JPS61102697A - 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 Industrial Application The present invention relates to a pattern comparison device that compares an input pattern with a plurality of types of standard patterns represented by a series of feature vectors and identifies the input pattern. The present invention relates to a pattern comparison device that can be applied to recognition of continuously uttered word sounds.

Conventional configuration and its problems 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 very large. 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 dynamic programming (hereinafter referred to as DP) two times, 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 perform extremely high-speed processing, making the device complex and expensive. It has to become something.

In order to eliminate the above drawbacks, O(nJ D P method, RCWDP method,
A pattern comparison device using a method has been proposed. However, implementing these methods in hardware requires a large capacity memory that can be accessed at high speed.

The present invention provides a pattern matching device that eliminates these drawbacks. This will be explained using the (n)DP method as an example. Before explaining the present invention, O(nJ
First, the DP method will be explained.

The pattern comparison device of the present invention can be used to recognize various input patterns, and will be described below using continuous word speech that is continuously uttered as an example.

The voices uttered by humans change from person to person and from time to time, and are non-linearly expanded and contracted in time with respect to a standard pattern that serves as a reference. In order to recognize the input speech by comparing this non-linearly expanded/contracted input pattern with the standard pattern, we need to non-linearly associate each feature vector of the input pattern with the standard pattern. Do you need to calculate which is the most similar? However, even though this input voice expands and contracts non-linearly, 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, we consider the validity determined by the physical properties of the input patterns. Comparisons should be made within the range that can be used.

The input audio signal is
゛Wavenumber analysis, LPG analysis, PARCOR analysis, correlation□
Through analysis or the like, it is converted into a series of several sets of numerical values (feature vectors), and each vector is compared with the feature vector 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 requires an enormous amount of calculation, and cannot be put to practical use as a pattern comparison device.

, considering a plane (hereinafter referred to as i-i plane) with the input pattern on one axis and the standard pattern on the other axis,
The combination of each vector of the input pattern and the standard pattern can be represented by each grid point (hereinafter simply referred to as a point) on the i-j plane. Therefore, calculating the distance between each vector for all the above combinations means calculating the distance between vectors at each point, and calculating the cumulative distance means calculating the feature vector of the input pattern and the feature of the standard pattern corresponding to it. This involves sequentially calculating and summing the distances between vectors. 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=a1.a2.・-, ai,・−
・, a), ai is the feature vector of the i-th frame, I is the number of frames of the input pattern Rn: n-th standard pattern (Rn=b?, ]lF2...
・, bt..., b7n) b9 is the j-th frame of the n-th standard pattern] Feature vector ■0 is the number of frames of the n-th standard pattern, where N is the total number of standard patterns, 1≦n≦N” ('* j): j-th frame of the n-th standard pattern.

Intervector distance D(i) between the feature vector bn of the input pattern and the feature vector a□ of the 8gi frame of the input pattern: the input pattern from the first to the i-th frame,
Distance between patterns with the combined pattern of the optimal combination of each standard pattern (hereinafter referred to as the terminal cumulative distance) N(i): Distance between the optimal combination of each standard pattern with respect to the input butter and Number indicating the last standard pattern that constitutes the combined pattern when the combined pattern / is calculated (hereinafter referred to as the last standard pattern name)
B(i): Number indicating the frame one frame before the starting point frame of N(i) (hereinafter referred to as back pointer)"('
+]) : Distance between the partial pattern from the 1st frame to the 1st frame of the input input butter and the partial pattern from the 1st frame to the i'th frame of Rn (hereinafter, partial cumulative distance D?' (i, j ) and D(i'-1)
The minimum value for i' of the sum of (hereinafter referred to as intermediate cumulative distance) , i.e. i'=argmin (D(i'-1)+M/ (i
. shows.

D"(i): is the intermediate cumulative distance when j = I n, and D"(i) = D" (i, I") 't'
67).

B"(i): This is the intermediate back pointer when n=1", and B"(i)=B"(i, J").

