JPS60130799A - 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 The present invention relates to a continuous speech recognition device, and more particularly to an improvement in a device that recognizes sentence speech continuously uttered according to grammar.

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. Particularly in continuous number recognition, if there is a restriction on the number of digits in the input speech, the recognition rate can be improved by regularizing the restriction.

A method of recognizing sentence sounds continuously uttered according to such a grammar is described in Japanese Patent Application No. 199098/1983. The principle behind this is as follows. We express the grammar as an automaton α, and define the automaton α as follows.

α=(K, Σ, Δ+ pal F) ・・(1)
Here, K: set of states p (plp=1.2...,
π) Σ: Set of input words n (n1n=1.2., , ,
N) Δ: State transition rule ((p, q, n)) Here, (p, q, n) means a state transition of p −> q.

po: initial state. Hereinafter, it will be indicated as pniO.

F: Final state set FcK Next, the sound bang A obtained by continuously uttering the word n6Σ according to the automaton α is A=1 + a+2
＋・・・、a、・・・、a■ ・・・・・・・・・(2
), and this is called an (unknown) input button. each word n
Standard slam B' === 1) for EΣ? , Ib; , , , , 1
b7 , , , , 1bsi1. ------1I:3)
This is called a word standard slam. A continuous speech standard pattern C2B0゛ is obtained by connecting this word standard batan according to the automaton α.

B1. ...Performs DP matching between Bnx and the input button A, calculates the amount representing the degree of mutual difference between the two buttons (hereinafter referred to as the degree of dissimilarity), and recognizes the word sequence that gives the minimum degree of dissimilarity. shall be.

Here, the minimum degree of dissimilarity is determined using the following dynamic programming method. Initial condition T(0,0)20'J) (m, q)=”, m40, q”
r.

() (p, n, j) = -... river (4), and sequentially from rrl = l to I, based on the boundary condition of equation (5), the recursion equation of equation (6) is (p, q , n) eΔ for all pairs (p, n). That is, the pair (p,
For n), the boundary condition G(p, n, o) = 'l'(m-1, 1)
) H (p, mouth, o) = m-1 (5),
Recurrence formula % formula %) (6) However, d (jl:= Dis (a-, lb'; )=Σl
a+nr 6'jr l...(b], )rr
l j = argmin G(p, n, j')
''-2≦3'≦j from j=x to J'' and store g(1, h(j) in G(p, n, j) and H(p, n, j), respectively. It means the minimum X. Next, as a minimization at the word boundary, if T (m, q) > G (p, n, Jn) ・---
(force then T(m, q) = GCp, n,
J”) N(m, q)=n P(m, q)=p L(m, q)=)((p, n, J”).

That is, in the recurrence formula of equation (6), as shown in FIG. 1, calculations in one vertical column along the standard batten axis are performed for each pair (p.

n), and move this vertical column of calculations # along the input button axis l until it reaches the terminal m = I.

The recognition result of the input button is determined by the following procedure as a determination process.

q = argmin (T(I, q)) ・
...(8) qεF Initial condition qniq, m-1...(9) Recognition single H
n=N(m, q) −・・−ao) word starting point lniL(
m, q) State transition pniP(m, q) If X>O, repeat U(2) with q=p and r11=4. If n=0, end. (11) In the method described above, all the standard buttons must be read out once in the calculation of the five recurrence formulas in one vertical column. It is also necessary to read %G(p, n, j) and store the calculated results.

When the numbers of p and n become large, the standard pattern or O(p.

n, j), )i(p, n, j) becomes clear, and the total memory access time (one read time minus one read time) becomes long. On the other hand, iterative calculation of DP matching.

can be executed by multiple arithmetic units, and 1. The computation time can be less than the total memory access time.

Furthermore, it is also possible to process memory read/write and performance B in parallel. In this case, l) the processing time for P matching is 7 times for all memos, and there is a drawback that as the number of words and states increases, the processing time for DP matching becomes longer. Simply put, the drawback was that the number of words and states that could be processed within a certain amount of time was small.

