JPS5898796A - 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 that automatically recognizes continuous speech in which one or more words are successively uttered.

As a means of voice recognition, the Narasa method has been tried in the past. Among them, the pattern matching method is the simplest and most effective method.

This method is a must-recognize W! A standard pattern (hereinafter referred to as the word standard pattern) is prepared for each word in the search, and the input unknown voice pattern (hereinafter referred to as the input pattern) is prepared in advance.
(hereinafter referred to as the degree of dissimilarity)
This is a method for determining that a word belongs to the same word as the word standard pattern with the least difference, that is, the minimum degree of difference.

JP-A-51-104204 describes the operating principle of a continuous speech recognition device that operates based on the pattern matching method described above. The principle behind this is as follows.

That is, a pattern obtained by connecting several word standard patterns in any permutation is considered as a continuous speech standard pattern (hereinafter referred to as a continuous speech standard pattern), and is matched with the entire input pattern. Recognition is performed by determining the number of word standard patterns and the permutation of word standard patterns so that the overall degree of difference is minimized.

In reality, the above minimization is divided into two stages: word-by-word minimization and overall minimization, and each minimization is performed using dynamic programming (hereinafter, the minimization is performed using dynamic programming). This matching is called DP matching).

In the minimization of each word using the device described in the above publication, the input pattern is divided into every possible word,
DP matching with word standard patterns is performed for all of them. That is, if the input pattern length is N and the number of word standard patterns is V, then M.V times of DP matching are required.

By the way, the number of times of the above DP matching is Lmax・V
(the maximum possible number of digits of the input pattern is Lmax) is based on IEEE
OUSTlCS, SPEECH, AND SIGNAL
PROCESSING, VOL ASSP-29NO.
2APRIL 1981, pages 284 to 297. Next, this method (hereinafter referred to as HLB method)
I will give an outline. Input pattern A and continuous voice standard pattern C
=Bv1, Bv2,..., Bvl,..., BvLm
The degree of difference with ax 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=nm (hereinafter referred to as time normalization function) as shown in Figure 1, and The distance d(m
, n) along the time normalization function is defined as the degree of dissimilarity 8 (A, C).

ζ Here, the distance d(m, n) can be determined, for example, using equation (8).

Equation (1) is minimized using the following dynamic programming method.

In other words, the initial condition D (0, 0) = O (4) D (m, Q) = ∞ m = 0 ~ M (5) F (
m, O) = m m = 0 ~ M ...... Based on (6), the recurrence formula is written for m = L (m) ~ U (m), m = 1 ~ M, that is, the first Find the shaded area in the figure.

U(m)=2m-1...(10)L(m)=(
m+1)/2...(11) Here argmin
y means x that minimizes y under the condition of xEX. In other words, equation (9) is D(
m-1, n') is set to be n'. In addition, equation (7) indicates that the minimum is selected from the three routes shown in Figure 2, and limiting the number of allowed routes to three causes the correspondence by the time normalization function to become unnecessarily distorted. This is to prevent this. Here, the path that selects the minimum value used when calculating the degree of dissimilarity is called the matching path, and F(m) in equation (8).

n) is called route information. The recurrence formula (7), t8), (
D(N, N) obtained by calculating 9) from the end M of the input pattern to the end N of the continuous speech standard pattern is the degree of dissimilarity S(A, C) in equation (1).

By the way, at a certain point (ml, nl) on the matching path (m, n(i)) obtained when calculating the overall minimum dissimilarity, the point (m
The partial dissimilarity obtained up to the point (ml, nl) is
nl) is the minimum value of partial dissimilarity obtained along all matching paths passing through nl). That is, a certain point (ml, nl
) The minimum value of the overall dissimilarity obtained along all matching paths passing through It is given by the sum of the minimum values of each degree.

That is, where (ml, nl) is the point on the matching path where S(A, C) was obtained. Now, considering the break points for each digit of the continuous speech standard pattern, the minimum partial dissimilarity can be found for each digit, and the minimum overall dissimilarity can be obtained as the sum of the results. Therefore, the minimum degree of difference between the input button A and the continuous voice standard pattern C=BVs, Bv2, . . . , BVl, . . ., B can be determined as follows. First, each word standard pattern in the first digit of the continuous speech standard pattern is matched with the input pattern, the minimum value of the degree of dissimilarity is found, and the result is used as the initial value for matching in the second digit. The standard pattern of each word in the digit is matched with the input pattern. After matching up to the Lmax digit,
The minimum value of the degree of difference for each digit at the terminal end M of the input pattern is determined to obtain the optimum digit aL. The degree of difference for the L-th digit is obtained, and the nine matching paths are traced in reverse to obtain the recognition category for each digit in sequence.