0 (n) D P method assumes that the input pattern ends at the 10th frame, and when the last pattern is Hn, the intermediate cumulative distance D '' (z o ) is calculated by setting i' to the starting frame. DP matching is performed between a partial pattern A (i'-1, io) of an input pattern whose end frame is l'O and a standard pattern Rn with a free start point and a fixed end point. The initial value Dn (i', 1) of the intermediate cumulative distance and the m direction B''(i', o) of the intermediate back pointer are expressed as D''(i', 1).
)=D(i'-1)Id"(i,1)B"(i'
, 0)=E"(i'-1), DP matching of partial pattern A(i'-1°10) and standard pattern Hn is performed as a continuation of DP matching from the first frame to the first frame. By doing this, for example, under the constraint conditions for matching paths as shown in FIG. 1, the possible matching paths are limited to the region P in FIG. ) The calculation of d''('+1)+''(i+t) needed to calculate d''('+1)+''(i+t) only needs to be performed once for each point in the area P.In Fig. 2, the horizontal axis is the input pattern, and the vertical axis is the input pattern. is the last standard pattern Hn.A region P is a straight line P1 with a slope H
This is an area surrounded by a straight line 2 with a slope of 2.

To find ``('+1), D''(i-2, j-1), as is clear from FIG.

D”(i-11j-1)ID”(f-11j-2)Id
Since it is only necessary to know “(i-1tj)1d” (t, +), the intermediate cumulative distance D” (io,
j) (button = 1.2..., In), the intermediate cumulative distance D''(i-1゜j)tD''(i-21j ) oh! and 1st
−17v−m, distance between vectors on the first frame”('
−'+1)+”(' r 3 ) (Tata L s=1.
2, +++, J”) Just store it in your memory.

At this time, D '' (io) can be obtained as D n (zo) = D n (io, J '').

As described above, each time frame 1 of the input pattern advances by one frame, the intermediate cumulative pressure 1ifID"(
i + i) (Tata L, j-=1.2.・+1"; n
= 1.2°..., N) is the intermediate cumulative distance D"(i-1+i), D"(i-2,j
),! :17v-distance between vectors before the frame and in the current frame d'' (n=1.j), d'' (i+i) CftJ
r:I, n=1.2. ....

J''; n==1.2.・IN), D(i)'=
=s+m(D”(i.

Jo)], the final cumulative distance to the first frame can be determined. It was obtained in this way” ('+t
) (fc, l=' + 2*"'*I"+ "=1e
2r...'+N) until it is no longer needed, that is, D''(i,
It is stored until the calculation of j) is completed.

Further, the back pointer for D(i) (the value obtained by subtracting 1 from the starting point for CD(i)) B(i) is determined as follows.

Set the intermediate back pointer for D"(i, j) to B"('
I]), 1) Dn(iti)=D"(i-2+1-1)+d"
(i-1+i)+d"(iti) then B"('+3)=""('-2+1-1)2)
”(' + + )=”('-' I ]-1)”(
' + ]) Nodo @ is B"('+1) = B"('
−11]−1)3) D”(t+ +)=D”D−1,
j-2)+dn(i l Dotoki is Bn(i,
j) = Bn (i-1, 5-2), then B(i) = B combination (i, 1 combination). Therefore, for B''(t, +), remember the one frame before and the one before two frames.
In the case of a route like the one shown in the figure, "('++)".

B''('l]) only needs to remember the value of one frame before.

A conventional example of a pattern comparison device using the principle explained above will be explained. FIG. 4 is a block diagram showing a conventional example in which a pattern comparison device based on the above principle is applied to continuous word recognition. In the figure, In is an input terminal for an audio signal, and 1 is a feature extraction unit composed of a filter bank, etc., which converts the input audio signal into a series A of feature vectors a.

2 is a word standard pattern storage unit, which stores the recognized word N
Each word is a standard pattern R''=b〒.

...l b'j' l..., b'i'n, (1
<n<N) and is registered in advance in the form of a feature vector. 3 is an inter-vector distance calculation unit, which calculates the distance d'' (i, i) between the feature vector a in the i-th frame of the input pattern and the feature vector b of the n°-th word standard pattern HH, 5= 1.2..., In, and store it until it is no longer needed.In this example, the distance between the vectors of the frame immediately before the frame for which the intermediate cumulative distance 1 is being calculated, 1', and the frame concerned. "(ill) is stored until the intermediate cumulative distance of the frame is calculated." (ill) is, for example, a, and bj
It can be defined as the city distance of In other words, let the vector b dimension be 1, and “i:(a il +”i2””i
l) l b j=(b jl +”j2+...
・, b〒1).

4 is a cumulative distance calculation unit which calculates an intermediate cumulative distance "('I))" and a terminal cumulative distance D(i) for the i-th frame.

