203 is a part that performs processing on the i-th frame of input.

Step 204 is a process of matching the first to Nth standard patterns and the input pattern for each grid point in the block. Step 206 determines the width of the block for the nth standard pattern. Step 206
determines from the state of the flag Gn (e) whether or not the end cumulative distance from the first To frame to the i-th frame has been determined in the input i-th frame, and if it has not been determined, calculates that block. This is a process in which the process skips to the process for n=n+1 when it has been found.

In step 2o, when it is found that the corresponding block has not been calculated yet, the block is calculated in the following step, so the flag G0 (e) is set to 0 for the frame included there. It is now in a calculated state. In step 208, for convenience of later calculations, the starting frame -1 of the block to be calculated is
is calculated as mo. In step 209, j=1.
200 (io, J) (however, 10=1~Eng.+1) is calculated. Step 211 determines the frame serial numbers from the first frame of the input pattern. Step 2
12 calculates the inter-vector distance between the m-th vector convergence of the input pattern and the first vector bn1 of the standard pattern. Based on this value, step 213 calculates D(to, i)+B('o) for j=1.

J) (however, 10=1~■.+1) is calculated.

Steps 214 and 215 are j=2, 1o=1~T o
The inter-vector distance d (io, 2) at +1 is calculated.

Based on this value, step 216 performs D(io
, i), hundred (io, i) (however, 10 = 1 ~ k. +1)
is being calculated. Step 217 performs the replacement d'(2)=d(J.+1.2) for use in the calculation of the next block.

Step 218 is D (10°k) at j=3~J0
, B (io, k) (however, 10=1~k).

+1) is calculated. Step 219 is D(50, k
), B(io.

k ) + d (t o , k )
In order to use J' cyclically, it is set according to the value of k and the value of ti. The steno knob 220 has its initial value set from the value of the immediately previous block. Step 221 calculates the distance between vectors in the third stage. In step 222, based on this value, D(10, J), B(10°)) (where 1°-1 to T o +1) in the ]th stage are rewritten. Step 223 replaces the blocks to be used in the calculation of the next block.

Since step 224 was obtained as described above, D (e), B (e), and N (e) are calculated according to the above explanation, and each time n changes, it is compared with the previously calculated value and the step The minimum value is found systematically.

When the process of step 203 is completed, determination process begins. That is, in step 230, i = I, and the word string to be found in step 231 is found in the reverse order according to the above explanation.

In this example, the maximum inclination of the path is 2 because the constraint conditions for the path as shown in Fig. 3 are used, so the width of the block is (Jn/2) + 1. There are various other constraints as well, such as:
In the case shown in FIG. 7, the maximum slope is 3, so the width of the block in this case is 2-4l-(In-1)/3]. Generally, when the maximum slope of the path is k, the block width W2-(1-,-(In-1)/) should be less than 'l' (if k = 2 (') then 1d , C
Jn/2) +1 t2-(1-(J"-1)/2) is the same as ■1 if it is an integer).

As described above, according to the present invention, like the ACDP method,
The method of performing DP matching continuously at high speed required high-speed, large-capacity memory, but now the number of memories that require high speed can be significantly reduced, making it possible to realize inexpensive equipment. This is what became.

[Brief explanation of drawings]

Figure 1 is a block diagram showing the principle of the present invention, Figure 2 is a DP
An explanatory diagram for explaining the principle of recognizing continuous patterns by matching, Figure 3 is an ACD to which the present invention is applied.
Fig. 4 is an explanatory drawing to explain the principle of the present invention, Fig. 5 is a detailed explanatory drawing to explain the present invention in detail, Fig. 6 is an explanatory drawing to explain the principle of the present invention. A processing procedure diagram for explaining the processing procedure of the invention in detail, FIG. 7 is an explanatory diagram for explaining another example of the constraint condition of the route. ...Word standard pattern storage section, 103... Inter-vector distance calculation section, 104... Intermediate cumulative distance calculation section, 105.
... End fine distance storage section, 106 ... Back pointer storage section, 107 ... Last word storage section, 108 ... Mark/voice section detection section, 109 ... Frame Number counting section, 110... Sedamentation seventh section, 112... Manonang calculation execution range determining section. Name of agent: Patent attorney Toshio Nakao Haga 1 person No. 4
Figure 5 Figure 6 (f/)su2DI, L-[Jll, ヮ/2] Aygo7 i Σ Figure 6 (+04) A “I” Ge230 i=I Figure 7

Claims (1)

[Claims] The input signal is converted into feature vectors a_1, a_2, ..., a_
a feature extraction means for converting into a series of i, ..., a_I;
Series of feature vectors b^n_1, b^n_2,...
A standard pattern storage means for storing a standard pattern R^n (where n=1 to N) consisting of b^n_j, ..., b^n_(J^n), and input frame i as the horizontal axis, In a lattice graph with the frame j of the standard pattern as the vertical axis, when matching with the standard pattern R^n, for the lattice points in the area sandwiched by straight lines separated by W frames from each other with respect to the frame of the input pattern, DP matching is performed with the width of W in the i-axis direction for each feature vector b^n_j of each frame j of the standard pattern j = 1 to J
DP matching means that performs the matching calculation over the entire range of the input pattern, such that the matching calculation is performed in the order of ^n, and when the calculation of one area is completed, the calculation of the next adjacent area similar to the above is performed in the same manner. A pattern comparison device comprising: