JPS59198A - Pattern comparator - 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 pattern comparison device that performs matching with a plurality of registered patterns, and particularly to a pattern comparison device that can be applied to recognition of continuously uttered word sounds.

In order for high-pitched voice, which is the most natural means of information generation for humans, to demonstrate its true value as an input means for human-machine systems, it is necessary to be able to recognize continuous normal conversational speech without limiting the speaker. is desirable.

FIG. 1 is a block diagram of a speech recognition device that uses words as recognition units. (1) is an audio signal input terminal, (2) is an input audio signal that performs frequency analysis, LPC analysis, and PAR(X).
(3) is an acoustic analysis unit that converts into a series of several sets of numerical values (feature vectors) by R analysis, correlation analysis, etc.; The pattern storage unit (4) compares the series of feature vectors for the input audio signal to be recognized analyzed by the acoustic analysis unit (2) with each of the standard patterns,
A pattern matching unit calculates the distance or similarity between the two by t+, and (5) is a determination unit that recognizes the input phonetic word based on the calculation result of the pattern matching unit (4) and determines it as a recognition result. 6) is an output terminal that outputs this recognition result. In the peak voice recognition device having such a configuration, an excellent pattern matching method is a method of performing matching (DP matching) using time axis nonlinear expansion/contraction using the dynamic 31-stroke method.

In continuous word recognition by the apparatus of the present invention, this I)P matching plays a central role. Next, the DP matching algorithm will be briefly explained.

of? The four-tone pattern has a feature vector a for each
, ,b,.

When the distance between vector al and , is d(i, j),
The weighted average of d(1,j) is obtained by examining various correspondences between the vectors constituting both the above-mentioned series, and the correspondence that minimizes the weighted average is taken as the optimal correspondence between both series, and then M
The Tn average is set as the distance 111fiD(A, B) between both series, and DP matching efficiently performs this procedure using the dynamic 31-stroke method. Note that d(i, j)
The normal vector a1 is the Euclidean distance of 11%, and the city distance is used.

Figure 2 shows this two-dimensionally.A.

The time correspondence of the 8-car pattern, that is, the time conversion function j(i
) is the grid point c(k)=(i(k),j
(k)) series F=c(1)c(2) ... c(k)-c(T(
) −(2) (i(K)=I, j(K)=J). At this time, D(A, B) is defined as follows.

Here, w(k) is a non-negative constant, and its value is the time transformation 1
It is determined by the method used to approximate the seasonal number j(i) using a sequence of points. Here, if the denominator of equation (3) is a constant M-Σw(k) that does not depend on F, then I) (A, B) is a dynamic 31
Both methods can be calculated more efficiently by -1. That is, =min[min [r' d(c(1)w(
/? ), :l publication (c(k)w(k)]c(k)c(+)
c(2)-c(k-1)'', so g(c(1)
)=g(1,1)=d(1,1), recurrence formula (4)
By solving, g(c(K))=g(1, J) is obtained, and from V, (A, B)= −I−g(1, J) Moo−−−°−−− (5) as I )(A, B) is obtained.

As a method to make the fraction of equation (3) constant, M=I+J
There is a method of setting K so that it becomes (symmetrical 11 flights) and a method of setting K so that M=Ij or J (asymmetrical 4!l). Third
Figure fal ~ (fl shows an example of the constraint conditions when selecting the point sequence F. 1) The route leading to the point (i, j) can only take the route shown by the arrow in the figure. The number shown above indicates the load w(k) when that line segment is selected as the path. (a) and (b) are the symmetrical example mentioned above, where M = 10 J, (cl~(f) is an example of nO^ self-asymmetric type, and M=1.

To recognize word sounds using such a matching method, proceed as follows. The word class to be recognized is n (n=1 to N), and its standard pattern is °B0.
Expressed as Distance NU I) between human power A and each standard pattern H'
. -L) (A, B”) in the above method Jl) Qshi, I)
11 o-min (Inn catty given class n. Let be the recognition result for A.