Next, the calculation procedure of the HLB method will be explained using FIGS. 3 to 6. Figure 3 is a diagram showing the progression order of difference degree calculation, Figure 4 is a diagram showing the difference degree calculation of the first digit, Figure 5 is digit route information FB (l, m), digit recognition fiber category Figure N6, which is a diagram showing the W(J, m)-wise determination calculation order, is a flowchart representing Algorithm 5 described on pages 289 to 290 of the cited document. Here, m is the time point of the input pattern, n is the time point of the standard pattern, V is the word, and l
is a digit, M is the end of the input pattern, Nv is the end of the Vth word standard pattern, ■ is the number of word standard patterns, Lmin is the minimum number of digits of the input pattern, and Lmax is the maximum number of digits. The dissimilarity calculation is performed using the initial conditions (4), (5), and (6) as the recurrence formula (γ), (8), and (9) in the region L shown in Fig. 3). In other words, each digit of the continuous speech standard pattern can be found in order. Setting of initial conditions is performed in block 1 of FIG. Next, the degree of difference calculation at the l-th digit is performed as follows.

The digit dissimilarity DB(l-1, m), which is the result of the previous digit, is set as the initial value of the dissimilarity D(m, 0) (this is done in block 2 of FIG. 6), (7), The recurrence formulas shown in equations (8) and (9) are converted into upper limit U(m) and lower limit L(m
) (carried out in block 3.4 of Figure 6). Here, the upper limit U(m) means the left line segment AB (shown in equation (10)) and the upper line segment BE in FIG. 4, and the lower limit L(m) means the lower line segment AC
and the line segment CE on the right (shown in equation (11)). Calculate up to the end NY of the word standard pattern,
The dissimilarity degree D (m, NY) at the terminal NY is the word dissimilarity degree D
(v, m) (performed in block 5 of Fig. 6.1
After calculating the word standard pattern, the word dissimilarity is (v,
m) 7) Find the minimum value and use the minimum value as digit dissimilarity DB ()
1m), the category v to which the word standard pattern whose minimum value was obtained belongs is defined as the digit recognition category W(l, m), and the matching path information F(v, m) from which the minimum value was obtained is the digit path information. FB(l,m) (block 6 in Fig. 6)
). In this way, the Lmax is calculated from the first digit.
After calculating the degree of difference up to the digit, the obtained digit path information F
The manual pattern is determined from B(l, m) and digit recognition category W(l, m). First, from the digit dissimilarity DB (l, N) of each digit at the end M of the input pattern, find the minimum value between the allowed digits, that is, between Lmin digit and Lmax digit. ), the obtained digit of the minimum value is finally determined as the nine-digit number of the input pattern. Subsequently, as shown in FIG. 5, the L-th digit recognition result R(L) is obtained from W(L, N), and the digit path information FB(L) at the terminal end N of the input pattern is obtained.

M) to obtain the end of the L-1st digit (performed in block 8 of FIG. 6). By repeating the above operations in order, a recognition result R(l) for each digit can be obtained.

As explained above, in the HLB method, it is only necessary to perform DP matching seven times for each digit, so there are a total of Lmax V times of DP matching.
P matching is required.

On the other hand, the method disclosed in Japanese Patent Application Laid-Open No. 51-104204
V times of DP matching are required. Normally, when the maximum number of digits Lmax of the input pattern is about 5, and assuming the frame period is 20 mg, the input pattern length M is 100.
This means that the amount of calculation for the HLB method is significantly smaller.

In a speech recognition device, the recognition response time is the time from when the end of speech is detected to when the recognition result is output. By the way, in the HLB method, after the input pattern necessary for matching the first digit is obtained, matching of the first line is started, and matching is performed sequentially up to the Lmax digit to obtain a recognition result. Regarding the lth digit in the middle, when the input pattern is obtained up to point E (m,, n,) in the upper right corner shown in Figure 4, that is, the maximum value up to the lth digit of the continuous speech standard pattern. If the pattern length is no and the maximum pattern length of the word standard pattern is Nmax, then from the relationship of the coordinates of point E, n, = L(m,) (14), so m, = 2.Nmax.l-1 (16), and when the input pattern is obtained up to 2.Nmax.l-1 points, matching of the l-th digit can be performed. Now, the input pattern has L digits, and the average word length of each digit is N, and L=Lma
If it is assumed that
The calculation can only proceed up to ax digits, and the remaining Lm
The ax digit will be calculated later, and this L
The calculation time for max digits becomes the recognition response time, resulting in a response with a large delay. On the other hand, in order to shorten this recognition response time, a complex high-speed arithmetic unit capable of parallel processing or pipeline processing is required.

