JPS5938600B2 - Renzokuonseininshikisouchi - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an improvement in an apparatus for automatically recognizing speech in which one or more words are uttered consecutively, so-called continuous speech.

Devices that automatically recognize speech, that is, speech recognition devices, have many advantages as a means of inputting data and programs to a computer, and as a means of inputting control data to various machines.

However, the speech recognition devices that have been developed to date cannot recognize continuous speech, and have the disadvantage that they do not work unless each word is uttered separately. As a result, when entering multi-digit numerical data, there is the inconvenience of having to separate each digit individually.
A device that can recognize continuous speech is highly desired. Various methods have been tried in the past as voice recognition techniques.

Among them, the pattern matching method is the simplest and most effective method. In this method, a standard pattern (hereinafter referred to as standard pattern) is prepared for each word of the word to be recognized, and the input unknown pattern (hereinafter referred to as input pattern) and each standard pattern are This method calculates the degree of matching (hereinafter referred to as similarity) by performing a comparison operation (that is, pattern matching), and determines that the most similar pattern, that is, the one with the maximum similarity, belongs to the same word as the standard pattern. be. It is difficult to directly apply this pattern matching method to continuous speech recognition. The reason is that it is difficult to classify continuous speech into words. In other words, since standard patterns are in word units, pattern matching must also be performed in word units.
For this reason, it is necessary to divide the input speech into words. However, it is extremely difficult to accurately determine boundaries between words in continuous speech by examining changes in voice level and pitch cycle. Patent Application No. 29891 filed in 1975 describes a method of extending the pattern matching method to continuous speech recognition.

The principle behind this citation is roughly as follows. A pattern obtained by connecting several standard patterns in any order is considered as a standard pattern of continuous speech, and matching with the input pattern is performed. Recognition is performed by finding the number of standard patterns and the permutation of the standard patterns that maximizes the overall similarity. This method does not require the input pattern to be divided into words in advance. In addition, by breaking down pattern matching into two steps: pattern matching for each word and pattern matching as a whole, and executing each step using dynamic programming, the amount of calculation can be reduced to a practical range. It's settled down. As described above, the method cited above is considered to be the most effective continuous speech recognition method to date, but on the other hand, it has the following drawbacks.

In other words, when connecting standard patterns and considering them as standard patterns for continuous audio, and performing pattern matching with input patterns,
We are trying to optimize not only the permutation of standard patterns but also the number of standard patterns. This causes the inconvenience as shown in FIG. In other words, 1 in Figure 1 has ``31'' (/Sa
This diagram schematically shows the voice pattern of uttering nitfi/), and the time axis is \ in the horizontal direction.

As shown in the diagram, the /n/ part of /San/ is long, and /I
If the i part at the beginning of the word tfi/ is long, it may be divided into time points shown as 3 and 4 in the figure, resulting in "321" (/Sanniitfi/) as shown in figure 2. be. In other words, the two-digit number "3" may be mistakenly recognized as the three-digit number "32F". Conversely, the three-digit number "321" may be mistakenly recognized as the two-digit number "31". A variety of similar examples can be considered for both numbers and non-numbers, but a major drawback of the method cited above is that the number of digits may be mistaken in this way. The present invention provides a continuous speech recognition device that operates accurately by improving the disadvantage of the cited method of incorrectly determining the number of words (hereinafter referred to as the number of digits in the case of numbers) included in an input pattern. It is intended to.

The continuous speech recognition device according to the present invention is expressed as a series of feature vectors A-a1, A2, etc.
Ai...Input pattern buffer holding a1 and standard pattern Bn-b of word n? , B.W.・・・・・・
bri...Standard pattern storage JJ section for storing Bnn, number of digits specification section for specifying the number K of words included in the input pattern, A from al+1 of the input pattern
Pattern matching is performed between A(1.m), which is a partial pattern up to m], and each standard pattern Bn, and the similarity S(A
(1.m), Bn), and partial similarity S<1. which is the maximum value of the similarity S(A(1, m), Bn) with respect to n. m〉 and a maximum value detection unit for calculating the partial judgment result n×1, m〉, which is n giving the maximum value, for the required partial pattern A(1, m), and furthermore, Dynamically performs a process of dividing into a series of partial patterns, the number specified by the number specification part, that is, K, and determining the division so that the sum of the partial similarities S1, m〉 corresponding to each partial pattern is maximized. A second-stage matching processing unit executed based on the planning method, and K partial judgment results n corresponding to K partial patterns determined thereby.
1, m> as a recognition result.

