JPH0134400B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a continuous speech recognition device that automatically recognizes continuous speech in which one or more words are successively uttered.

Various methods have been tried in the past as voice recognition means. Among them, the pattern matching method is the simplest and most effective method.

In this method, a standard pattern (hereinafter referred to as word standard pattern) is prepared for each word of the word to be recognized, and a comparison operation is performed between it and an unknown input speech pattern (hereinafter referred to as input pattern). (in other words, pattern matching) to calculate the amount representing the degree of mutual difference (hereinafter referred to as the degree of dissimilarity). This is a method of determining.

A patent application filed by the same applicant (December 9, 1981 Patent Application (23)) describes the operating principle of a continuous speech recognition device that operates based on the above pattern matching method. The principle is roughly as follows. That is, word standard pattern B ^vl = (〓 ^vl ₁ , 〓 ^vl ₂ ..., 〓 ^vl _o
,
..., 〓 ^vl _Nvl ) in any permutation, the standard pattern of continuous speech (hereinafter referred to as continuous speech standard pattern) C=B ^v1 ,
Considering B ^v2 , ..., B ^vl , ..., B ^vLmax , matching with the input pattern A=〓 ₁ , 〓 _{2 ,} ..., 〓 _n , . Recognition is performed by determining the permutation of word standard patterns so that the overall degree of difference S(A, C) is minimized.

Input pattern A and continuous voice standard pattern C = B ^v1 ,
The degree of difference from B ^v2 , ..., B ^vl , ..., B ^vLmax is determined as follows. The time point m of the input pattern and the time point n of the continuous speech standard pattern are correlated using an optimal monotonically increasing nonlinear function n=n(m) (hereinafter referred to as time normalization function) as shown in Figure 1. The sum of the distances d(m, n) between the feature vectors at the associated time points along the time normalization function is defined as the degree of dissimilarity S(A, C).

S(A, C) = min ^n(m) _M 〓 ^m=1 d(m, n(m)) ……(1) n=n(m) ……(2) Here, the distance number d(m, n) can be determined using equation (3), for example.

d(m, n)=Dis(〓 _n , 〓 ^v _o )
= _R 〓 ^r=1 ｜a _nr − ^v _or ｜ ...(3) where 〓 _n = (a _n1 , a _n2 , ..., a _nR ) 〓 ^v _o = (b ^v _o1 , b ^v _o2 , ..., b ^v _oR ) Equation (1) is minimized using the following dynamic programming method. Here, v is a word, V is the number of words, ^Nv is the end of the vth word standard pattern, M is the end of the input pattern, l is a digit, and Lmax is the maximum number of digits.

Initial conditions D(l, v, n)=〓 ……(4) l=1~Lmax, v=1~V, n=1~N ^v DB(l, m)=〓 ……(5) l= 0 ~ Lmax, m = 0 ~ M DB (o, o) = 0 ... (6) Based on, m = 1 to M are shown below in order (7) (8)
Find the recurrence formulas of equations (9), (10), and (11) using the set values from equations. That is, D(l, v, o)=DB(l-1, m-1)...(7) l=1~Lmax, v=1~V F(l, v, o)=m-1... (8) Set l = 1 ~ Lmax, v = 1 ~ V, and use the recurrence formula D (l, v, n) = d (n) + D (l, v, n^)
...(9) F (l, v, n) = F (l, v, n^) ... (10) where n^ = argmin [D (l, v, n')] ... (11) n -2n'n is determined for n=1 to ^Nv , v=1 to V, l=1 to Lmax, that is, for one vertical column shown in the shaded area in FIG.

Here, ^argmin _xEX y means x that minimizes y under the condition of xEX. In other words, equation (11) is n-2
Minimize D(m-1, n') under n'n
n' is set to n^. Also, equation (9) is expressed as 3 shown in Figure 2.
indicates that the smallest path is selected from among the two paths,
The reason for limiting the number of allowed routes to three is to prevent the correspondence by the time normalization function from being unnecessarily distorted. Here, the route from which the minimum value used when calculating the degree of dissimilarity is selected is called a matching route, and F(l, v, n) in equation (10) is called route information.