The purpose of the present invention is to create a standard pattern, G(p,
It is an object of the present invention to provide a continuous speech recognition device that can reduce the number of accesses to H(p, , n, j), shorten the processing time for DP matching, and process a large number of words and states.

The continuous speech recognition device according to the present invention uses input speech as a time series of characteristics A2aI1. ab,..., all,..., a)
I and each car 4 in the instep of the word set Σ=(n)! Standard butter 7 B″=lb′ for n;
, lb'; , -・-, +b7, -.

lb5', a state transition table Δ which is a group of rules (pe q+ n) which means that a state transition p −+ q occurs due to the input of word n, and a group of employment termination states. an automaton storage section storing F, and a table memory 'I' (m, q), P(m
, q), N (m, q) and L (m, q), state p, immediate word designation n, and standard slam time j and l, thereby table memory () (p, n, j) and H(p
, n, j) and each time m of the i-th block of the manual slam.
= (i-1)・IL+1. ..., each feature vector 1b7 of the standard pattern and the feature vector a- of the input button corresponding to the Togo designation signal n sequentially specified in 凰/IL
A distance calculation unit that calculates the distance d(rn, j) from
The table memories G(p, n, j) and H(p,
n, j) and the table storage T(m-1, q)
DP matching recurrence formula value g (m.

j) and the path value h(m, j) as m = (i-1) −
From IL-t-1 to i・IL, and from J-1 to Jn
In summary, the table memories G(p, n, j) and H(p
, n.

j) and the DP matching unit stored in 1) above, and write the recurrence formula value g'(m, J'') calculated by the P matching unit to T(m, q)I to find the minimum value with respect to state q. , word name n that gave this minimum, state specification p.

Route value h(m, J”) is converted to N(m, q), P(m, q)
A comparison circuit that writes L(m, q) and G respectively, the input time m, and the input block number.

A control unit for generating signals such as a word designation n, a state designation p, qi6, and a standard slam time j, and a final time m=
In I, the table memory T (m, q) is determined to be the minimum state q included in the final state group F, and the table memory P (m, q) is determined based on the minimum q and time m. , q ), N(m, q ), and L(m, q ) to determine the recognition result.

Next, the principle of the present invention will be explained. In the method according to the present invention, as shown in FIG. 2, the recurrence formula (6) is calculated by dividing the IL frame of the input button into blocks.

In other words, it can be done using the following dynamic programming method. Initial condition T (0, 0) = O... @T (m, q) =
Flash 1m~o, q~0 G (p, n, j) = ■, and sequentially u from i = l to I/IL (Here, I/If is divisible for simplicity of explanation).
Based on the boundary conditions of the ui equation, the recurrence equation of the 1-step equation is (p, q, n
)εΔ for all pairs (p, n). ′
That is, the boundary condition g(m-1iQ)=T(m-1,p) m:m, , -・
-, m6-print h (m-1, Q) = m-1m=m
, , −・−, m.

However, m, = (j-1) * IL-1, me=m
e ILg (m, -1, j) = G (p, n, j) J =
1. −, J”−114h(m, −1, j)=H(p, n
, j) j=1. ---, J', recurrence formula % formula % ) where d (m, j) = Dis (ay, b7) j = arg
min g (m-1, j')1-2≦j'<j from tn = m, calculate up to m6, and then calculate from J=1 to Jn. After completing one block of calculation, g, (m
a.

j) and h(town, j) respectively as G(p, n, j
).

Store in H(p, n, j).

Next, as a minimization at the word boundary, if T(m, q)>g(m, J”) --------
u*tllen T(m, q)=g(m, J”)N(m
, q) 2n P(rn, q)=p L(m, q)=h(m, J') However, m=m, , ・−, mlI is calculated. The recognition result of the input button is (8 ) , (9) , (
IG, (lυ) is obtained in the same manner as in the conventional method.