If M is set to 1 in the asymmetric DP matching described above, M becomes 1%, which is related only to the input cover turn length, and M is constant for any standard pattern in equation (5). , can be defined as . Hereinafter, it is assumed that the distance between patterns ~t is based on equation (6). Under the constraint conditions shown in Figure 3(C), the formula (
6), it is best to calculate the following recurrence formula (7).0 Initial condition: g(1,1)=d(1,1)Next, continuous word recognition will be explained. Continuous word speech recognition can be formulated as follows: 1, 1X words q(1), q(2
), ...high IJ1 when uttering q(x) continuously
The pattern is represented by A.

A,=aa...al...al...
... (8) 2 Kogo qcx) palm pattern by 1%, == 1], qLx) 1) 2qtx)...bIq
When (x)・bzQ(x) + HHH+ (9), X words Bq(D ``'q(2bi'' Bq
The palm pattern ■ obtained by connecting (x) is B=Bq, q→BQ(21■...Φ, prisoner(x)
・・・・・・・・・ (10=1rrLl)1〕'j(
+),,, b, qtD l)? (2斥E""' by
+(2) ・4. q (x also, qLx) ,...23y
, represented by seven. Here, the mountains represent the connections of the patterns.

Therefore, continuous rli candy speech recognition performs 1) P matching between this hundred and input speech pattern A, and then calculates X and q(x
)(x=1.2...,x). In other words, it is sufficient to calculate R1 and find the conditions under which 'l' becomes the minimum.

Equation (1υ) would require a huge amount of measurement if we were to calculate it in minutes.In other words, if the maximum number of consecutively uttered words in the input speech pattern is K, and the number of word mark patterns is N, then , N times of R1 calculations are executed.Therefore, in the 1st Song Dynasty, this problem was reduced to solving the following recurrence formula.

In input audio pattern A, 1=f4-1 to i=m
The partial section up to is defined by partial pattern A (8m).

A, (l+m)=a a...radius...
...Φ...@6+I ff+2 At this time, the distance between the patterns is defined by equation (6), Vf, and the following can be changed.

I) (A, B, ωB2) = min [DcA(o, m)
,B, )+D(A(m,I),B2))・(
131 Using this fact, the equation 0υ can be solved as follows.

The meanings of the symbols used hereinafter are summarized in Table 1.

- Below is the remainder - Table 1 1) When the number of input words Once the solution is found, the recognition results are sequentially obtained from the last Fv tlt-word name and segmentation result of the X word string to the first Jlt sn name and segmentation result according to the flowchart shown in FIG.

11) When the number of input words X is unknown = min[D(m)+D”(m+1:i))
・・・・・・ (lrNN(i)−−+ B(+)−
From the solution of the recurrence formula m (where n and m satisfy equation a9), the υ recognition result is obtained according to the flowchart of FIG.

A two-stage DP method has been proposed to realize the above idea. Next, the outline of the two-stage DP method will be explained.

In the two-stage DP method, first, s(s:t) is obtained by DP for every combination of s and t, and then I)(i) is
)P, and is characterized by having two stages of DP.

Although a forward algorithm and a backward algorithm have been proposed as this two-stage DI) method, the backward algorithm will be explained here.

(p Person Kabataan frame 1-1VC row, D(
i-1), N(i-]), and 13(i-1) are calculated as 1.

(2) Perform DP munching on the standard pattern and input cover turn of term n(n-1, 2, -, N) in the reverse time direction with Io as the starting point. Therefore, the hookworm condition for the route is the third
Corresponding to Figures (cl, (dl, (el, f)), Figure 7 (at, (b), (cl, (dl)).

Although matching may be performed within the matching window width R, here it is assumed that matching is performed within the range of slope L to 2 (within the slope limit, the shaded area in FIG. 6).

This matching is performed with a free termination. As a result, I
):(s:i) is found. However, i −2J ”+1
3SS, i-(J/2)J'.

■ Find 戊(1)(i), N(i), and B(i).

(4) Return to C) with i = i old.

