JPS62161200A - Continuous voice 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 [Field of Industrial Application] The present invention relates to a continuous speech recognition device, and more particularly to a continuous speech recognition device with a high recognition rate that rapidly recognizes sentence speech continuously uttered according to grammar.

[Conventional technology]

Among voice recognition devices, devices that recognize sentence sounds uttered according to grammar can recognize computer programs, texts for limited business use, commands for air traffic control and control of various devices, and have a wide range of applications. It is known in principle that when grammatical constraints are given, misrecognition can be prevented by using the grammatical rules. In particular, when there is a restriction on the number of digits in input speech in continuous number recognition, the recognition rate can be improved by regularizing the restriction.

A method for recognizing sentence sounds continuously uttered according to such a grammar is proposed in Japanese Patent Application No. 58-239303 by the same applicant.
It is described in the specification ``Continuous speech recognition device''. The blockwise DP matching method based on this principle is as follows. The grammar is expressed by an automan α, and the automan α is defined as follows.

α=<K, Σ, Δ+ PO+ ”> ・・・・・・
...(1) Here, K: set of states p (p l p
=l, 2. ...π) Σ: Input word n (7) set (n
l n =1.2. -N) Δ: State transition rule ((p,
q, n)) Here, (p, q, n) means a state transition from p to q.

Po: initial state, hereinafter expressed as p=Q.

F: Final state set FCK Next, the speech pattern A obtained by continuously uttering the word nεΣ according to the automaton α is A =: a1.
a2. ...ai, ...al...
... is indicated by ■ and is called an (unknown) input pattern.

Standard pattern Bn=b\b2 for each word nεΣ. . ...bjrI,...
・・b−・・・・(3) is prepared and is called a word standard pattern. Continuous speech standard pattern C=Bn 1 nI'12 . . . B obtained by connecting this word standard pattern Bn according to the automaton α
Perform DP matching between n .

Here, the minimum degree of dissimilarity is found using the following dynamic programming method. Set the initial condition, and proceed sequentially from b=1 to I/BL (here, I/BL is assumed to be divisible for ease of explanation) (5).

Based on the boundary condition of equation (6), the recurrence equations of equations (7) and (8) are calculated as (p).

Calculate for all pairs (p, n) such that q, n) εΔ. That is, the boundary condition with state p is i=i.

...About ducks and the boundary condition with block b-1 is j=1. ....

For each standard pattern time j called Jn, the acclimatization formula % formula % (7) [ ]
.

j) which gives (i, j).

Input pattern time i=1. Calculate up to ie. Here, d(i, j) is the feature vector ai at time i of the input pattern and the feature vector b at time j of the standard pattern.
, is the distance between . 'd(i, j)=Di
s(ai, bρ...(9) Store the boundary value of the calculation result for calculation of the next block.

After the calculations of equations (6), (7), (8), and (10) are completed up to the standard pattern time j=Jn, the minimization end at the word boundary is calculated.

The recurrence formula calculation of equations (7) and (8) described above is shown in Figure 2 (
As shown in the explanatory diagram), the BL frame portion of the input pattern is converted into a processor and executed for each block.

Finally, the recognition result of the input pattern is determined by the following procedure as a determination process.

Initial condition Q= argmin CT (1, q)
) −(12) qεF q=min lm=1 (13) is determined.

If minutes>O, repeat equation (14) with q=minutes 2m=minutes. If all = 0, end.

The recurrence formula (7) of the method described above is calculated along the path shown in FIG. 3(a). Therefore, the correspondence between the input pattern and the standard pattern on the time axis is limited to expansion/contraction from 1/2 to 2 times. This is called "tilt restriction" and eliminates locally unnatural time axis correspondences. By using this slope restriction, it is possible to increase the recognition rate.

However, when using recurrence formula (7), the work memory G1 (p, n, j), H□ (p, n, j)
, o2(p, n, j).

H2(pl nl J) is required. Furthermore, data transfer (corresponding to equations (6) and (10)) between the calculation unit and the work memory is required. That is, when using the recurrence formula (7), there is a drawback that the work memory becomes large and data transfer increases.

On the other hand, it is possible to use a recurrence formula without slope restrictions in the blockwise DP matching method. In this case, the calculation is performed along the path shown in Fig. 3 υ). The calculation procedure for this DP matching is performed in the same manner as the above equations (4) to (14), but as the initial moxibustion condition, instead of equation (4), (
Using equation (15), using equation (16) instead of equation (6) as the boundary condition with block b-1, using equation (17) instead of equation ('7:) as the recurrence equation,
Formula (18) is used instead of formula (10).