As explained above, the amount of calculation of the method according to the present invention is the same as that of the conventional method, but G(p, n, j).

The number of accesses to H(p, n, j) becomes l/IL, and the access time is reduced. Furthermore, if the data from the m1 frame of the input button to the town frame is stored in the register or high-speed memory of the arithmetic unit, d (m ) for reading the standard button 1 frame.

j) m = m, ', . . . up to 1me can be calculated.

Therefore, the standard button only needs to be read once for every l block calculation, and compared to the conventional method, the access time is 1
/IL becomes. Since the calculation of DP matching involves many repeated calculations, parallel processing is possible and it is easy to construct a high-speed calculation section, but it is difficult to make a large-capacity memory section such as a slam memory high-speed. Therefore, the calculation time is short and the memory access time is long. The method according to the present invention has a memory access time of 1 compared to the conventional method.
/IL, the processing time for DP matching becomes close to 1/IL, and the number of words and states that can be processed within a certain amount of time can be increased.

Next, a first embodiment of the present invention will be described. FIG. 3 is a block diagram showing the first embodiment of the present invention.

The standard pattern B'' of the word n included in the word set Σ is stored in the standard pattern storage unit 130. The automaton storage unit 230 stores state transition rules and final state designation information. In this embodiment, state transition rules are stored as shown in FIGS. 9(a) and (b).
A) is called a state specification table and specifies a set of p allowed as state transition p −+ q by word n.

In the example shown in the figure, p corresponding to word n is pl, ..., p
,..., it is exemplified that p6 can occur. 9th
Figure tb) is called a transition table, and the state p is defined by the word n.
A set of states q that can be transitioned from is stored. The example shown in the figure shows that qz can be expressed as qz e qz *q3 *G4.

When an unknown input voice is input from the microphone 100, the human power section 110 performs frequency analysis, converts it into a vector ai representing characteristics, and sequentially sends it to the input bang buffer 120i. Furthermore, the human power unit 110 is 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 controls the rise time l of this voice section signal S.
At this point, an initialization pulse IT 1 is generated. As a result, initialization corresponding to block 10 in FIG. 8 is performed on the G memory 150 and the T memory 200.

When the above initialization is completed, the input slam block signal i is changed to 1, 2, etc. in synchronization with the input of the input feature a1 for the subsequent processing frames. It is counted as... In this manual slam block i, the word designation signal n from the control section 240 changes from 1 to down. At each n, the state table in the automaton storage section 230 is referred to, and the state designation signal p is changed from pl to p8. Also, that n and p
A state designation signal q defined by is also output.

Next, the operation within one cycle of word n1 state p will be described below. The calculations shown in the shaded areas in FIG. 2 are executed through this one cycle. That is, U.

Based on the boundary conditions of the I equation (1), calculate the QQ equation. First, the set of values corresponding to blocks 11 and 12 in FIG. This is performed for the g memory 143 and the h memory 144.Next, the standard slam time signal j from the control unit 240 changes from 1 to J''.For each j,
Input slam time signal m it m, (m, = (i-
1) changes from −1L+1) to m, (me=i −IL).

In the processing section J40, m frames of the input button and j frames of the standard button of the n-th word are first read out, and the distance between the R-dimensional beams shown in equation (6-1) is calculated by the distance calculation section 141. Demeru. In other words, the human-powered slam buffer 120
Then, l pieces of data are sequentially read out from the pattern memory unit 130 according to the signal.

The absolute value of the difference is determined by the absolute value circuit 1411, and the adder 141
2 and the accumulator 1413, Dis (a, b;) is obtained. On the other hand, the three minimum values of dissimilarity can be found by distance calculation and parallelism. In other words, g(m-t*j)eg(In-
1, j-1), g(tn-1, j-2) and h(m-1, j), h(m-xnj-B) from the h memory 144.

Read h(xn-1,j　2) and register pile.

