JPS63226693A - Pattern matching system for voice recognition - 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] (Industrial Application Field) The present invention is a voice vt@ that automatically recognizes human voice.
This paper relates to a pattern matching method.

(Prior art) Many technologies have been developed regarding pattern matching methods for speech recognition, but one of the most established and heavily used among them is the one described in "Journal of the Acoustical Society of Japan, Vol. 42, No. 9. (Published September 1986), page 725, there is a DP matching method. This is considered to be extremely effective as a method for matching time axis distortion of audio. In addition, as an extension of the DP matching method to one continuous word recognition, see @p. 727 of the above document and Japanese Patent Application No. 56-19.
Clockwise DP methods are known, such as those described in No. 9098. Although this method has been described as a continuous word recognition method with syntactic control, it can of course also be applied to discrete word recognition. Here, to simplify the explanation, the main part of the clockwise DP method will be explained using the formation of discrete word recognition.

Assuming that a word name is specified by reference n, a word set (nln==1.2.---N) is set as a recognition target. Standard pattern Bn=l) for each word? , kg * ”” ”j・' 11
Consider W n. Here, j indicates time, and 1. means the characteristic of the standard pattern Bn at time j. Input voice pattern 'frlilFi stock A=&I JL! -! i・”11.

Speech recognition is performed by determining the inter-pattern distance L) (A, B) i between the input cover pattern A and the standard pattern Bn, and fixing the minimum nt as the recognition result. In DP matching, this distance is calculated using the following recurrence formula.

O initial condition gn (1, 1) = dn (1, 1) ・−・ (1) O
The recurrence formula %J) = II a 1-b 3 II. On the other hand, gn(i, j) calculated by equation (2) is called the optimal cumulative distance.

Initially, this DP matching process was executed for each word, but in the clock quiz DP method, it has been improved to a format in which it is executed for each word in parallel.

FIG. 5(a), Φ) is a diagram explaining the conventional method.

That is, in the space where i, j, and n are related as shown in FIG. n specified.

Calculate the optimal cumulative distance gn(iej) for j,
After that, time i is advanced and processing is continued.

The actual calculation is gn(j) = gn(i, j)
, ■n(j)=gn(i-11J), and to hold these, two columns of memories indicated by reference numeral l in FIG. 5(a) are prepared. Between these memories,
tIXs figure) Naru performance 3! Then, the recursion equation Xt of equation (2) is included.

Here, dn(j)=dn(i,j)−=(5). For all combinations of n + J, execute the calculation of equation (4), advance the end time and time i by one clock, and calculate the plan n(
The above process is repeated by setting j) to g ''(j).In this way, when the process up to i・t is completed, the distance between patterns is set as D(A tBn)=gn(J''), and for each n can be found in parallel. Clock (Problem to be Solved by the Invention) This method allows processing to proceed in synchronization with time i of the input pattern, so processing can proceed in parallel with utterance, and has good real-time performance. It is said that but,
When this method is applied to speech recognition of large words, there is a problem in that the amount of calculation becomes large. That is, the recurrence formula calculation of equation (4) must be performed for all combinations of j and n. The standard pattern is Jn=30,
If you try to recognize 1000 words, 3X10' points (4
) will calculate the formula. Even if it is executed at 10 μs per point, it takes 3 ootns. In normal f-voice recognition, the period for sampling the input pattern Q''i is less than 2Qms, so real-time execution is impossible with such dog language.

The object of the present invention is to improve the above-mentioned disadvantages of clockwise temperature 1) P matching, such as the large amount of calculation, and to provide a pattern matching method for a speech recognition device that is high speed and inexpensive. It is.