The method using the recurrence formula (17) without slope restriction described above can reduce the work memory to 1/2 compared to the method using the recurrence formula (7) with slope control, and data transfer can also be reduced to 1/2. /2.

However, since there is no slope limit, the correspondence between the time axes of the manual pattern and the standard pattern is allowed to become unnatural, causing misrecognition.

[Problem that the invention seeks to solve]

The above-mentioned conventional continuous speech recognition apparatus has the disadvantage that when using a recurrence formula with slope restriction, the work memory is large and data transfer is frequent. On the other hand, when using a recurrence formula without slope restrictions, there is a drawback that there are many erroneous recognitions.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a continuous speech recognition device which eliminates the above-mentioned drawbacks, has a small work memory, requires little data transfer, and has a high recognition rate.

[Means for solving problems]

The continuous speech recognition device of the present invention includes an input pattern memory section that stores input patterns, a standard pattern memory section that stores standard patterns, and a state transition p −+ according to the input of word n.
A rule that means that q occurs (p.

q, n) group, which is a state transition table Δ and a final state group F, and a state transition designated by a state p and a word n, which is determined by an input pattern time i and a standard pattern time j'.・Manual pattern BL within the shellfish area
a distance calculation unit that calculates the distance d(i, j) between the feature vector of the input pattern and the feature vector of the standard pattern at each point (i, j) in block b having the width of a frame; a boundary DP calculation unit that calculates a DP matching calculation at a boundary portion with the previous block b-1 using a first recurrence formula; and a boundary DP calculation unit that calculates a DP matching calculation inside the block b using a second recurrence formula. The present invention includes a main body DP calculating section for calculating the body DP, and a determining section that determines a combination of words that yields the minimum cumulative distance calculated by the boundary DP calculating section and the main body DP calculating section.

[Effect]

Next, the operation of the present invention will be explained with reference to FIG.

FIG. 4 is a detailed illustration of the present invention.

The calculation of DP matching according to the principle of the present invention proceeds by using a recurrence formula without a slope restriction at the boundary of a block, and using a recurrence formula with a slope restriction inside the block.

The initial conditions are set as the following (20
), based on the boundary condition of equation (21), (22) to (25)
A recurrence formula of the equation is calculated for the pair (p, n) such that (p, q, n) εΔ. In other words, the boundary condition with state p is 1=i3
．． ..., but for 1e, part (b-1)・BL+ 1, r e=b
-BL, and the boundary condition with block b-1 is j=1.・
....

The first recurrence formula % formula % for calculating the block boundary part is given as standard pattern time j.
(22) (provided that it is small) is (i, j) that gives the minimum on the right side of equation (22).

and then the second... which calculates the body of the block.
(24) h(i, j)=h(仝,仝) ・−・(2
5) However, 仝 and 仝) are (i, j) that give the minimum on the right side of equation (24).

t = 1. Calculate +1 to Cie.

The boundary value of the calculation result is stored for calculation of the next block.

Calculation of formulas (21) to (25) is standard pattern time j = J
After terminating at n-j, i as a minimization at a word boundary
=i..., calculate BL for Ie. Finally, the recognition result of the input pattern is determined using the same procedure as the conventional method, namely (12),
It is obtained from equations (13) and (14).

〔Example〕

Next, the present invention will be explained in detail using FIGS. 1, 5, to 10.

FIG. 1 shows an embodiment of the continuous speech recognition device of the present invention.

FIG. 5, FIG. 6, FIG. 7, and FIG. 8 are partial block diagrams each showing a detailed configuration of a part of the embodiment in FIG. 1, and FIG. 9 is a block diagram showing the operation time in FIG. 1. A time chart showing the relationship, FIG. 10(a).

-(d) are flowcharts showing the flow of operations in FIG.

The standard pattern memory section 130 stores the standard pattern B of the word n included in the word set Σ, and the automaton storage section 230 stores the state transition rule (p.

q, n) and the designation information of the final state F are stored.

When an unknown input voice is input from the microphone 100, the input section 110 performs frequency analysis, converts it into a vector a1 representing characteristics, and sequentially sends it to the large capata/memory section 120. The input unit 110 is also provided with a function of determining a voice section by detecting the voice level, and generates a voice section signal S that is "1" during the voice section and "0" otherwise. The control unit 240 generates the initialization pulse 5ETI at the time when the voice section signal S rises. As a result, initialization corresponding to block 1° in FIG. 10(a) is performed on the G memory 150 and the T memory 200.