The purpose of the present invention is to improve the above HLB method, and to
It is an object of the present invention to provide an economical continuous speech recognition device by shortening the recognition response time and reducing the overall amount of calculation.

Therefore, the calculation order of the HLB method is changed to derive a new calculation principle called the VLB method, which is the principle of the present invention. H
In the LB method, as shown in the flowchart of Figure 6, the degree of dissimilarity between each word standard pattern and the input pattern is calculated,
First, calculations are performed using the standard button and input pattern for each word, and then the same calculation is performed by increasing the digit by one. That is, the sixth
The loose order of calculations shown in blocks 3 and 4 of the figure is, from the innermost side, the time point n of the word standard pattern, the time point m of the input pattern, the number v of the word standard pattern, and the digit number l. Here, it is shown that it is possible to change the order of the calculation loop and set the outermost point to time point m of the input pattern, as shown in the flowchart of FIG. Since the time normalization function n- of DP matching is a monotonically increasing function, the initial value DH (l-1, m-1) of the lth digit is determined by the data before time point m-1 of the input pattern. ing.

That is, if the degree of dissimilarity calculation has been completed at all points before the time point m-1 of the input pattern, the degree of dissimilarity calculation at the time point m of the input pattern can be performed. That is, as shown in the shaded area in FIG. 8, it is possible to calculate the degree of dissimilarity in one vertical column that is parallel to the n-axis and includes each digit. To calculate the dissimilarity of one vertical column including each digit of jin, use the initial value DB (
l-1, m-1) and the degree of dissimilarity D(
m-1, n) are required, and these are obtained by calculation at the m-1 point. However, it is necessary to store the degree of dissimilarity D (mt, n) at point m-1 and the route information F (m-1, n) for each digit V and each word V. Therefore, the degree of dissimilarity D (m-1, n) and F at digit l and word V
(m-le”) respectively D(l, v, n) and F(
l, v, n). This D(l, v, n) and F(l,
v, n) is shown in FIG. 15.

The calculation procedure of the VLB method in which the calculation order is changed in this way will be explained using FIGS. 7 and 8. The dissimilarity calculation is to obtain the recurrence formula of the dynamic programming in the order of the time axis m of the input pattern within the region between the upper limit U (e) and the lower limit L-, as shown in FIG. The initial condition for dissimilarity calculation in the VLB method is D(l, v, n)∞ (17) l = 1 ~ Lm
ax, v=1~V, n=1~NvDB(l, m)=∞・
...(18) l = 0 to Lmax, m = 0 to M DB (0,0) = 0 (18), and the seventh
This is done in block l of the figure. Next, the time point m of the input pattern
The dissimilarity calculation for one vertical column parallel to the n-axis is performed as follows. First, the vector distance between the feature vector am at time point m of the input pattern and the Vth word standard pattern by is calculated using equation (8) (block 2 in Figure 7).
). Next, a dissimilarity calculation is performed in one vertical column for each digit. In this vertical column dissimilarity calculation, the initial value is D(l
,v,0)=DB(l-1,m-1)...(2
0) F(l, v, 0) = m-1... (21) (performed in block 3 of Figure 7), the recurrence formula is expressed as U(m) and L(m) (This is done in block 4 of FIG. 7.) As shown in Figure 2, the calculations for (m, n) points are (m-1, n), (m-1
, n-1) and (m-t, n-2). The calculation for the next (m, n-1) point is (m-t, n
-1), (m-1, n-2), (m-1, n-3) 3
It is calculated from the dissimilarity of the points (m-1, n).Since the dissimilarity of the point is not used, the calculation result of the (m, n) point is (m-1, n)
Even if it is stored in the point (m, n-1), it does not affect the calculation of the 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 clarification formula calculation in one column, the dissimilarity D (l, v, Nv) at the end Nv of the word standard pattern and the digit dissimilarity DB which is the minimum word dissimilarity calculated so far
(l, m), and obtained dissimilarity D(l, v, Nv
) is smaller, its dissimilarity D(l, v, NY)
Let be the digit dissimilarity degree DB (l, m), let the category V to which the word standard pattern belongs be the digit recognition category W (l, m),
The matching route information F(l, v, Nv) from which the degree of dissimilarity D(l, v, NY) was obtained is converted into the digit route information FB(l, m)
(performed in block 5 of FIG. 7). Dissimilarity calculation for one vertical column performed in this way (block 2 in Figure 7)
, 3.4.5) are performed for V word standard patterns. Next, the time point m of the input pattern is increased by one, and the same dissimilarity calculation in one vertical column is performed for seven word standard patterns to obtain the end point M of the input pattern. Finally, the input pattern is determined based on the digit path information FB (l, m) and the digit recognition category W (l, m).

