JPH0193798A - Division labelling apparatus - Google Patents

Abstract

PURPOSE: To improve the accuracy of pattern matching and the accuracy of divided labeling by statistically processing variation in the section length of divided sections at the time of pattern matching. CONSTITUTION: A file device 1 stores K word voice time sequence patterns, a file device 2 stores K groups of vector label strings and section label strings and a file device 7 stores a division result obtained by m-th repeat. An initial division part 3 reads out a section length label string corresponding to the time sequence pattern of each vector stored in the device 1 from the device 2, executes time division so that each section length is proportional to the weight of a label sort and sends the result to the device 7. A statistic part 4 reads out the newest division result found out by the initial division part 3 or the best division part 6 from the device 7 and finds out the average and dispersion of cepstrum coefficients as the statistic value of vectors and finds out the average and dispersion as the statistical value of section length in each label and stores the obtained results in a file device 5. The best division is repeated in accordance with the contents of the device 5.

Description

[Detailed Description of the Invention] [Field of Industrial Application] This invention is an improvement of a device that divides a time series pattern of a vector represented by a voice button according to a given label string and labels these divided sections. Regarding.

[Conventional technology]

Figure 2 is from March 1986. Patent application filed on US day 1986-0
64952, "Divided labeling device", is a functional block diagram of a conventional device of this kind. In the figure,
11) is a file device for storing time-series patterns of vectors; (2) is a file device for storing label strings to be attached to time-series patterns of vectors in file device+11;

(31 is the initial division section, (41 is the statistics section, (5) is a file device for storing statistics for each label output by the statistics section (4), (6) is the best division section, (7) is This is a file device for storing division results.

Here, the initial dividing section (3) or the best dividing section (6) is.

Time series pattern of each vector in file device III.

It divides the file device (2) according to the corresponding label string, and outputs the division results to the file device (7). On the other hand, the statistics section (4) calculates statistics such as the average or variance of the vectors included in the nine divided sections from these division results for each label, and outputs them to the file device (5).

By the way, when dividing a certain time series pattern, the best dividing unit (6) reads the statistics of vectors for each label from the file device (5) in the order of the corresponding label sequence, and uses the sequence in which these are arranged as a template. As such, we use a slam matching technique based on dynamic programming to match this template with the time series pattern to be divided.

The likelihood of matching between the baton and the template is maximum 6
The time division state of the time series pattern is stored in the file device (7) as the best division result.

The best division result at a certain point in time is used again by the statistics unit (4) to obtain the latest statistics for each label, and these are used again for the best division. By repeating the best division and the process of obtaining statistics for each label, a successively improved division result is obtained.

With this kind of device that works on this principle.

In order to perform accurate segmentation labeling, the content of the statistics used in the best segmentation unit and the technique of matching between the model version and one time series pattern based on the statistics are important.

[Problem that the invention seeks to solve]

As a technique for more accurately matching the template to the time series pattern of vectors, the cited document (Transactions of the Institute of Electronics and Communication Engineers (A), Volume 69-A, No. 2, pp 261-270)
As described in , when dividing a time series pattern of vectors, there is a method of determining upper and lower limits for the interval length of each divided interval and finding the best consistency condition under the limits of the allowable range of these interval lengths. be.

However, in this method, the allowable range of the interval length of the divided interval was determined and given by nine people based on the experience of observing a limited amount of data, and what was given once was given a constant width from then on, so the amount of data This method does not take into account the true distribution of the section lengths of the 9-section sections, which is thought to become clear only when the number of sections is increased (there is no guarantee that correct matching can be achieved).

Further, there is a problem in that when there is a large variation in the section length that deviates from the allowable range, an incorrect matching state is given, resulting in a decrease in the accuracy of one division.

This invention was made to solve such drawbacks,
An object of the present invention is to provide an apparatus that improves the accuracy of one-segment labeling by statistically capturing fluctuations in section lengths, which are a problem during baton matching.

[Means for solving problems]

The segmented labeling device according to the present invention includes means for determining, in the statistics section, not only vector statistics for each label but also statistics on the section length of segmented sections for each label from the latest segmented labeling state. Along with using.

Even in the best segmentation section, a method is used to perform slam matching using both the statistics of vectors for each label and the statistics of the section length of segmented sections for each label to find the best segmentation labeling state. This is what I did.

[Function] The divided labeling device according to the present invention prevents fluctuations in the length of one divided section, which is a problem when slamming.

Since it can be grasped statistically, the accuracy of baton matching is improved, and as a result, an apparatus with improved accuracy of 9-division labeling is realized.

