JPS6313099A - Continuous voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a continuous speech recognition device, and more particularly to an improvement in a continuous speech recognition device that recognizes sentence speech '(r) continuously uttered according to grammar.

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

A method of recognizing sentence sounds continuously uttered according to such a grammar is described in Japanese Patent Application No. 199098/1983. This method is called the CWDP method, and the principle is as follows. The grammar is expressed by an automaton α, and the automaton α is defined as follows.

α=(K, Σ, Δ*po+F) ・・・・・・−
...(1) Here, K: set of states p (plp=1
*2y..., π)Σ: Set of input words n (nln=1
．． 2. -, N) Δ: State transition rule ((p, q, n) Here, (psqtn) means a state transition of pxLq.

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

F: Rt final state set FCK Next, according to the automaton α, the sound bang obtained by continuously uttering the word n (Σ) AiA= a, , a2
．． ---, ai, ・-, a, ・-・----
--(2), and this is called the (unknown) input button.

For each word n(Σ, the standard bang B=N, to, ..., bj, ..., bjn
・・・・・・・・・(3) is prepared and this is called a word standard slam. DP matching is performed between the continuous speech standard pattern ♂1B,...,B obtained by connecting these standard word batons according to the automaton α and the input baton A, and the degree of mutual difference between the two batons is expressed by (referred to as the degree of dissimilarity) is calculated, and the word sequence that provides the minimum degree of dissimilarity is taken as the recognition result.

Here, the minimum degree of dissimilarity is found using the following dynamic programming method. The initial conditions are T(0,0)=O T(i, q)=oo, i〆O+ Q〆0G(q
*nti)==(1) ・・・・・・・・・
...(4), and the recurrence formula of formula (6) is (pt q * n) based on the boundary condition of formula (5) sequentially until i = 1.
'Calculate for all pairs (p+n) of Δ. That is, for the pair (p, n), the boundary conditions G(ptnto)=T(i-19p) H(p, n, o)=i-1--(5
), where j is the minimum G(p) on the right side of equation (6).

”+J).

from j=1 (from the start to the end of the nth standard pattern), and calculate go)h(j) as G(p
+”pJ)+H(p+npJ).Here, d
(j) If the feature vector ai at input pattern time iK
and the feature vector bj at the nth standard slam time j
It is the distance between

Next, as a minimization at word boundaries, if T(i, q)') CT(p, n, J)
・・・・・・−=・(9) then T(i, q)−G
Calculate (p, n, J) N(i・q)=n P(i, q)=p L(ieq)=H(psn+Jn).

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

△ vinegar. If p=Q, the process ends.

In the method explained above, the word dissimilarity q(j) is a value proportional to the input pattern time length i, as shown in Figure 2 (al) and equation (6). It is an explanatory diagram showing a path. That is, a feature vector distance d(j)el is added for each input pattern time.

In continuous speech recognition, the cases in which the degree of word dissimilarity is proportional to the time length of the input pattern are as follows. , For example, if P path is used as shown in FIG. 2(b), the word dissimilarity g(i, j) becomes W1=W,=1. W,=V2, and this degree of word dissimilarity is proportional to the sum of the input pattern time length and the standard pattern time length. That is, it is proportional to (i+j). In this case, the degree of difference becomes a smaller value as the standard pattern time length becomes shorter. In other words, the shorter the length of the connected standard patterns, the more advantageous it is, and the more likely words will be omitted.

Therefore, in continuous speech recognition, it is necessary to use a recurrence formula that does not depend on the time length of the standard pattern, such as formula (6).

(Problems to be Solved by the Invention) However, with the recurrence formula proportional to the input pattern time length, it is not possible to take a direction perpendicular to the input pattern time axis as shown in FIG. 2(b). Therefore, the degree of expansion and contraction of the input pattern is limited.

That is, the input pattern can only be reduced to 54, which is the standard pattern, using the recurrence formula shown in equation (a).

'Therefore, there is a pattern in which there is a lot of expansion and contraction in one part.