Let us consider applying this idea to the recognition of continuous monosyllabic speech. A monosyllabic voice has a consonant plus a vowel, and the consonant part is shorter than the vowel part. However, in particular, monosyllabic speech that has the same vowel part must be distinguished by subtle sequences of consonant parts. Therefore, when matching each of the input Qt 1-syllable speech and standard bata-n monosyllabic speech with the entire middle syllable speech in the pattern matching described above, the influence of the vowel part on the matching result is large, and the influence of the vowel part on the matching result is large. It becomes difficult to distinguish subtle differences.

The present invention compensates for this drawback and provides a pattern comparison device that performs matching with emphasis on consonant parts.

That is, in order to actively introduce prior knowledge and perform more accurate matching, it is necessary to introduce weights to each frame of the standard pattern and input pattern. Even if a car is randomly introduced into each frame of the input pattern, all the algorithms up until now will still hold true. However, if a weight is introduced into the standard pattern, the cumulative matching distance will depend on the standard pattern length, etc., and the recurrence formula (141 (1G) will no longer hold.

Next, the reason will be explained. For example, the standard pattern has a weight of 39 people (...).An example is the symmetric! In order to evaluate whether the two patterns fit best, it was necessary to divide (normalize) the cumulative matching distance between the two patterns by 411, which is the human cover pattern length and the standard pattern length, as described in Id Ril.

Now, suppose that the standard pattern that best fits partial pattern A (0, m) of human cover turn A is B1, its length is J, and any other standard pattern is B2, its length is B2, then the following formula is obtained. holds true.

1) Is (P, Q) the sum of pattern P and pattern Q before normalization? is used to represent the matching distance.

At the time of the i-th friendship, Shikien, a! Let us consider the case of searching for the bank pointer and the last Fc word (single word) based on '9 (assuming, of course, that it is normalized by the input pattern length and the $2% pattern length). The final immediate word is X, and its dissonance is x.
, pack pointer In , l! : (When R is set,
The cumulative matching distance of the standard pattern combining B, and In order to search m and X by Equation 04) QFD, α
Naturally, the value of the modulus must also be small.

That is, if β<α holds, D at 2o9
This is because D(m) obtained at the m-th frame cannot be used as (m).

However, α and β generally do not hold. For example, D(A
(0, m) , B, ) = 10 , D(A(0,
m), B2)=20rr+=20, b,=10
, If b2=20, then in formula Qej, left side = 10/(20+10)=1/3 right side = 20
/(20+20)=1/2, and the value in F is the formula O
Satisfy Q. But 1=40, x=10.1)(
A (m+l, i),
)=7/6β-(20-1-60)/(40+20+1
0)-8/7, so α>β, and equation θ6 is no longer satisfied.

However, in the case of the asymmetric I)P method that depends only on the human cover turn length, then 48 (0, m million B) - tI) (A (m + 1.i), X
) is clear, so Equation 04)(, 1~ can be used without contradiction.

'' In the recognition of one-syllable speech, it is better to increase the weight of the consonant part in order to place emphasis on the consonant part, but if this is simply done, the problems described above will occur.

The present invention is characterized by a method of photographing car matching that eliminates this drawback and enables matching with emphasis on the child's portion.

The above problem can be solved so that the sum of the weights for each frame j, [of the standard pattern is constant for any standard pattern. In other words, if the weight of the nth mtit pattern in the jth frame is W'(j), by determining W''(j) as Since it depends only on the number of monosyllables, when the number of monosyllables is specified, it does not depend only on the length of the input cover turn, and two-stage DP matching can be used.

FIGS. 8 and 9 show an example of a weighting method for each matching path.

Therefore, it would be sufficient to set rJ and the weight (=1) as shown in FIG. 10 in line with the hookworm of the matching path as shown in FIG. 7 (dl).