02, G3. Hl, 1-12.83 respectively. Comparison circuit 1422 has three registers oi, G2
, 03, and the minimum value is stored in h(m, j) of the register memory 164 from which the minimum value was obtained. Furthermore, the minimum value g (m-1°J) output from the comparison circuit 1422 is the distance d (tn,
j) and the adder 1421, and the g memo IJ
143 g(m,j). By changing the input slam time m from m to nl, the processing for the standard slam time jI is completed.

Furthermore, after the "standard slam time" becomes the terminal Jn, the acclimatization formula (1 warm result g(In++, j)", 11 according to the signal 5ET3 as shown in block 14 of FIG.
Crne, j) as table memory G(p,
n, j), and stored in H(p, n, j). Subsequently, the comparison shown in block 15 of FIG. 8 is performed in accordance with the signal m2 (m, which changes from m to m8) issued by the control section. That is, as shown in FIG.
T(m, q) is read from the table memory 200 and g(m, J') is read from the g memory 143 according to
, J”), then wp times the signal is emitted, and g(m, J'),
n, p, h (m, J”) is table memory T (m, q)
, N(m, q).

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

With the above operations, the processing of the state p and single word n large entry kabatan block is completed. Furthermore, the state designation signal p is changed from pl to p6, 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 slam block i.

After the input slam block signal becomes I/IL, (8
)t9) The determination process shown in equations (10) and (11) is started. The determination unit 220 is configured as shown in FIG. 6, and first, as the process of block 16 in FIG. 200
T(1,q) is read from the comparison circuit 221° and the minimum value register 222. Using the status register 223, the minimum T(
1, q) is given.

Next, in the process of FIG. 8, the judgment restriction part 2
27 issues an address signal m3゜q to the table memory 190, 210.180 as q:q, m:engine, N(m.

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

These L(m, q) and P(m, q) become the following (7) m and q, and N(m, q,) is output as the recognition result.

By repeating this process until l reaches 70, recognition results can be obtained in sequence.

In the first embodiment described above, in the state p, from the input button time Ins of the single candy n, m. up to (input slam block i
) part calculation result g(ms, Jn) ~ g(me, J“
) are obtained all at once. In order to use this result in the calculation of the initial value of the same human-powered slam block 1 in the next state p' (block 14 in Figure 5), it is necessary to proceed with the < it calculation from the initial state of the automaton to the final state. . On the other hand, if the state transition of the automaton includes a loop (well, the inconvenience arises that the calculation result of state pi box A is used to calculate the initial box of its own state p, and the correct calculation result cannot be obtained.

In the second embodiment, the above-mentioned drawbacks are eliminated, and correct results can be obtained even when a loop is included in the state transition of the automaton. In other words, g (me-J'
) is used to calculate the initial value of the state q of the input button i+1.

The calculation procedure in the second embodiment is almost the same as the method described in detail of the present invention, but the boundary condition of the equation ([, 1 -p)
31.1. -Q7)h (m, 1, 0)==m=
1 g (mt 0 ) =oo m=m, ,..., change to dust, and furthermore, the minimization of equation u6) at the word boundary is iff T(m,, q)>g(me, J') − -us
then T(ms, q)=g(meJ”)N (m,
I q )20 P(m,, q)=p L(m,, q)=h(me, Jn).

The configuration of the device t in the second embodiment is the same as the first embodiment, but the calculation method 11ffi is different because the boundary condition (2) equation and the minimization US equation at the boundary of car candy are used.

The operation time r of the second embodiment will be shown below, and only the different parts will be explained. The boundary condition is (from the first floor formula (17)
According to the control signals SL and T2, the 11th
The boxes shown in blocks 11 and 12 of the figure are g memo IJ 1
43.

h memory 144. Furthermore, the minimization at word boundaries has been changed from the tttp method to the 0-barrel method, and the comparison process shown in block 15 of FIG. 11 is performed in accordance with the control signal m2.

