JPS59196A - 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 provides a pattern comparison device that compares each standard pattern registered as a series of feature vectors with an input cover pattern converted into a series of feature vectors, and recognizes an input cover pattern. The present invention relates to a pattern comparison device suitable for recognizing a standard pattern in which only a part of the pattern is different and other pattern parts are very similar.

There are great expectations that voice, the most natural means of generating information for humans, will be realized as an input means for human-machine systems. In recent years, there has been interest in combining word processor input with kana-kanji conversion technology and performing monosyllabic voice input instead of input using kana keys. In this case, each monosyllable is produced in isolation, but "-ta",
"Pa".

In the case of an input pattern such as "-shiki" which differs only in part but is very similar as a whole (hereinafter referred to as similar pattern), a problem of misrecognition occurs.

By the way, when performing speech recognition by comparing a standard pattern and an input pattern and calculating their similarity and distance, the dynamic 31-stroke method (hereinafter referred to as DP method) is used to calculate the similarity and distance. . In order to solve the problem of erroneous recognition of similar patterns in speech recognition using dynamic programming, it is necessary to increase the number of feature vectors per single syllable,
It is necessary to increase the dimension of the vector. However, increasing the number and dimensions of feature vectors dramatically increases the amount of calculation, making it impossible to put it into practical use.

In addition, in recognition of continuous monosyllables, rather than isolated monosyllables, there is the problem of separating the boundaries between monosyllables in addition to the marking problem. , this also affects the increase in the amount of calculation.

Such an increase in the number of h]N kites becomes a major obstacle when trying to perform speech recognition in real time.

In view of the above points, it is an object of the present invention to provide a pattern comparison device that can recognize similar patterns without substantially increasing the amount of calculation, and can recognize human cover patterns in real time.

In order to achieve this object, the present invention, when performing pattern matching calculations using dynamic programming, matches all pattern parts of 9 quasi-patterns with human cover patterns to find the cumulative distance. In calculating the cumulative distance, find the partial cumulative distance that can be obtained at the same time, that is, the cumulative distance between the specific part of the standard pattern and the incoming cover pattern, ^
The system is configured to recognize input patterns using two types of cumulative distances.

The principle of the present invention will be explained below with reference to FIG. 1, taking continuous monosyllable recognition as an example.

FIG. 1 shows the state of matching when the first frame of the introductory pattern is used as the starting point and is matched with the standard pattern monosyllable n.

That is, each of the 41 openings is a straight line with an inclination of L-'2, 2, and matching is performed in an area surrounded by these two straight lines. This area is determined in consideration of the required recognition accuracy and speed change speed. In the case of FIG. 1, it is determined that the human cover turn is allowed to vary within a range of 0.5 to 2 with respect to the standard pattern length.

C shows a matching path when the first frame is the starting point and the first frame is the ending point.

in is a frame that separates the monosyllable consonant part and the I-sound part of the standard pattern n, and is registered at the same time as this consonant part is registered as the standard pattern. Therefore, from the first frame of α recording 04 syllable n to the ++
The consonant part of the monosyllable up to Fushi -no\ is the consonant part of the monosyllable, and the part 1 (fU, 'ff part of one syllable) from the j''+i frame to the JIIth fret\\. This is a frame of a human kataan corresponding to j”′.

Here, the distance between the input cover patterns from 1st to 1st Freno\ and the combined pattern of the optimal combination of each standard pattern (hereinafter referred to as the minimum cumulative distance)
Let be I)(1). Therefore, D(i) is used as a measure of the degree of similarity between the two, and is defined by the following equation.

Note that I)(i'-1) is the minimum cumulative distance to the i'-1st frame of the input pattern, and D:(i':i) is the minimum cumulative distance to the i'-1 frame of the input pattern for the standard pattern n. The distance #E (hereinafter referred to as partial distance) between Freno\'s pattern (hereinafter referred to as partial input cover 1)
It is.

In a pattern comparison device using the conventional DP method, D(i'-1)l-11;(i
i′(-′r′) and n(□′+ma) that minimize ′:i) were found for N, but in the pattern comparison device of the present invention, only D8(i′:l) is required. 1) (1), n(i
), N(1). In addition, B(1)−♀
'1eN(i)=It. ■)(1) in the present invention
) include, for example, the following: In the quantity-j plane of Figure 1, D? Let o(1'tj) be the starting point (
c, j'') to point (1', j), when L''(I', j) is the number of input frames from the first frame to the first frame. , D′io(h, 1
)/L'' (b, 1) and ■3' (1) weighted sum (J so,
Find the monosyllable 11(-n) that minimizes the bottom margin), and calculate
=D(b1)+D5(b:+)・・・・・・・・・・・・
......(3)B(1)i-b-] N(i)=-9.

In the equation (1), n is fixed and only i' is reduced by 1, and i' at that time is set as b.

D(i) can also be defined as follows.

B(1)=i'-1 N(i)=n However, IIn is i' and n that satisfy equation (31).

Examples of the present invention using the above principle will be described below.

FIG. 2 is a block diagram showing an embodiment in which the pattern comparison device of the present invention is applied to recognition of continuous monosyllabic speech.

In the figure, the audio signal input to the audio signal input terminal 01 is processed by the feature extraction unit Qυ into a series of feature vectors A''=al
a2...converted to al. 0 堠 is a vowel recognition unit that recognizes each feature vector as a vowel, and 0 葎 is a vowel recognition unit 02 that stores a discrimination function necessary for recognizing the feature vector as a vowel. This is a function storage section. For recognition using a discrimination function, a conventional well-known pattern recognition method is used as is.

0→ is a vowel recognition result storage unit that stores the result recognized as a vowel as described above by the vowel recognition unit 0, together with a frame number. θe is a monosyllabic standard pattern storage unit, and each monosyllable is stored in the form of a series of feature vectors. 0yu is the nth monosyllable bn, br...1)? stored in the monosyllable standard pattern storage unit 09 in the i-th frame of the input audio signal. The vector distance "d"("*j) (
1-2J"-1-1<1'<i) for the portion surrounded by 42 straight lines in Fig. 1. This is a vector-to-vector 11"J4111ff calculation unit that calculates H1 for the portion surrounded by 42 straight lines in Fig. 1 and temporarily stores the calculation result. . Vector al'=(altad2.-*atb
s) and vector b'', = (l bow) e162, .

*16M,! :(7J distance d'(itJ)c Example 7Li,
fd”(IIj)−1′ frame to i frame vector sequence a+′a+′++′−′−at(i−2
J′ ruled line 1<+′<+−−J”) and the nth
! Ji syllable bnbr:,...b''i, matching distance D: (i':I) and matching distance m of only the consonant part
+r7o (H', 1) and the number of frames 1r (i', 1) in the input direction of the consonant part are calculated from the output of the vector distance calculation unit 0, and the results are temporarily stored. This is the calculation section. Naoriiku (i':I) is calculated as follows.

When using the matching path constraints shown in Figure 8, the distance Djl'(1'l, l) between the patterns from the starting point (i, J...) to (i', j) is as follows. It becomes like this. In other words, the range of frame i' of the human cover turn corresponding to the j-th frame of standard pattern n is as shown in FIG. −H, j′))−=
--(41"<,j'-,j+2 D':(+'*J)"=Dri(i'+le3)+(1
”(+*3)...The recurrence formula (5) is Dt(i, Jn)-dn(i9.■n)...
... Solve (6) using the initial value. Here, the notation argmin() means X that minimizes the value in (2). Therefore, D3(*':i)-L)S'(+', 1)-=---- (
7).

Also, when performing this 41 calculation, Dl at the same time.

Also find C+'ej, l, I-'(+'9J,). That is, 1) j=j" (or j=J-1 and j=J1-
In the case of 1), ■) to (i', j) - 4n (i', j)...
・・・・・・・・・・・・(8) I-”(1′$3)=
′・・・・・・・・・・・・・・・・・・(9)2)■)
'C is 1)70(i', j) for j<j"-1 other than
：=DYo口′H,j)+d(i′,j)−01L”(
i', j)-L'(i'll, j)11=...=-
・(++) is calculated. Therefore, DS′0
(i'ej) is found as the cumulative matching distance of only the consonant part from j=jn or J=j"-1 to j along the path when I)ro(1':l) is calculated, and L""C.
+'*j) is determined as the number of frames in the input direction of the path from j=j' or j=j'-1 to j. The matching result of only the consonant part of the input kataan from the i'th frame to the ith frame for the monosyllable n is Di(+
'e1) * Item 1 means that the starting point of the consonant part is jt'' when the matching path passes through jn, and jn-1 when it does not pass through n.

Frame 0 is the cumulative distance D''(i) from i=1 when the final monosyllable is n, assuming that the i-th frame is the final frame.
is the cumulative distance 81 calculator that calculates the back pointer B" (1) of the monosyllable n by d1 and stores them, that is, i' = argmin (1) (i'-1) month JS (i':
i)) - (II7 λ″) 1)'(1)=IN+1)+D3(+:l)...
......(11B"(1)=i'-1...-...
・・・・・・・・・・・・(Determined as +4.

OI is the last monosyllabic vowel frame 81, and is 0.
→, α, (g) calculations for n=1, 2.・・・・・・
, N, the last monosyllable is n=a
rgminj-D"(i)II, and also find the midpoint i' (corresponding to the stationary part of the vowel) between the input frame corresponding to jn and the first frame for the matching path for a single syllable. As a simple way to determine i', a certain number of frames from the first frame can be determined as rich frames.

? 1 is a candidate monosyllabic generator, which has already generated the 1st
Ql) generates a monosyllable having that vowel as the vowel part from the output corresponding to the i'th frame of the vowel recognition result storage section θ→, which stores which vowel the frame is. When the candidate monosyllable is k, 1), o(B
(i) -1-t, +), T-(R(i)l-1*1) are read from the partial cumulative distance calculation section Q, 1) k(i) is read from the cumulative distance calculation section 0 This is the final monosyllable determining unit that calculates #I, when ut, WI, and w2 are weighting coefficients. In this way, the last syllable r( in the i-th frame is determined to be ζ).

(c) is the last + 11 syllable storage unit, and N(i) =
・It is memorized as k.

Moreover, when the last monosyllable is determined to be 'i', Dk(1)91',k stored in the cumulative distance calculation unit 0 barrel
(i) are stored in the cumulative distance storage section (c) and back-up ink storage section (c), respectively. D(1'-1), which is being calculated by the formula (c) obtained in the cumulative distance calculation section (11), is already stored in the axis of this cumulative distance storage section as described above.However, D( 0)=0.

(2) is a voice section detecting section, which determines the voice section from the magnitude of the input signal, etc. When this voice section detecting section (c) detects the beginning of the voice input, it detects the frame. The number counter (c) starts counting every frame. The process from the vowel recognition to the determination of the last monosyllable was the process for the first frame, and the old value of the frame number counter (c) is used to set this i. Therefore, the same processing as described above is performed every time the frame advances by one. The frame number counter (to) starts counting when a voice section is detected, and is reset when the voice section ends. Therefore, N(]) and B(i) are 1 in the last monosyllable storage unit (A) and back pointer storage unit (C).
=1°2, ..., ■ will be stored.

The segmentation department (foundation) is the back pointer storage department (
For c), an instruction to read a predetermined back pointer is issued. That is, when the segmentation section) sends the value i to the back pointer storage section (c), the back pointer B(i) is read out from the back pointer storage section (c). When the segmentation section) receives the value B(1) from the back pointer storage section (c), it issues the same value to the back pointer storage section (c). Therefore, when the voice section detecting section @ detects the end of the voice input, the final value ■ of ) in the frame number counter is supplied to the segmentation section (product), and the segmentation section (product) first stores the value I in the back pointer. Send to the department (c). Thereafter, according to the operation described above, the back pointer storage section (c) stores B(1), B(B(1)), B
(B(B(1))),...

Outputs of 0 will be sequentially obtained. These values are
The frame at the end of the penultimate monosyllable, the frame at the end of the eighth monosyllable, the frame at the end of the fourth monosyllable, etc.
..., and since N(1) is a monosyllable ending in i-frame, if we input this value as it is to the last monosyllable storage section), the recognition results will be obtained in reverse order starting from the last monosyllable. To avoid this, either convert this order to the output of the back pointer storage section (c), or convert the output of the last monosyllable storage section (a). Just go.

FIG. 4 is a flowchart when the functions of the apparatus of the embodiment are realized by software, and will be explained below in conjunction with the operations of each part of the apparatus of the embodiment.

Step 101 is initial value setting, and indicates that D(0)=O in the cumulative distance storage section (to).

Step 102 is the input frame number, 1=1
．． 2.・・・・・・Step 1 for each of ■
Processes 08 to 121 are performed. In reality, this value is set to the number of frames (81).

Step 10B is a part that recognizes each frame i as a vowel, and is an operation performed by the 19-sound recognition unit (IJ) and the vowel recognition result storage unit α4.

Step 105 is the monosyllable n=1.2° to be recognized.
This indicates that the processing of steps 105 to 117 'C is performed for N.

Steps 106, 10'l, and 108 show the operation of the inter-vector distance calculation unit 0Q, and calculate the distance between the corresponding feature vectors of the standard pattern and the input pattern for the part surrounded by the straight line 49 in FIG. demand.

Steps 109 to 116 are the operations of the partial cumulative range meter Minoh, where the range defined by r'AI is 1 equation (i':i),
T):.

This is the part that calculates "e')*7-'(1'@I).

Step 117 shows the operation of the cumulative distance calculating section θ, which calculates the cumulative distance D''(i) when the last term is n and the back pointer Bn(i) at that time.

Steps 118 to 119 are operations of the final monosyllabic vowel frame calculation unit OII, which determines the optimal final monosyllable from the result of step 117, and selects the frame i' that appears to be the vowel at that time. This is the part found using the method.

Step 120 is the part to find the optimal one for the monosyllable k of the vowel stage, since it can be known which vowel the i' frame obtained in step 119 corresponds to by referring to the result of step 108. is the last monosyllable determining part 6! Step 121 performed in p is N(i) for the optimal monosyllable k found in step 120.
= in the last monosyllable memory), D(1) - D↑
As (1), B(()2Bk(
It is shown that they are stored in the back pointer storage unit (c) as i).

Steps 122 to 125 are steps for obtaining monosyllable recognition results in the reverse order from N(i) and B(1) obtained as described above.

As described above, in this embodiment, continuously uttered speech can be converted into a monosyllable string with high precision, and it can be applied to voice typewriters and the like.

In addition, in this embodiment, the vowel part is recognized first, and the results are used to recognize the vowel part, A-dan, I-dan, Tsu-dan, and Ni-dan.

Since it is determined which monosyllable in the o-stage, there is no need to calculate the sum of weights for all patterns, and the amount of calculation is reduced to 115. Furthermore, since a highly accurate method can be adopted as a method for recognizing vowel parts, the recognition rate can be improved.

