JPS58137899A - Pattern recognition equipment - 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] The present invention relates to a slam recognition device.

A bang recognition device is a device that performs homology recognition by matching time-series bangs expressed as a time series of feature vectors, such as voice, with standard bangs. For example, a speech recognition device has a standard button for each speech to be recognized.9. It is determined that one input voice bang is the same voice as the standard bang by comparing the input voice bang with each of the above-mentioned standard bangs, that is, performing matching, and determining the standard bang that provides the best match. Human speech can be converted into machine code through this type of bang recognition.

The above-mentioned standard clicks are manually inputted by the speaker in advance, and conventionally, various manually operated clicks are stored in a memory as they are and used as standard clicks. Therefore, in the case of a word recognition device, for example, one word requires a memory capacity equivalent to the length of Hekabatan, and an additional capacity for the number of words. Therefore, a huge memory capacity is required. In order to reduce memory capacity, there is a method of extracting feature vectors of input buttons at regular time intervals as a representative vector sequence and using them as standard buttons, but Jin's method results in missing features of welfare input buttons, making it difficult to recognize them. The disadvantage is that the rate decreases. Pattern recognition is not limited to the recognition of voice slams, but can be applied to any pattern that is expressed as a time series of feature vectors. Regarding the huge memory capacity or the reduction in recognition rate when the memory capacity is reduced,
The same is true.

An object of the present invention is to solve the above-mentioned conventional drawbacks, and to provide a button recognition device f capable of compressing a standard button without reducing the button recognition rate and reducing the capacity of a standard button memory. .

The bang recognition device of the present invention has the following characteristics:
．． 2. ..., J), and a breakpoint sequence (Ak) that divides the vector string stored in the button buffer into sections t*<J) = 41+ txt..., ZA.

...I LK+1 is the center position of each section ('k + '
The error It bQ -b)' between k6h and each vector b in the interval is calculated using the dynamic programming method so that It is minimized in the entire interval. part and the breakpoint sequence obtained by the above operation <
A determining unit determines and holds Lh't, and a central vector bIkt-compressed pattern C=e1. which is read from the button buffer using an address signal indicating the center position of each section outputted by the determining unit. c, . . . C1; and a recognition processing unit that compares an arbitrary manual punch with the compressed button.

Next, the present invention will be explained in detail with reference to one drawing. Sample points 11.1. ,...* "is..., tl
The vectors fa□, a2. ..., a8. ..., al, then the person Kabatan is A ``it a go''' s ``ze''' e ''1
It is expressed as The standard baton B to be compared is s=b□
, b2. ..., bJ, ..., bJ. In the present invention, as shown in FIG.
: , b ;,... Compress the column of bKJ with the button C
shall be. The vector number at the center of each division is the vector bαk +LA+x' at the center position of the 9 division points Lk and Lk+.
4 is extracted as a representative vector of the section.

That is, bk = b(Ll, +lk+1)/.

As shown in FIG. 2, the breakpoint sequence (tt) is the magnitude of the vector error with respect to the vector b included in the interval between each breakpoint tk and tk+□, as shown in FIG.
When b′ is expressed as −bj II, the following (1)
It is assumed that the minimization condition of Eq.

......(11) The calculation of the minimization problem of formula (1) will be described later, but through this calculation, +□ is determined for the nine division points t1 to 1.The division point sequence (convergence) is , (-)=x□, L3...e 'Ig*...e Z
It is expressed as K+1. If the center O solid 1 torque between these break points is rewritten as C4, then compression
C=C11C,t…*Ck,”・, The oK represented by C1 is of course smaller than J, and the compressed button C can be stored in a memory with a smaller capacity than the standard button B0 (1 ) can be calculated using the dynamic programming method.Now, we have an integer book that takes a 1-JO image, k that takes a value of +1 for i~, and rlL-δ$t≦rIL. Consider the following recurrence formula (%) (2) determined by L equal to -1.Here, δFi is a parameter that determines the dynamic stroke calculation area, and the adjacent breakpoints ZA and 'bit
This is an integer that is set as appropriate to regulate the maximum interval between. (The recurrence formula of the J formula is the initial condition 'r(0,0)=
It can be determined from 0 by sequential calculation. Then, when standard pattern B is divided into a breakpoint sequence (th>), a breakpoint sequence (th ) can be found as the optimal breakpoint sequence.
This is performed within the range of %m, and k is performed within the range of 1 to +1 for the value of %m. Furthermore, the calculation is performed by changing l in the range from −δ to m −1 for each value. And for each value, Σ ” b(L+m > 7
1-b) II
. . . (3) No L calculation is performed, and then T (t-1, k-1) t (31
By adding to the equation ('A, the value in brackets on the right side of the equation is found for the match. As mentioned above, Z#/'im
Since it is varied in the range of -δ to quasi-1, [] in equation (2)
δ values are found for any (''*k). The minimum value among the above six values is T(m.