FIG. 11 shows an embodiment of the present invention. 01 is an input terminal for audio signals. (1j) is a feature extractor that converts the input audio signal into a series of feature vectors, which is composed of a filter puncture, etc. 09 is a feature extraction unit that converts the input audio signal into a series of feature vectors. This is the pre-registered monosyllabic standard pattern H self-inertia part.Here, the weight W'(j) of 11J is also recorded for each monosyllable and each frame.θ(
Reference numeral 1 denotes a vector distance meter leg, and for each i-frame, a vector l'al' forming the input cover turn in the shaded area shown in FIG. 6, and a vector 1) forming the standard pattern n, 7) 1111 t7J) Distance N
f d'(i, j) (n=1, 2 +"'+N
F j = 1.2, -0J") is calculated and stored. Here, the distance d" (i, j) is, for example, a city distance of 111fl. Zunawajia'(=(ai'l l ai1
2 +"'ha'ρ, b7- (b'4 H, L)'j2
,...,l)'17) When d'(i, j)-
Σl a i h L) x k lk・1 and [7] can be defined.

Q71 is the vector distance calculation unit (149 output d"(i
, j) (n=1゜2,..., N+ j=1.2s
..., J') To, M' if fm I! Quasi-Hatan (weighting coefficient W” (j) stored in the 1m section
From, human cover 11- :/ (7) i'(i'=i
-2J"+1~1-2-J") Kara i 7 v -J-
Cumulative matching distance between the partial pattern and the standard pattern at 1
) is a part of the partial cumulative range meter that calculates and stores (i': i ), where Dro (i': i ) is calculated from the following recurrence formula.

That is, the first JJJI value 1)', (i, J')=d"(
1, Jn) To L, Te Kei Yasuru. However, the constraint conditions for route selection are as shown in FIG. The result of this calculation is D'', (i', 1)
is temporarily stored in the next cumulative distance calculation unit O barrel as D:(i', j). In formula (a), D7 (+', j
), the range of frame i' of the human cover turn corresponding to the j-th frame of standard pattern n is i2J''
I+2J, 4i"; i-↓J-chohito1+1j2 22 Therefore, in this range of i', j=J", Jff
This is calculated for i, . . . , 1.

θ8) assumes that the i-th frame is the final frame,
Cumulative distance from i=1 when the final monosyllable is n1):(
i) and pack pointer B of monosyllable n: Compute (i),
This is a cumulative distance calculation unit that stores them. That is, x=1.2. ..., find for X as B (i)=i' △ △ (n, i'i where n, i' satisfies the formula υ). Is X here input J11? f is the number of clauses.

The cumulative NfD(+), the barak pointer Bx(i), and the last immediate syllable Nx(+) obtained as above are stored in the cumulative distance storage unit (2), the pack pointer storage unit Stored in the tail monosyllable storage section (A). formula? υ
Dx (iLl) is the value obtained previously, and the cumulative distance record (c) is a voice section detection unit that determines the voice section from the magnitude of the human signal, etc., and this drunkenness zone detection When the part (c) detects that audio input has started, the frame number counter starts counting every frame. In the vowel recognition process, the process up to the determination of the last jlt box is rarely performed on the i-th frame, and the count value of this frame number counter is the one that sets this i. Therefore, the same processing as described above is performed every time the frame advances by one. The frame number counter (c) starts counting when a voice section is detected, and is reset when the voice section ends. Therefore, N(i) and B(i) are i
=1.2. ..., ■ will be stored.

The segmentation unit (a) is the pack pointer storage unit (
For c), an instruction to read a predetermined pack pointer is issued. That is, when the segmentation unit (A) issues the value i to the pack pointer storage unit (C), the pack pointer B (1) is retrieved from the pack pointer storage unit (C). When the segmentation section (c) receives the value B(i) from the self-initiation section (c) in the pack pointer, it lights the same bean paste into the pack pointer storage section (2). Therefore, when the voice section detecting section (c) detects the end of the voice input, the final value 1 of the frame number counter I9 is added to the segmentation section (a) K+1, and the segmentation section (a) first calculates the value I. Issued to the back pointer storage section (c). Thereafter, according to the operation described above, J3(I) B(
13(1)), B(1', (B(1))),...
Outputs of 0 will be sequentially obtained. , Nirera no Arihakuke, the frame at the end of the penultimate rlt syllable, the frame at the end of the third rlt syllable, the frame at the end of the rlt syllable 4],... ) &yo Since it was a monosyllable ending in an i frame, we simply changed this n^ to 5. Then, from the last middle H clause, the reverse l
Recognition results will be obtained in the order of 1lNi. If you want to obtain results in the normal order, you can perform this order conversion on the output of the pack pointer storage section (c) or on the output of the last monosyllable storage section @.