Next, the minimum degree of difference with respect to v at the terminal end ^Nv of the word standard pattern is determined. i.e. However, n^=argmin[D(l,v, ^Nv )]...(15) Find 1vV. Here, DB (l, m) is the order of magnitude difference, FB
(l, m) is called digit path information, and W(l, m) is called digit recognition category.

In this way, by proceeding with the calculation of dissimilarity in one vertical column shown in equations (7) to (15) while increasing m, and finding the minimum dissimilarity with respect to l at the terminal M of the input pattern, (1 ) is obtained. Furthermore, the recognition result of the input pattern is obtained as follows. The digits allowed by the digit dissimilarity DB(l, M) of each digit at the terminal M of the input pattern, that is,
The minimum value is found between the Lmin digit and the Lmax digit, and the digit L where the minimum value is obtained is the number of digits of the input pattern. Furthermore, the recognition result R(L) of the Lth digit is obtained from W(L,M), and the termination of the L-1st digit is obtained from the digit path information FB(L,M). By repeating the above operations in order, the recognition result R(l) for each digit can be obtained.

However, in the aforementioned patent application, the degree of dissimilarity D
The storage capacity of (l, v, n) and the route information F(l, v, n) increases in proportion to the number of digits l. Also, (9),
The amount of calculation for the recurrence formulas (10) and (11) also increases in proportion to the number of digits l. On the other hand, in many cases, such as when recognizing a string of consecutively uttered numbers, there is no need to limit the number of digits.

The purpose of the present invention is to increase the maximum number of digits of input audio by removing the restriction on the number of digits included in input audio.
It is an object of the present invention to provide a continuous speech recognition device that reduces both the storage amount and calculation amount to 1/Lmax of the conventional one when Lmax is used.

Next, the principle of the present invention will be explained using FIG. 4.
Now, a certain point (m ₁ , n ₁ ) on the matching path (m, n (m)) obtained when calculating the overall minimum dissimilarity
The partial dissimilarity obtained from the starting point to the point (m ₁ , n ₁ ) along the matching path is
It is the minimum value of the partial dissimilarities obtained along all matching paths passing through the point (m ₁ , n ₁ ). In other words, the minimum value of the overall dissimilarity obtained along all matching paths passing through a certain point (m ₁ , n ₁ ) is the partial dissimilarity from the starting edge to that point (m ₁ , n ₁ ) and that point ( m ₁ , n ₁ ) to the terminal end.

That is, S(A, C) = min ^n(m) _M 〓 ^m=1 d(m, n) = min ^n(m) _n 〓 ^m=1 d(m, n)＋min ^n(m) _M 〓 ^{m =m1+1} d(m, n)...(16) However, (m ₁ , n ₁ ) is a point on n(m).

Thereby, the minimization from a certain point (m ₁ , n ₁ ) to the terminal can be performed independently of the minimization from the starting end to the point (m ₁ , n ₁ ). Now, consider the minimization from the point X _l to the terminal E _l+L shown in Figure 4. This partial dissimilarity S ^l _L (A′, C ^l _L ) is the partial input pattern A′ (〓 _n+
1 , 〓 _n1+2 , . . . 〓 _M ) and C ^l _L =B ^vl , B ^vl+1 , . . . , B ^vl+L , which are partial patterns of the continuous speech standard pattern. That is, S ^l _L (A′, C ^l _L )=min ^n(m) _M 〓 ^m=m1+1 d(m, n) (17). Here, if there is no limit on the number of digits, change L and find the minimum value of the partial dissimilarity S ^l _L (A′, C ^l _L ), and the number of digits L^ for which the minimum value is obtained is the partial It is the number of digits of input pattern A′. i.e. is given by