The characteristic principle of the present invention is as follows.

In other words, it is equivalent to have a means to specify the number of digits of the input pattern, and to determine the pattern that best matches the input pattern by connecting the number of standard patterns specified as the number of digits in any permutation. It is important to have the means to carry out the necessary actions. The configuration of the present invention will be described in detail below, but to simplify terminology, the objects of recognition will be limited to numbers. The input speech pattern is expressed as a time series of feature vectors Ai.

Here, I represents the number of feature vectors, and since this corresponds to the temporal length of the human pattern, it is called the input pattern length.

Various quantities can be considered as the feature vector Ai, but one example is the output level a of the P channel analysis filter.
One possible method is to use 1q.

The numbers (0, 1, 2J...tan) are expressed by the time series of the feature vector Br! Here, b is expressed as in the same way as in equation (2).

Jn is called the standard pattern length like I above. Further, in the following, since the same process is often repeated for each standard pattern Bn, the subscript n will be omitted as necessary and the standard pattern B, feature vector b, standard pattern length J, etc. will be referred to.

It is assumed that the number of numbers included in input pattern A, that is, the number of digits, is K.

K standard patterns Bn(1), Bn(2),...-
-Bn(k), . . . It is shown as a pattern in which Bn(k) are connected and is called a K-continuous standard pattern. Here? The operator indicates a standard pattern of connection.

In general, a pattern Bn that connects Bn and Bm? Bm is defined as -.

Let us denote the degree of similarity between input pattern A and standard pattern B as follows.

In the present invention, K numbers n(1), n(2), which maximize the similarity between input pattern A and K-continuous standard pattern B,
・・・・・・ By calculating n(k), the input pattern is the number n(1), n(2), ・・・・・・, n(k)
It is determined that

Each of n(1), n(2),...n(k) is a number from 0 to 9, so the total is 10k.
Types of k-continuous patterns r are considered. For example, in the case of a 4-digit number, 104 = 10,000th type, and it is difficult to calculate the similarity S (A.B) for all of them in terms of the amount of calculation. Therefore, by improving and applying the method cited earlier, we can find the optimal n(1), n(2) with less calculation amount.
, . . . n(k) is determined. Partial pattern A (1.m) of input pattern A is defined as follows.

This l is called the starting point of partial pattern A (1.m), and m is called the ending point of partial pattern A (1.m).

Using such partial patterns, input pattern A is decomposed into K partial patterns that are interconnected and do not overlap. Here, l(0) = 0, 1(k) = I (input pattern length)
It is assumed that

l(x) is called a breakpoint. Thus, input pattern A
The similarity between the K-continuous standard pattern B and the K-continuous standard pattern B is further transformed as follows, assuming that the similarity is made additive.

Therefore, optimally determining n(1), n(2), . . . n(k) means calculating the following maximum problem.
This problem consists of maximization with respect to n(1), n(2),...-n(D and l(1), l(2),..., l
(k-1) is decomposed into a maximal form.

(Maximization of equation 10 is performed in two stages: first-stage matching and second-stage matching.

First perform inner maximization. Therefore, for a required combination of l and m (1.m) where l<m, the degree of similarity with the standard pattern Bn is calculated for all numbers n.

Next, find . Here, the operation Argmax means calculating n that maximizes the value in [ ]. Round type S<1. Since m> is the optimal similarity of partial pattern A (1.m), it is called partial similarity. Further, since n<l, m> is the result of recognizing the partial pattern A (1.m), it is referred to as a partial determination result. The above procedure is called first-stage matching. By this procedure, all required 1 such that l<m
、． Combination of m<1. m>, the partial similarity S<1.
m> and the partial judgment result n<1, m>・In addition, (
Various methods can be considered for defining the degree of similarity appearing in the equation 6) and the O (branch) equation.