FIG. 12 is a flowchart when the functions of the 11th century embodiment device are realized by software.

Steps 100 to 103 are the initialization portion.

Steps 106 to 108 are steps for calculating the inter-vector distance between the feature vector of the standard pattern and the feature vector of the human kataan in the shaded area in FIG.
This corresponds to the process in step 11.

Steps 109 to 114 are partial accumulation part M D6 (+'
: +), and corresponds to the processing in the partial cumulative deviation calculating section 07).

Steps 115 to 116 include the cumulative IH)x(i), the last H monosyllable Nx(i), and the pack pointer Bx(+)
This is the part that calculates and memorizes each, and the previous i?
8 cumulative distance t1 part 0 frame, cumulative distance storage section (c), last monosyllable storage section (a), pack pointer storage section (c)
This is the process in.

Steps 117-120 are for i=1.2. ..., Nx(i), Bx(
i) to obtain the final α・J recognition result,
This corresponds to the processing performed between the url entry pointer storage unit), the segmentation unit (a), and the last monosyllable ml unit (2).

As shown in 1 above, according to the device of the present invention, in continuous monosyllable recognition using the recurrence formula R1a of formula 09, by introducing weights, the verification that emphasizes the consonant part is used as a guideline, and the recognition rate is significantly increased. This is an improvement.

Although this embodiment has been explained using the recognition of monosyllabic speech as an example, general word speech may also be used, and especially when the mutually sweet single 11ha is included in the pleading word 粱, the characteristic parts of the speech may be used. The effect can be improved by adding a large ηf to . Of course, the present invention is applicable not only to voice recognition but also to recognition of other patterns.

[Brief explanation of the drawing]

Figure 1 is a block diagram of a conventional speech recognition processor. is an example of one constraint condition for selecting lattice points on the i-j plane.
and Figure 5 are respectively when the number of human words is known,
A flowchart showing the procedure for determining the secumendation and opening words in the unknown case of continuous unit ^ (l? 4 doors). , Figures 7(al to d) are diagrams showing examples of constraint conditions when selecting grid points on the 1-j plane, and Figures 8 to 10 are examples of weight adjustment for matching paths. FIG. 1I is a block diagram of an embodiment of the present invention;
Figure 12 shows the functions of the device in this embodiment realized by software (
7 is a flowchart. qD... Feature extraction unit, (1~... Monosyllabic standard pattern storage unit, OQ... Inter-vector distance calculation unit, 07)・
... Partial cumulative distance calculation section, 08) ... Cumulative distance calculation section, @ ... Last (at 111-Syllable storage section, (c) ...
Cumulative distance storage unit, (c)...Pack pointer storage unit,
(c)...Voice section detection unit, 9...frame &
Meter & Equipment, VH...Segmentation Department Agent
Figure Yoshihiro Morimoto (j> tbl (0 <d> Figure 1 Figure 2 Figure 70 Figure 7+ +'UJ''(J' Figure 11 Figure 12

Claims (1)

[Claims] 1 Sequence of feature vectors aI ”2...a
Feature extraction means for converting into l and feature vector series 1)
Standard pattern B consisting of I r'b, n...b';n
A standard pattern storage means for storing '(just...W'(J'')) and a feature vector aIa2 constituting a pattern in which the distance between the human cover turn and the standard pattern R' are entered in 11. ...a, and the feature vector b+ 1)2...b;n that constitutes the standard pattern R'
,! : As a function consisting of the weighting coefficient W'(j),
A pattern comparison device characterized by having means for minimizing by dynamic programming.