k). If L oil (that is, the frame number serving as a break point) corresponding to this T (711, A) is L (folding, k), then m = 1 - J, = 1 to +
1, an L(m,k) table is created. By referring to the table, the optimum breakpoint sequence (Lh) can be determined. That is, in the upper [L table, the end break point ZIC+, = 1. (J, +1)
Therefore, the optimal breakpoint sequence (Lh) is found by sequentially going back to the starting end of the breakpoint t□ using the reverse recurrence formula 'k'' L (Z &+11 k+1).The optimal breakpoint sequence (Lk) is Once determined, the center vector CA between each break point (Ll+Lh+x)/@
I want it! ll, the compressed button C that is appropriately compressed = b is found. Since the compression pattern C obtained by the above calculation matches the pattern most similar to the standard pattern B before compression, the pattern can be compressed without reducing the recognition rate.

FIG. 3 is a block diagram showing one embodiment of the present invention.

That is, the output signal sFi of the microphone 10 is converted into a time-series bang of the spectrum envelope by the voice analysis section 20. The bangs are stored in the audio bang pan 7a30.

Now, when creating a standard pattern from human voice, the voice pattern buffer 30 stores a standard pattern B. The dynamic programming calculation unit (DP calculation unit) 40 calculates the vector value bjf output from the audio bang buffer 30 as human power (T(m, A)t− of the force formula, and calculates the corresponding L(
m, k) find f (details will be described later).

The control unit 200 controls the sound pattern buffer 30 when five standard button sequences are stored in the sound pattern buffer 30.
It receives the signal J output from the DP calculation unit 40 and outputs a clear signal CLR, a signal M for controlling switching of the multiplexer, and numerical signals m and k necessary for calculation by the DP calculation unit 40. The numerical votes are output in order from 1 to J, and the signal k is output from 1 to +1 for the value of 6. Malteplek? 50
U and signal M are used to switch the address signal to be given to the audio slam buffer 73o. In other words, the signal M is “0”
", it is the DP calculation mode, and each vector bj of the uncompressed standard pattern B stored in the audio pattern buffer 30 is specified. The address j of bj is determined by 1) P in the calculation process of the DP calculation unit 40. It is output from the calculation unit 40. When the signal M is “1”, it is the mode for determining the compression pattern Ci, and specifies the central vector bi extracted from the button B. The address of the flag is output from the determination unit 80. It is given by the signal t.

T memory 60 is Dr calculation W & 40 is calculation suru T ('11
Similarly, the L memory 70 is a memory that holds the L(m,k) table obtained by the DP calculation unit 40. The 0 determination unit 80 is held in the L memory 70. The optimum breakpoint sequence (Lk) is determined by referring to the L(m,k) table, and the address signal L=(ZA
+Z! +□)/2 and outputs it to the multiplexer 50.
The central vector b', that is, bz, is read from the corresponding address of the audio bang buffer 30 via the central vector b'. This vector is written into the compression button memory 90 as the vector C1 of the compression button C.

The compression pattern memory 90 is a memory that stores each of the above-mentioned vectors ckt- and holds the compression pattern C.

Recognition processing [810G approximates the vector in each section by the vector Ck output from the compression button memory 90,
It is compared with the voice button A output from the voice button buffer 30, and recognition is performed by a well-known bang matching method.

The DP calculation unit 40 is configured as shown in FIG. 4, for example. That is, the DP control section 400 controls the control section 2
From 00 to manually input signals m and k, (L+m)/2=
7. Calculate and output the address and send it to the multiplexer 50.
The temporary central vector bQ is read out from the corresponding address of the audio buffer 30 via 0 The read temporary central vector b
Q is stored in buffer 410. Next, distance calculation section 4
20 calculates the distance Il bk −bj 11 between the vector b read from the address j of the audio bang buffer and the central vector bQ, and sends the calculation result to the accumulator 430 . The accumulator 430 goes up from L to the folding screen. The adder 440Fi receives T(Z-x, -1) read from the T memory 60 and the output of the accumulator 430.

Since the above T(j, k) is calculated for each t from -δ to m-1, δ pieces are obtained. The minimum detection unit 450 determines the minimum value among the δ T(j, A), and converts the minimum value into T(m).

0 to be output and stored in the T memory 60 as k), or l, (Z, k) corresponding to the above minimum value T(m, A) as L(
m, k) and stored in the L memory 70. Next, the time chart shown in FIG.
The operation of this embodiment will be explained with reference to the figures.1.
When a standard batten B with a pattern length J is input to the buffer 30, the pattern length J is given to the control unit 200 (see FIG. 5(a)).
)), the control unit 200 determines the number of compressed frames K (K<J) based on the pattern length J, and sets the signal M to "0" based on the determination of the number of compressed frames. Fifth
Figure (C)). In addition, the clear signal CLRI is output (fifth
Figure (d)) is sent to the T memory 60, the T memory 60 is cleared and T (0,0)=0 is initialized. Thereafter, as shown in Fig. 5 (el), the output value m is outputted sequentially from 1 to J. For each wind value, the output value k is outputted from 1 to +
The data is sequentially output up to 1 ((f) in the same figure). Furthermore, in the DP control unit 400 of FIG. 4 for each value of l,
-δ to m-1 (Fig. 5(g)). Then, the DpH1JIi section 400 first creates an address signal with j=(L-Jm)/2 for each value of l (see FIG. 5(h)) and sends it to the multiplexer 50. The multiplexer 50 selectively outputs the address signal and sends it to the audio button buffer 30.