〔Example〕

FIG. 1 is a functional block diagram showing an embodiment of the present invention. Each functional block is realized by a program running on a general-purpose minicomputer.

In the figure, 11) is a file device that stores time-series patterns of (feature) vectors (specifically, 16-dimensional cepstral coefficient vectors) obtained by analyzing word audio data, and +21 is a file device that stores label sequences. (3) is the initial division section, (14) is the statistics section, (5) is the file device for storing statistics for each label, and (6) is the file device for storing statistics for each label.
) is the best division unit, and (7) is a file device for storing the division results.

In this embodiment, the time-series pattern of word sounds in the file device Ill is labeled into one segment according to the label string in the file device (2).

Hereinafter, as a label string, the component labels are a label component for identifying vector statistics (hereinafter referred to as a vector label) and a label component for identifying a 2-interval length statistical part (hereinafter referred to as a 1-interval length label). ).

Also, as vector statistics, the mean and variance of cepstral coefficients, and as statistics for one interval length.

The case where the average and variance are used will be explained.

Furthermore, the number of repetitions of one division (hereinafter expressed as 2m) is defined so that 9m=1 for the result of the initial division.

Furthermore, it is assumed that the file device +11 stores time-series patterns of the word sounds of KI[Fj.

(((C(k, i, n), O≦n≦+s), I
≦i≦I(k)), and 1≦≦K). (Here, the symbol (X) represents the sequence X, and this notation will be used hereinafter without further notice.

C(k, i, n) is the third-order cepthrum coefficient in the first frame obtained by analyzing the th word,
[1 is the length of one time series pattern (number of analysis frames). ) Furthermore, it is assumed that the file device (2) stores a set of vector label strings and a section length label string, and these are defined as the vector label string.

(il?s (k, j), 1≦j≦J(k)), 1≦
≦K) (where l R3(k, j) represents a vector label component attached to the j-th divided section of the th-th time series pattern, and J(k) represents the length of the label string. )and,
Also, the one section length label string is:

(tRt (k, D, l≦j≦J(k)), *≦to≦
K) (However, Rt (k, j) represents the section chief label component attached to the j-th division interval of the th-th time series pattern.) 1
The sets of different labels included in these label strings are called a vector label set and an interval length label set, respectively. Vector label set is.

(QS (qs), l≦qs≦NQs) (However,
NQs represents the total number of different vector labels. ), and the 9 interval length label set is.

(Qt (qt), t≦qt≦NQt) (However, N
Qt represents the total number of different interval length labels. ).

Furthermore, the division result at the m-th repetition is stored in the file device (7).

11e (m, k j), 1≦j≦J(k)−H),
1≦ and ≦K). (Here, e (m
, ni, j) is.

It represents the frame number at the start end of the j-th division section when dividing the time-series pattern of the th vector in the m-th repetition, and satisfies the first-order equation.

=e (m, ni, 1)≦e (m, ni, 2)≦...
・≦e (m, ni, J (published by kl) = I(k)-z...
...11) As a result, the frame range of the j-th divided section of the th time-series pattern in the m-th repetition is as follows.

e(m, ni, j)≦ice (m, ni, j +1) (
i is a frame number). ) With the above preparations in place, the operation of each part will be explained.

The initial dividing unit (3) reads the corresponding interval length label string from the file device (2 ribs) for the time series pattern of each vector in the file device (1), and calculates the interval length of each divided interval as follows. Nine time-series bumps are time-divided in proportion to the weight predetermined according to the type of section length label attached to the section, and one division result is output to the file device (7). The initial division result output to the file device +71, ie, the first division.

(m=1) division result in repetition.

1(e(z, ni, j), + ≦j≦J(k)+1
), l ≦ to ≦K) (where 1≦j≦J(k)
+ 1) ・・・・・・(2) (However, the symbol (X
) shall represent the largest integer not exceeding X only in the second equation. ) is required. (Here, W(
Rt(k,j)) is.

It represents the weight given to the interval length label Rt (k,j). ) Also, the statistics section (4) reads the latest division results obtained by the initial division section (3) or the best division section (6) from the file device (7), and calculates the cepstral coefficients as statistics regarding the vector. The mean and variance are obtained for each label as statistics for the length of 9 sections, and these statistics for each label are stored in the file device (5). Here, the statistics for each label are given as 9th order in the m-th repetition. That is, the qs-th vector label Qs (qs
) for the average cepstral coefficient (hereinafter Hs (
m, qs, n)) and variance (hereinafter, vS(m,
qS, n) are respectively.