This determination method is similar to the determination method of the HLB method. First, digit difference degree DB (
l, N), that is, Lmin digit, Lma
The minimum value is found among x digits (this is done in block 6 of FIG. 7), and the digit L for which the minimum value is obtained is the number of digits of the input pattern. Furthermore, the recognition result R(L) of the L-th 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) (block 7 in FIG. ). By repeating the above operations in order, the recognition result R(
l) is obtained.

Since the continuous speech recognition device of the present invention is a device that executes the above-mentioned VLB method, it requires the following parts. That is, input pattern A and continuous voice standard pattern C = Bvl, Bv
2, ..., Bvl, ..., BvLmax, and the signal m that misses the time point of the input pattern for each part below is varied from 1 to M, and the signal V indicating a word for each m is varied from 1 to M. ■ A signal l indicating the digit for each V
from 1 to Lmax and a signal n indicating the time point of the standard pattern by varying it from 1 to Nvt, and a dissimilarity memory section D (addressed by the signals l, v, n of the control section). l, v, n) and the route information memory section F(
l, v, n), and word standard patterns b, v, n of the word V that are sequentially specified by the control unit at each time point m.
= 1~Nv and the input pattern am vector distance d(a
m, bnv) n=l~N7; a distance memory unit d(n) that stores this distance; and at each time point m, initial conditions are first set for each digit l and each word v. Digit dissimilarity DB (l-1.
m-1) and the digit path information FB (l-1, m-1), and the distance d(n) and the degree of dissimilarity D at the time point m-1
(l, v, n) and the route information F(l, v, n) to calculate the recurrence formula of the dynamic program and calculate the dissimilarity degree D at time point m.
(l, v, n) and route information F(l, v, n) are sequentially obtained, word dissimilarity degree D(l, V, NY) and word route information F(l
, v, Nv); at each time point m, from each word dissimilarity degree D(l, v, NY) calculated by the recurrence formula calculation unit for each digit l; Find the minimum, set this as digit dissimilarity DB(l,m), and set the corresponding word path information F(l, v, NY) as digit path information FB(l,m
), and each word v for which the minimum value was obtained is defined as a digit recognition category W
(l, m); a digit difference memory unit DB (l, m) for storing these, a digit path information memory unit FB (l, m), and a digit recognition category. Memory part W (
l, m), and a determination unit that determines and outputs the category of each digit of the input pattern in reverse order based on the digit path information FB (l, m) and the digit g, m categories W (l, m). ing.

In this way, by using the VLB method, which is the principle of the present invention, the dissimilarity calculation can proceed in the time axis direction of the input pattern. As soon as a voice input is detected due to a break, calculations can be started, and the calculations can be performed sequentially in synchronization with the voice input, so it is possible to start the determination process of blocks 6 and 7 in FIG. 7 at the same time as the voice ends. can. Therefore, compared to the conventional HLB method, the recognition response time can be shortened. In addition, distance calculation is surrounded by a loop of n, m, v, l as shown in block 3 in Figure 6 in the HLB method, but n, as shown in block 3 in diagram 7 in the VLB method. ,
It is surrounded by a loop of v and m. That is, the number of distance calculations in the HLB method is Nv.

M.■.Lmax, and the number of distance calculations in the TLB method is N'-V,M. Therefore, compared to the conventional HLB method, the amount of distance calculation can be reduced to 1/Lmax.

Next, the specific configuration of the apparatus of the present invention will be explained with reference to the drawings. FIG. 9 is a block diagram showing a configuration example of the present invention, and FIG. 10 is a time chart of control command signals. The control unit 10 is m1.

It has a function of controlling other parts by issuing control command signals such as n1 and v1 as shown in FIG. 10, but the details will be explained each time in relation to the operation of the other parts. The human power section 11 analyzes the input speech given by the signal Spee-chin and outputs a feature vector at regular intervals. This audio analysis includes, for example, frequency analysis using a filter bank composed of multi-channel filters. The input unit 11 also monitors the level of input audio,
It has a function of detecting the start and end of audio, and transmits the detected time to the control unit 10 by a signal SP.

The input pattern buffer 12 stores the feature vector am given from the input unit 11 in accordance with the signal m3 after the start of the voice is detected. Signal m3 is time point m of the input pattern
This is a signal corresponding to The standard pattern memory section 13 has 7
word standard patterns B1, B2. ...Remember Bv,
The bran standard pattern length memory section 14 stores the length Nv of the word standard pattern Bv.