What you need is partial pattern A (1.m) and standard pattern B.
It is necessary to have the ability to correct the deviation of the time axis between n.
It is described in the round formula of 91). (s(i, j(i))
indicates the degree of similarity between Ai and Bj...) is considered as the most appropriate example.

In the second stage matching process, the outer maximization of the UO formula is executed.

For this purpose, calculate the optimal l(1), l(2),
...It is necessary to find l(k-1).

That is, the input pattern is divided into K partial patterns, and the division is optimized so that the sum of the partial similarities of the respective partial patterns is maximized. In the case of the cited example, the number of digits K was free, but in the present invention, l(1), l(2),...l( k
-1) is characterized by optimizing. Therefore, it has K or more registers each having the number of words equal to or more than I words (input pattern length). One word storage element of this register is denoted by Tx (RT). X indicates the register number, and m indicates the address within the register. Similarly another register I
X(IT) will be established. The next dynamic programming calculation is executed using these two sets of registers. This means calculating 1 for the initial condition recurrence formula.

In the following, Tx(... is called a recurrence formula value, and 1x(... is called a temporary division point.

Q9), calculate the formula (to) x=K. After completing the calculation up to m=I, the determination result is determined by the following procedure.

Initially x=K. Assuming m=I, the following procedure is repeated until k=1. As a result, the result of determining input pattern A is n
(1), n(2),..., n(x),...
. . , n(k) are determined in reverse order.

In the (company) procedure, l=l(x) is given by (a),
n(x) is found by (b), and both are determined in parallel.

Alternatively, 1(x) may be determined from x=1 to x=K first, and then n<l(x-1), 1(x)> may be read out to determine the determination result. In the description of the configuration of the invention described above, the process up to the determination of 1(x) was described as second-stage matching processing. In the embodiment described later, the determination of l(x) is performed by the determination unit. Considering that l(x) is uniquely determined and stored in Lx (RI), these are simply problems of functional division, and are not particularly contradictory.According to the above principle, Number n (1) to be determined
,...The number of n(k), that is, the number of digits, is necessarily K
Therefore, the drawbacks of the cited method, namely 2 as shown in Figure 1.
The possibility of a digit number being mistakenly recognized as a 3-digit number is eliminated.

FIG. 2 is a block diagram illustrating one embodiment of the principles of the present invention.

20 in the figure is a control section; l, M. Control signals such as tl are issued to control the operations of other parts. Reference numeral 21 denotes an input pattern buffer, which temporarily holds input patterns in the form of equations (1) and (2).

22 is a standard pattern storage unit which stores standard patterns in the formats (3) and (4).

23 is a first stage matching unit which calculates the similarity S(
Calculate A(1.m), Bn).

24 is a maximum value detection circuit, and the first stage matching section 2
Similarity S(A
(1.m), Bn), calculate the partial similarity Sku1,m〉 and the partial similarity nku1,m〉, and use them as signals S1 and n1, respectively. Output.

25 is a partial similarity table in which the required partial similarity S<1. m> is temporarily held.

Reference numeral 26 is a partial determination result table that temporarily holds necessary partial determination results n<l, m>.

27 is the second stage matching section, (self), (self),
Execute the dynamic programming operation (20).

Reference numeral 28 is a recurrence formula value register (support) that stores the necessary number of recurrence formula values TX (m) necessary for calculating the σe formula.

29 is a temporary dividing point register. (Temporary dividing point 1x(Rr) obtained by executing Equation 2I is held temporarily.

The second-stage matching section 27, the recurrence formula value register 28, and the temporary segmentation point register 29 are involved in the second-stage matching process described in the explanation of the configuration and principle described above. Reference numeral 30, collectively referred to as a second-stage matching processing section, is a digit number specifying section, which generates a signal k specifying the number of digits K. 31 is a determination unit which executes the procedures (a), (BXc), and (d) of the Co., Ltd. formula and produces determination results n(1), n(2), . . .
...outputs n(k).

Next, the configuration and organic operation of each of the above sections will be explained in detail. 1st
The configuration of the step matching section is based on the similarity S(A(1, m), Bn
) depends on the definition and scale adopted. In short, the partial pattern A (1.m
) and the standard pattern Bn such as equation (3) can be calculated. As an example (self)
When based on the definition of the formula, for example, Patent Application No. 627 of 1972
The structure can be similar to the device shown in FIG. 5 of No. 82. The control unit 20 receives the signal N. The standard pattern Bn is specified by the signal 1. and M. Specify starting points 1 and m of partial section A (1.m) by . At this time, the first stage matching section 23
Then, pattern matching between A(1, m) and Bn is performed, and the similarity S(A(1, m), Bn) is the signal S. is output as The maximum value detection section is configured as shown in FIG. 3, for example. In the figure, 241 and 242 are S-
S at the time when the register and the n-register are first set to O by the clear signal Cl and then the write pulse WP1 is applied. and N. remember. 243 is a comparator and a signal S.

and S1 by numerically comparing the magnitudes of S1 and S1. >S1, the pulse WP1 is emitted. Therefore, the control section 2
The signal N given by 0. The similarity S(A(1, m), Bn) changes according to the change of n specified by the signal S. , the S-register 241 stores the maximum value of similarity up to that point, and the n-register 242 stores n, which gives the maximum value of similarity up to that point. Yotsute N. When is changed in the range of 0 to 9, (1
4), (to) the equation is executed, the partial similarity S1, m> becomes the signal S1, and the partial judgment result n<1sm> becomes the signal n1.
It will be output as These are signals 1 from the control unit of the partial similarity table 25 and the partial determination result table 26, respectively. and M. is stored in the location specified by . Next 1. Or M. By changing the partial similarity S<1. m> and partial judgment result n<1. m> is calculated and entered in all partial similarity tables 25 and partial determination result tables 26. The required combination of 1 and m1. m>, the partial similarity S<1. m〉 and partial judgment result n〈1, m
> are all calculated and stored in the partial similarity table 25 and partial judgment result table 26. When the process is finished, the control unit 2
A timing signal t1 is issued from 0 to instruct the second stage matching section to start operation.

Also, the timing signal t1 is the content TX of the recurrence formula value register.
(It also has the function of setting m to all O.The configuration and operation of the subsequent second-stage matching unit 27, recurrence formula value register 28, and temporary segmentation point register 29 are explained below.) This will be explained using FIG. 4 which shows one configuration example.The broken line blocks 27, 28, 29 in FIG.
They correspond to the stage matching unit 27, recurrence formula value register 28, and temporary segmentation point register 29, respectively.

270 is a second stage matching control section, 271 is an addition circuit,
272 is a maximum value detection unit, 281 is a recurrence formula value register reading circuit, 282 is a recurrence formula value register writing circuit, 283
291 is a temporary division point reading circuit, 292 is a temporary division point register writing circuit, and 293 is a storage circuit.

When the second stage matching section control section receives the operation start command signal t1 from the control section 20 shown in FIG. 2, it sets the signal X2 indicating x of the σ9 equation to 1, and thereafter increases it sequentially up to the number of digits K. Note that the number of digits K is the signal k from the number of digits designation section 30 in FIG.
specified by. For each value of x, the value of m specified by signal M2 is increased from 1 to input pattern length 1. The recurrence formula σ9) and formula (to) are calculated once for each value of m. For this purpose, the value of l specified by the signal 12 is sequentially changed from O to (m1), and from the partial similarity table 25 of FIG. 2, S<1. m> as the signal S2, and the recurrence formula value Tx-1(1) is read from the recurrence formula value register 28. These sums are calculated by the addition circuit 271. As a result, a signal S25 is obtained, which is a value in brackets [ ] in the equation σ9). This signal S2' is sent to the maximum value detection section 272 together with the signal 12 indicating l. The maximum value detection section 272 is configured similarly to the maximum value detection section 24 shown in FIG.
The maximum value of S2' given sequentially from 71 and 12 corresponding thereto are determined and outputted as signals Tx and 12', respectively. The signal Tx is written as the recurrence formula value Tx (Rr) into the storage circuit 283 via the recurrence formula value register write circuit 282. The signal is written as a temporary dividing point 1x(m) into the storage circuit 293 via the temporary dividing point register writing circuit 292. This means that the calculations of the σ9 expression and (to) expression have been executed for the specified x and m. Note that the storage circuit 283 has sufficient capacity to specify up to K digits for x, and up to 1 input pattern length for m, and the recurrence formula value register readout circuit 281 specifies the signal X.
It has a function of calculating an address based on 2 and 12 and reading out Tx-1 (1) from the storage circuit 283 as a signal TX-1. The recurrence formula value register writing circuit 282 has a function of calculating an address based on the signals X2 and M2 and writing the signal given as Tx into the storage circuit 283 as Tx... Further, the configuration and functions of the temporary segmentation point register read circuit 291, the temporary segmentation point register write circuit 292, and the storage circuit 293 are the same as those of the recurrence formula value register readout circuit 281, the recurrence formula value register write circuit 282, and the storage circuit 283, respectively. Suppose that However, it is assumed that the temporary division point register reading circuit 291 outputs 1X(m) when x is specified by the signal X3 and m is specified by the signal M3.
Let us return to Figure 2 and continue the explanation.