In the second embodiment described above, a loop may exist in the state transition of the automaton, but since the word boundaries are minimized only at the point munii・IL, the word The degree of resilience has deteriorated slightly. On the other hand, table storage T(m, q).

Since the part used for m in N(m, q), L(m, q), and P(m, q) is only me=i・IL, the table storage in the first embodiment is It is possible to have a small storage capacity of iL.

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. In particular, in the present invention, the simplest equation (17) is used as the recurrence equation for DP matching, but it is also possible to use other higher-performance equations.

For example, one possible method is to use Naru. However, in this case, the recurrence formula value is up to time (m-2), and the distance is also up to time (m-2).
Since it is necessary to save up to 1), table memory G (
It is necessary to add something similar to p, n, j). Further, in this explanation, recognition was performed based on distance, but a method of recognition based on correlation may also be considered. In this case, the magnitude relationship is reversed, so it is necessary to replace all unexplained maximum value detection functions with minimum value detection functions.Also.

In this description, table storage 0 (p, n, j) and H (p, n, j) are three-dimensional arrays used in Fortran, but C may be a memory designated by p, n, j. In other words, Chiful memory is p+”
+ All ranges of J (table size” pm-x
Instead of having the range (X rlmax This covers the scope of rights.

[Brief explanation of drawings]

Fig. 1 is a diagram for explaining the prior art, Fig. 2 is a diagram for explaining the present invention in detail, and Figs. FIG. 10 is a memory configuration diagram of an implementation of the present invention--a tomato storage unit. In the drawings, 100...microphone, 110...input section, 12
0... Input button buffer, 130... Standard button storage section, 140... DP matching section, 141...
Distance calculation section, 150...Q memory, 160...H memory, 170...comparison circuit, 180...P memory, 190...1N memory, 200...T memory. 210... L memo, 220... Judgment section, 23
0: automaton storage unit, 240: control unit. To input 〇tan, 11 m Fig. 2 Cover 69 A m, i Fig. 5 Fig. 67 Fig. 7 r - "Sat l Niyo j -" Fig. 8 (A-1) Fig. 9 (b) Figure to: -M100 Figure 11 (Child 1) No. 110 (A 2)

Claims (1)

[Claims] Time series A=all, a! featuring input speech. ,...,
ai. ...e "I" input section and word set Σ=
For each word n in (n), the standard button B”=l)i
, lb; . ..., 呵..., +b:;・, and a state that is a group of rules (p, q, n) that means that state transition p-+q occurs due to the input of word n. An automaton storage unit that stores a transition table Δ and a group F of final states, and a table storage T(m, q) whose address is specified by an input slam time m and a state q. P(m, q), N(m, q) and L(m, q),
A table memory G(p. 1-1)・IL+1
．． .... 1-Distance d (m, j) between each feature vector 7 of the standard pattern corresponding to the word designation signal n sequentially specified in LL (IL is a positive integer) and the feature vector 4 of the input button.
and the table memory G (p , n, j) and H(p, n, j), the value of the table storage T(m-1, q) is a recursion boundary condition, and L(m 1*q) is a path boundary condition. As DP matching recurrence formula value g(m, j) and path value h(m,
m=(1-1) ・From IL+1, m=1-IL
and from j=1 to j=J', the DP matching unit stores the table memories G(p, n, j) and H(p, n, j), and the DP matching unit calculates Write the recurrence formula value g(m, J'') into T(m, q) to find the minimum value with respect to state q, and write the word name n that gave this minimum,
State specification p, route value h(m, J”) as N(m, q), P
(m, q), L(rn. q), and signals such as input kabatan time m, input kabatan block number i, word designation n, state designation p, q, standard baton time j, etc. and a control unit for generating the table memory T(m, q
) is determined as the minimum, and based on the minimum value q and time m, the table storage P(m, q), N(
a determination unit that determines a recognition result with reference to L(m, q) and L(m, q).