A (temporary) central vector bQ is read from the (Z+Ffi)/21# location of the audio bang buffer 30 and stored in the buffer 410 in FIG. Next, the address signal
The value of is changed sequentially from l to wind (see FIG. 5(h)), and each time the vector b is changed to the sound bang buffer 30.
1lbA"7" in the distance calculation unit 420.
is calculated. Each calculation result is integrated by the accumulator 430, and the second term in brackets on the right side of equation (2) is calculated. The adder 440 adds the output value of the accumulator 430 and 'r(4-1°h-1) read from the T memory 60 to obtain T(L,k).
The calculated result is sent to the minimum detection unit 4rO. For δ values of t from -δ to wL-1, the above T(A, A
) are calculated and stored in the minimum detection unit 450, respectively,
Minimum detection unit 45 (1, the above a@O'r (L, k
) (T memory 6 as the minimum value iT(m, A) in D
o (FIG. 5(i)). On the other hand, the corresponding t
is written to the L memory 70 as L(w*,k) (
5th heir (j)).

When the above processing ends with m=J4, L of L memory 70 is
The (m, k) table is completed, and the optimal breakpoint sequence +Lh) is determined in decision s80 by referring to the table. That is, the end break point is +1 at ', and 1, (J) read from the L table above.

Since K+1) indicates LY, the address, LK-1 can be obtained by reading L(ZK, K) from the L table.

In this way, 'k '' L (Lk + s + '' + 1
), the starting end break point L1 is sequentially determined by the reverse recurrence formula.
By hollowing out the hesaka, an optimal nine-point sequence t th ) is obtained. At this point, the signal M output from the control unit 200 in FIG. 3 becomes slm, which is determined by the sequence of dividing points Move to compression button read mode. That is, the determining unit 80 outputs an address indicating the intermediate point between each breakpoint as "=('h + Lh+1)/2," and reads b-lt- from the corresponding address of the audio button buffer 30 via the multiplexer 50. , this is stored in the compression button memory 90 as lc4. That is, the compression button memory 90
A compression tank C is stored in . Each vector coFi of the compressed pattern is taken as a representative vector within each section, and the recognition processing unit 100 executes a well-known bang matching method between the vector coFi and the human-powered button A that will be manually operated later.◇In this embodiment, the standard button is However, compression can be performed using representative vectors selected so as to reduce approximation errors, and the memory capacity can be reduced compared to the conventional method. Furthermore, the error caused by the above compression is small and does not reduce the recognition rate. Note that with a similar configuration, it is possible to similarly reduce the memory capacity for feature vector sequences other than voice bangs without reducing the recognition rate.

[Brief explanation of the drawing]

Fig. 1 is a diagram showing the relationship between the center vector of each section obtained by dividing a standard batten into sections by a 9-point sequence (lh) and the compression pattern, and Fig. 2 is a diagram showing the relationship between the center vector bk and each vector within each section. Fig. 3 is a block diagram showing an embodiment of the present invention, Fig. 4 is a block diagram showing the configuration of the DP calculation section 1fl11 of the above embodiment, and Fig. 5 is a block diagram showing the above embodiment. FIG. 2 is a time chart for explaining the operation of FIG. In the figure, 10...Microphone, 20...Speech analysis unit, 30...Sound slam buffer, 40...D
P calculation unit, 50... multiplexer, 6o... T memory, 70... L memory, 80... decision unit', 90
...Compression bag/memory, 100...Recognition processing section, 2
00...! 111I41 part. 400...DP control unit, 410...buffer, 42
0... Distance calculation section, 430... Accumulation section, 440...
- Addition section, 450... Minimum detection section. Agent Patent Attorney 1) Toshi Sou

Claims (1)

[Claims] Vector sequence S, ()=1.2. ..., J) for holding a standard baton, and
The vector sequence stored in the button buffer, t-K
Break point sequence (convergence) for dividing into sections (K<J) = L1
s t, ..., th, ...*' +1 to the center position of each section (Lk '''h+x)/2 Vector b; error between c and each vector tono in the section II b
A calculation section that calculates by a dynamic programming method so that λ-bj II is minimized over the entire interval, a determination section that determines and holds the breakpoint sequence (Zt) obtained by the above calculation, and each output of the determination section. The central position of the interval tt--the central vector bQ read from the button buffer according to the address signal indicating the compressed button C=c. A pattern yil characterized by comprising a compressed button memory for storing C1, .
&! recognition device.