Intermediate bank pointer Bn (1, j), back pointer B
(i) as 1 = 1+ 2+”'+ J n: 1l=
=1t 2 , . . . + N is determined, and N(i) indicating the last word is determined. The matching path constraint conditions shown in Fig. 1 are adopted...0) Initial conditions, D"(i, 1) = D(i-1) + d"(i,
1) DCi)=min (D” (i, I”)
)g''(i, j), B(t) is oBn obtained from the following o formula
(iti) is Bn(i, 0) = Bn(i-1), the initial condition is 1
) ”('+])=”('2*+-')+
d"('l])+dn(i-1,0, then B"(ill)=B"(t-1+1-2) -
...----- (4), and B (i) is determined as n if the word number that satisfies equation (1) is B (i) = B" (i l 1") ---
...(6). Also, the terminal cumulative distance D(i) obtained as 0 or more where N(i) = n
= Dn (i, J") is the end JRi distance storage sK, back pointer B (i) = Bn (i, T") is stored in the back pointer storage 6, last word number N (i) = n is the last It is stored in the tail word storage section 7.

Naori” (i+i), Bn(i+i) (however, i=1.
2゜..., T''; n = 1.2..., N) are temporarily stored in the cumulative distance calculation unit 14 until they are no longer needed. The intermediate cumulative distances of the frames immediately before and two frames before the current frame are stored until the intermediate cumulative distance of the frame is calculated.

Also, the terminal cumulative distance D stored in the terminal cumulative distance storage unit 6
(i) is necessary to find the initial condition of equation (1)7z
Moro DYl'6 Ri, D (iJK Tsuiteha D"
It suffices if it is stored until (i +1.1) is found.

Reference numeral 8 denotes a voice section detecting section, which determines a voice section from the magnitude of the input signal and the like. The voice section detection unit 8
When it is detected that audio input has started, the frame number counting section 9 starts counting each frame. The above process is
Although the processing was performed on frames, the count value of the frame number counting unit 9 sets i. Therefore, the same processing as described above is performed every time the frame advances by one. The frame number counting section 9 starts counting when a voice section is detected, and is reset when the voice section ends. Last word storage unit 7. Therefore, in the back pointer storage unit 6, N (i) and B (i) are i;1.2.
..., 1 will be stored.

The segmentation unit 10 issues a command to the back pointer storage unit 6 to read a predetermined back pointer. That is, the segmentation unit 10
When the value B (i) is issued to the back pointer storage section 6, the back pointer B (i) is read out from the back pointer storage section e. When the segmentation unit 10 receives the value B(i) from the back pointer storage unit 6, it issues the same value to the back pointer storage unit 6. Therefore, when the voice section detecting section 8 detects the end of the voice input, the final value of the frame number counting section 9 is supplied to the segmentation section 1°, and the segmentation section 1Q first detects the end of the voice input. A value is issued to the back pointer storage section e. Thereafter, according to the operation described above, B(I), B are stored from the back pointer storage unit, etc.
(B(I) ) , B(B(B(I) ) ) ,・
..., outputs of 0 will be obtained sequentially. These values are the last frame of the second to last word, the third last frame of the last word, the fourth last frame of the last word, etc.
..., and since N(i) is a word that ends in the i frame, if this value is given as is to the last word storage section 7, the recognition results will be obtained in the reverse order starting from the last word. Note that in order to prevent recognition results from being obtained in the reverse order, this order conversion is performed using the back pointer storage unit 6.
or the output of the last word storage section 7.0 The total number of calculations up to the final frame in this conventional example device is determined by the distance between the vectors d( at each point in the i-j space) '＋
1) + Intermediate Cumulative Distance ('+I) is calculated only once for each point, so if the average number of complaints is 1, then all of them are NIl.The total calculation in the device using the conventional two-stage DP method Compared to the number of times Nl-!12, the number of calculations becomes 1/(-T).The length of 7 rams is now 10 m5e.
If the average word length is 0n5 sec in c, the average number of frames I of the standard pattern is J=50, so the number of calculations is 1/37.5. Also, compared to the number of calculations NI IR when a window of width R (calculation restriction area) is provided, it is 1
/R and Nashi, R is approximately 20th place, so the number of calculations is 1/
It becomes 2o.

As described above, the 0(b)DP method can recognize continuously uttered speech with far fewer calculations than the two-stage DP method, realizing a continuous speech recognition device with high recognition speed. Since processing can be performed for each input frame, a recognition device that operates in real time is possible.