The signal vl is a signal corresponding to the word v of the continuous speech standard pattern, and the control unit 10 reads the length Nv of the word standard pattern Bv from the standard pattern length memory unit 14 according to the signal v1, and calculates the time point of the word standard pattern. signal n1 corresponding to n
occurs. Input pattern buffer 12 according to signal n1
The feature vector am of the input pattern is read out from the standard pattern memory section, and b1v, b2v, ..., bvNv are read out from the standard pattern memory section.
are sequentially read out, the distance calculation section 15 calculates the equation (8), and the distance d(n), n=1.2...Nv, is stored in the distance memory section 16.

In the distance calculation section 15, as shown in FIG. The values are calculated and added by the adder 152, and the distance Dia of equation (8) is obtained.
(am, bnv) is determined by the accurator 153, and this distance is output to the distance memory section 16.

The initial value for the recurrence formula calculation is set by the signal CL of the control unit 10 before the voice is input, and the initial value is set by the signal CL of the control unit 10.
, to the digit difference memory unit 21 (17), (18), (19
) is set.

The recurrence formula calculation unit 17 is a part that performs block 4 in FIG. 7, and executes recurrence formulas (22), (23), and (24).

That is, the recurrence formula calculation unit 17 has three dissimilarity registers D1, D2 . D3 and its three registers D1, D2 . A comparison circuit 171 that calculates the minimum value of D3, an adder 172, three path registers F1,
Consists of F2 and F3. The dissimilarity memory section 18 and the route memory section 1 are controlled by the signal n2 issued from the control section 10.
9, the three dissimilarities D(l, v, n), D(l, v, n
-1), D(l, v, n-2) and three route information F(l
, v, n), F(l, v, n-1), F(l, v, n-
2) are read out and set in the respective difference registers D1. D2. D
3 and stored in route registers F1, F2, and F3. The comparison circuit 171 includes three dissimilarity registers DI, D2 . The minimum value is detected from D3, and a gate signal n is generated to select the path register Fn corresponding to the dissimilarity register Dn (n is any one of 1, 2, or 3) from which the minimum value was obtained. The contents of the route register Fn selected by the gate signal n are stored in F(l, v, m) of the route memory section 19. Further, the minimum value D(l
, v, n) are added to the distance d(n) read from the distance memory unit 16 by an adder 172, and stored in the dissimilarity memory unit 18.

This recurrence formula calculation is calculated from n=U(m) to L(m), and the resulting word dissimilarity D(l, v, Nv) is calculated for each v and each l.

The digit dissimilarity calculation unit 20 is a part that performs block 5 in FIG. 7, and sequentially calculates the minimum value of V word dissimilarities D(l, v, Nv). That is, the digit difference calculation unit 20
As shown in the figure, the comparison circuit 201 and the word dissimilarity degree D(l
, v, Nv), and a register 203 that holds the category v to which the standard word pattern belongs.
Register 204 that holds route information F (l, v, Nv)
It consists of

The signal l1 is generated Lmax times in one section of the signal vl. This signal l1 is a signal corresponding to digit l of the continuous voice standard pattern. According to the signal l1 issued from the control unit 10, the word dissimilarity degree D (l, v, Nv) and the word route information F (l, v, N
v) are read out and stored in registers 203 and 205, respectively, and the category v to which the word standard pattern belongs is stored in register 204. On the other hand, the comparison circuit 201 compares the word dissimilarity degree D(l, v, Nv) with the digit dissimilarity degree DB(l, m) read out from the digit dissimilarity memory unit 21, and determines the word dissimilarity degree. If it is determined that D(l, v, Nv) is smaller, a gate signal v is generated. Word dissimilarity degree D held in registers 202, 203, and 204 according to gate signal v
(l, v, Nv), category v word path information F(l, v
, Nv) are respectively DB(l, Nv) of the digit difference memory unit 21.
m), are stored in W(l,m) of the digit recognition category memory section 22, and FB(l,m) of the digit path memory section 23. Furthermore, a signal l is issued from the control unit 10 following the signal l1.
12, the initial setting shown in equations (20) and (21) of the dissimilarity calculation for one vertical column, which is the part performed in block 3 of FIG. 7, is performed. In other words, the digit difference memory section 21
DB(l, m-1) is read out, stored in D(l, v, 0) of the dissimilarity memory section 18, and stored in F of the path memory section.
m-1 is stored in (l, v, 0). The determination unit 24
This is the part that performs block 6.7 in FIG. 7, and outputs the recognition result R(l) of each digit of the input pattern from the digit path information FB(l,m) and the digit recognition category W(l,m). That is, as shown in FIG. 14, the determination unit 24 includes a comparison circuit 241, a register 242 that holds the minimum digit difference, a register 243 that holds the number of digits, and digit path information F (l, m). register 244 to hold the recognition result and register 2 to hold the recognition result.
It consists of 45 pieces. When the end of the voice is detected, the input section 11 notifies the control section 10 by a signal SP, and the control section 10 then issues a signal m1 to the determination section 24, and the determination section 24 starts determination processing. The determination control unit 246 receives the signal m
1, he issues a signal l3 to the digit difference memory section 21.