(Means for resolving problems) The pattern matching method for speech recognition according to the present invention uses a time series B n = 6 characterized by a standard pattern of valley words n.
means for storing it as n...lIn...111n;
means for temporarily holding the feature i of the input speech pattern, and the distance dn (i, j) (7) maximum JX shape (agn (i ,j)
In a method having means for calculating by a recurrence formula of tkh-like programming, M! up to time (i, 1)! Cumulative value g(i
−1, j) k means for storing in the form △ g n(j) corresponding to each word n and time j, and gn(j) corresponding to the optimal cumulative value gn(i, j) up to time l; At time i, the value of n(j) satisfies a predetermined pruning condition. d”(j)=dn
Perform the first process of fully adding (i, j) to obtain a new numerical value g, and compare the magnitude of with g '' (j) corresponding to j in the vicinity of j, and find that g < g n(j >, a second process is performed to transfer g to a new g as j), and each time i advances by one clock, g n(j) is transferred to the new g.
In this method, the first and second processes are performed as n(j).

(Operation) The first feature of the present invention is that g n(j) is smaller than a certain criterion by using equation (4) (equation constraint (formula 27)).
The point is that the idea of pruning is introduced by executing the dynamic programming recurrence formula. Since the dynamic programming recurrence formula (2) or (4) for DP matching is of the form of searching for the complete small [, n(j) is large, which means that j
This means that there is a low possibility that the path exists on the optimal route.

Therefore, we introduced the idea of ignoring such g n(j). As a result, the calculation of equation (4) can be performed in the vicinity of g '' (j) shown with a hatch in Fig. 1, and a significant reduction in the amount of calculation is expected. However, the calculation of equation (4) Executing the recurrence formula in its original form will have little effect on pruning.Because (
2) Calculation of formula can be omitted by n(j) on the right side,
This is the case where both gn(j-1) and gn(j-2) are large. Therefore, it is necessary to determine the magnitude of each of the three values, and the probability that the logical product of the condition that all three values are large is satisfied becomes small.

Therefore, in the present invention, the calculation of equation (4) is performed as a forward conditional assignment process as described below.
The characteristics of FIG. 2 is a diagram for explaining the processing. Here ■n(j) is defined as % d('Ij
) corresponds to the cumulative optimal distance up to one time point (i-1), which is not added. Prior to the following processing, gn(
It is assumed that j) is initially set to a sufficiently large i value ■. The processing for a specific "+J" is as follows.

... (7) When the procedure of jin is repeated from (j-2) to j, g
The following processing has been performed on n(j).

... (8) △ gn (j-2) + dn (j-2) = gn (i, j-
2) △ gn (j-1) + an (j-1) = gn (i, j-1
)△ gnCj) + d”(j) = gn(i, j), and if we combine equation (8), we get that i in equation (6) is calculated with one clock advance. Become.

To summarize this, we can say that the recurrence formula calculation, which is a rule of the method, has been achieved. From the definition of equation (6), this △ g n (j) i, j) = d"(j) ten n(j), so find n(j) (The process of the method is to find gn(i, J) (2 E is equivalent to the calculation of .

In this way, when the dynamic programming method is executed forward in the plan 'm', the next rule for pruning is determined.

The determination of pruning is efficient because it is executed by testing only n(j) in method (17). If pruning is determined by this, (2), (3-1).

The processes (3-2) and (3-3) necessary for calculating the recurrence formula can all be omitted.

Owing to the above first and second features, the pattern matching method according to the present invention becomes extremely efficient.

The present invention performs the procedure of equation (7) at each time i for j in each word n.
The method is to execute each cycle up to =x, 2, . . . Jn, and advance the cycle at time i. (Method '(1)'s pruning conditions can be considered to have many variations. The most frequent example is the optimal cumulative distance as i increases!■n
(j) increases, and θ(i) is a linear function (monotonically increasing) of i (method (1) is applied. It is also possible to define gmin, add a margin α to this, and apply equation (7) and (1) as θ(i) = gmin + α.Another example is to apply the equation (7) to the total g, instead of using a threshold. j)
You may perform pruning under any number of conditions, starting from the smallest.

(Embodiment) FIG. 3 is a block diagram of a word speech recognition device implementing the present invention. The audio signal inputted from the microphone 10 is frequency-analyzed by the analyzer 20, then sampled, digitized, and sent to the microprocessor 30 as a feature vector Ai.