Next, consider the minimization from the point X _l+1 to the terminal E _l+1+L . This partial dissimilarity S ^l+1 _L (A′, C ^l+1 _L ) is the partial input pattern (〓 _n1+1 , 〓 _n1+2 ,...〓 _M ) and the continuous speech standard pattern C ^l+1 _L = B〓 ^l+1 , B〓 ^l+2 , ..., B〓 ^l+1+L
required between. Here, C ^l+1 _L is a pattern in which L word standard patterns are connected in any permutation.
It is the same as C ^l _L. Therefore, S ^l+1 _L (A′, C ^l+1 _L )=S ^l+1 _L (A′, C ^l _L )=S ^l _L (A′, C ^l _L )
...(21) Then, the minimization from the point X _l+1 to the terminal E _l+1+L is
The result is the same as minimizing from point X _l to terminal E _l+L .

If there is no limit on the number of digits, each point
The partial dissimilarities _from X ₁ , X ₂ , ..., ) is given by Cl ∧ L in the equation.

On the other hand, the overall dissimilarity is the sum of the partial dissimilarity from the starting point to point X _l and the partial dissimilarity from point X _l to the terminal, as shown in equation (16). is given by Therefore, the input pattern time m ₁
Find the minimum value of the dissimilarity of the points X ₁ , _{X 2} , _... The overall degree of dissimilarity S(A, C) can be determined by:

Next, the degree of dissimilarity S(A, C) which is the basic operation of the present invention
The calculation procedure will be explained using FIGS. 5 and 6.
The dissimilarity calculation is to obtain the recurrence formula of the dynamic program in the order of the time axis m of the input pattern under the initial conditions.

The initial conditions are D(v, n)=〓 ……(23) v=1~V, n=1~N _v DB(o)=o ……(24) DB(m)=〓 ……(25) m=1 to M, and is performed in block 1 of FIG. Next, the dissimilarity calculation for one vertical column parallel to the n-axis at time point m of the input pattern is performed as follows. Set the initial value as D(v,o)=DB(m-1)...(26) F(v,o)=m-1...(27) (This is done in block 2 of Figure 6) Gradually formula is calculated in the direction of decreasing n (performed in block 4 of FIG. 6). Here, d(〓m, 〓 ^v _o ) is the distance between the feature vector 〓 m at time point m of the input pattern and the v-th word standard pattern 〓 ^v _o (3)
(carried out in block 3 of FIG. 6).

As shown in Figure 2, the calculation at (m, n) point is (m
-1, n), (m-1, n-1), (m-1, n-
It is obtained from the degree of dissimilarity of the three points in 2). Next (m,
The calculation for point n-1) is (m-1, n-1), (m-
1, n-2), (m-1, n-3), and the dissimilarity of the (m-1, n) point is not used, so the (m, n) point is calculated. The result is (m-1,
Even if it is stored at point n), it does not affect the calculation at point (m, n-1). Therefore, if the calculation proceeds in the direction of decreasing n, the storage area for the dissimilarity of the m-1 point and the dissimilarity of the m point can be shared. After executing the above recurrence formula calculation in one column, we obtain the dissimilarity D (v, N ^v ) at the end N ^v of the word standard pattern and the digit dissimilarity that is the minimum word dissimilarity calculated so far. D.B.
(m), if the obtained dissimilarity D(v, N ^v ) is smaller, that dissimilarity D(v, N ^v ) is taken as the digit dissimilarity DB(m), and Let the category v to which the standard word pattern belongs be the digit recognition category W(m),
The matching path information F (v, N ^v ) from which the degree of dissimilarity D (v, N ^v ) is obtained is defined as the digit path information FB (m) (this is carried out in block 5 of FIG. 6). The dissimilarity calculations in one vertical column (calculations in blocks 2, 3, 4, and 5 in FIG. 6) performed in this way are performed for V standard word patterns. Next, the time point m of the input pattern
is incremented by one, and the same dissimilarity calculation in one vertical column is performed for V standard word patterns to obtain the terminal end M of the input pattern. Finally, the input pattern is determined based on the digit path information FB(m) and the digit recognition category W(m). This determination method is as shown in Figure 5.
First, the recognition result F(L) at the end M of the input pattern
is obtained from W(M), and then the digit route information FB(M)
Then, the end of the L-1st digit is obtained (this is done in the block of FIG. 6). By repeating the above operations in order, a recognition result R(l) for each digit l can be obtained.