However, the distance between the vectors d"(t, j) and the intermediate cumulative distance"'(is i ) are 21=1 for every 77-A i
．． 2°..., N; j=1.2. ..., In, a standard pattern storage memory is required to have very high speed access. for example,
Frame period 6ms, standard putter y number N m3ec
, the average number of 7 frames of the standard pattern J = s.

Then, the average number of grid points to be calculated per frame of the input pattern is NT = 9000, and if the number of dimensions of the feature vector is 16, then the access to the standard pattern storage memory per frame of the input pattern is NT = 9000. The number of times is 90
00X15=135ooo times".This is 5m
To perform this within 5eC, an access time of 37 ns or less is required. In addition, the amount of memory required to store the standard pattern is 136KB, assuming that one element of the feature vector is expressed in one byte, which is 37fiSeC.
This means that 135 KB of high-speed memory, which can be accessed below, is required, making the device extremely expensive.

OBJECTS OF THE INVENTION It is an object of the present invention to significantly reduce the number of memories that require high-speed access in a real-time sequential pattern comparison device according to the method described above.

Structure of the Invention The present invention converts an input signal into a feature vector 81r a 2 e
...+"i+..., a feature extraction means for converting into a series of features vectors bn, bn,...+
Standard pattern R consisting of b"j l..., b"in
n (however, n=1.2...

In a lattice graph with the input frame i as the horizontal axis and the standard pattern frame) as the vertical axis, when matching with the standard pattern Rn, the slope k is used to match the frame of the input pattern n. On the other hand, for the grid points sandwiched by straight lines separated by W(≦(-T-) +1) frames, the feature vector b of each frame j of the standard pattern? for j=1.2
．． ..., DP matching is performed in the i-axis direction in the order of In, with a slope restriction where the maximum slope of the matching path is k, 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 way. and DP matching means for performing the matching calculation over the entire range of the input pattern.

DESCRIPTION OF EMBODIMENTS FIG. 5 is a diagram showing the principle of the present invention. Choose the path restriction conditions as shown in Figure 1 or Figure 3: 24.25 is a straight line with slope 2, 20.21 is = 7 n and straight line 24
．． Intersection with 26, 22.23, j=1 and straight line 24.25
, and if W is J”/2, then the terminal cumulative distance of the grid points included in 20 to 21 is the terminal cumulative distance to each grid point included in 22 to 23, and the intermediate cumulative distance on straight line 24. If the distance is known, it can be calculated as 0.In addition, intermediate calculations can be performed horizontally from bottom to top as shown by arrow 26.In this way, the frame period is set to T. Then, during time WT, the calculation for the shaded area A is performed as described above, and during the next WT, the calculation for the shaded area B is performed as described above, and so on.
You can proceed step by step.

In this way, the jth frame of the standard turn is
Since it can be kept constant while checking the input turn of the W frame, the access time required for the standard turn memory can be slowed down.

In other words, the number of grid points in each shaded area is WT, and calculations for each grid point need to be performed during TW (
T is the period of the frame), then the calculation time allowed for one frame of the standard pattern is TW/IN
This takes a lot of time to read out data for one hour of the standard number turn, and if the order of the feature vector is set to 15 as before, then the elements of one vector of the standard pattern can be read out. Therefore, the access time allowed is 8.3 μsec/15 = 553 n5ec, which is about 16 times the value required in the conventional example. However, this time, high speed is required of the memory for storing the input pattern (for the j-th frame of the standard pattern, reading must be performed NW times in a short period of time).

TW For the above example, /MW=T/NJ=5mSE
/]-300X30==0.58μ It is necessary to read out 15 feature vectors. Therefore, the access time allowed to read one element of the vector is 37 n
5ec), the required elements are on average 16·W=15×30=450 bytes for the above example, leading to a significant cost reduction.

The above is the principle of the present invention, but in a simple concept where W is fixed, W≦(min (I n)/2) +1
Therefore, the value of W is determined by the shortest pattern among the standard patterns, which may not be very effective. In order to compensate for this drawback, it is conceivable to vary W accordingly each time In changes (W''-(
J"/2) +1). The principle of this further principle comparison device according to the invention will now be described.