A standard pattern storage section 40 and a G mesori 5o are connected as memories to this microprocessor. The standard pattern storage unit 40 stores standard patterns Bn or feature vectors bn of each word n as a time series. The G mesori 50 is g n (j) g shown in Figs. 1 and 2.
``(j).These standard pattern storage units 40.G memory 50 may be defined as a distinct memory area within the main memory of the microprocessor.

The recognition process is executed by the program of the microprocessor 30. When the first vector &l of the input pattern is input, the following initial setting is performed for each n.

This corresponds to the initial condition of equation (1).

Thereafter, each time the feature vector of the input pattern is inputted, the microprocessor 30 executes the process shown in the 7CI-chart of FIG. 101 in the diagram
The block corresponds to +27 pruning decision in equation (7). Similarly, the two blocks 102 correspond to the process (1), and the blocks 103 correspond to the processes (3-1), (3-2), and (3-3), respectively.

Block 104 stores all the values of gn(j) as ■n(j)
9 means to switch. This process is gn(j
) and the area where n(j) is stored. Block 105 is ■n
This means resetting area (j) with a sufficiently large value ■.

When the above processing is completed, the input pattern time is incremented by one clock, and the same processing is performed after waiting for the input of the next feature vector a. When the voice pattern ends and the last feature vector &□ is input, the following processing is performed inside the microprocessor 30. At this point, it is stored in the G memory 50. gn (Jn)
Since is the same as equation (6), D (A, Bn) = g (I , J) = gn (Jn) 10d (I , J) (13) For each word n The distance between patterns D(A.

B") 'l. By performing these t) 11th-order comparisons, the minimum value is determined, and the corresponding l1l=n is output as the recognition result.

Although the present invention has been described above based on Examples, these descriptions do not limit the scope of the present invention.

In particular, in the above description, the similarity between patterns is evaluated by distance, but it is also possible to evaluate the similarity by a quantity having the opposite magnitude relationship. In that case, the determination of the magnitude relationship of (7) Shikigin will be performed in reverse. Furthermore, in this embodiment, the features ai, bj
is a vector, but it may also be a scalar quantity such as a number specifying a vector in a codebook when vector quantization is performed. Further, for the sake of simplicity, an example has been described in which the pattern matching method is applied to a word recognition device, but it is obvious that the present pattern matching method can also be used in a Clockwise 92 type continuous speech recognition device.

(Beginning of the Invention) With the above-described effects of pruning and the present invention in which recurrence formula calculation is executed by forward conditional assignment processing,
It has become possible to significantly reduce the processing time for DP matching, and to realize a compact and low-cost audio recognition device.

[Brief explanation of the drawing]

Figures 11 and 2 are diagrams explaining the principle of the present invention, Figures 3 and 3 are block diagrams showing an embodiment of the present invention, Figure 4 is a flowchart showing the processing of the part, and Figure 5 (a)) is a diagram illustrating the prior art. 10...Microphone, 20...Analysis section, 30...Microprocessor, 40...
・・Standard turn memory section, 50-・・・・G memory. L 7sf S D vomit (and 12n ~ l

Claims (1)

[Claims] Time series featuring the standard pattern of each word n = b^n
_1...b^n_j...b^n_jn, means for temporarily holding the feature a_i of the input speech pattern, and storing the feature a_i and b^n corresponding to each word n.
The optimal cumulative value g^n(i
, j) by a recurrence formula of dynamic programming, the optimal cumulative value g^n(i, 1, j) up to time (i-1) is Means for storing in the form ■^n(j) corresponding to word n and each time j, and optimal cumulative value g^n(i, j) up to time i
and means for storing a quantity g^n(j) corresponding to , and at time i, the value of ■^n(j) satisfies a predetermined pruning condition for the set (n, j). Tenomi■^n(j)
The first process is to add d^n(j) = d^n(i, j) to a new value g, and then calculate the difference between g^n(j) corresponding to j in the vicinity of this j. Compare the size and find g<g^n
When (j), a second process is performed to transfer g as a new g^n(j), and every time time i advances by one clock, g^n(j) is transferred as a new ■^n(j). A pattern matching method for speech recognition characterized by proceeding with the first and second processing.