When the above initialization is completed, the input pattern block signals are counted as 1, 2, . . . in synchronization with the input of the input features ai for the subsequent BL frames. In this input pattern block b, the word designation signal n from the control section 240 changes from 1 to N. For each word n, the state table in the automaton storage section 230 is referred to, and a state designation signal q defined by the word n and state p is also output.

Next, the operation within one cycle of word n and state p will be explained. In this one cycle, calculations shown in the shaded area in FIG. 2 are executed. That is (20).

Based on the boundary condition of equation (21), equations (22) to (25) are calculated. First, a signal 5ET2 from the control unit 240 causes the values corresponding to blocks 11.12 in FIG. 10(a) to be set in the g memory 330 and h memory 340 of the DP matching work memory.

Subsequently, the standard pattern time signal j from the control section 240 changes from 1 to J. At each standard pattern time J, the input pattern time signal i□ is r s (t s = (bl)
・BL + 1) to F) changes to 1e (ie = b - BL).

In frame i, the calculation shown in block 13 in FIG.

FIG. 5 is a block diagram showing an example of a boundary DP calculating section. First, the i frame of the input pattern and the j frame of the standard pattern of the ir1 word are successively extracted, and the inter-vector distance d(i, j) shown in equation (9) is determined by the distance calculation unit 300.

On the other hand, the boundary DP calculation unit 310 calculates the minimum value of the three dissimilarities in parallel with the distance calculation. That is, from the g memory 330, g(iS'!J) + g(
1゜-1,j-1,g<i,-1,j-2) and h memory 340 h(is 11 J) Th(i5
L J-1) + h('S-1, j-2) is read out and registers Gl, G2゜G3 and Hl, H2, H3
are stored in each.

The comparison circuit 312 includes three registers Gl, G2. G3
Detect the minimum value from the register G from which the minimum value was obtained.
A gate signal is issued to select the register Hn corresponding to n (n is one of 1, 2, or 3).

The contents of the register Hn selected by the gate signal number are written to h (transport, j) of the h memory 340. Furthermore, the minimum value g(i, -1,
) is the distance d(-,
j) and the adder 311, and the g memory 330
It is added to 0g(-,j).

Subsequently, at each input pattern time i (i changes from i+1 to ie), the calculation shown in block 14 in FIG. 10 is performed in the main body DP calculation section 320.

FIG. 6 is a block diagram showing an example of the main body DP calculating section. First, the i-frame of the input pattern is read out and (9
) The distance between the vectors shown in the equation is calculated by the distance calculation unit 300, and d(i, j) is obtained.

On the other hand, the main body DP calculation unit 320 performs (
24) The minimum value of the right-hand side of the second recurrence formula is found.

That is, g(i-2
, j-1) and the distance d(i-1, j) delayed by one time in the delay circuit 324 are added in the adder 323 and stored in the register G1.

Also, from the g memory 330, g(i-1, j-1) and g(
i-1, j-2) are read out and register G2. Each is stored in G3. Furthermore, from the h memory 340, h(i
-2, j-1), h(i-1゜j-1), h(i-
1, j-2) are read out and stored in registers H1, H2, and H3, respectively.

The comparison circuit 322 detects the minimum value from the three registers G1, G2 and G3, and selects the register Go from which the minimum value was obtained.
(O corresponds to 1. 2. 3) △
△ Generates a gate signal n that selects the selected register Hnft. The contents of the register H selected by the gate signal n are written to h(i, j) of the h memory 340. Further, the minimum value output from the comparator circuit 322 is added to the distance d(i, j) obtained by the distance calculating section 300 by the adder 321, and
) is written to.

Input pattern time i is! By changing from +1 to ie, the processing for standard pattern time j ends.

Furthermore, after the standard pattern time j becomes the terminal J, the 10th
The result g(ie,j) of the recurrence equation (24) according to the signal SET 3 as shown in block 15 of figure (b).

h(ie, j) in table memory G(p, n, j).

Store in H(p, n, j). Subsequently, the comparison shown in block 16 of FIG. 10(C) is performed according to the signal 12 (complexed from i to Ie) issued by the control section 240. That is, as shown in FIG. 7, T(i
, q) and g(i, J> are read from the g memory 330, and compared by the comparison circuit 170, T(i, q)>g
If (i,, r) then wp times the signal is emitted and g(i, J
n), n, I), h(i, Jn) cuff-7'
k) Mori T(i, q), N(i,
q), p(i, q).

written to L(i,q).