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

Now, processing when the input pattern is the i-th frame will be explained. Each of the diagonal lines A, B, etc. will be called a block. At this time, if Dn(i-(')) is not found, the case indicated by O in the figure is a lattice point at j=1゜2,...,In of the block immediately before this block. The intermediate cumulative distances at these points are the initial values of the intermediate cumulative distances in this block. In this case, the constraint conditions for the matching path are as shown in FIG. When using the route constraint conditions shown in FIG. 1, it is also necessary to store the intermediate cumulative distance at 0. When Dn(i-[ko]) has been determined, the calculation for the standard pattern corresponding to n is skipped.

As described above, in order to determine that each standard pattern Hn is calculated by setting Wn = [work n/2] + 1 for each n, first set D・(i--2) to 0, Alternatively, if it is initialized to ■, it is possible to determine since it should be a finite value other than 0 when it has been determined.

FIG. 7 is a block diagram showing an embodiment of the present invention based on the above principle. Reference numeral 101 denotes a feature extractor which performs the same operation as 1 in FIG. 4, 102 a word standard pattern storage unit similar to that shown in FIG.

The inter-vector distance calculation unit is similar to that in 3, but in the actual 13th example fUi calculation, m = - for the vector b of the j-th frame of the 1° standard pattern Rn of [-!
−(1″−j+1) ] + dn(m11) is determined for the input pattern in the range of i.

104 is a cumulative distance calculation unit, which calculates the above-mentioned d''(m, j)
Find the intermediate cumulative distance and final cumulative distance in the same range as .

112 is the m
The intervector distance calculation unit 103 and the cumulative distance calculation unit 10 calculate d” (m, j) and Dn (m, j) within the range of
4. 105 to 107 perform the same operations as those in FIG. 4, 5 to 7. Reference numeral 111 is an output terminal for the recognition result, which is exactly the same as that obtained in FIG. 4, 11.

In order to more clearly explain the embodiment shown in FIG. 7, FIG. 8 explains its operation in terms of software, and this example may also be followed when the present invention is implemented by software.

In this embodiment, the conditions shown in FIG. 1 are used as the path constraint conditions.

201 is a step in which, prior to starting the matching path, initialization is performed such as substituting flashes into intermediate cumulative distances, final cumulative distances, and the like. 202 is a step in which the flag is set to '0' for processing after the voice section ends.2
03 is a step of detecting the start of a voice section, and corresponds to the operation shown in FIG. 7 108. Steps 201 and 202 are executed until a voice section is detected. When a voice section is detected, steps 204 to 230 are executed for each frame of input. Steps 205 to 224 are performed for each standard pattern in frame l. 206 is a step of determining whether matching calculation should be performed for standard pattern n in frame i. It is determined that matching calculation should be performed when the following three conditions are met.

D·(i-[pi])-1+n 1-(-)<, I+n 1-2(-)≧1 This corresponds to the operation of the matching calculation execution range determining section 112 in FIG. Steps 207 to 7 210 set the initial value of the recurrence formula for the matching calculation for the standard pattern n before executing the matching calculation in the IT frame. -1
), B''(m-1, O)=m-1 The steps 211 to 217 calculate the back pointer between j; 1..., 101 for the standard pattern n. 213 is its recurrence formula, This formula has exactly the same form except that i in formula (1) has been changed to m, and the order of calculation is j for each i in formula (1).
The difference is that this formula is calculated by changing -C'(1''-1+1)) + t ° for each j, whereas it was calculated by changing it from 1 to In.Step 218
~Step 222 calculates the measured distance Dn (m
, In) is the previously calculated standard pattern n/
The replacement shown in step 220 is performed when the intermediate cumulative distance D'' (m, J'') is smaller than the minimum value. As described above, steps 206 to 2
When 24 operations are completed, the minimum value for Hn calculated for each frame before l will be in D(i). The above operations are mainly performed from 103 to 103 in FIG.
107. 225 is a step of detecting the end of the voice section. The operations from step 206 to step 225 described above are performed until the end of the voice section is detected, and the end cumulative distance D(i), end back pointer B(i), and last word N(1) are calculated in sequence. , are stored in a predetermined location. Step 226 - Step 22
9, even if it is detected that one voice section has ended after performing the above calculation, processing is continued for each block, so the D (i), B (i)
. However, 'many max (T town, I is the frame where the voice section ends. Step 231
is the terminal back pointer B(i
), the input word example is obtained from the last word N(i) in the same manner as in the conventional example.

Step 232 is a part that outputs the results, which are printed out as characters or output as commands to operate other devices.