When m (specified by signal M2) is changed to the input pattern length for some x (specified by signal X2), X is incremented by 1 and m is reset to 1. Thus, x is changed to the number K of digits specified by the signal k.
By changing m to the input pattern length 1, all the l(to) expressions are completed. When the above processing of the second stage matching section 27 is completed, the timing signal T2 is issued from the second stage matching section 27, and the determining section 31 is instructed to start operation.

The determination unit 31 calculates the formulas (a), (BXc), and (d), and obtains K determination results n(1), n(2), . . .
...n(k), and its operation will be explained with reference to FIG. 5, which shows an example of the configuration of the determination section. In FIG. 5a, 310 is a determination control unit, 311 is a digit counter, 312 is a first register, 313 is a second register,
314 is an output register, and 315 is a zero determination circuit. FIG. 5b shows a time chart of the signals of each part mentioned above. That is, at the same time when the timing signal T2 is initially applied, the number of digits K sent as the signal k from the number of digits specifying section 30 is set in the number of digits counter 311, and the input pattern length 1 is set in the second register 313. . Thereafter, the determination control unit 310 uses timing signals Tll, tl2, tl.
Generates 3. First, m is specified by the content 3 of the second register 313, x is specified by the content X3 of the number of digits counter, and the temporary partition point 1x (
11) is read out as signal 13 and timing signal T,2
It is written to the first register 312 in synchronization with . This means that (a) of the company's formula has been executed. Next, m is specified by the content 4 (same as M3) of the second register 313, 1 is specified by the content 14 of the first register 312, and the partial determination result table 26 (FIG. 2) Judgment result n<1. m> is read out as a signal N2 and output to the output register 314 in synchronization with the timing signal Tl3.
written to. As a result, formula (b) of the company is executed to obtain the determination result n(x), which is output as the determination result N3. Next, in synchronization with the timing signal Tl, the first
The contents of register 312 are moved to a second register 313;
(c) of the (company) formula is executed, and at the same time the number of digits counter 31
The content X3 of 1 is counted down by 1, and (d) of the (company) formula is executed. The above operation is performed by the digit counter 311
is repeated until the content X3 becomes O, and the judgment result is n(k)
, n(k-1), . . . n(0) are output as signal N3. Finally, when the content X3 of the number of digits counter becomes O, this is detected by the zero judgment circuit 315 and an end signal e is issued, whereby the operation of the judgment control section is stopped and the control section 20 ( 2) is notified that all determination operations have been completed. After that, the next input voice can be recognized. Note that the signals 14, M4, and M3 are shown as an example in the case where the number of digits K in the time chart of FIG. 5(b) is 5 and the input pattern length 1 is equal to 60.
Various configurations of the number of digits designation unit 30 are possible.

