JPS634299A - Pattern matching - 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 pattern matching technology, particularly to improvement of pattern matching technology in speech recognition using syllables as recognition units.

(Prior Art) A method of performing speech recognition using basic phonetic units such as syllables is suitable for large language word recognition because it does not require a standard pattern for each word to be recognized. In this case, the recognition process includes two processes: a syllable recognition process that converts the input speech into a series of syllable candidates, and a word recognition process that matches this syllable series with each word in the word dictionary expressed in syllables. Consists of. The latter word recognition process can be realized by pattern matching between symbol sequences using syllables as symbol units. As such a pattern matching method, "Information Processing J Vol. 17 No. 7 pp. 650-658
(July 1976) is a typical method for matching between symbol strings.

In this method, the distance between symbols is determined in advance, and when two symbol string patterns are temporally associated, the distance between each associated symbol is determined so that the cumulative amount of distance accumulated over the entire pattern is the minimum. Matching is performed using dynamic programming to achieve this.

(Problem to be solved by the invention) Here, the distance between symbols is the word self sign, that is, the closeness between syllables,
Alternatively, it is resolved to indicate the probability that a certain syllable will be observed as a different syllable instead of the correct answer as a recognition result. If there are omissions or insertions in the syllable candidate series, one syllable will be associated with a plurality of syllables as shown in FIG. Syllable dropouts and insertions often occur in specific syllable connections due to the inherent properties of phonemes.

Therefore, when determining the distance between syllables, it is important to consider phonological transformations that depend on the context (connection before and after syllables).

However, in the above-mentioned matching using the inter-symbol distance, since the inter-syllable distance is given without regard to the context, such context-dependent transformations are not taken into consideration.

For example, consider the transformation /kiyo/-/kyo/. ``distance'' between the syllable /ki/ and the syllable /kyo/, the syllable /yo
Such transformations are typical even though the "distance" between / and the syllable /kyo/ is not small. However, in the above matching, the syllable string /kiyo/
The distance when / and the syllable /kyo/ are made to correspond does not become small.

The present invention provides means for performing matching with high accuracy even when syllable dropouts, insertions, etc. exist in a syllable candidate series or network by describing phonological transformation rules using distances in a table format. An object of the present invention is to thereby realize a speech recognition device with high recognition accuracy.

(Structure of the Invention) When comparing two patterns expressed as a syllable series or network, the present invention provides the syllable X1 and the syllable sequence X2 The distance d(X
x; X2, X3).

Furthermore, the present invention uses the distance d between the syllable x1 and the syllable string X2X3 when comparing two patterns expressed as a syllable series or network, and
are the two types of inter-syllable distances dF(X1, X2) and dB(X1, X2) given in advance for each syllable combination.
s).

Furthermore, the present invention uses the distance d between the syllable x1 and the syllable sequence Three types of inter-syllable distance dF (X1, X2), dB (X1,
a) and d' (X2.Xa).

(Operation) The operation of the present invention will be explained by taking as an example the case where two patterns are both expressed as syllable sequences.

Two patterns A and B respectively A=(a(1),·,a(i),−,a(I)).

Let B=(b(1), . . . , bψ, . . . , b(J)). Here, a(i) and b(j) are syllable symbols (specifically, numbers). Using a method similar to the above-mentioned document, by introducing the distance dO (X1, X2) between syllables 8 and 8, matching of breads A and H can be realized using dynamic programming as follows.

Initial conditions: g(0,0)=O g(Oj)=co, j=1. ..., J recurrence formula: g
(i,j)=dO(a(i),b(j))minlg(
i-1j) '1(1)・g(ij-1)'.

i=1. ..., ■ j=1. ..., J However, g(i, 0) =■, i=1. ..., ■ Recurrence formula (
The matching path of 1) is as shown in FIG. 2(A).

Next, the distance d(Xl;
Xs.

Define Xa). This distance is from syllable X1 to syllable string X2
3 to indicate the accuracy of observation. (
Conversely, the accuracy with which the syllable sequence X2X3 is observed replacing the syllable X1 can be defined separately, but in the following explanation it will be the same.) For the newly defined distance d, see Figure 2 ( A recurrence formula for P can be used as shown in B).

By employing such a distance d, it is possible to obtain a matching distance between syllable sequences that takes into account the ease with which syllables are deformed depending on the concatenation.

The first aspect of the present invention has been described above, and since the distance d is three-dimensional array data as described above, the amount of data is large. Number of syllables in Japanese - 100 = 100
Requires storage capacity of 3=106 words. Second of the present invention,
The third aspect is intended to reduce this amount of memory.

Here, the distance between the syllable X1 and the syllable string X2X3 is described by the following approximate expression.

d(Xl;Xs,Xs) dF(Xt,Xs) + d
B(XI,Xa) (3)d(Xl;Xs,X
s)”dF(X1,Xs)+dB(Xx,Xs)+
d'(Xs, Xs) (4) Formula (3)
Then, d is the distance dF(
X1, Xs) and the distance d between the second half of syllable x2 and syllable X3
It is given as the sum of B(X1, Xs). At this time, the storage capacity is 1002X2:2X104, which is a significant decrease in storage capacity.

Equation (4) has the third term d'(Xs,
Xs) is added, which means that the syllable string X2X3 is 1
The accuracy with which a syllable is observed is expressed by distance.

do and d in equation (2) or dF in equations (3) and (4),
dB and dJ may be given in advance by any method. It is possible to obtain it from phonological transformation rules or from real speech.

For example, the following methods can be considered.

(1), d(Xl;Xs,
Find Xs).

r [C]ya-[C]iya; [C] = Prepare a set of transformation rules for any consonant 1, etc., and add xl-X2 to this set.
If X3 is included, d=1.0; otherwise, d=α($-1,0).

(2) The distance is calculated from the matching score obtained by actually performing matching between the sound bangs using real speech (in this case, the time axis normalized matching method using the DP method can be used). Since it is not efficient to perform such an operation for all syllable combinations, it is practical to perform it only for combinations that have a certain degree of possibility of deformation from a phonological point of view.

(Embodiment) A block diagram showing an embodiment according to the present invention is shown in FIG. In the figure, 1 and 2 are the baton buffers, which store the batons A=(a(1),..., a(I)) B=(b(1),..., b(J)), respectively. has been done. 3 is a recurrence formula calculation unit, which sequentially reads symbols a(i) and b(j) from buffers 1 and 2, executes calculation of the recurrence formula, and calculates the integral amount g(
ij) is held in the integral amount buffer 4. A distance table memory 5 receives symbols a(i), b(i), etc. as address inputs, and outputs distance values do and d in equation (2).

The distance table memory stores the following contents corresponding to the three aspects of the present invention.

(i) Three-dimensional array table d(X1, Xs, Xs) and two-dimensional array table d'(X1, Xs) (ii) Three three-dimensional array table dF(X1,
s), dB(X1, Xs).

do(X1, Xs) (iii) Four three-dimensional array tables dF(XI,
s), dB(XI, Xs).

d' (X1, Xs), do (X1, x2) The recurrence formula calculation unit 2 executes formula (2) in case (i),
In the case of (iii), the following equations (2'Xs") are executed by substituting equations (3) and (4) into equation (2).

Finally, the integral amount g(I, J) obtained by 1=Ij=J is output as a matching distance.

The principle of the present invention has been explained above using the case where the pattern is expressed as a syllable sequence, but the principle of the present invention is as follows.
It can be applied directly to not only syllables but also other phonetic units. Such a unit is phon
eme (phoneme), phone (a unit even finer than a phoneme), diphone (from one phone to the next ph)
one unit), demi
-syllable (one unit is an interval from one syllable to the next syllable), etc.

Further, the principles of the present invention can be directly applied even when a pattern is represented not as a symbol sequence but as a network of symbols.

For example, suppose two patterns A and B are respectively A=[(a(1
), ·, a(I)); (t(i', i)), i, i'
=1.・, IIB=[(b(1),...,b(J
)); (S(j', j))jj' = 1. ...
, Jl, where I and J are the number of nodes in the network, and t(i, i') and 5(jj') indicate the presence or absence of a path from node i to i' and from node j to j', respectively. represent.

When t(i', 1) = 1, path i' -i exists Q II II L not s(j'
When =)=1, there is a pass completion-j Q n
When n is not present, the following recurrence formula can be used instead of the recurrence formula (2).

(Effects of the Invention) As described above, according to the present invention, it is possible to perform matching at the syllable symbol level, taking into account the deformation of syllables depending on the environment before and after the syllable, and to prepare patterns for each word. It becomes possible to provide a means for performing high-precision matching without having to perform a large amount of speech recognition.

[Brief explanation of the drawing]

Figure 1 shows an example of matching symbol strings, Figure 2 (
A) and (B) are diagrams explaining the principle of recurrence formula in matching, and FIG. 3 is a block diagram showing an embodiment according to the present invention. In the figure, 1.2...Pattern buffer 3...Integral amount buffer 4...Recurrence formula calculation unit 5...Distance table memory storage (word 3 and a half) i-
2j-/Me

Claims (3)

[Claims] (1) When comparing two patterns expressed as syllable sequences or networks, the distance d (d) between the syllable X_1 and the syllable sequence X_2 A pattern matching method characterized by using (X_1; X_2, X_3). (2) When matching two patterns expressed as sequences or networks regarding syllables, syllable X_1
and the syllable string X_2X_3, the distance d
is the two types of inter-syllable distances d^F(X_1, X_2) and d^B( given in advance for each syllable combination.
X_1, X_3). (3) When comparing two patterns expressed as sequences or networks regarding syllables, syllable X_1
and the syllable string X_2X_3, the distance d
is the 3 given in advance for each syllable combination.
Seed intersyllable distance d^F(X_1, X_2), d^B(X
_1, X_3), d^J(X_2, X_3).