In addition, in this example, for convenience of explanation, point (
ir + ), intermediate cumulative distance,
Intermediate back pointers etc. are d” (ilj) and D” respectively.
(ilj) IB"(t, j) etc., but these storage areas do not need to be prepared over the entire range of ir1, and D"(i, j), B"(
ilj)+7) As is clear from the recurrence formula shown in FIG. Cumulative m distance mD”(l+
1-1) ,”(ir1-2L intermediate hack;'f:i7
Bn(i, i-1), Bn(i, j-2)
All you need to remember is Dn.
(ir1-2), B" (ir1-2) are discarded, "(ilj t), Bn(i, 1
*) and the contents of nD"(i, j-2) and B"(i,
j-2) (D storage area K, "(' + +) lB" (
11]) O contents fDn(ir 1-1), Bn(ir
1-1) O memory area W2F1m Sase, Dn (ir + )
The calculation result of the next stage may be stored in the storage area of IB" (111). Also, the width of each block may be at most Vvma x maw(T")/2) +I, so The buffer 1 memory for calculating the recurrence formula consists of one stage of W-frame backer memory for the input pattern, and one Wmax-frame backer memory for the intermediate m&X cumulative distance and intermediate back pointer. For storing three stages and the distance between vectors, in the case of the path restriction conditions shown in Figure 1, it is sufficient to store only dn(i-1,j), so prepare only one frame. That's a good thing. In the case of the route restriction conditions shown in FIG. 3, the distance between vectors is used each time it is calculated, so there is no need to specifically store it.

In addition, in this example, since the constraint conditions for the path as shown in FIG. 1 or 3 are used, the maximum slope of the path is 2, so the width of the block is [engine n/2] + 1. However, there are various other constraint conditions for the path.For example, in the case shown in Figure 9, the maximum inclination is 3, so when the maximum inclination of the path in this shell is k, is the same as that of the block if Jn is an integer)0 Furthermore, in this embodiment, the intermediate cumulative distance is initialized to ~, etc., but for input frame 1, i Of course, it is also possible to determine whether or not is divisible by Wn, and if it is divisible, perform the matching calculation, and if not, skip the matching calculation.

Effects of the Invention As described above, according to the present invention, the CWDP method and 0 (n
Methods such as the JDP method that perform 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. This makes it possible to realize an inexpensive device.

[Brief explanation of the drawing]

Figure 1 shows the constraint conditions of the matching calculation path, Figure 2
The figure shows the area in which matching calculations are performed, Fig. 3 shows another constraint condition for the matching calculation path, Fig. 4 is a block diagram showing a conventional example of a speech recognition device, and Fig. 6 shows the principle of the present invention. 6 is a diagram explaining the details of the principle of the present invention, FIG. 7 is a block diagram of a speech recognition device according to an embodiment of the present invention, and FIG. 8 is a diagram showing the functions of the device according to the embodiment. FIG. 9 is a flowchart when this is realized by software, and is a diagram showing another example of constraint conditions for the matching calculation path. 1
:'101...Feature extraction section, 102...
・Word standard pattern storage unit, 103... Inter-vector distance calculation unit, 104... Cumulative distance calculation unit,
106...Terminal cumulative distance storage unit, 106...
... Back pointer storage section, 107 ... Last word storage section, 108 ... Voice section detection section, 1
09... Frame number counting section, 110...
- Segmentation unit, 112...Matching calculation execution range determination unit. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 1
Figure 2 Turn to λ force Figure 3 Figure 9 (1-2°

Claims (1)

[Claims] The input signal is converted into feature vectors a_1, a_2, ..., a_i,
..., a_I, and a feature vector sequence b^n_1, b^n_2, ..., b^n_j,
...standard pattern R^n consisting of b^n_Jn (however,
standard pattern storage means for storing n=1, 2,...,N), and standard pattern frame j with input frame i on the horizontal axis.
In a lattice graph whose vertical axis is the standard pattern R^n
When matching with , the slope is k, and W(≦2-[(1-J^n-1)/k]) for the frame of the input pattern.
For the grid points sandwiched by straight lines separated by frames, the matching radius is calculated in the i-axis direction in the order of j = 1, 2, ..., J^n for the feature vector b^n_j of each frame j of the standard pattern. The maximum slope of is calculated over the entire range of the input pattern by performing DP matching with a slope limit of k, and when the calculation of one region is completed, the calculation of the next adjacent region similar to the above is performed in the same way. D performing the matching calculation
A pattern comparison device comprising: P matching means.