In addition, in the above embodiment, the consonant plus the vowel
Although the explanation was based on monosyllables with YuruCv as a unit, VC
The present invention can also be applied when V (vowels, ten consonants, and ten vowels) is used as a unit. In this case, as shown in FIG. 5, when j"tjn2 is the boundary between the consonant part and the vowel part of the standard pattern, in the matching path screen, j-j02~
The value obtained by subtracting the cumulative distance of the portion j-j'I to 1 from the cumulative distance of the portion 1 may be evaluated as the cumulative distance of jr+I□jn2. The cumulative distance of the part j'n2 to 1' or jilI can be determined by replacing jn in the case of Cv with jn2 or j71, so all that is required is the following ii change in the above embodiment. That is,
In the above example, the cumulative distance between jnl~1 is j
n as jn1, Ln(i', J) as Ln1(i', j),
All you have to do is replace Dn0(i',j) with Dn10(+'#3), and the cumulative distance between jn2~1 is jo
j′pL′(PI))L′′(Lj)e′:o(t
'sj) may be replaced with DH)(PpJ). Therefore, the cumulative distance of only the consonant part DI″o(b.

1), the length L'(b,'1) in the input shaft direction is as follows.

D;'O(be1)-Dlo(be1)Dbo(L1)
Ln (b, 1) = L (b, 1) - L (b, 1) = g - f
If the partial cumulative distance calculating section α in the embodiment shown in FIG. 2 is replaced with one having the function of performing the above operations, the embodiment can be used as it is when vCV is used as the standard pattern.

In addition, although the present invention has been described as an example of recognizing continuously uttered monosyllables, it can also be used to recognize isolated uttered monosyllables, and the matching result of the entire monosyllable and the consonant The final recognition result can be obtained from both the matching results for only the parts. That is, if it is assumed that the input speech is just one monosyllable, the present invention can be directly applied to the recognition of isolated monosyllable speech.

Further, although an embodiment using a two-stage DP method has been described as a matching method, a matching method using other DP matching methods may be used.

Furthermore, although monosyllabic speech has been described in the embodiment, the present invention can be used for matching any pattern expressed by a sequence of vectors.

As described above, the pattern comparison device of the present invention compares the matching result of the standard pattern and input pattern as a whole by dynamic programming, and the part of the standard pattern obtained in the process of calculating the matching result as a whole. Since the system is configured to identify the standard pattern closest to the input pattern using the matching result of a part of the input pattern, it is possible to reduce the amount of calculation required to identify input patterns that are similar to the silkworm body but have only a slightly different part. This can be done with almost no increase.

[Brief explanation of drawings]

FIG. 1 is a diagram for explaining the present invention in detail, FIG. 2 is a block diagram of an embodiment of a pattern comparison device according to the present invention, FIG. 8 is a diagram showing constraint conditions for matching paths, and FIG. A flowchart of the software that realizes the functions of the embodiment device shown in Fig. 2, and Fig. 5 shows that the standard pattern is VCV.
It is an explanatory diagram in the case of a syllable. 0υ...Feature extraction section, αQ:...
・・Monosyllabic standard pattern storage unit, 0Q・・・・・・・・
・Vector distance calculation unit, θno... Partial cumulative distance calculation unit, 0→... Cumulative distance calculation unit, shi0...・Final monosyllable determining part, kan・
...Last single syllable memory section, slope......cumulative distance memory section, (c)......back pointer storage section, (c)... ...Voice section detection unit, (E) ...Frame number counter, (Foundation) ...
...Segmentation department. Agent Yoshihiro Morimoto Figure 1 Figure 3 Figure 5 Senate yl, <

Claims (1)

[Claims] 1. The human input signal is converted into a feature vector al (i=1.2...
.... ■) A feature extraction means for converting into a time series, and a standard pattern l1=Jb2...I)Jn(n==1, 2.-...
・−. a standard pattern storage means for storing N); and a matching calculation means for performing a matching result using a dynamic 81 stroke method between each of the standard patterns and the input cover pattern represented by the series of feature vectors; Based on the overall matching result of the standard pattern and the input cover pattern, and the matching result of the part of each standard pattern specified in advance at the time of registration of each standard pattern and the corresponding part of the input cover pattern, find the one closest to the input cover pattern. A pattern comparison device comprising: means for identifying a standard pattern.