According to the signal l3, the digit disparity DB(l, N) at the end M of the input pattern is sequentially read out from the digit disparity memory unit 21, and the minimum value is sequentially determined by the comparator circuit 241 and stored in the register D. The number of digits l at that time is stored in register L. After Lmax digit differences are read out according to signal l3, the contents of register L indicate the number of digits of the input pattern. The determination control unit 246 issues an address signal m2 to the digit path memory unit 23 and digit recognition category memory unit 22 with l=L and m=M, and FB (L, M) and W (L, M) are read out. Stored in register F and register R.

The contents of register R are output as the recognition result.

Furthermore, the determination control unit 246 issues an address signal m2 to the digit path memory unit 23 and the digit recognition ability category memory unit 22 with l=l− and m=F, and FB(l, m) and W(l, m) are read out. and stored in register F and register. By repeating this process sequentially from L to 1, the recognition result of L digits 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, in the explanation of the VLB method, which is the principle of the present invention, the order of the calculation loops is n・l, v, m from the innermost, but l,
It is also possible to use n, v, and m for the same reason as the reason for deriving the VLB method.

In addition, digit difference degree DB (l, m), digit route information FB (l,
m), in the explanation of the part in which the input pattern is determined from the digit recognition category W (l, m), the minimum value of DB (l, M) is determined and the number of digits of the input pattern is determined.
TRANSACTIONS ON ACUSTICS,
SPEECH, AND SIGNAL PROCESSI
NG.

It is also possible to determine the number of digits of an input pattern under constraints such as those described in VOL ASSP-27, DECEMBER 1979, pages 588 to 595.

Furthermore, although the distance between the input pattern am and the standard pattern bnv was explained using a distance measure such as equation (8), instead of this, the Euclidean distance as shown in equation (25),
You may use an inner product, etc. as shown in Eq.

Also, the recurrence formula for calculating the degree of dissimilarity is (22), (2
In addition to the forms of equations 3) and (24), various forms have been prepared, and this (2)
2), (23), and (24) instead of Special Publication 56-2
It is clear that the forms described in No. 8278 can also be used.

[Brief explanation of drawings]