Since the continuous speech recognition device of the present invention is a device that executes the above-described dissimilarity calculation procedure, it requires the following sections. That is, input pattern A=〓 ₁ , 〓
₂ , ..., 〓 _n , ..., 〓 _M and V pre-memorized word standard patterns B ^v = 〓 ^v ₁ , 〓 ^v ₂ , ..., 〓
^v _o , ..., 〓 ^v _Nv (v = 1, 2, ..., V), the signal m indicating the time point of the input pattern is changed from 1 to M, and the signal v indicating the word for each m is set to 1.
a control section that varies the signal n from 1 to N v, and further varies the signal n indicating the time point of the standard pattern for each v from 1 to N ^v , and a difference address specified by the signals v and n of the control section; It has a route information memory section D (v, n) and a route information memory section F (v, n), and at m, v, n specified by the control section, input pattern 〓 _n and word standard pattern of word v. ^〓 _Intervector distance d(〓
_n , 〓 ^v _o ), and the initial condition for each word v at each time point m is the word dissimilarity which is the result of the time point m-1.
The distance d (〓 _n , 〓 ^v _o ) given by DB (m-1) and time point m-1, the degree of dissimilarity D (v, n) at time point m-1, and the route information F (v, n) a recurrence formula calculation unit that sequentially calculates the degree of dissimilarity D (v, n) and route information F (v, n) at time point m by calculating a recurrence formula of the dynamic program with reference to m, find the minimum among the dissimilarities D(v, N ^v ) at the end of each word obtained by the recurrence formula calculation unit, and use this as the digit dissimilarity DB(m), and calculate the path corresponding to this. The information F(v, ^Nv ) is the digit route information
A digit dissimilarity calculating unit which takes the word name v for which the minimum value is obtained as FB(m) as a digit recognition category W(m), and a digit dissimilarity memory unit DB(m) for storing these.
, digit path information memory section FB (m), digit recognition category memory section W (m), digit path information FB (m) and digit recognition category W (m) in reverse order for each digit of the input pattern. and a determination section that determines and outputs the category.

In the present invention, by eliminating the limit on the number of digits of the input pattern, the degree of dissimilarity calculation is not performed for each digit,
It is possible to perform calculations all in one place. That is, in dissimilarity calculation, calculation for finding the minimum value of an input pattern passing through a certain time can be divided into the first half of minimization and the second half of minimization based on the maximization principle. Here, if the digit limit of the input pattern is eliminated, the latter half can be minimized regardless of the degree of difference between the number of digits in the first half from 1 to X. In minimizing this latter half, the number of digits is determined so that the degree of difference is minimized. The minimization of this latter half is the same regardless of the number of digits in the first half, and therefore the minimization of the latter half can be calculated all at once. By applying this principle starting from the first digit, it becomes possible to consolidate the entire dissimilarity calculation into one digit.

In this way, in the present invention, by eliminating the limit on the number of digits of the input pattern, it is possible to calculate the degree of dissimilarity all at one place without having to perform the calculation for each digit, which reduces the amount of calculation and memory. Both are conventional 1/Lmax
(Lmax is the maximum number of digits of the input audio).

Next, the specific configuration of the apparatus of the present invention will be explained with reference to the drawings. FIG. 7 is a block diagram showing a configuration example of the present invention, and FIG. 8 is a time chart of control command signals. The control unit 10 has m1, n
It has a function of controlling other parts by issuing control command signals such as 1 and v1 as shown in FIG. 8, but the details will be explained each time in relation to the operation of the other parts.

The input section 11 analyzes the input section given by the signal Speech in, and outputs a feature vector at regular intervals. This continuous analysis includes, for example, frequency analysis using a filter bank composed of multi-channel filters. The input unit 11 also has a function of monitoring the level of input audio and detecting the start and end of the audio, and transmits the detected time to the control unit 10 by a signal SP. The input pattern buffer 12 is
After the start of the voice is detected, the feature vector am given from the input section 11 is stored in accordance with the signal m3. Signal m3 is a signal corresponding to time point m of the input pattern. The standard pattern memory unit 13 stores V word standard patterns B ¹ , B ² , . . . B ^v , and the standard pattern length memory unit 14 stores the length N ^v of the word standard pattern B ^v .

The control unit 10 issues a signal v1 instructing the word v of the standard pattern to the standard pattern length memory unit 14, reads out the length ^Nv of the word standard pattern ^Bv , and outputs the signal n1 corresponding to the time point n of the word standard pattern. Occur.
According to the signal n1, the feature vector 〓m of the input pattern is read out from the input pattern buffer 12, ^〓vo _is sequentially read out from the standard pattern memory section, and the distance calculation section 15 calculates the equation (3).

In the distance calculation section 15, as shown in FIG. 11, the accumulator 153 is first cleared by the signal Cl153, R pieces of data are read from the input pattern buffer 12 and the standard pattern memory section 13 according to the signal r1, and the absolute value circuit The absolute value is obtained in step 151 and added in adder 152, and the distance d in equation (3) is obtained.
(〓 _n , 〓 ^v _n ) is found in the accumulator 153,
This distance is input to the recurrence formula calculation section 17.

The dissimilarity memory section 18 and the route information memory section 19 have a two-dimensional configuration as shown in FIG. This is a one-dimensional configuration shown in .

The initial set of the recurrence formula calculation is performed by the signal CL from the control unit 10 before the voice is input, and the data is sent to the dissimilarity memory unit 18, digit dissimilarity memory 21 (23), (24),
The value shown by equation (25) is set.

The recurrence formula calculation unit 17 is a part that performs block 4 in FIG. 6, and executes recurrence formulas (28), (29), and (30). That is, the recurrence formula calculation unit 17 operates on three difference degree registers D1, D2, and D as shown in FIG.
3, a comparison circuit 171 that calculates the minimum value of the three registers D1, D2, and D3, and an adder 172.
and three route registers F1, F2, and F3. The signal n1 issued from the control unit 10,
Three dissimilarities D(v, n), D
(v, n-1), D(v, n-2) and three route information F(v, n), F(v-n-1), F(v, n-2)
are read out and the difference registers D1, D2,
D3 and path registers F1, F2, and F3. The comparison circuit 171 has three dissimilarity registers D.
The minimum value is detected from 1, D2, and D3, and the degree of dissimilarity Dn^ (where n^ is 1, 2, or 3) is the minimum value obtained.
A gate signal n^ is generated that selects the path register Fn^ corresponding to the path register Fn^. The contents of the route register Fn^ selected by the gate n^ are stored in the route memory section 19 F.
(v, n). Further, the minimum value D(v, n^) of the degree of dissimilarity outputted from the comparison circuit 171 is the distance d(〓m, 〓 ^v _o ) outputted from the distance calculation section 15.
is added by the adder 172 and stored in the dissimilarity memory unit 18.

This recurrence formula calculation is calculated from n=1 to N ^v ,
The degree of dissimilarity D(v, N ^v ) which is the result is calculated for each v.

The digit dissimilarity calculation unit 20 is a part that performs block 5 in FIG. 6, and sequentially finds the minimum value of V dissimilarities D(v, N ^v ). In other words, the digit difference calculation unit 2
0, as shown in FIG. 13, the comparison circuit 201,
Register 202 that holds the degree of dissimilarity D (v, N ^v )
, a register 203 that holds the category v to which the standard word pattern belongs, and a register 204 that holds the route information F(v, N ^v ). Control part 1
According to the signal n1 emitted from 0, the dissimilarity memory section 18 and the route memory section 19 calculate the dissimilarity degree D(v,
N ^v ) and route information F (v, N ^v ) are read out and stored in registers 202 and 204, respectively, and the category v to which the word standard pattern belongs is stored in register 203.
is stored in On the other hand, the comparison circuit 201 compares the degree of dissimilarity D(v, ^Nv ) with the degree of digit dissimilarity DB(m) read out from the digit dissimilarity memory section 21, and calculates the degree of dissimilarity D
Gate signal v^ that determines that (v, N ^v ) is smaller
occurs. Register 20 according to gate signal v^
Dissimilarity D held at 2,203,204
(v, N ^v ), category v, route information F (v, N ^v )
are stored in the DB(m) of the digit dissimilarity memory unit 21, the digit recognition category memory unit 22W(m), and the digit path memory unit 23FB(m), respectively. Furthermore, the control unit 10
(26) of the dissimilarity calculation unit in one vertical column, which is the part performed in block 2 of FIG. 5 by the signal Cl2.
The initial set shown in equation (27) is performed. That is, DB(m-1) is read from the digit dissimilarity memory section 21, stored in D(v, o) of the dissimilarity memory section 18, and stored in F(v, o) of the path memory section 19.
-1 is stored. The determining unit 24 is a part that performs block 6 in FIG. 5, and outputs the recognition result R(l) of each digit of the input pattern from the digit path information FB(m) and the digit recognition category W(m). That is, the determination unit 24
As shown in FIG. 14, it is composed of a register 244 that holds digit path information F(m) and a register 245 that holds recognition results. When the end of the audio is detected, the input unit 11 notifies the control unit 10 by the signal SP, and then the control unit 10
The determination unit 24 starts processing. After receiving the signal m1, the determination control unit 246 determines m
= M, and sends the address signal m2 to the digit path memory section 23 and digit recognition category memory section 22, and
(M) and W(M) are read and stored in register 244 and register 245. The contents of the register 245 are output as the recognition result. Furthermore, the determination control unit 246 issues an address signal m2 as m=(content of the register 244) to the digit path memory unit 23 and the digit recognition category memory unit 22, and outputs FB(m) and W
(m) is read and register 244 and register 2
45. This process is performed sequentially from m=M to m
By repeating this until =0, the recognition result is output from register R.

Although the principle of the present invention and one configuration example thereof have been explained above, these descriptions do not limit the scope of the present invention. In particular, input pattern 〓 _n and standard pattern 〓 ^v _o
Although the distance between

d(m, n)= _R 〓 ^r=1 (a _nr −b ^v _or ) ² …(31) d(m, n)= _R 〓 ^r=1 (a _nr ×b ^v _nr ) ……(32 ) Also, the recurrence formula for calculating the degree of dissimilarity is (28),
Various forms other than formulas (29) and (30) can be considered, and instead of formulas (28), (29), and (30), Japanese Patent Publication No. 56-28278
It is clear that the forms described in this issue can also be used.

[Brief explanation of drawings]