In this case, the time axis correspondence between the input pattern and the standard pattern cannot be accurately achieved, which causes misrecognition. On the other hand, in order to increase the degree of expansion and contraction using a recursion formula proportional to the input pattern time length, it is necessary to
It is necessary to use a path υ, which can be obtained using the recurrence formula shown in equation α4 or equation α9. However, the α chrysanthemum formula αQ formula is fb1
The disadvantage is that the amount of calculation increases compared to the formula.

An object of the present invention is to eliminate the above-mentioned drawbacks and provide a continuous speech recognition device with a high recognition rate and a small amount of calculation.

(Means for solving the problem) The continuous speech recognition device according to the present invention uses the degree of dissimilarity T(i, p) proportional to the time length of the input pattern as a word boundary value in each state P of a finite state automaton. a DP matching unit that calculates the degree of word dissimilarity between the input pattern and the standard pattern in a format that is proportional to the time length of the input pattern;
a dissimilarity degree conversion section that converts the word dissimilarity degree obtained by the P matching section into a dissimilarity degree proportional to the input pattern time length; and a dissimilarity degree conversion section that converts the word dissimilarity degree obtained by the P matching section, and Find the minimum degree of dissimilarity in state q and use this as the word boundary value T
(itq).

(Function) Next, the function of the present invention will be explained. In the present invention, the digestion formula used to calculate the DP matching between the input pattern and the standard pattern is of a type that is not proportional to the input pattern time length.

For example, 03 is substituted for the formula (6) used in the conventional method.
Use the formula. This has the DP path shown in Figure 2 (bl). If we rewrite this α4 recurrence formula into the form of the CWDP method, which is calculated for each frame of the input pattern, we get h(j) = f(( psnsj)or H(psnsj-
1) or h(j-1) (17) However, in response to the minimum selection of the α0 equation, the right side of the 09 equation is selected.

becomes. Calculation of DP matching using the αe and αη formulas is as follows.

Let the initial condition be equation (4), and sequentially from i-1 to 9I (5
) Based on the boundary conditions of equation 00, the recurrence equation is (p, q, n)
(Calculate for all pairs (p, n) where Δ.

For each pair (p e n) and i, from j=1 to Jn or 1
) based on the boundary conditions of the equations αG and αη, and the results g(j) and
'+'tj).

Next, the degree of dissimilarity at the word boundary is converted into a degree of dissimilarity proportional to the input pattern time length according to equation 09.

α=G(psn,”)×i/(i×J')
＋＋＋＋＋＋＋＋α Then as a minimization at the word boundary, l f T (ir q)>”
・・・・・・・・・Value ■then T(i, q)
=G' N(i, q)=n P(i, q)=p L('-q)=H(pyn,'') Calculate.

The recognition result of the input pattern is obtained by the determination process of equation (11, (Ill, 02).

In the above method, the degree of word dissimilarity is not proportional to the input pattern time length, but is converted to a value proportional to the input pattern time length at the word boundary to obtain the overall degree of dissimilarity. As a result, the degree of dissimilarity between the input pattern and the connected standard pattern can be determined as in the conventional method. Furthermore, since a recurrence formula such as Equation 10 can be used for the degree of dissimilarity within a word, the degree of expansion and contraction in the time axis correspondence between the input pattern and the standard pattern can be increased.

(Example) Next, the present invention will be described in detail with reference to the drawings. 1UA is a block diagram showing a practical example of the present invention, and FIG. 3 is a time chart showing the time relationship of operations in the embodiment of FIG. 3 is a flowchart showing the flow of operations in the embodiment.

The standard pattern Bn of the word n included in the word set Σ is stored in the standard button memory 1130, and the state transition rule (p*Q+'') and the designation information of the final state F are stored in the automaton storage unit 230. There is.

