JPS6358400A - Continuous word voice recognition equipment - Google Patents

Abstract

(57) [Abstract] 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 word speech recognition device that realizes recognition of a string of words continuously uttered by a specific speaker.

(Prior Art) Conventionally, this type of continuous word speech recognition device (hereinafter abbreviated as a recognition device) first requires a user to separate the words to be recognized in advance and utter them individually (hereinafter referred to as isolated words). , the speech pattern for each word is stored in the recognition device as a standard pattern (the recording operation is called standard pattern registration),
Next, a comparison operation (pattern matching) is performed between the human-generated continuous 41-word speech (hereinafter referred to as a human-generated pattern) with each standard pattern, and the degree of agreement (similarity) between the two is examined. , a method was used in which the combination of standard patterns that yielded the maximum match was determined, and the combination was determined to belong to the same word. Dynamic programming (hereinafter referred to as dynamic programming) is a method to implement this method efficiently and with good granularity.
Recognition technology using DP (abbreviated as DP) (patent application 1984-1:1)
2003 and l: ] 22004 (hereinafter referred to as cited documents) are known. This cited document describes the operating principle of a recognition device using the above pattern matching method. The outline of this principle is as follows. For an input pattern in which several consecutive words are connected, the pattern obtained by connecting several standard patterns in any permutation is considered to be the standard human pattern, and is matched with the entire large Kapatan. . Recognition is performed by determining the number and permutation combination of standard patterns that maximizes the similarity obtained as a result. In reality, the above-mentioned maximization is divided into word-by-word maximization processing and overall-level maximization processing, and each maximization process is executed using DP, thereby reducing the amount of processing and making it more practical. Achieving speed. The method of citing cited documents described above is considered to be the most effective recognition method to date.

[Problem to be solved by the invention] For example, when the number "3", rSanJ, is uttered, the frequency structure of rnJ of rSanJ changes greatly depending on the number that follows, or the energy of "n" becomes extremely large. Sometimes it becomes low (devoiced). However, rsanJ's r Sa
Compared to "n", J has less frequency fluctuation and a stable high energy value. However, in the conventional apparatus, since 'SaJ and rnJ are matched with the same weight, the matching effect is not sufficient, resulting in erroneous recognition or inability to recognize (the matching value is extremely poor).

The conventional recognition device described above performs a comparison operation between the standard pattern and the large pattern without using the characteristics of the standard pattern and the characteristics of the large pattern, so when put into practical use, misrecognition may occur due to various factors. There is a drawback.

[Means for solving problems]

The continuous Qi-word speech recognition device of the present invention retains speech patterns uttered by dividing into single 3Δ as isolated word patterns, converts consecutive Q'-word cross patterns uttered continuously, and rearranges the isolated word patterns in all permutations. , perform a comparison operation between the connected isolated song word pattern and the continuous word pattern, check the degree of matching between the two, and determine the combination of isolated word patterns that yields the maximum match. continuous! ...In a continuous four-speech recognition device that recognizes word speech, the characteristics of time-feature information possessed by each isolated word pattern to be combined are
During the orthography comparison operation, the weight function storage unit stores weights to be emphasized and time-feature information held by the continuous word pattern to be recognized. and a weighting function generating section that generates weights that emphasize the characteristics of.

That is, in the present invention, the characteristic frequency part of each word to be recognized is checked in advance, and the result is prepared as a weighting function for each word in the same way as a standard pattern. By similarly investigating and preparing a weight function, and obtaining a matching result according to the weight function described above when matching a large Kapatan with an arbitrary standard pattern, it is possible to improve the matching accuracy on a word-by-word basis and improve the continuous Q1 word recognition level. This improves the performance of the overall matching.

〔Example〕

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the continuous nine-word and voice recognition device of the present invention.

In this embodiment, a human voice 111 from the microphone 10 is used.
an analysis unit 11 that performs frequency analysis of the code; a large pattern buffer 12 that temporarily stores the output of the analysis unit 11 as a time series of feature vectors; and a standard pattern storage unit 13 that stores the isolated A word pattern bj as a standard pattern. Standard pattern storage section 1
Weight W that emphasizes the characteristics of the isolated rIi word cross pattern in 3
(j); a weighting function generating unit 23 that generates an IR model W(i) that emphasizes the features of the continuous word pattern held in the manual pattern buffer 12; The weight W(j) of the standard pattern storage unit 1
The weighting function generating unit 23
The weight W(i) from W(i) is added to the continuous rlL '5f) pattern from the input pattern buffer 12, and the added isolated word pattern and the continuous word pattern are compared to obtain a partial pattern of the continuous word pattern. A first matching unit 14 performs a partial determination of the partial similarity of each isolated word pattern to which isolated word pattern the partial pattern is determined to be, and stores the partial similarity in correspondence with the start and end of the partial pattern. The data of the partial similarity storage section 15 that stores partial judgments, the partial judgment result storage section 16 that stores partial judgments, the recurrence formula value storage section 18, the partial similarity storage section 15, and the partial judgment result storage section 16 is inputted. , based on the value input from the number of words setting terminal 1fk, the recurrence formula value storage unit 18 sets the maximum value of the sum of the partial similarities of the continuous word pattern up to the first time point as the first recurrence formula value.
The maximum value of the sum of the recurrence formula value and the partial similarity from the first time point to the second time point is stored in the recurrence formula value storage unit 18 as the recurrence formula value of 'f,2. The first time point is memorized, the first time point is changed sequentially, the record operation is repeated, the first time point is output as a temporary dividing point, and the partial judgment result from the first time point to the second time point is obtained. A second matching unit 17 that outputs a provisional determination result, a provisional division point storage unit 19 that stores the provisional division point output from the second matching unit 17, and a provisional determination result storage unit 20 that stores the provisional determination result. and temporary segmentation point storage section j9
The judgment unit 21 determines the segmentation point and word name of each word by referring to the tentative segmentation point in the table and the tentative judgment result in the tentative f'lJ determination result storage unit 20.

Next, the operating principle of this embodiment will be explained.

The operating principle executed by the apparatus of this embodiment can be expressed mathematically as follows. The audio signal manually input by the microphone IO is analyzed by the analysis unit 1 and can be expressed as a time series pattern A of a multidimensional feature vector a1 representing the frequency structure and the like.

k=al+a2.・-, aH, ・-, aj −
---(]) On the other hand, each word uttered singly (isolated 1...
The pattern B is also analyzed in the same way and can be expressed as a time series pattern B.

B” = br, b2.---, b'7, ・=
-----(2) where n is a subscript for identifying words, and k is the number of words included in consecutive words, the maximum problem T =
(m(k) (S(＾9口(1) ■ a n (
2) ■・・・■Bn(” ) ) )
----Calculate (3), find the optimal parameter (word name) n (k) = n (k) (k = 1.2.-, K) and simultaneously find the division point 1 (k) . Here, ■ is an operation f representing the connection of patterns. For example, Bn■am is B" Ctl "" = br+ br, """ 1)
Ul, b7', b';'.

b″J1 ----(4) Calculating the maximization of equation (3) using the brute force method regarding k and n(k) requires a huge amount of calculation, but as in the cited document, equation (3) Practical processing speed is made possible by dividing the maximization calculation into two stages: processing in units of lli words and processing as a whole.In other words, 1 = 11 of the human pattern A expressed by the formula (+) Therefore, a partial pattern A(1, m) is defined as a partial interval up to i=m.

A(fl, m)=a,,1. x+2. ,,,,
B.

Hereinafter, l will be referred to as the starting point and m will be referred to as the ending point. Now there are K-1 segmentation points fi (1), l (
2) , "1(k)..., ff1(K-1) are provided, and 1<fi(+)<ρ(2)<...<1 (
Assuming that K-1) < ll(K) =■, the large Kapatan A is divided into partial patterns.

A = A (+, f (1)) ΦA (j2 (+), j
2 (2))■・・・■A(,12(k-1)!! (
k) ) ■-...-0+8(R(K-1), I
), On the other hand, if we define the time axis normalized similarity between patterns, the similarity S (8, B) has the following properties regarding the connection decomposition of patterns.

S(A・8"■8pa)"T°゛(耶)su・,x,3nido(T Substitute equation (5) into equation (3), and then repeatedly apply the relationship of equation (6). When rearranged, it becomes, and the maximization problem of equation (7) can be decomposed and calculated as follows.

[11 Similar Jff S (8 (42, m), [1
(8) is calculated for all partial intervals A(n, +n) where Q<m and the isolated, Qt-packed pattern Bτ1.

[21 Partial similarity S (Q, m) = max (S (8(Qlm
), Rn )) Partial judgment result N (11, m)=arg max (S(8<l
, m), B'') 3 is calculated and stored in the table.Here, the symbol arg-max [·1 means that a, calculate the variable n that serves the maximum of [].