The simplest example is equipped with a rotary switch or a selective push button switch, which allows the number of digits K to be specified manually, and outputs a signal with a code corresponding to the number of digits k as the signal k. The configuration is such that Another possible configuration is to predetermine the format of the number of digits of data to be input by voice, and generate the signal k in accordance with the format. For example, first set the format as 2 digits, then 4 digits, then 1 digit,
Numerical codes corresponding to 2, 4, and 1 are recursively generated as the signal k. Furthermore, a configuration is also possible in which the number of digits of data to be input next can be controlled based on the recognized result. The point is that it is a means for specifying the number of words included in input speech as the number-of-digits specification signal k, and its specific configuration is not directly related to the content of the present invention. The gist of the present invention is to have means for specifying the number of words (number of digits) to be included in an input pattern, and means for performing pattern matching between a partial pattern of the input pattern and a standard pattern to calculate the degree of similarity. It has a means for calculating partial similarities and partial judgment results for all partial patterns, and it has a means for calculating partial similarities and partial judgment results for all partial patterns, and a specified number of sequences of partial patterns for each part so that the sum of partial similarities for 5 patterns is maximized. and outputs the partial determination result for each partial pattern as a recognition result.

The descriptions from FIG. 2 to FIG. 5 show examples of the above principle, and do not limit the scope of the present invention. In particular, various embodiments are possible regarding the method of defining the degree of similarity S (A(1.m), Bn) and the configuration of the first-stage matching section and the second-stage matching section. Also,
The recurrence formula value table 28 (Fig. 2) contains the recurrence formula value Tx(n
It is not necessary to memorize all of 1) from x-1 to x-1K (number of digits). For calculation of the σ9 formula, it is sufficient to store only the value of x one step before, that is, Tx-, (1) (1=1 to I). Furthermore, in the above explanation, the second stage matching is executed after all the first stage matching processes are completed, but the partial similarity S
1, m> and the partial judgment result n<1. When m> is all set for a specific m, the formula (to) is executable, so it is also possible to execute the first-stage matching process and the second-stage matching process in parallel in a time-sharing manner. be. Furthermore, in the above description, the degree of similarity is defined as a quantity that becomes larger as the two patterns become more similar; however, it is also possible to use a quantity that becomes smaller as the patterns become more similar, such as distance.
In this case, all maximization operations such as (self) expression and (self) expression should be replaced with minimization.

[Brief explanation of the drawing]

Figure 1 is a diagram showing the necessity of the present invention.
/Sanitfi/) is continuously uttered 63 times.
21゛(/Sanniitfi/) may be mistakenly recognized by one digit more. FIG. 2 is a block diagram showing an embodiment of the present invention, in which 20 is a control section, 21 is an input pattern buffer, 22 is a standard pattern storage section, 23 is a first stage matching section, 24 is a maximum value detection section, and 25 is a 26 is a partial similarity table; 27 is a second-stage matching section; 28 is a recurrence formula register; 29 is a provisional breakpoint table; 30 is a digit number specification section; 3
1 is a determination section. FIG. 3 is a diagram showing an example of the configuration of the maximum value detection section 24 in FIG. 2, in which 241 is an S-register, 242 is an N-register, and 243 is a comparator.

Claims (1)

[Claims] 1 Input pattern A expressed as a series of feature vectors
=a_1, a_2...a_i...a_
An input pattern buffer holding I and a standard pattern B^n of word n = b^n_1, b^n_2, ......b^
A standard pattern storage unit for storing n_j...b^n_J_n and the number K of words included in the input pattern
The number of digits specification part for specifying and the a_l_ of the input pattern
A(l, m) is a partial pattern from +_1 to a_m
and each standard pattern B^n to calculate the similarity S(A(l, m), B^n); , m),
Partial similarity s<l, which is the maximum value with respect to n of B^n),
m> and the partial judgment result n<l, m which gives the maximum value
> for a required partial pattern A(l, m), and further divides the input pattern into a series of K partial patterns, the number specified by the number of digits designation section. The partial similarity S<l corresponding to each partial pattern
, a second-stage matching unit that executes a process of determining divisions based on dynamic programming so that the sum of m> is maximized;
A continuous speech recognition device comprising: a determination unit that outputs K partial determination results n<l, m> corresponding to the K partial patterns determined thereby as recognition results.