When unknown input audio is input from microphone 100,
Frequency analysis is performed by the input unit 110, and the vector at is converted into a characteristic vector at, which is sequentially sent to the large-capacity memory unit 120. Further, the input section 110 is provided with a function of determining a voice section by detecting the voice level, and generates a voice section signal SP which is "1" during the voice section and rOJ at other times. The control unit 240 generates an initialization pulse 5ET at the time of rise of the voice section signal SP.
1 (Figure 3). As a result, the memory 200 corresponding to equation (4) and block 10 in FIG. 4(a) is initialized.

The calculation of the degree of difference is performed along the input pattern time i.

At each time i, the state transition rule r(pt''*q) is read from the automaton storage unit 230.

Subsequently, the signal 5ET2 from the control section 240 causes the fourth (a
) Boundary conditions corresponding to block 11 and equation (5) are set.

Subsequently, the standard pattern time signal j changes from 1 to Jn, and the DP matching section 310 calculates the recurrence formula corresponding to block 12 in FIG. 4(b) and the αe, αη formulas.

When the standard pattern time reaches Jn, the work memory corresponding to prop 13 in FIG. 4(b) is updated. Subsequently, the dissimilarity conversion unit 350 performs dissimilarity conversion corresponding to block 14 in FIG. 4(b) and the α ladder formula. Finally, word boundary minimization corresponding to block 15 and formula a9 in FIG. 4(b) is performed using comparison circuit 170.

The above process is performed for all state transition rules r, and the input pattern time i is determined from 1 to It. After that, the minimum dissimilarity at the terminal point corresponding to the block 16 in FIG. The result determination unit 220 performs a process of determining the recognition result corresponding to the formula.

Although the present invention has been described above based on examples, these descriptions do not limit the scope of the rights of the present invention. Various recurrence formulas other than the 00 formula can be used in the present invention.

For example, Wl=W2=W, =1 may be used.

In this case, the diagonal direction is somewhat advantageous, but the calculation is simpler.

Furthermore, although this embodiment is based on the CWDP method as described in Japanese Patent Application No. 56-199098, various similar automaton-controlled DP matching methods (for example, Japanese Patent Application No. 56-1999)
4-104669. The principles of the present invention can also be applied to Japanese Patent Application No. 61-031179).

(Effects of the Invention) As explained above, according to the present invention, in a continuous speech recognition device, it is possible to increase the degree of expansion and contraction in the time axis correspondence between the input pattern and the standard pattern without complicating the recurrence formula of DP matching. This has the effect of significantly increasing the recognition rate and significantly reducing the amount of calculation.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention, FIG. 2 is an explanatory diagram showing a DP machining bus, and the third factor is a time chart showing the time relationship of operations in the embodiment of FIG.
4(a) to i) are flowcharts showing the flow of operations in the embodiment of FIG. 1. 110... Input section, 120... Input button memory section, 130... Standard button memory section, 170... Comparison circuit, 180 to 210 memory, 22
0...Result determination unit, 230...Automaton storage unit, 240...Control unit, 310...
...DP matching section, 320°330...
・Work memory, 350...Difference conversion unit, 4
00... Termination determination section. αυ (b) (C) (
Figure 4 (a) Figure 4 (b) Section 4 (C) Figure 2

Claims (1)

[Claims] Continuously uttered speech is combined with a standard pattern concatenation pattern specified by a finite state automaton and DP (Dyn
In a continuous speech recognition device that performs matching (dynamic programming (amic programming)) and recognizes a series of standard patterns that yield the minimum degree of dissimilarity, the time length of the input pattern is determined in each state P of a finite state automaton. A DP matching unit that uses a proportional degree of dissimilarity T (i, p) as a word boundary value and calculates the degree of word dissimilarity between the input pattern and the standard pattern in a format that is not proportional to the time length of the input pattern, and the DP matching unit. a dissimilarity converter that converts the word dissimilarity obtained into a dissimilarity proportional to the input pattern time length; 1. A continuous speech recognition device comprising: a minimum dissimilarity calculating section which obtains a dissimilarity and uses this as a word boundary value T(i, q).