−−−(+1) Calculate the maximum problem and find the optimal parameters (+7° point) fl(k) = l(k), k= 1.2.・
..., find K.

The maximum problem of equation (11) can be calculated using the following recurrence formula.

Initial value To(fl,)=O,R=1,2. ...,
■.

k=1.2. ..., the recurrence formula m=l, 2. −． 1, k=I, 2. ...,
The hypothetical equinox L' (m) = arg max (Tk-' (M)
+S (,Q, m))! −−−−(+3) Temporary judgment result N'(111) = N (L is m), m)
----(14) (+2), (+3), Calculation of the (th) formula is performed in the direction of increasing m.

After completing the above-42 JA filling, the formula (13) is
), the segmentation point fl (x) is determined as follows.

ffi (k) = L”' Q (k+1) ),
(k = 1.2..., -1), the hypothetical equinox L
Accordingly, the judgment result n (k) is calculated as n (k) = N '(fl (k)), (k =
1.2. -, K) ----(15).

By the following operations, the segmentation point and C1i word name of each word constituting the continuous single word are dan(k), (k=+, 2...,
ni-1).

n (k), (k = I, 2..., k).

Next, the operation of this embodiment will be explained. The voices uttered in units of 11Δ are input to the analysis unit 11 as an isolated -1iΔ pattern from the microphone 1o.

The solitary L last word pattern frequency-analyzed by the analysis unit 11 is (2
) To generate the time series of the feature vector b shown by the formula 〒
The symbol i+ (is stored in the pattern storage unit 13 as a quasi-pattern lln through the human cover pattern buffer 12.The continuously uttered 7g voice passes through the microphone 1o and the analysis unit 11, and has the characteristic shown by equation (1). The time series of vector ai is 44
- is stored in the input pattern buffer 12 as a continuous 'l' +EΔ pattern A. Also, (weight function wnU expressing the characteristics of standard pattern 11 Tl of urchin words) (j
・1.2. :II,...1. J) Kaichi Miseki 2 friends 8 members 22 members remember 11 members. Continuous m1: Δ pattern (man-powered pattern) Emphasize the features (each ?11-word) of A that have significant feature components 1-ru rl'i, mi ?
fi Wn(i) (i=1.2.3....,
I) 'ti + l The first function generated by the function generator 23
The data is sent to the matching section 14.

FIG. 2 is a diagram showing an example of the slope function Wn(,i).

This figure shows the standard pattern of J to rO58K [for lI+! This is an example of an n-shape function, which lowers the matching effect of the Iυff-1 consonant junction, and reduces the matching effect of the Iυff-1 consonant junction,
The weighting function W(j) is determined so as to enhance the matching effect of the break'Q:'r). In this example, in this way, the characteristics are determined only in the j direction of l[mifunction WnU) or time'pHl, but the characteristics are statistically adjusted not only in the direction of time 111 kb but also in the frequency axis direction f. However, the method for determining the market function Wn(j, f) is not limited to the example shown in FIG. 2.

゛) Margin below J Table 1 Table 1 is an example of a feature-weight value table that is referred to when the weighting function generation unit 23 generates the weighting function Wn(i).

For example, if the feature of the i-th frame of the human pattern is the fricative of S, the weighting function W(i) = 1.9 is selected when the matching target is Osaka, and the weighting function W(i) = 1.9 is selected when the matching target is Osaka.
(i) Select either -1 or 0 and calculate the similarity according to the recurrence formula of equation (111). Also, the gradient function Wn(i)
Similarly to the weighting function Wn(j), it is also possible to determine the weighting function Wn(i, f) by statistically examining the characteristics in the frequency axis direction f. The first matching unit 14 calculates the recurrence formula defined by the following equation for each isolated word pattern Bn and partial pattern A(ffi, m) of pattern A, each time a large Kapatan vector al11 is manually generated using equation (8). Calculate the similarity S of . That is, the initial condition g(!, j”) = S
(am, bp) i = m = Oi mim −−−− (+7) Recurrence formula % formula % (18) Calculate the recurrence formula as j = j”, j” −I,
jn-2゜..., 1, and the similarity S (A(Q, m), B'') = gC(1+1
．． 1)-(20) m-J”-r≦2≦m-J”
+r --- (21) Calculated in the range.