Figure 1 is a diagram showing an example of matching paths, Figure 2 is a diagram showing matching paths allowed in the recurrence formula, and Figure 3 is a diagram showing the order of dissimilarity calculation. Fig. 4 is a diagram showing the principle of the present invention, Fig. 5 is a diagram showing the calculation order of the judgment process, and Figs. 6-1 and 6-2 are diagrams showing the calculation procedure of the present invention. FIG. 7 is a configuration diagram of an embodiment of the present invention, FIG. 8 is a time chart for explaining the operation of the embodiment of the present invention, and FIG. 9 is a flowchart showing the difference degree. FIG. 10 is a configuration diagram of a memory unit and a route information memory unit; FIG. 10 is a configuration diagram of a word dissimilarity memory unit, a word route information memory unit, and a word recognition category memory unit; FIG. 11 is a configuration diagram of one configuration of the present invention. FIG. 12 is a configuration diagram of a distance calculation section which is one of the elements, FIG. 12 is a configuration diagram of a recurrence formula calculation section, and FIG. 13 is a configuration diagram of a digit dissimilarity calculation section.
FIG. 14 is a configuration diagram of the determination section. Figure 7, 1st
1, 12, 13, and 14, 10...control section, 11...input section, 12...input pattern buffer, 13...standard pattern memory section, 14...standard pattern length memory section , 15... Distance calculation section, 17... Recurrence equation calculation section, 18... Dissimilarity degree memory section, 19... Route information memory section, 20...
... Digit difference calculation unit, 21... Digit difference memory unit,
22... Digit recognition category memory unit, 23... Digit path information memory unit, 24... Judgment unit, 151... Absolute value circuit, 152... Adder, 153... Accumulator, 171... Comparison circuit, 172... ...Adder, D1, D2, D3...Register that holds the degree of dissimilarity, F1, F2, F3...Register that holds the path, 201...Comparison circuit, 202...Register that holds the degree of word dissimilarity, 203...Register for holding categories, 204...Register for holding route information, 244...Register for holding word route information, 245...Register for holding and outputting recognition results, 246...Determination control unit.

Claims (1)

[Claims] 1. Input pattern A consisting of one or more words that is a time series of feature vectors = ₁ , ₂ , ..., _n , ...,
〓 _M and V word standard patterns stored in advance B ^v = 〓 ^v ₁ , 〓 ^v ₂ , ..., 〓 ^v _o , ..., 〓 ^v _N v (v =
1, 2, ..., V) connected in any order C = B ^v1 , B ^v2 , ..., B ^vl , ...,
B A time function n that makes the time axis m of the input pattern correspond to the time axis n of the continuous audio standard pattern between ^vLmax
Regarding the operation to find the minimum value of the degree of dissimilarity defined as the sum of the vector distance d (〓 _n , 〓 _o ) between the input pattern 〓 _n and the continuous speech standard pattern 〓 _o on (m), the time of the input pattern Digit dissimilarity DB(m) indicating the minimum cumulative amount of distance between vectors given by the optimal time function n(m) between the part of the input pattern with point m as the end of the word and the word standard pattern and,
Digit path information indicating the first time point of this time function
In a continuous speech recognition device that sequentially obtains FB(m) and digit recognition category W(m) for time points m of an input pattern, and finally determines the number of digits of the input pattern and the recognition result of each digit, The signal m indicating the time point of the input pattern is varied from 1 to M, the signal v indicating the word is varied from 1 to V for each m, and the signal n indicating the time point of the standard pattern is varied from 1 to N for each v. a control section that changes the information up to ^v , a dissimilarity memory section D ( v , n) whose address is specified by the signals v and n of the control section, and a route information memory section F (v, n). and at m, v, and _n specified by the control unit , _the vector distance ^d (〓
_n , 〓 ^v _o ), and the initial condition for each word v at each time point m is the word dissimilarity which is the result of the time point m-1.
Given by DB(m-1) and time point m-1, the distance d(〓 _n , 〓 ^v _o and the degree of dissimilarity D at time point m-1
(v, n) and the route information F(v, n), calculate the recurrence formula of the dynamic program, and calculate the dissimilarity degree D(v, n) at time point m and the route information F(v, n). ), and a recurrence formula calculation unit that sequentially calculates the difference D(v, N ^v ) at the end of each word calculated by the recurrence formula calculation unit at each time m. Set the degree of dissimilarity DB (m) to the corresponding route information F (v, N ^v ) as the digit route information FB.
(m), and a digit dissimilarity calculation unit which sets the word name v for which the minimum value was obtained as a digit recognition category W(m), and a digit dissimilarity memory unit DB(m) for storing these;
The category of each digit of the input pattern is determined in reverse order based on the digit path information memory section FB (m), the digit recognition category memory section W (m), the digit path information FB (m), and the digit recognition category W (m). A continuous speech recognition device characterized by having a determination unit that makes a determination and outputs the result.