With the above operations, the processing of the state p and single word n large input Kapatan block is completed. Further, the state designation signal p is changed from po to π, thereby completing the processing for the word designation signal n. Furthermore, the word designation signal n is changed from 1 to N, thereby completing the processing for the input pattern block b. After the input pattern block signal becomes I/BL, (12).

The determination process shown in equations (13) and (14) is started.

The determination unit 220 is configured as shown in FIG. 8, and first, as the process of block 17 in FIG. , T from T memory 200
(I, q) is read and the comparison circuit 221. Minimum value register 222. Using the state register 223, the minimum T(I,
q) obtains the state command given. Next, Figure 10(d)
As the process of block 18, the determination control unit 227 determines qJ,
Address signal m3. q is N memory 190°L memory 210. P is sent to memory 180 and N (m.

q), L(m, q), and P(m, q).

These L(m, q) and P(m, q) are the next time m and state q
Then, N(m, q) is output as the recognition result. By repeating this process until the minutes become zero, recognition results can be obtained sequentially.

The configuration of the present invention has been described above based on examples, but
These descriptions do not limit the scope of the present invention.

This example is explained based on the blockwise DP matching method as described in the specification of Japanese Patent Application No. 58-239303, but DP matching is executed by grouping multiple frames of the input pattern into blocks. If it is a method, it is possible to implement the block boundary portion and the block body portion, which are the principles of the present invention, using different recurrence formulas.

In other words, the blockwise Dl) method having the modified DP matching calculation section described in the specification of Japanese Patent Application No. 59-067116, the block described in the specification of Japanese Patent Application No. 59-068015 is diagonally relative to the standard pattern axis. The block boundary portion, which is the principle of the present invention, can also be applied to the diagonal block width DP method tilted to A first acclimation formula can be used, and the block body portion can use a second acclimation formula.

Furthermore, as the first recurrence formula for calculating the block boundary part (
Although Equation 22) is used, any form may be used as long as it is a recurrence equation that uses only input pattern times i and i-1. For example, or again.

Also, as the second recurrence formula for calculating the block body, (24
) formula is used, but it is clear that, for example, ... (30) ... (31) etc. may also be used.

〔Effect of the invention〕

As explained above, the present invention can reduce the work memories G and H by using only the input pattern times i and i-1 for the adaptation equation of the boundary part in the calculation of each block of DP matching, Furthermore, the amount of data transferred can also be reduced. Furthermore, the recurrence formula for the main body of each block takes a form with tilt control, which has the effect of reducing the false recognition rate.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the continuous speech recognition device of the present invention, FIGS. 2, 3, and 4 are diagrams for explaining the invention in detail, and FIGS. 5 and 6 7 and 8 are partial block diagrams showing a part of the detailed configuration of the embodiment shown in FIG. 1, FIG. 9 is a time chart showing the time relationship of the operations in FIG. 1, and FIGS. (d) is a flowchart showing a series of operations in FIG. 1. 100...Microphone, 110...
Input section, 120...Input pattern memory section, 13
0...Standard pattern memory section, 150...
・G memory, 160...N memory, 170...
...Comparison circuit, 180...P memory, 19
0...N memory, 200...T memo I
J, 210...L memo IJ, 220...
- Judgment unit, 230... Automaton storage unit, 2
40...Control unit, 300...Distance calculation unit, 310...Boundary DP measurement unit 320...
...Main body DP meter section 330...g memory,
340・・・・・・h memory〇 agent patent attorney
Hara 1. Coverage with %z diagram A n, L Figure 3
4th. Father-in-law (skin) ￥ left leap % times) η 7 buoy) / A(, I 'jsLs+1., 6 6., 'place C/
θ t21 law) ￥710 figure (b u

Claims (1)

[Claims] In a continuous speech recognition device that recognizes speech in which a string of words specified by a finite state automaton is continuously uttered by performing DP matching with a standard pattern, there is provided an input pattern memory section that stores an input pattern, and a standard pattern memory section that stores the standard pattern. a memory unit, an automaton storage unit that stores a state transition table Δ and a final state group F, which are a group of rules (p, q, n) that mean that a state transition p → q occurs due to the input of a word n; Input pattern time i and standard pattern time j at state transition specified by p and word n
The distance d(i, j
) and the DP matching calculation for the boundary between the block b and the previous block b-1.
a boundary DP calculation unit that calculates the internal DP of the block b using a recurrence formula; a main body DP calculation unit that calculates the internal DP matching of the block b using a second recurrence formula different from the first recurrence formula; A continuous speech recognition device comprising: the boundary DP calculation unit; and a determination unit that determines a combination of words that yields the minimum cumulative distance determined by the main body DP calculation unit.