As a result of the above method, the partial similarity S(1,+n) shown by equation (9) and the partial judgment result N(fi, m) shown by equation (10) are respectively stored in the partial similarity storage unit 1.
51 is output to the partial determination result storage section 16. The second matching unit 17 reads out the partial similarity S(f, m) from the partial similarity storage unit 15, and at the same time reads out the partial similarity S(f, m) from the recurrence formula value storage unit 15.
8, the recurrence formula value Tkl(2) of equation (12) where l<m
, with k being constant, the recurrence formula value T'(m
) is calculated and output to the recurrence formula value storage section 18. Similarly, the temporary equinox L'(m) is calculated using equation (13) and output to the temporary equinox storage section 19. The provisional judgment result N'(m) is (14
) is calculated by referring to the partial determination result N(42, m) and the temporary equinox L′(m), and is output to the temporary determination result storage unit 20.

In the matching unit 17 of 7J, 2, the above operation is started from =1 based on the value input from the word count setting terminal %ilk, and k
Execute while increasing k sequentially until =K.

In the device configured as described above, the value L'(m) regarding the segmentation point can be obtained by manually performing the above-mentioned operations sequentially from the start point a1 to the end point a of the known continuous word pattern A of the word series. The value N which determines the word name (m) is for all m = (+, 2.-, I) +' = (
1, 2, ..., K). In the determination unit 21, the temporary equinox L′(m
) and the provisional determination result N′(m) in the provisional determination result storage unit 20
1 (k-]), j2 (k-
2).

..., +2(1) is determined. Similarly, each word name n (k-1), n (k-2) according to equation (16)
,・”, n (]) is determined.

Although practical examples of the present invention have been described above, these descriptions do not limit the scope of the present invention. For example, in this specification, the operation has been described based on the degree of similarity, but similar processing is possible using a measure in which the magnitude relationship is reversed, such as distance. Furthermore, although the portion to be extracted has been described as a word, a word or phrase consisting of a plurality of a-clauses can also be processed in the same way. Furthermore, although the degree of similarity between the input speech pattern and the standard pattern has been explained using dynamic programming, the present invention is not limited to dynamic programming.

We have explained the method for recognizing continuous words from north to north, but under the constraint of equation (19), the weighting function Wn(
i), by executing the similarity calculation of equation (20) with Wn(j), weights are applied to the vicinity of the maximum values of the weighting functions Wn(i) and Wn(j), that is, to the frequency interval that characterizes the isolated word. The performance of matching at the level of isolated words can be improved because pattern matching is performed by attaching words to words, and the matching effect of relatively unstable parts at the intersection of vowel parts and consonants (vowel parts) can be reduced. This makes it possible to achieve continuous word recognition with high accuracy.

〔Effect of the invention〕

As explained above, when recognizing a continuous word pattern as a connected isolated word pattern, the present invention applies weights to emphasize features to the time-feature + l11 information of the single-word pattern and the continuous word pattern. +1 by adding and comparing! -Word? This has the effect of increasing the matching granularity at the forest level and improving the overall matching performance at the continuous word recognition level.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the continuous word speech recognition device of the present invention, and FIG. 2 is a weighting function W (
FIG. 6 is a diagram showing an example of j). 10...Microphone, 11...Analysis department, 1
2... Human pattern buffer, 13... Standard pattern storage section, 14... First matching section, 15... Partial similarity storage section, 16... Partial judgment result storage section, 17... - Second matching section, 18... Recurrence equation value storage section, 19... Temporary equinox storage section, 20... Temporary judgment result storage section, 21... Judgment section, 22...
Weighting function storage section, 23...Weighting function generation section, Wk.
・Word count setting terminal.

Claims (1)

[Claims] A speech pattern uttered by dividing each word is held as an isolated word pattern, and the isolated word pattern is connected in any permutation to a continuous word pattern uttered continuously. Continuous word speech recognition that performs a comparison operation between isolated word patterns and continuous word patterns, checks the degree of agreement between the two, determines the combination of isolated word patterns that yields the maximum match, and recognizes continuous word speech. The apparatus includes: a weighting function storage unit that stores weights that emphasize, during the comparison operation, the characteristics of the time-feature information of each of the isolated word patterns to be combined; and the time-features of the continuous word patterns to be recognized. A continuous word speech recognition device comprising: a weighting function generator that generates a weight that emphasizes the characteristics of time-feature information of the continuous word pattern during the comparison operation according to the information.