Hs (m, qs, n) = K J (kl Σ Σ ΣC(k, i, n)k
=l j =1. Rs(k,j)=Qs(qs) e
(m, ni, j)≦i(e(mkj+1)K J
lkl Σ Σ (e(m, to, j+1)−e(m, to,
j))k=+j=t,Rs(k,j)=
Qs (qs) (1≦qs≦NQs, 0≦n≦1
5) -・-+a+ and.

Vs (m, qs, n) = (However, I≦qs≦NQs, Q≦n≦15) ・-
---It is given by (4). On the other hand, the average interval length for the qt-th interval length label Qt (qt) of the set of 1 interval length labels (hereinafter referred to as Ht (m, qt))
and variance (hereinafter referred to as vt (m, qt >), respectively).

Σ Σ 1 = l j = +, Rt (k, j) = Qt (qt) (however, 1≦qt≦NQt) ...
...(5) and.

Vt (m qt ) = Σ Σ 1 = 1 j = 1. Rt (k, j) = Qt (qt
) (However, 1≦qt≦NQt) ・
...It is given by (6).

Furthermore, the best dividing unit (6) rearranges the label-specific statistics read from the file device (5) for each time series pattern in the file device +11 according to the corresponding label string in the file device (2). By doing so, a template is constructed between one time series and the template.

Based on dynamic programming, find the partitioning state that maximizes the likelihood of matching, and use this as the best partitioning result.

Output to file device (7). Here, it is determined based on the statistics for each label in the m-th repetition.

The likelihood of the best match (hereafter denoted S ” (m, k)) is:

S ” (m, k) = , and the best division state at this time (hereinafter +
te” (m*kej)+ 1≦j≦Jlkl+
1) is written as ).

(e ” (m, ni, j), 1≦j≦J(k)+1
)= is defined. Here, S (m, k) in the seventh and eighth equations is calculated based on the statistics for each label obtained in the m-th iteration.
e (m, to, j), is the likelihood of matching determined for 1≦j≦J(kl-z), and is the sum of the likelihood of matching for the vector label string and the likelihood of matching for the 1-interval length label string. As.

S (m, k) = Ls (m, k) + Lt (m, k
) −Wt (+≦≦K) (9) Defined as follows. Here, Wt is a weighting coefficient assigned to the likelihood of matching the interval length label string.

Ls (m, k) and Lt (m, t) are, respectively,
These are the likelihood of matching for a vector label string and the likelihood of matching for a one-section length label string.

The likelihood of matching for a vector label sequence is.

Ls (m, k)=j=1. e(m,ni,j)p! e (m, ni, j+ t
)・・・・・・(1(1 (However, + Is (e (m, ni, j), e(m, ni,
j+1), qs) corresponds to the qs-th vector label Qs(qs) of the vector label set. is the likelihood of the jth interval.

ls (e (m, ni, j), e (m, ni, j-)
−1), qs)=. ) Also, the likelihood of matching for a one-section length label string is:

Lt (m, k) = ......-3 (However, It (e (m, to, j), e (m
, j+1), qt) corresponds to the qt-th interval length label Qt (qt) of the one interval length label set. jth
with the likelihood of the interval.

It (x, y, qt) = It is expressed as

Now, the interval division method 1e (m, to,
D.

The maximization regarding 1≦j≦J(kl−Ml) becomes a kind of combinatorial optimization problem, and is a 9th order recurrence formula based on the principle of dynamic programming.

j=2. a, ..., J (about kl-z.

And 1=23. ..., about I(k) + +.

Gs (i, j) = Gs (i-bi, j-+) + 1s
(i-?, i, Qs-' (Rs(k,j)))G
t(i,j)=Gt(ir”,j-1)+1s(i-
τ'', i, Qt'(Rt(k,j)))Ns(i,
j)=Ns(i-r", j-1)+u(i-ge+')B
(i, j)=i-ge...
...a4 (here + Gs (i, j), Gt
(ij)e Ns(i, j), B(i, j) are state variables in dynamic programming and have the following meanings.

Gs(i,j): Cumulative value of matching likelihood for vector labels.

Gt(i,j): Cumulative value of matching likelihood for interval length labels.

Ns(i,j): cumulative number of matching likelihoods for vector labels.

B(i,j): Pock pointer.

still.

Qs' (Rs): A function that returns the vector label number qs that satisfies Rs=Qs (qs).

Qt (Rt): A relationship that returns the section length label number qt that satisfies Rt=Qt (qt).

1s (x, y, q): Likelihood of the interval for the vector label (Equation 11) It (x, y, q): Likelihood of the interval for the interval length label (Equation 13) u(x, y): This is a function that satisfies the following equation. ) as the initial condition, i.e.