Figure 1 is a diagram showing the range for dissimilarity calculation and examples of matching paths, Figure 2 is a diagram showing matching paths allowed in recurrence formulas, and Figure 3 is a diagram showing HL
FIG. 4 is a diagram showing the calculation order of the first digit in the HLB method, and FIG. 5 is a diagram showing the calculation order of the determination process. , Figure 6 (1)
and (2) are flowcharts showing the calculation procedure of the HLB method, and FIGS. 7(1) and (2) are flowcharts showing the calculation procedure of the VLB method which is the principle of the present invention,
FIG. 8 is a diagram showing the calculation order of the VLB method, FIG. 9 is a block diagram of an embodiment of the present invention, and FIG. 10 is a time chart for explaining the operation of the embodiment of the present invention. Fig. 11 is a block diagram of the distance calculation section which is one of the components of the present invention, Fig. 12 is a block diagram of the recurrence formula calculation section, and Fig. 13 is a block diagram of the digit dissimilarity calculation section. This is a configuration diagram of the first part.
Fig. 4 is a block diagram of the determination section, Fig. 15 is a block diagram of the dissimilarity degree memory section and route information memory section, and Fig. 16 is a block diagram of the digit dissimilarity degree memory section, digit route information memory section, and digit introduction section. FIG. 2 is a configuration diagram of an organ category memory section. In FIG. 9, FIG. 11, FIG. 12, FIG. 13, and FIG. 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, 16...
... Distance memory section, 17 ... Recurrence formula calculation section, 18 ... Dissimilarity memory section, 19 ...
route information memory section, 20... digit difference calculation section,
21... Digit difference memory section, 22...
Digit recognition category memory section, 23... Digit route information memory section, 24... Judgment section, 151...
・Absolute value circuit, 152... Adder, 153...
. . . Accumulator, 171 . . . Comparison circuit, 172 . . . Adder, D1, D2. D3...
...Registers that hold the degree of difference, F1, F2, F3・
...Register that holds the route, 201...
Comparison circuit, 202...Register for holding word similarity, 203...Register for holding category, 204...Register for holding route information, 241... ... Comparison circuit, 242 ... Register that holds the minimum digit difference, 243 ... Register that holds the number of digits, 244 ... Holds digit route information 245...Register for holding and outputting recognition results, 246...Determination control unit. Fig. 1 Fig. 2 Fig. 3 Input pattern Fig. 4 556- Fig. 5 Input pattern Fig. 10 y+4'IJ (21%) -- L −−−−−−−−−−−−−−−−−−−−" 1st
5 Figure 1 □V-V Figure 16 0-1---◆7FLM Procedural amendment (spontaneous) 58.2゜2 Date Showa Date Mr. Commissioner of the Japan Patent Office 1, Indication of the case Patent application filed in Showa s6 No. 1971141 2
, Name of the invention Continuous speech recognition device 3, Person making the amendment Relationship to the case Applicant: 5-33-1-4, Shiba 5-chome, Minato-ku, Tokyo, Agent: Sumitomo Mita, 5-37-8 Shiba, Minato-ku, Tokyo 108 Building NEC Co., Ltd. (6591) Patent Attorney Susumu Uchihara Telephone Tokyo (03) 456-3111 (Main Representative) 5, "Claims" and "Detailed Description of the Invention" columns of the specification to be amended and "drawing". 6. Contents of the amendment: "Claims" column As shown in the attached sheet @ "Detailed Description of the Invention" column (2) Page 15, line 5, "To make it shorter. To make it shorter. ,” is changed to “In order to shorten it,
” he corrected. (3) In the 11th line of page 17, rn(4Y,n)"J is corrected to rD(1,Y,n)=a■". (4) In the 14th line of page 22, "rW(4m)" is corrected to "rW(4m);". (5) In the 7th line of page 25, the phrase "first with signal nl" is corrected to "first with signal Ct153." (6) On the 9th line of page 25, the phrase "r from 13" is corrected to "from 13, from 1 according to signal rl." (7) In the 1st θ line of page 25, the phrase "calculate the absolute value" is corrected to "calculate the absolute value of the difference." (8) In the fifth line of page 26, the phrase "by signal n2" is corrected to "by signal n2.n21.n22." (9) On page 26, line 17 of li, correct rF(4v, m)J to "F(4ma, n)". (10) In the 20th line of page 26, correct "memory section 18J" to "D(t,ma,n) of memory section 18". (11) Correct "203 and 205" in the 19th line of page 27 to "202 and 204." 021) On page 26, line 20, replace [204J with “
Corrected to 203J. (13) In the 11th line of page 28, it says, ``A stone that corrects the signal t2J that is emitted following the signal t1 as the signal Cl2J. (1) Correct "register D" on page 29, line 15 to "register 242.""RegisterL" on page 29, line 16 of α is replaced with "
The text "4" on the 18th line of page 29 is corrected to [243J]. Bit 4゜■ “m=” is written on page 29, line 20 of α
2”. (19) In the second line of page 30, "register F and register 几" is corrected to "register 244 and register 245." (1) On the third line of page 30, replace “Register B” with “
register 245”. (21) rt-1+, tn= on the 4th line of page 30 of 111E
The one that says FJ is [L=1-1. m=(contents of register 244)". (2@Page 30, line 7, "register F and register" is corrected to "register 244 and register 245". (qr drawings) - Figures 7-1 and 9 of the drawings attached to this application , Fig. 10, 1st
Replace Figures 1, 12, 13, and 14 with the attached drawings. Attachment 2, time series of claim feature vectors consisting of one or more words = -□, J8. ...+"ms..."+7M
7 standard word patterns BY that are pre-memorized as
=1)? ,br,...,Tol,...,ToXv(
Ma = 1, 2. ..., ■) Maximum LH obtained by all combinations
Continuous voice standard pattern of ax digits: /C-Bvl, Bvl, -
, BVl, ., BYLaax, the vector between the input pattern J11m on the time function n- that makes the time axis m of the input pattern correspond to the time axis n of the continuous voice standard pattern and the continuous voice standard pattern To. In order to find the minimum value of the degree of dissimilarity defined as the sum of distances' (am, bn), continuous speech standard pattern C==Bvl. B1. ..., BA, ...I31xa*x is divided into each digit. Digit dissimilarity DB (4
to), the digit path information FB (4m) indicating the first time point of this time function, and the digit recognition category W (j, m) which is the word name V that gave the minimum cumulative distance on this time function. In a continuous speech recognition device that sequentially calculates the digit t and the time point m of the 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 calculated from 1 to M. For each 11, the signal ma indicating a word is changed from 1 to V, and the signal t indicating a digit for each ma is changed from 1 to L1! a control unit that varies signals up to Inx and from itl to NY indicating time points of the standard pattern; and a dissimilarity memory unit D (4v, n) whose address is designated by the signals t, ma, and n of the control unit; Path information memory unit F(zt't''); inter-vector distance between the word standard pattern N, H=l-Jiv of the word matrix sequentially specified by the control unit at each time point m and the input pattern -1'(a,,bnv); a distance calculation unit that calculates n=1%NY; a distance memory unit d(n) that stores this distance;
At each time point m. For each digit t and each word mark, first set the initial condition to digit dissimilarity DB (t-1゜m-1) which is the result of time point m-1.
is given from the digit route information FB (L-1, va-1)K. The distance dIn) and the degree of dissimilarity D(4) at the time point m-1
゜v, n) and route information F(4v, n) to calculate the recurrence formula of the dynamic program and calculate the dissimilarity degree D(4v, n) at time point m.
. m, each word dissimilarity degree D(4
v, NY), and calculate this as the digit dissimilarity DB (
4m), and the corresponding word path information F(4v, N
Y) is the digit route information FB (4m), and the word name V for which the minimum value is obtained is the digit recognition category! A digit dissimilarity calculating section for calculating JW (4m); and a digit dissimilarity memory section D for storing these.
B (4m), digit route information memory section FB (4m), digit recognition category memory section W (4m), and j digit route information FB (4m).
m) and a determination unit that determines and outputs the category of each digit of an input pattern in reverse order based on the digit recognition category W (4m). 1st Z nz, nzt, ηZ2