Gs(1,j)=O(t≦j≦J fk) + 1)G
t (1,0= o (+≦j≦J fk) + 1
)Nt(+, D=o (+≦j≦J(k)+1)G
s (i, 1) =oo (2≦゛i≦I(k)+
°+)Gt (i, 1) = Flash (2≦i≦I(k)
+1) Nt (i, 1) - 1 (2≦i≦I(k)+l)
Solve under ・・・・・・−ms, the likelihood of the best match S
′″(m, k).

S” (m, k) = Ns (I(k)+ +, J(k)+ +)
It can be solved by putting J(k)...ae, and it is the best division state.

(e” (m, ni, j) + t≦j≦J(k)+l)
is found by tracing the back pointer in the opposite direction. Namely, first.

e” (m, to j(k)+t)=I(k)+1
-...-6n and place 1st order.

j = J(k), J[kl-+,..., 3.2.
About l.

e”(m, ni, j)=B(e”(m, ni, j+1), j
+1) =-...UFj. This result, ie.

The division result at the m+1st iteration.

((e(m+l, ni, j), 1≦j≦J [k)+
1), l≦ and ≦K) in the file device (7),
e (m+l, j)=e” (m.

k, j) and output.

It was requested in this way. The division result at the m+1st iteration is used again to calculate the statistics for each label at the m+th iteration, and this statistic is further used to determine the best division at the m+2nd iteration of the first order. It is also used to obtain results.

By the way, from the explanation above in 9, what is the state of the latest division labeling determined by the best division unit (6)?

Maximizes the likelihood of matching S(m, k) defined by the ninth equation, and the likelihood of matching for the interval length label string included as the second term of the ninth equation, Lt (
m, k).

Furthermore, the likelihood of matching for this interval length label sequence.

Lt (m, k) is the statistic for each label of the interval length as shown in equations 12 and 13, that is, its average (Ht (m, q
t), 1≦qt≦NQt), and variance (Vt (m
qt), and 1≦qt≦NQt). According to this embodiment, it is possible to perform the matching that captures the statistical fluctuations of nine section lengths.

Since the section length statistics used at this time are always obtained from the latest division results, it can be said that this embodiment performs slam matching using statistics close to the true distribution of the section length data.

〔Effect of the invention〕

As explained above, in the segmented labeling device according to the present invention, in addition to the statistics for each label of vectors, means for obtaining the statistics for each label of the section length of 9 segmented sections are added to the statistical means in conventional devices of this type. In addition, since the best segmentation method uses a means to determine the best segment labeling state based on the label-specific statistics of these vectors and the label-specific statistics of the interval length of one segment, it is possible to obtain the best segmentation. It is possible to statistically capture the fluctuations in the length of the 9-division section during the baton matching, and the accuracy of the baton matching is improved.

As a result, this has the effect of improving the accuracy of one-segment labeling.

[Brief explanation of the drawing]

FIG. 1 is a functional block diagram of a divided labeling apparatus according to an embodiment of the present invention, and FIG. 2 is a functional block diagram of a conventional apparatus. In the figure, (1) is a file device that stores time-series patterns of vectors. (21 is the file device that stores the label string, (3) is the initial division section, (41 is the statistics section, (5) is the file device that stores the statistics for each label, (6) is the best division section, (7 )
is a file device that stores the division results. Note that the same or corresponding parts in FIG. 1 are indicated with the same reference numerals. Figure 1 7: Core 4 Discipline Figure 2 Procedural Amendment (Voluntary)

Claims (2)

[Claims] (1) A device that divides a time-series pattern of a vector according to a pair of label sequences and labels the components of this label sequence on these divided intervals, and is an initial device for determining the initial state of division labeling. A dividing means, a statistical means for calculating statistics of vectors for each label according to the latest divided labeling state, and optimal dividing labeling based on the statistics of vectors for each label obtained by the statistical means. starting from the initial dividing state determined by the initial dividing means, and determining the best dividing state based on the latest vector statistics for each label determined by the statistical means. In the segmented labeling device, the best segmented labeling state determined by the segmenting means is replaced with the latest segmented labeling state, and the statistical means calculates the statistics of vectors for each label from the latest segmented labeling state. and the statistics of the interval lengths of the division intervals for each label, and the best division means is a means for calculating the statistics of the vectors for each label and the statistics of the interval length of the division intervals for each label. A divided labeling device characterized in that it is a means for determining the best divided labeling state based on the following. (2) The component labels of the label string consist of a label component for identifying vector statistics and a label component for identifying interval length statistics, and the two components do not necessarily match. A divided labeling apparatus according to claim 1.