Claims (1)

[Claims] Input pattern A of one or more words that is a time series of feature vectors A = a1, a2, ..., am, ..., aM
V word standard patterns Bv stored in advance as
:=bv,'*b2v...,bH,...,bto(v
= 1, 2, ..., V) maximum Lma obtained by combining
Continuous voice standard pattern of x digits C=Bvl, Hvl,...
, Bvl, as+, BvLmax, the vector distance between the input pattern am and the continuous speech standard pattern bm on the time function n- that makes the time axis m of the input pattern correspond to the time axis n of the continuous speech standard pattern. In order to find the minimum value of the degree of dissimilarity defined as the sum of d(am, bn), continuous speech standard pattern C=BV1°By2..., Bvl,...
, BVLmax is divided into each digit, and the digit dissimilarity DB(l, m) indicating the minimum cumulative amount of inter-vector distance given by the optimal time function n(m) at the l-th digit and this time function are calculated. Digit path information FB(l, m
) and the digit recognition category W(l, m) which is the word name V that gave the minimum cumulative distance on this time function.
In a continuous speech recognition device that sequentially calculates the time point m of the 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 changed from 1 to M. For each m, the signal V indicating the word is varied from 1 to V, and for each V, the signal l indicating the digit is varied from 1 to Lmax, and the signal n indicating the time point of the standard pattern is varied from 1 to Nv. a dissimilarity memory section D (l, v, n) and a route information memory section F (l, v, n) whose addresses are specified by the signals l, v, n of the control section; a word standard pattern b of the word V sequentially designated by the control unit at each time point m;
a distance calculation section that calculates the intervector distance d(am, bl, v) n=1 to NV between v, n=1 to Nv and the input pattern am; a distance memory section d(n) that stores this distance; At each time point m, for each digit l and each word V, first set the initial condition to the digit dissimilarity DB (
l-1゜m-1) and digit route information FB (l-1, m-1)
given the distance d(n), the degree of dissimilarity D(l, v, n) at the time point m-1, and the route information F(l, v, n). The equation is calculated and the degree of dissimilarity D (l, v, n) and route information F (l, v, n) at time point m are sequentially obtained, and the degree of dissimilarity D (l, v, NY) and the word route are calculated. a recurrence formula calculation unit that calculates information F (l, v, Nv); at each time point m, round can word dissimilarity degree D (l, v, Nv) calculated by the recurrence formula calculation unit for each digit l; ), set this as the digit dissimilarity DB (J, m), and set the corresponding word path information F (l, v, Nv) as the digit path information FB.
(J, m), and the word name V for which the minimum value was obtained is the digit recognition category W(l, m); a digit dissimilarity calculation unit; a digit dissimilarity memory unit DB for storing these; l,m), digit route information memory section FB(l,m), digit recognition category memory section W(l,m), digit route information FB(l,m) and digit recognition category W(l,m ), and a determination unit that determines and outputs the category of each digit of an input pattern in reverse order based on the following: