JP2836968B2 - Signal analyzer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a signal analysis device such as a speech recognition device.

BACKGROUND ART HMM (Hidden Markov Mod) is a method for performing speech recognition.
A method using el) and a method using DP matching are known. Both are widely used as basic techniques for speech recognition, but in these methods, how to reduce the amount of computation without deteriorating performance is important for realizing large vocabulary and continuous speech recognition. One of the problems. As a solution to this problem, one using vector quantization has already been proposed. The present invention relates to this improvement. So, before going into the subject, first, HMM and D
A general description of P matching and how the vector quantization technique is used will be described.

The HMM is considered to be one of models that generate a time-series signal according to a certain stochastic property. Speech recognition using an HMM is performed by an HMM corresponding to a recognition unit of a word, a syllable, a phoneme, or the like to be recognized (hereinafter referred to as a word) r (= 1,..., R).
r, a vector series Y = (y ₁ , y ₂ ,..., y _T )
When (y _t : a vector observed at time t) is observed, the degree of occurrence of Y is calculated from each of the HMMs r, and the word corresponding to the HMM with the maximum degree is used as the recognition result.

FIG. 1 shows an example of the HMM.は indicates the state of the system to be modeled by the HMM, → indicates the direction of state transition, and q ₁ indicates state i. It is assumed that the transition from the state i to the state j occurs with a probability a _ij . If only states and their transition probabilities are defined, they are called Markov chains.
Further, it is assumed that a vector is generated with each state transition, and the degree ω _ij (y) at which the vector y is generated is defined according to the state transition q _i → q _j . If y occurs with the state, not with the state transition,
_ij (y) = ω _ij (y) = ω _i (y) or ω _ij (y)
= Ω _ij (y) = ω _j (y) in many cases. In the present application, a description will be given assuming that y occurs according to the state. Here, the structure, state transition probability, and vector occurrence probability of the HMM are determined so that the behavior of the object to be modeled by the HMM (a voice pattern such as a word when used for voice recognition) can be described as faithfully as possible. Is done. FIG. 2 shows an example of the configuration of an HMM often used in speech recognition.

When a certain HMM is defined, the degree L (Y | λ) of the observation vector sequence Y generated from a certain model (named λ) can be calculated as follows.

Here, X = (x ₁ , x ₂ ,..., X _{T + 1} ) is a state sequence, and π _i is t
= 1 and the probability of state i. In this model, x _t ∈
{1,2, ..., J, J + 1}, where _{xT + 1} = J + 1 is the final state. In the final state, it is assumed that only a transition to that occurs, and that no vector is generated there.

HMMs are broadly divided into continuous and discrete types. The continuous type is a continuous function of y such as ω _i (Y) probability density function,
generation degree of y _t is given as the value of .omega.i (y) when the y = y _t. A parameter that defines ω _i (y) is defined for each state i, and the degree of occurrence of y _{t in} state i is calculated by substituting y _t into ω _i (y). For example,
If ω _i (y) is given by a multidimensional normal distribution, Where the parameters of the function defined by state i are μ _i and Σ _i .

In the discrete type, an occurrence probability b _{im of a} label m {1,2, ..., M} to which y _t is to be transformed by vector quantization is stored as a table for each state i. generation degree of y _t, when transformed label y _t is m, it is an b _im.

Vector quantization is performed using a codebook. When the size of the codebook is M, the feature vectors collected as learning samples are clustered into clusters of 1, 2,..., M, and the cluster m (= 1, 2,. representative vectors is obtained by storing (mean vector, centroid, also called code vector such) mu _m with a searchable form with label m. LB as a clustering method
G. A method called an algorithm is well known. y _t
Is performed by transforming y _t to the nearest centroid label. Therefore, the degree of occurrence of y _t in state i is mathematically expressed as Given by d (y _t , μ _m ) is the distance between y _t and μ _m, and various types including the Euclidean distance can be considered.

FIG. 3 is a block diagram of a speech recognition device using a discrete HMM. Reference numeral 301 denotes a feature extraction unit which converts an input audio signal into a predetermined time interval (called a frame), for example, 10 ms, by a known method such as filter bank, Fourier transform, LPC analysis, or the like.
Convert to a feature vector for each ec. Therefore, the input speech signal is converted into a sequence of feature vectors Y = (y ₁ , y ₂ ,..., Y _T ). T is the number of frames. Reference numeral 302 denotes a codebook, which holds a representative vector corresponding to each label in a form that can be searched by the label. Reference numeral 303 denotes a vector quantization unit which replaces (encodes) each vector of the vector series Y with a label corresponding to the closest vector corresponding to a representative vector registered in the codebook. Reference numeral 304 denotes a parameter estimating unit, which is an HM corresponding to each word that is a recognized vocabulary from the learning sample.
It estimates the parameters of M. That is, in order to create an HMM corresponding to the word r, first, the structure of the HMM (the number of states and its transition rules) is appropriately determined, and then the label sequence obtained by uttering the word r a number of times is used. The state transition probability in the model and the probability of occurrence of a label generated in accordance with the state are determined so that the degree of occurrence of the label sequence is as high as possible. An HMM storage unit 305 stores the HMM obtained as described above for each word. A likelihood calculation unit 306 calculates the likelihood of each model stored in the HMM storage unit 305 with respect to the label sequence for the label sequence of the unknown input speech to be recognized. 307 is a determination unit and a likelihood calculation unit 306
The word corresponding to the model giving the maximum likelihood of each model obtained in step (1) is determined as the recognition result. In FIG. 3, a broken line indicates a signal flow when the HMM is created.

In continuous HMM, the degree of occurrence of the observation vector in each state is given by the probability density function defined there,
The accuracy is higher than the discrete type, but there is a problem that a large amount of calculation is required. On the other hand, the discrete HMM calculates the label m in each state in calculating the likelihood of the model for the observed label sequence.
(= 1, ..., M) the probability of occurrence of b _im is there is an advantage that very small amount of calculation because can be executed by reading from the storage device stored in advance in connection with the label, due to quantization There is a disadvantage that the recognition accuracy is worse than that of the continuous type due to an error. To avoid this, it is necessary to increase the number of labels M (increase the codebook size),
With the increase, the number of learning samples required for learning the model becomes enormous. If the number of learning samples is insufficient, there is the estimated value of the b _im is frequently 0, can not be the correct estimate.

This estimation error is, for example, as follows. Now, let us consider a case where a word corresponding to "Osaka" is present in the recognized vocabulary and a model corresponding to the word is created. Voice samples corresponding to the word "Osaka" uttered by many speakers are converted into a feature vector sequence, and each feature vector is converted into a label as described above. Thus, each audio sample for the "Osaka"
It is converted into a label sequence corresponding to each. By estimating the HMM parameters {a _ij , b _im } from the obtained label sequences so that the likelihood for those label sequences is maximized, a discrete HMM corresponding to the word “Osaka” is completed. The well-known Baum-Welch method or the like can be used for this estimation.

In this case, the label sequence of the learning sample corresponding to the word “Osaka” does not necessarily include all the labels existing in the codebook. The occurrence probability of a label that does not appear in the label sequence of this learning sample is estimated to be “0” in the learning process in the model corresponding to “Osaka”. Therefore, in the label sequence corresponding to the word voice “Osaka” uttered at the time of recognition, a label that happens to be not included in the label sequence used for creating the model of “Osaka” exists ( When the number of learning samples is small, this is satisfactorily possible.) When the label sequence of “Osaka” uttered during this recognition is generated from the learned “Osaka” model, the degree becomes “0”. However, even in such a case, even if the label is different, at the stage of the feature vector before being converted to the label, it is quite close to the voice sample used for learning the model, and if it is seen at the vector stage, it is enough "Osaka" May be recognized as appropriate. They should be similar at the vector stage because they originally uttered the same word, but if the vectors are near the boundaries of the clusters of labels to be transformed, there will be a slight difference at the vector stage. However, it is quite possible that the label will be converted to a completely different label. It is easy to imagine that this adversely affects recognition accuracy. Such a problem occurs more frequently as the codebook size M increases and the number of learning samples decreases.

One of the methods for eliminating this defect is an HMM (FVQ / HMM) based on fuzzy vector quantization. Among them, the synergistic FVQ / HMM described in IEICE Technical Report SP93-27 (June 1993) is notable for showing excellent performance.

FIG. 4 is a block diagram illustrating the general principle of FVQ / HMM. In the same figure, the broken line indicates the signal flow when creating the HMM. Reference numeral 401 denotes a feature extraction unit, which is similar to 301 in FIG. Reference numeral 402 denotes a code book, which is similar to 302 in FIG. Reference numeral 403 denotes a membership calculation unit, which converts the feature vector into a membership vector. Membership vector, the feature vector at each time point, a vector with the membership elements for each cluster, the feature vector at time t y _t, the cluster C _1, .., C _M, the y _t
If the degree of membership in C _m is u _tm , y _t is the degree of membership vector u
_t = ( _ut1 , ..., _utM ) Converted to ^T. Hereinafter, in the present application, a vector is a vertical vector, and T at the right shoulder represents transposition. There are various possible definitions of u _tm , for example, Can be defined as (JGBezdek: “Pattern Recognition w
ith Fuzzy Objective Function Algorithm ", Plenum Pre
ss, New York (1981). ). d (x, y) is the distance between the vectors x and y. In this equation, F> 1 is called fuzziness, It is. Where δ _ij is the Kronecker delta, i =
When j, δ _ij = 1, and when 1 ≠ j, δ _ij = 0. F →
When 1 y _t when the label of the cluster corresponding to the closest centroid to it and o _t, 1 degree of belonging to the cluster, because the degree of belonging to other cluster becomes 0, this is normal vector (Equation 5) means that when F → ∞, the membership of y _t is 1 / M for any cluster, and the ambiguity is maximized. If the posterior probability of C _m with respect to y _t can be calculated using a neural network or other means, the posterior probability can be defined as the posterior probability. Are referred to as "degree of membership").

In practice, for the reasons described below, the degree of membership u _tm
Is not calculated for all clusters,
d (y _t , μ _m ) is calculated for the K-th smallest cluster (K-nearest neighbor) from the smallest cluster.
That is, the elements forming the membership vector u _t is the terms cluster large upper K degree of membership is computed values (Equation 4), the other is zero. 404 is a parameter estimation unit. An HMM storage unit 405 stores an HMM corresponding to each recognition unit such as a word or a syllable to be recognized. Reference numeral 406 denotes a likelihood calculation unit that calculates each of the HMMs from the membership vector series obtained at the output of the vector quantization unit.
Likelihood for the input speech, i.e., sequence y ₁ of the feature vector, ..., y _T is the HMM r (r = 1, ... , R) is adapted to calculate the degree L ^r generated from each is there. 407 is a determination unit, Is calculated, and r ^* is a recognition result.

Likelihood calculating section 406 calculates likelihood L ^r corresponding to recognition unit ^r by r
= 1, ..., but is intended to calculate the R according to equation (1), wherein the omega _i various HMM by way of definition _{(y t)}
Is defined. Synergistic FVQ / HMM featured here
Defines ω _i (y _t ) in principle as follows. Here, b _im is the occurrence probability of cluster m in state i of the HMM.

As described above, the addition or multiplication of m in (Equation 7) is actually performed only in the K clusters with the highest degree of membership. In this case, (Equation 7) becomes (Equation 8) ( Hereinafter, the description will be made in the form of addition.) Here, h (k) is a cluster name in which y _t has the k-th highest degree of belonging. When the degree of membership is defined by (Equation 4), (Equation 4) may be calculated for d (y _t , μ _m ) up to the K-th order from the smallest. In this case, u _{t, h (1)} +... + U
_{t, h (K)} = 1, ut _{, h (K + 1)} =... = ut _{, h (M)} =
It becomes 0. The reason why the addition in (Expression 7) as in (Expression 8) is performed only in the K clusters with the highest degree of membership is due to the following reason, not to mention reducing the amount of calculation.

The reason why the FVQ type shows a higher recognition rate than the discrete type is due to the interpolation effect of the learning sample at the time of parameter estimation. This complementary effect works in the following manner, for example. For example, consider a case where the probability that clusters A and B occur in state i is estimated from a learning sample. In the case of the discrete type, a vector to be quantized is classified as A if it is on the A side even if it is close to B, but it is classified as B if it is on the B side. Therefore, even if A and B are included in the same proportion as the population, there is a bias in the learning sample, and especially because many vectors near the boundary between A and B happen to be included in A,
It is possible that the probability of occurrence of A is estimated to be larger than the probability of occurrence of B. If the number of training data with respect to the codebook size is small, such bias of the training data is likely to occur. Deteriorate.

On the other hand, in the case of the FVQ type, A
Not only that, but also the occurrence of B will be calculated assuming that B has also occurred, so for the above learning sample, even if the occurrence probability of A is estimated to be somewhat higher, The occurrence probability of B is also estimated in accordance with the degree of belonging, and no extreme estimation error occurs as in the discrete type.
This means that interpolation is performed on the learning samples by using the FVQ type, in other words, the number of learning samples is approximately increased. This is the reason why the recognition rate of the FVQ type exceeds the recognition rate of the discrete type particularly in a case where the codebook size is large.

However, even though the FVQ type interpolates the lack of training samples, this means that the number of training samples is increased apparently from the given training sample itself. It's a little different from increasing. Therefore, the number of learning samples is relatively increased for each cluster codebook size is reduced, the accuracy of estimation of b _im come up sufficiently,
Depending on the method of interpolation, it is quite possible that the discrete type that does not perform interpolation will have a higher or similar recognition rate than the FVQ type, rather than performing poor interpolation.

The degree of this interpolation is affected by the value of K as well as the codebook size and fuzziness. As K = 1 approaches, that is, as the discrete type approaches, the effect of complementation decreases, and as K increases, the effect of interpolation increases. Therefore, when fuzziness is fixed, the degree of interpolation can be controlled by K. In other words, it is not good to increase K indiscriminately, and in the sense that the improvement of the recognition rate by the FVQ type with respect to the discrete type is maximized, the K has an optimal value K ₀ according to the codebook size. I do. According to the experiment, in recognizing 100 city names by an unspecified speaker, the optimal value is K = 6 for a codebook size of 256 and the optimal value is K = 3 for a codebook size of 16.

Thus, the FVQ type needs to calculate (Equation 8) at the time of recognition as compared with the discrete type, so that the calculation of K times of membership and the product-sum operation of K times increase, but the recognition rate is discrete. It is better than the case of the continuous type, is equal to or more than the case of the continuous type, and the calculation amount is considerably reduced as compared with the case of the continuous type.

As a method of calculating (Equation 1), a method called the Forward-Backward method is used.
The terbi method is often used, and is usually used in the form of logarithmic addition. That is, Is calculated, and L ′ is set as the likelihood. (Equation 9) can be efficiently calculated by the dynamic calculation method. That is, L ' Is assumed to be φ _i (1) = log π _i , and t = 2,. Is required. This is called the Viterbi method. Since there is no great difference between using L and L 'as the recognition result, the Baum-Welch method (For
ward-Backward method), and the Viterbi method is often used for recognition. For synergistic FVQ / HMM,
When the Viterbi method is used for recognition, b _im is log b _im
Do not used only in the form of, instead of directly storing the b _im, by storing the log b _im, (7) or calculation of the equation (8), the logarithmic operation is only required a sum of products I can do it.

Next, DP matching will be described. Most basically, there is a method based on pattern matching between feature vector strings. FIG. 5 shows a conventional example. Reference numeral 51 denotes a feature extraction unit, which is the same as 301 in FIG. Reference numeral 53 denotes a standard pattern storage unit that stores standard patterns corresponding to words. This standard pattern is registered in advance in the standard pattern storage unit as a result of conversion into a feature vector sequence by the feature extraction unit 51, corresponding to the word to be recognized. The broken line in FIG. 5 indicates the connection used for the registration, and the connection indicated by the broken line is released at the time of recognition. 52 is a pattern matching unit,
A matching calculation between each standard pattern stored in the standard pattern storage unit 53 and the input pattern is performed, and a distance (or similarity) between the input pattern and each standard pattern is calculated. Reference numeral 54 denotes a determination unit that finds a word corresponding to a standard pattern that gives a minimum value (maximum value) of the distance (or similarity) between the input pattern and each standard pattern.

The following is a more specific explanation. In this example, a description will be given assuming that a “distance” between patterns is obtained.
(If based on "similarity,""distance" is replaced by "similarity,
Just replace "minimum" with "maximum.") Now
The feature vector output at the time point t in the feature extraction unit 51 is y _t , and the series of input patterns is Y = (y ₁ , y ₂ ,.
y _T ), the standard pattern corresponding to the word r And then, the distance to Y of ^{Y (r) D (r)} , when the distance between y _t and y ^(r) _j and ^{d (r) (t, j} ) ( where each when expressed in multiplication format Is represented by D ₂ ^(r) , d ₂ ^(r) (t, j), and D ₁ ^(r) , d ₁ ^(r) (t, j) when expressed in addition form), And calculate Is the recognition result. Where x (k) in (Equation 13)
= (T (k), j (k)) is the k-th grid point on the matching path X between Y and Y ^(r) in the grid graph (t, j), and w (x (k)) Is a weight coefficient for weighting the distance at the grid point x (k).

Thereafter, a parallel discussion holds in both the multiplication and addition forms, and it is easy to convert to a multiplication form expression if necessary (d ₁ ^(r) (t, j) = log d ₂ ^(r) ( t, j), D ₁ ^(r) = log
D ₂ ^(r) ), since it is generally used in the addition form, the description will be mainly made in the addition form here (therefore, the suffixes 1 and 2 are omitted), and the multiplication form is used as necessary. write.

x (k ₁₎ from x (k ₂₎ up to the point sequence _{x (k 1), ...,} x
Let (k ₂ ) be X (k ₁ , k ₂ ) and x (K) = (t (K), J
If (K)) = (T, J), the meaning of (Equation 13) means that the characteristic of each of the input pattern Y and the standard pattern y ^(r) associated with the point sequence X (1, K) The minimum value of the accumulation of the weighted distances between the vectors with respect to X (1, K) is defined as the distance D ^(r) between Y and Y ^(r) . The calculation of (Equation 13) is
If the weighting factor w (x (k)) is properly selected, dynamic programming (Dy
It can be efficiently executed using dynamic programming, and is called DP matching.

In order to perform DP, the principle of optimality must be established. In other words, it must be said that "a sub-policy of an optimal policy is a sub-policy and also an optimal policy".
If this can be said, For The following recurrence formula holds, and the amount of calculation is greatly reduced.

The optimal strategy from point x (1) to point p ₀ = x (k) is
When the weighted cumulative distance along the point sequence X (1, k) = (x (1),..., X (k) = p ₀ ) is φ ′ (p ₀ , X (1, k)) ,
This is to find a point sequence (optimal point sequence) that minimizes φ ′ (p ₀ , X (1, k)). X ^* (1, k) = (x
^{* (1), ..., x} * (k-1), x * (k) = p 0) and then,
If φ ′ (p ₀ , X ^* (1, k)) is φ (p ₀ ), the fact that the principle of the above-mentioned optimality is satisfied means that the point x (1) is converted to the point x (1).
^* (K-1) optimal point sequence up is that it coincides with the point sequence ^{X *} (1, k) on the column points from the point ^x * (1) to the point ^x * (k-1) is there. In other words, in the optimal point sequence starting at x (1) and ending at x (k-1), φ (x
(K-1)) + w (p 0) d (r) (p 0) X * the column point where the minimum (1, k-1) = (x * (1), ..., x * ( k-1))
, The point sequence from x (1) to x (k−1) in the optimal point sequence from x (k) = p ₀ is X ^* (1, k−1)
Matches. Therefore, the optimal point sequence starting at various x (1) and ending at various x (k-1) is known, and therefore φ (x (k-1)) for various x (k-1) If known, the optimal point sequence from various x (1) to a specific x (k) = p ₀ and the weighted cumulative distance along it can be calculated by (Equation 16). That is, the weighted minimum cumulative distance φ (x (k)) from the point x (1) to the point x (k) is calculated using the weighted minimum cumulative distance φ (x (k−1)) as number
16), and φ (x
Since (1)) = w (x (1)) d ^(r) (x (1)) as an initial value, D ^(r) = φ (x (K)) is recursively obtained.
The weighted minimum cumulative distance can be obtained with much less calculation amount than calculating the cumulative distance in all allowable paths by brute force.

Here, as an example of a weighting factor that can satisfy Equation 16, And so on. That is, if the weighting coefficient is (Equation 17) or the like, the principle of optimality holds, and the dynamic programming can be applied. (1) if the sum of the weighting factors is equal to the length of the input pattern (the number of frames), (2)
Is the case where the sum of the weighting factors is equal to the length of the standard pattern, and (3) is the case where the sum of the weighting factors is equal to the sum of the lengths of the input pattern and the standard pattern.

If Expression (1) of Expression (17) is used, Expression (18) can be considered as one specific example of the recurrence expression of Expression (16).

By sequentially calculating (Equation 18) for t = 1,..., T, j = 1,..., J, (Equation 13), that is, D ^(r) can be calculated. In this case, the path that can lead to x (k) is shown in FIG.
Will be constrained as follows. That is, the path leading to the point (t, j) is as follows: point (t−2, j−1) → point (t−1, j) → point (t, j), point (t−1, j−1) → Point (t, j), point (t−1, j
-1) → point (t, j), and passes through only one of the three ways, and the numerical value on the path indicates the weight coefficient when each path is selected. In this case, w (x (1)) + ...
+ W (x (K)) is equal to the number T of input frames. Accordingly, in this case, the denominator of (Equation 14) is constant irrespective of the standard pattern. Therefore, when calculating which standard pattern is closest to the input pattern, w (x (1)) +.
There is no need to normalize with w (x (K)). In this case, d
^{(r) As the} (t, j), the Euclidean distance or the city distance as a more simplified one is often used.

The most calculation amount in the matching calculation is
This is distance calculation or similarity calculation between feature vectors.
In particular, as the number of words increases, the amount of calculation increases in proportion to the number of words, and it takes a long time to respond, which poses a practical problem. To reduce this, there is a so-called “SPLIT method” that uses vector quantization (SPLIT: Wor
d Recognition System Using Strings of Phoneme−Lik
e Templates). (Sugamura, Furui, "Large Vocabulary Word Speech Recognition Using Onomatopoeia Standard Patterns", IEICE (D), J65-D, 8, pp.104
1-1048 (Showa 57-08). FIG. 7 is a block diagram showing a conventional example. The feature extraction unit 71 is the same as that in FIG. Reference numeral 73 denotes a codebook in which M labeled representative vectors are stored in a form that can be searched by labels. Reference numeral 74 denotes a vector quantization unit that converts the output feature vector y _t of the feature extraction unit 71 into a label of a cluster having a centroid closest to y _t using the codebook 73. Reference numeral 77 denotes a word dictionary which stores a standard pattern of a word voice to be recognized, which is converted into a label sequence by the above-described operation. This label is also called onomatopoeia. Assuming that the pseudophoneme of the k-th frame of the word r as the standard pattern is s ^(r) _k , a word to be recognized in the form shown in FIG. J ^(r) is the last frame (and thus the number of frames) of the standard pattern of word r. The broken line in the figure indicates a connection used only during the operation of registering a recognized word. Reference numeral 72 denotes a distance matrix calculation unit that obtains the distance of each output vector of the feature extraction unit 71 to the centroid of each cluster, converts the distance into a vector having the distance as an element, and converts a feature vector sequence into a distance vector sequence. That is, it is converted into a distance matrix. For example, the distance matrix is as shown in 75, and the feature vector y _t of the frame _t
Of, cluster C _m distance d (y _t of the centroid μ _m of, μ
_m ) (denoted as _{dTM in} FIG. 7) as distance elements (d (y _t , μ ₁ ), d (y ₁ , μ ₂ ),..., d
(Y _1, μ M) is converted into y _t ^{in) T.} For example, when using the city distance Can be defined as Where y _tk is the k-th element of the vector y _t ,
μ _mk is the k-th element of the centroid vector μ _m of C _m . A matching unit 76 matches the distance matrix output from the distance matrix calculation unit 62 with each word in the word dictionary, and calculates the distance therebetween. Specifically, when s ^(r) _j = C _m , the distance d between y _t and s ^(r) _j
^(r) (t, j) the (number ^{20) d (r) (t} , j) = d (y t, μ m) as will compute the (number 18). That is, FIG. 7 shows a case where d ^(r) (t, j) in the conventional example of FIG. 5 is replaced with d which is calculated in advance by referring to a distance matrix.
The difference is that (y _t , μ _m ) is used, and the calculation can be performed in exactly the same way using DP. Reference numeral 78 denotes a judgment unit,
4) is calculated and finally the recognition result is obtained. In this case, the denominator of (Equation 14) has the same value as in FIG.
For the same reason as described in the embodiment of FIG. 5, w (x
Since (1)) +... + W (x (K)) = T, there is no need to normalize this.

In the case of the conventional example of FIG. 5, the distance calculation of y _t and y ^(r) _j increases with the number of recognized words, but in the case of the conventional example of FIG. 7, the distance matrix 75 is calculated once. Once
Since the distance between y _t and the onomatopoeia only needs to refer to the distance matrix 75, the amount of calculation of d ^(r) (t, j) remains unchanged regardless of how many words are added.

For example, consider the case of recognizing 100 words with an average of 50 frames per word and a 10-dimensional feature vector.
In the case of, y _t and the number of standard pattern vectors for which the distance calculation is to be performed are of the order of 50 × 100 = 5000, and if the distance is the Euclidean distance, the number of multiplications is multiplied by 10 to 5000.
It becomes 0 times. In the case of FIG. 7, the calculation of y _t and distance is
Since each centroid vector in the codebook is the same, if the number of clusters is M = 256, 256 distance calculations are required regardless of the number of recognized words, and the number of multiplications is 2
560, which means that the latter is about 1/20 of the former.

Here, although the input feature vector sequence has been described as being converted into a distance matrix, in fact, the distance vector _{(d t1, ..., d tM} ) T is the standard pattern onomatopoeic rhyme s ^(r) _j
(R = 1,..., R; j = 1,..., J ^(r) ). If the calculation of the formula is performed for all the standard patterns, d (y _t , μ _j ) does not need to be stored as a matrix. It is sufficient to store the distance vector for each minute, and the storage amount is actually smaller.

The FVQ / HMM shows a recognition rate equal to or higher than that of the continuous HMM, and the calculation amount is much smaller than that of the continuous HMM. However, when performing word spotting, the definition of ω _i (y _t ) is defined by the FVQ / HMM. It cannot be said that it is the same as HMM.

The SPLIT method requires much less computation than a method of directly matching spectra, but has a problem in that recognition accuracy is deteriorated.

DISCLOSURE OF THE INVENTION The first invention of the present application has solved this problem. The second invention relates to an improvement of the SPLIT method, and applies the concept of FVQ to DP matching. The third invention relates to a reduction in the amount of storage and the amount of calculation in the HHM and DP. The fourth invention is
Particularly, in the HMM, the amount of calculation at the time of recognition is further reduced.

(1) The first invention assumes that a system to be analyzed has a plurality of states, clusters a feature vector space, and stores a representative vector of each cluster in a form searchable by its label. Cluster occurrence probability storage means for storing the occurrence probability of each label in each state (accordingly, the occurrence probability of each cluster); and the degree of belonging of an observation vector to each cluster (the Membership calculation means for calculating the posterior probability of the observation vector), the logarithmic value of the degree of membership of the observation vector to each calculated cluster, and the occurrence probability of each cluster stored in the cluster occurrence probability storage means Calculate the sum of products of the observation vector and the equivalent amount, and generate the observation vector as the degree of occurrence of the observation vector in each state of the system. And a degree calculation means.

(2) The second invention clusters a feature vector space, and stores a codebook in which a representative vector of each cluster is stored in a searchable form by its label, and a degree of membership of each cluster of observation vectors or each of the above Means for calculating a posterior probability of the cluster with respect to the observation vector (both are hereinafter referred to as the degree of membership), and calculating a degree of membership vector having the degree of membership of the observation vector with respect to each cluster as an element; Standard pattern storage means for storing a standard pattern represented by the membership degree vector,
A matching means for matching an input pattern consisting of the membership vector converted from the observation vector obtained as an output of the membership calculation means with the standard pattern;

(3) The third invention clusters a feature vector space, and stores a codebook in which a representative vector of each cluster is stored in a form searchable by its label, a probability of occurrence of a cluster m in the state i of HHM, or DP The degree of belonging of the feature vector of the i-th frame of the standard pattern vector to the cluster m in the matching is b _im , and the number of clusters is M
Then, N pieces of b _{i, g (i, 1)} , b _{i, g (i, 2)} , ..., b _i taken in order of size from b _i1 , ..., b _{iM , g (i, N)} (g
(I, n) is the label of the nth largest cluster) is the value as it is or its logarithmic value log b _{i, g (i, 1)} , log b
_{i, g (i, 2)} , ..., log b _{i, g (i, N)} and store the remaining b _{i, g (i, N + 1)} , ..., b _{i, g ( i, M)} includes a cluster occurrence probability storage unit or a membership standard pattern storage unit that stores a constant value.

(4) According to a fourth aspect of the present invention, a feature vector space is clustered, and a code book in which a representative vector of each cluster is stored in a form searchable by its label, and a probability of occurrence of each label in each state (accordingly, Cluster occurrence probability storage means for storing cluster occurrence probabilities; and membership degree calculation means for calculating the degree of membership of an observation vector to each cluster (posterior probability of each cluster for the observation vector) using the codebook And calculating the sum of products of the calculated degree of belonging of the observation vector to each cluster and the logarithmic value of the occurrence probability of each cluster stored in the cluster occurrence probability storage means or an equivalent amount thereof,
An observation vector occurrence degree calculating means for calculating the occurrence degree of the observation vector in each state of the system, and estimating the occurrence probability of each cluster in each state,
It includes means for calculating using the observation vector occurrence degree calculating means, and at the time of recognition, calculating the degree of belonging of the observation vector such that the maximum degree of belonging is 1 and all other degrees of belonging are zero.

 The operation of the present invention will be described below.

(1) In the first invention, a system to be analyzed assumes a plurality of states, clusters a feature vector space, and stores a representative vector of each cluster in a form searchable by its label. The probability of occurrence of each label in each state (accordingly, the probability of occurrence of each cluster) is stored by cluster occurrence probability storage means, and each cluster of observation vectors is stored by the membership degree calculating means using the codebook. (The posterior probability of each cluster with respect to the observed vector) is calculated, and the logarithmic value of the calculated degree of belonging of the observed vector to each cluster and each cluster stored in the cluster occurrence probability storage unit are calculated. The sum of products with the cluster occurrence probability or an equivalent amount is calculated by the observation vector occurrence degree calculation means, and the observation vector is calculated. To the calculated occurrence rate in each state of the system.

(2) In the second invention, the input signal is converted into a sequence of feature vectors by the feature extraction unit, and
For each vector of the vector series, the degree of membership of each vector stored in the cluster storage means to each cluster to be classified is calculated, and the degree of membership of each vector to each cluster is defined as an element by the standard pattern storage means. An input pattern consisting of a sequence of membership degree vectors obtained as an output of the membership degree calculation means by a matching means is calculated by calculating a membership degree vector to be recognized, each recognition unit to be recognized is represented by a membership degree vector sequence, and stored. And the standard pattern.

(3) In the third invention, the HMM includes cluster occurrence probability storage means. When the occurrence probability of cluster m in state i is b _im and the number of clusters is M, the HMM includes b _i1 , R, b _{i, g (i, 1)} , b _{i, g (i, 2)} , ..., b _{i, g (i, R)} taken in order of size from ..., b _iM (G (i,
r) is the label of the r-th largest cluster) as is or the logarithmic value log b _{i, g (i, 1)} , log b
_{i, g (i, 2)} , ..., log b _{i, g (i, R)} and store the remaining b _{i, g (i, R + 1)} , ..., b _{i, g ( i, M)} stores a constant value,
The feature extraction means converts an input signal into a sequence of feature vectors, the cluster storage means stores clusters to which the vectors are to be classified, and the membership calculation means calculates the respective clusters of each vector of the feature vector series. The feature vector generating means calculates the degree of belonging to each cluster of the feature vector and the probability of occurrence of each cluster in each state of the HHM based on the degree of belonging to each cluster of the feature vector. Calculating the degree, the vector series occurrence degree calculating means calculates the degree of occurrence of the feature vector sequence from the HMM using the output of the feature vector occurrence degree calculating means, and the feature vector occurrence degree calculating means, 2. A method according to claim 1, wherein the first and second clusters having the highest degree of membership correspond to the respective cluster occurrence probabilities. It calculates the degree of occurrence of the feature vector.

(4) In the fourth invention, the feature vector space is clustered, and a code book in which a representative vector of each cluster is stored in a form that can be searched by its label is provided. The probability of occurrence of each label (accordingly, the probability of occurrence of each cluster) is stored, and the degree of membership of the observation vector to each cluster (the posterior probability of each cluster with respect to the observation vector) using the codebook is calculated by the degree of membership calculation means. )
The observed vector occurrence degree calculating means is a product sum of the calculated degree of belonging of the observed vector to each cluster and a logarithmic value of the occurrence probability of each cluster stored in the cluster occurrence probability storage means or is equivalent thereto. And calculating the degree of occurrence of the observation vector in each state of the system, and estimating the probability of occurrence of each cluster in each state using the observation vector occurrence degree calculation means. At this time, the degree of membership of the observation vector is calculated such that the maximum degree of membership is 1 and all other degrees of membership are zero.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is an explanatory diagram of an HMM.

FIG. 2 is an exemplary diagram of an HMM frequently used in speech recognition.

FIG. 3 is a block diagram showing a conventional example of a speech recognition device using a discrete HMM.

FIG. 4 is a block diagram showing a conventional example of a speech recognition apparatus using an HMM based on fuzzy vector quantization and an embodiment of the present invention.

FIG. 5 is a block diagram of a conventional example of a speech recognition apparatus using pattern matching.

FIG. 6 is an explanatory diagram showing an example of a constraint condition of a matching path in the input pattern axis-dependent DP matching.

FIG. 7 is a block diagram showing a conventional example of a speech recognition apparatus using vector quantization.

FIG. 8 is an explanatory diagram of one method of word spotting.

FIG. 9 is a block diagram showing an embodiment of a speech recognition apparatus using DP matching according to the present invention based on fuzzy vector quantization.

FIG. 10 is an explanatory diagram showing an example of a constraint condition of a matching path in DP matching of an input pattern axis dependent type.

FIG. 11 is an explanatory diagram showing an example of a constraint condition of a matching path in the input pattern axis-dependent DP matching.

FIG. 12 is an explanatory diagram showing an example of a constraint condition of a matching path in the input pattern axis-dependent DP matching.

FIG. 13 is an explanatory diagram showing an example of a constraint condition of a matching path in the input pattern axis-dependent DP matching.

FIG. 14 is an explanatory diagram showing an example of a constraint condition of a matching path of the input pattern axis-dependent DP matching.

FIG. 15 is an explanatory diagram showing an example of a constraint condition of a matching path in the standard pattern axis-dependent DP matching.

FIG. 16 is an explanatory diagram showing an example of a constraint condition of a matching path in the standard pattern axis-dependent DP matching.

FIG. 17 is an explanatory diagram showing an example of a constraint condition of a matching path in the standard pattern axis-dependent DP matching.

FIG. 18 is an explanatory diagram showing an example of a constraint condition of a matching path in the standard pattern axis-dependent DP matching.

FIG. 19 is an explanatory diagram showing an example of a constraint condition of a matching path in the standard pattern axis-dependent DP matching.

FIG. 20 is an explanatory diagram showing an example of a constraint condition of a matching path of a frequently used standard pattern axis-dependent DP matching.

FIG. 21 is an explanatory diagram illustrating a method of storing the probability of occurrence of a cluster in each state in the HMM according to the present invention or the degree of membership of a feature vector in a standard pattern in a DP matching according to the present invention.

FIG. 22 is an explanatory diagram for explaining a method of storing the probability of occurrence of a cluster in each state in the HMM according to the present invention or the degree of belonging of a feature vector to a cluster in a standard pattern in DP matching according to the present invention.

FIG. 23 is an explanatory diagram for explaining a method of storing the degree of membership of a feature vector to a cluster in an input pattern in an HMM according to the present invention or in DP matching according to the present invention.

FIG. 24 is an explanatory diagram for explaining a method of storing the degree of membership of a feature vector to a cluster in an input pattern in an HMM according to the present invention or in DP matching according to the present invention.

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(Equation 7) is distribution u _t = { _{ut 1} , ..., _{ut M} } and distribution b
Kull back−Leibler divergence of _i = ｛ _i1 , ..., b _iM ｝
(Hereinafter abbreviated as KLD) (Reference: IEICE Technical Report SP93-27 (June 1993)). That is, if the degree of deviation of u _t from b _i is D (ut _t b _i ), KLD
Is Given by This occurs difficulty of u _t from the population b _i, in other words, is representative of the resulting difficulty of u _t in state i. _{Therefore, log ψ i (y t)} = - D if put the _{(u t ‖b i), ψ} i (y t) will be expressed and the resulting ease of u _t in state _i ω i (y _t) Can be used as Therefore, if ω _i (y _t ) in (Equation 1) is ψ _i (y _t ), And Is a quantity determined only by the input, regardless of the model,
In the case where the model from which the input pattern is likely to be generated is compared using the value of (Equation 22), this can be omitted. So new Can be defined as That is, (Equation 24) becomes (Equation 1) It is defined as (Equation 7) can be derived in this way. This is the principle of the synergistic FVQ / HMM.

However, this can be said because all the vectors forming the input pattern Y occur once and always once for any state sequence X in each HMM. . On the other hand, when the input pattern is considered to be a concatenation of vector sequences generated by several models, the situation is different when searching for which sub-interval has a high degree of occurrence from a specific model. In this case, the input pattern frame corresponding to the first state 1 of the state sequence of the model to be matched is represented by S
(X), if the input pattern frame corresponding to the final state J is E (X), in principle, Is calculated for every X , And S (X ^* ) to E (X ^* ) can be set as a partial section of the input voice pattern.

If this is calculated properly, E (X), Z
Would be calculated for every combination of (X),
The amount of calculation becomes enormous. Moreover, in this case, (Equation 23) is X
It depends, so we can't omit it. Therefore, we consider solving the optimization problem of (Equation 27) by dynamic programming. Assuming that frame s is the starting frame,
The end frame t is changed in the range of t ′ ± v, and the degree of occurrence of the input partial patterns y _s ,..., y _t from the HMMλ is calculated by dynamic programming. In this case, can be applied. (Number 10), the input pattern _{(y s, .., y t'} -v, ..., (y s, ..., occurrence rate for each maximum y _{t '+ v)} Is obtained by dynamic programming, that is, by moving the end frame t within an appropriate range and selecting a further optimum one from each end, the optimum end starting from s is obtained. The end range should be checked against, for example, the start s
A value set in advance from the average length of the vector sequence in which the HMM occurs can be used. In this case, every time t changes t
It is necessary to normalize by −s + 1. If the same operation is performed while changing s, the optimum start and end, that is, a desired partial section is eventually obtained. In this case, it is not necessary to calculate (Equation 10) every time t is changed for a certain s. That is, when the model of FIG. 2 is used, for a certain s, the correspondence between the input feature vector and the state of the HMM is limited to the range of the diagonal line in FIG. The range of m (i) is And the range of m (i + 1) is (Equation 29) According to (Equation 10), φ over the range of (Equation 29)
_{i + 1} (m (i + 1)) is calculated as a continuation of φ _i (m (i)) over the range of m (i) in (Equation 28). Therefore, for each of 0i = 1,..., J + 1, φ _i (m
If (i)) is calculated in the range of (Equation 28), i = J +
For one ΦJ _{+ 1} (m (J + 1)) obtained in the range of
Are matching results for the end frames t ′ ± v when the start frame of the input pattern is s. That is, in this way, the result for the end frame t '± v for one start frame s is obtained at a time. However, even in this method, the calculation must be performed for the range within the oblique line by changing the starting end for each frame, and the amount of calculation is extremely large. In order to further reduce the calculation amount, the starting frame may be automatically determined by the dynamic programming. To do this, the recurrence equation (Equation 10) is transformed as follows.

Now, the optimal state sequence corresponding to y _s , ..., y _t is X ^* = x
^* _S , x ^* _{s + 1} , ..., x ^* _t . In order to apply the dynamic programming, for an input frame m ′ at X ^* , s <
m '<t with ^{_{x * m'-1 = i,}} x * m' if a = j (j ≠ i), the partial pattern y _s, .., state series _{optimal y m 'corresponding} in , X ^* _s , ..., x ^* _{m '} . That is, Φ _j (m ′) is
Partial pattern for state sequence x ^* _s , ..., x ^* _{m '}
y _s, .., in the degree to which y _{m 'is} generated, the section length of the state i in the time (number of frames) is a z (i), dur _i (z) =
Let bur _i (z (i)) be the degree to which state i lasts for z frames given that, Where, for m '<m <t, x ^* _m-1 = j, x
^* = H If so, Must hold. Where W _i , W _i ′, w _di , w _ai , w
_bi (k), w _di ′, w _ai ′, w _bi (k) ′ (i = 1,..., J) are weighting factors associated with the state sequence or the sum thereof, _{ｉ i} (m)
Is the state sequence _{xs '} , ..., _{xmz (i)} ( ₎ x ^* _s , ..., x ^*
_{mz (i)} ), the degree of occurrence of the input partial pattern y _s ,..., y _{mz (i)} . W _i is the state sequence x ^* _s , ..., x ^*
The sum of the weighting factors along _{mz (i)} , W _i ′, is the sum of the weighting factors along the state sequence _{xs ′} ,..., xm _{−z (i) ′} . Here, if these weighting factors are properly selected, (Equation 31) to (Equation 34) can always be established regardless of the state sequence. For example, obviously, W _i = W _i ′, w _di ′ = w
_di , w _ai ′ = w _ai , w _bi (1) ′ +... + w _bi (z (i) ′) ′
= W _bi (1) +... + W _bi (z (i)) satisfies the above condition. That is, regarding a state sequence that changes from another state to a state i in an input frame m, the sum of the weighting factors along the state sequence up to that point regardless of how to take the state to m, the starting frame s, and the point (m, i). What is necessary is just to make it constant. As a specific numerical value, for i = 1,..., J, w _di = w _ai = 1, w _bi (k) = 1 / z (i).

From the above discussion, assuming that the state in the input frame t has now changed to j, the optimal degree of occurrence of the partial section up to the point (t, j) is determined by the number of frames in the input partial section corresponding to the state i. Assuming that z is obtained from the following recurrence formula.

At this time, i, z that satisfies (Equation 35) is i ^* , z ^* , Are stored simultaneously, word spotting can be performed in the next step.

(1) Initialization Φ ₁ (t) = 0 for t = 1,..., T (π ₁ = 1, π _i ≠ 0
∀i ≠ 1) Φ _i (0) = − ∞ for i = 1, ..., JB ₁ (t) = 0 for t = 1, ..., T (2) t = 1, ..., Execute (3) and (4) for T + 1 (3) Execute (Equation 26) and (Equation 27) for i = 1, .., J + 1 (4) Φ (t) = Φ _{J + 1} (t), B ( t) = B _{J + 1} (t−z
^* ) (5) Partial pattern detection End frame: In this way, the calculation of Φ _i (t) in (Equation 35) need only be performed once for each (t, j), and log ω _i
The sum of (y _{tz-1 + k} ) over k = 1 to z does not need to be calculated every time z changes.
(M), the sum up to z (m) is s (m + 1) = s
Since the calculation of (m) + log ω _i (y _{t−z + z (m)} ) is sufficient, the calculation amount is significantly reduced.

Here, in the present invention, It is a thing to keep. At this time, according to (Equation 35) The following recurrence formula is defined. When using the model of FIG. 2, Becomes Now, assuming that the optimal state sequence has been obtained, the length of state i at that time is z (i) ^* , and the following is set for simplicity.

And It is. Therefore, when using the recurrence formula of (Equation 31), Φ
_{J + 1} (t) includes the sum of the last term on the right-hand side and the second term from the last term on the right-hand side, regardless of the state sequence, regardless of the input pattern. Since it is a determined amount and is not related to the problem of maximizing this equation, it is unnecessary when finding the optimal section of the input pattern. Therefore, as ω _i (y _t ), one redefined as the following equation can be adopted.

Next, a method of applying the concept of FVQ, which is the second invention of the present application, to DP matching will be described. FIG. 9 is a block diagram showing the principle of the present invention. 91 and 93 are 71 and 73 in FIG.
Performs the same operation as. 92 is a membership degree matrix calculation unit,
Although this corresponds to the distance matrix calculation unit 72 in FIG. 7, in the present embodiment, a membership matrix is calculated. That is, the degree of membership u _{tm of} the feature vector y _{t to} the cluster C _m (m = 1, ..,
M, and calculates the _{u t1 + u t2 + ... +} u rM = 1). As the degree of membership, a membership function or the like used in fuzzy theory or the like can be used. In this case, the FVQ /
The same one used in the HMM is used. Reference numeral 95 denotes a word dictionary, which comprises a membership degree matrix registered in advance corresponding to each word to be recognized. That is, the standard pattern of the r-th word is registered as the membership matrix obtained by the membership matrix calculation unit 92 for the utterance. In FIG. 9, the degree of belonging of the word r to the cluster m of the j-th frame in the standard pattern is represented by b ^(r) _jm . 96 is the word 1, ..., R
Is a membership matrix for.

At this time, the frame t of the input pattern at the time of recognition
Membership vector similarity between frame j of reference pattern r and _{u t = (u t1, u} t2, ..., u tM) T and _{^{b (r) j = (b}} (r) j1, b
^(r) _j2 , .., b ^(r) _jM ) is given as the similarity of ^T. Here, _utm ≧ 0, _ut1 +... + _UtM = 1, b ^(r) _jm ≧ 0, b ^(r) _j1
+ ... + b ^(r) _jM = 1, so that both vectors can be regarded as a probability distribution vector (when u _tm and b ^(r) _jm are posterior probabilities, it is exactly the probability distribution itself).
As in the case of the HMM, the similarity can be given by Kullback-Letbler bivergence known as the distance between probability distributions. That is, the distributions (q ₁ , .., q _M ) and (p ₁ , ..,. P
_M ) Given by

Using this, the following three definitions as d ^(r) (t, j) are possible.

(Equation 45) (1) is a useful distance definition when the weighting coefficients of (Equation 17) and (1) are adopted, and (Equation 45) and (2) are (Equation 17)
(2) is a useful distance definition when employing the weighting coefficients, equation (45) (3) When the problem symmetry process ^{(d (r) (t,} j) = d (r) ( j, t) is a useful distance definition. Of course, in these definitions, a constant multiplication can be used in the case of the addition type, and a constant multiplication can be used in the case of the multiplication type.

A matching unit 94 performs DP matching between each of the membership matrices corresponding to each word registered in the word dictionary 95 and the membership matrix obtained from the input pattern. That is, the distance between frames shown in (Equation 45)
Based on d ^(r) (t, j), calculate the recurrence formula (Equation 18),
(Equation 13) The cumulative distance D ^(r) defined in (a ⁾ is calculated.
Reference numeral 97 denotes a determination unit that calculates (Equation 14) and obtains a recognition result.

In this way, instead of replacing each feature vector of the feature vector sequence uttered to construct a word dictionary to be matched with only one onomatopoeia, multiple onomatopoeia correspond to each frame with degree of belonging Therefore, the adverse effect of the quantization error of the conventional example can be reduced. As is clear from the above description, the present invention has a feature that the calculation of the degree of belonging and the calculation of the distance between each frame of the standard pattern and the input pattern are determined based on a mathematically clear definition. is there.

 Next, further improvements of the present invention will be described.

First, a case where (Equation 45) (1) is used as a distance scale will be described. In this case, the distance between frames is , And using this as a weighting factor (Equation 17) (1),
13) Where t ≦ (k) −t (k−n) = 1 and 1 ≦ n ≦
k-1 exists (there is no jump for the input pattern frame in the matching path), and for this n,
The sum of the weighting factors along the path from x (kn) to x (k) is 1, that is, w (kn + 1) + w (kn + 2) +
.. + W (k) = 1, for example, in the case of FIGS. Becomes In the example of FIGS. 10 to 14, x (k) = (t + j), k−1
For ≧ n ≧ 1, FIG. 10 shows that x (k−1) = (t−1,
j) or (t−1, j−n); FIGS. 11 and 14 show x (k−1 = (t
−1, j) of (t−1, j−1), m = 2, .., n, x
(Km) = (t−1, j−m); FIGS. 12 and 13 show x (k−m)
1) = (t−1, j), (t−1, j−1) or (t, j−1), m
= (2, .., n-1), x (km) = (t, j-m),
This is the case where x (kn) = (t−1, j−n). The numerical values attached to the side of the path in each figure are examples of weighting factors along the path in each case.

At this time, the first term on the right side of (Equation 48) is independent of any of the path selection method and the standard pattern, and is an amount determined only by the input pattern. Therefore, this can be omitted when only the magnitude relation of the comparison result between each standard pattern and the input pattern is considered. Therefore, by omitting this term and changing the sign, Can be the similarity between the patterns. In this case, the inter-frame similarity between the input frame t and the frame j of the standard pattern r is It can be.

Here, furthermore, t (k) -t (k-1) = 1 (in the matching path, there is no jump to the input pattern frame and no overlap ... (FIG. 6) or (applicable in the case of FIG. 10). if, Becomes Here, j = j (t) is a function representing a matching path on the tj plane, and t = t (k), j =
It is obtained by eliminating k from j (k). In the case of using a path of FIG. 6, the matching unit 94 on the basis of the equation (50) at the indicated inter-frame similarity s ^(r) (t, j), for example, Calculates the recurrence formula, and calculates the cumulative similarity defined by (Equation 51)
S ^(r) will be calculated. The judgment unit 97 Is calculated to obtain a recognition result.

There exists 1 ≦ n ≦ k−1 such that t (k) −t (k−n) = 1 (there is no jump to the input pattern frame in the matching path).
The sum of weighting factors along the path from n) to x (k) is 1,
That is, w (kn + 1) + w (kn + 2) + ... + w
The above method of setting (k) = 1 is useful in continuous word speech recognition and the like. In other words, by doing so, the problem of finding the most suitable connected pattern of individually registered recognized word standard patterns, which is most similar to the continuously uttered input word voice pattern, is solved by a known two-step method. DP
This is because the calculation can be efficiently performed using the above. The inter-frame similarity proposed here can be applied in such a case, and can provide high recognition performance with a simple calculation.

Next, a case where (Equation 45) and (2) are used as a distance scale will be described. In this case, the distance between frames is , And the weighting coefficient is expressed as (Equation 17) (2) (Equation 17).
13) Where 1 ≦ n ≦ k, which is j (k) −j (kn) 1
-1 exists (there is no jump to the standard pattern frame in the matching path), and for this n, x
The sum of the weighting factors along the path from (kn) to x (k) is 1, that is, w (k−1 + 1) + w (kn−2) +.
In the case of + w (k) = 1, for example, in the case of FIGS. Becomes 15 to 19, x (k) = (t, j), k−1 ≧
For n ≧ 1, FIG. 15 shows x (k−1) = (t, j−1) or
(T−n, j−1); FIGS. 16 and 19 show x (k−1) = (t, j−).
1) For or (t−1, j−1), m = 2, ..., n, x (k
−m) = (tm, j−1); FIGS. 17 and 18 show x (k−1)
= (T, j-1), (t-1, j-1) or (t-1, j), m =
For 2, .., n-1, x (km) = (tm, j), x
This is the case where (kn) = (t-n, j-1). The numerical values attached to the side of the path in each figure are examples of weighting factors along the path in each case.

At this time, the first term on the right side of (Equation 56) is independent of any path selection method and any section of the input pattern, and is an amount determined only by the standard pattern (standard pattern r
This amount relative to C ^(r) ). Therefore, it can be omitted when only the magnitude relation of the comparison result between a certain standard pattern and various sections of the input continuous word voice pattern or various input patterns is considered. So we omit this section,
If you change the sign, Can be the similarity between the patterns. In this case, the inter-frame similarity between the input frame t and the frame j of the standard pattern r is It can be.

When using the definition of the similarity between frames to determine which standard pattern the input pattern is close to,
^(r) −C ^(r) ) / J ^(r) and find the largest one.

Here, further, j (k) -j (k-1) = 1 (in the matching path, there is no jump to the standard pattern frame and no overlap ... (FIG. 20) or (applicable in the case of FIG. 15)). if, Becomes Here, t = t (j) is a function representing the matching path on the tj plane, and t = t (k), j = j
It is obtained by eliminating k from (k). When using the paths shown in FIGS. 15 to 19, the matching unit 94 calculates the inter-frame similarity s ^(r) (t, j) shown in (Equation 58). Calculates the recurrence formula, and calculates the cumulative similarity defined by (Equation 59)
S ^(r) will be calculated.

There exists 1 ≦ n ≦ k−1 such that j (k) −j (k−n) = 1 (there is no jump to the standard pattern frame in the matching path).
The sum of weighting factors along the path from n) to x (k) is 1,
That is, w (kn + 1) + w (kn + 2) + ... + w
The above method of setting (k) = 1 is useful when performing so-called word spotting for identifying a partial section that best matches a certain standard pattern from an input pattern of continuous word speech in which words are uttered continuously. In this case, assuming that the standard pattern to be compared is r regardless of the length of the section of the input pattern, it is only necessary to compare S ^(r) in each section. That is, in this way, the word spotting problem can be efficiently calculated by applying the dynamic programming method in the following steps. The inter-frame similarity proposed here can be applied in such a case, and can provide high recognition performance with a simple calculation. For example, when word spotting is performed on a certain word using the route restriction condition in FIG.

(1) Initialization f (0, f) = f (−1, j) = − ∞ for i = −1,0,1, ...,
J f (0,0) = 0 (2) Execute (3) to (6) for t = 1,..., T + 1 (3) f (t, 0) = − ∞ (4) f (t, 1) = s (t, 1) (5) B (t, 1) = t-1 (6) Calculate the following recurrence formula for j = 2, ..., J (7) D (t) = f (t, J), B (t) = B (t, J) (8) Partial pattern detection End frame: Next, a method for reducing the amount of storage and the amount of calculation in the HMM and DP according to the third invention will be described.

The basic idea is that when the memory amount is reduced, the degree of membership of the standard pattern is stored for the top N <M, and when the amount of calculation is reduced, the degree of membership of the input pattern is calculated only for the top K <M. Is based on In this case, it should be noted that the probability distribution (p ₁ , ..., p _M ) and the probability distribution (q ₁ , ...,
q _M ) Where p _i = 0∃i∈ ｛1, ..., M｝ is possible, but q _i > 0∀i∈ ｛1, ..., M｝, and q _i is 0 It cannot be. Therefore, when calculating or storing only the top N of q _i, for the remaining q _i, determined to be the common value for _{q 1 + ... + q M =} 1, so that use that value. _{Accordingly, q i (i = 1,} ..., M) in this case amount of storage required for _{the, q g (1), ...} , q for _{g (N) N,} q _{g (N + 1)} ,..., Q _{g (M)} is 1. Here, g (n) is a subscript of the n-th largest q in {q ₁ ,..., Q _M }. p _i is divided into upper K and below K + 1
q Same as _i (K does not have to be the same as N)
However, since these can be 0, h (k) is expressed as ｛p ₁ , ..., p
If the subscript of the k-th largest p in _M ｝, then _{ph (1)}
+ ... + _{ph (K)} = 1, _{ph (K + 1)} + ... + _{ph (M)} = 0. In this case, the amount of storage required for p _i (i = 1,..., M) is only _K for _{ph (1)} ,.

Ω ^(r) _i (y _t ) in the synergistic FVQ / HMM (the ω _i
(Y _t ), _bim , a _ij, etc., are particularly related to the word r, where (r) is attached to the right shoulder), and s ^(r) in the synergistic FVQ / DP (t , j) have the form of (Equation 63), and the same can be said for the reduction of the memory amount and the calculation amount. Therefore, the following description is for the case of the synergistic FVQ / HMM, that is, ω ^{(r )} An embodiment will be described for _i (y _t ). In this case, the state j in the HMM, the j-th frame of the standard pattern in the DP, the occurrence probability b ^(r) _im of the cluster m in the state j of the HMM are assigned to the cluster m in the j-th frame of the standard pattern r in the DP matching. In other words, the same argument as HMM holds in the case of DP matching.

As a definition of ω ^(r) _j (y), the following method is considered as a method for reducing the storage amount for each of u _tm and b ^(r) _jm . Here, the subscript g (r, j, n) means a cluster name (number) in which the occurrence probability of the cluster in the j-th state of the HMM r is the n-th, and b ^(r) _{i, g (r, j, n)} is the probability of occurrence of the cluster g (r, j, n) in the j-th state of HMMr, and h (t, k) is the degree of membership of the feature vector of the t-th frame of the input pattern is k.
U _{t, h (t, k)} means the degree of belonging of y _t to cluster h (t, k).

[First method] And However, for b ^(r) _{j, g (r, j, n)} , 1 ≦ n ≦
For N, the estimated values for n = 1,.
For + 1 ≦ n ≦ M And For u _tm , (1.1) m in 1 ≦ m ≦ M
Use all estimates of
_{As for t, h (t, k)} , when 1 ≦ k ≦ K In K + 1 ≦ k ≦ M, it may be estimated such that u _{t, h (t, k)} = 0. In the case of (1.2), the calculation of the degree of membership is also reduced (described later).

[Second method] , For b ^(r) _{r, g (r, j, n)} , for 1 ≦ n ≦ N And for N + 1 ≦ n ≦ M, b ^(r) _{j, g (r, j, n)} =
It is assumed that it has been estimated to be zero. For ut _{, h (t, k)} , use (2.1) all estimated values of ut _{, h (t, k)} at 1 ≦ k ≦ M or (2.2) at 1 ≦ k ≦ K Uses u _{t, h (t, k)} similar to the above, and K + 1 ≦ k
For ≤M And In the case of (2.2), the calculation of the degree of membership is also reduced (described later).

[Third method] , With respect to u _tm , when 1 ≦ k ≦ K
_{t, h (t, k)} is the estimated value of k = 1, ..., K, K + 1 ≦
For k ≦ M, And ^{For (r)} _{j, g (r, j, n)} , (3.1) 1 ≦ n ≦
For M, use all estimates for n = 1, ..., M, or (3.2) for 1 ≦ n ≦ N And for N + 1 ≦ n ≦ M, b ^(r) _{j, g (r, j, n)} =
It may be estimated to be zero. In the case of (3.2), the amount of memory is reduced at the same time.

[Fourth method] , For ut _{, h (t, k)} , when 1 ≦ k ≦ K _{Ut, h (t, k)} = 0 when K + 1 ≦ k ≦ M
It is estimated to be b ^(r) _{j, g (r, j, n)}
(4.1) For 1 ≦ n ≦ M, use all estimated values of n, or (4.2) For 1 ≦ n ≦ N, b ^(r)
_{j, g (r, j, n)} is an estimated value as it is, and when N + 1 ≦ n ≦ M, May be defined as In the case of (4.2), the amount of memory is also reduced at the same time.

In the first method, the second method, the third method (3.2), and the fourth method (4.2), the occurrence probability of the cluster in each state of the HMM is calculated by the following equation. ) Instead of remembering every cluster, HM
For each state of M (each frame of the standard pattern of DP), the labels and probabilities (degrees of membership) are stored for clusters having probabilities (degrees of membership) up to the Nth rank of clusters with a high probability (degree of membership). Things. For example, the HMM (standard pattern) for the r-th word is shown as in FIG. 21 or FIG. FIG. 21 can be used to define the similarity by (Equation 67) and (Equation 70), and FIG. 22 can be used to define the similarity by (Equation 64) and (Equation 73).

In the first method (1.2), the second method (2, 2), the third method, and the fourth method, instead of calculating the membership matrix as an input pattern for all clusters, the For each frame of the pattern, the degree of belonging of the cluster having the high degree of belonging to the K-th rank is calculated. For example, the input pattern is shown as in FIG. 23 or FIG. FIG. 23 shows a case where the similarity is defined by (Equation 64) and (Equation 73).
Can be used to define the similarity in (Equation 67) and (Equation 70).

In the case of (Equation 64) and (Equation 73), in FIG. 22, instead of b ^(r) _{j, g (r, j, n)} as the membership matrix of the standard pattern,
By storing log b ^(r) _{j, g (t, j, n)} (not shown),
This calculation may be a product-sum operation. At this time, the amount of calculation that increases compared to the conventional example of FIG.
In 73), if K = 3, each grid point multiplies 3
The number of multiplications is 25
That is, 60 + 3 × 50 × 100 = 4060, which is certainly increased as compared with the conventional example of FIG. 7, but is much smaller in calculation amount as compared with the case of FIG. 5, and further, as compared with the conventional example of FIG. And high recognition accuracy can be obtained.

(Number 67), if the (number 70), as membership matrix of the input pattern, in FIG. 24, u _t, log instead of _{h (t, k)}
If u _{t, h (t, k)} is stored (not shown), this calculation may be a product-sum operation. At this time, if the amount of calculation which increases compared to the conventional example of FIG. Therefore, the number of multiplications is 2560 + 3 ×
That is, 50 × 100 = 4069, which is certainly increased as compared with the conventional example of FIG. 7, but is much smaller in calculation amount than in the case of FIG. The recognition accuracy can be obtained. In this case, compared to the case of storing the log b _im in the previous section, log u _t for each frame of the input _pattern, it is necessary to operation of the _{h (t, k).} But,
If K = 3, this is only 3 times per frame and u
_{Since t, h (t, k)} can only take a value between 0 and 1, 0 ≦
If log x is tabulated for x ≦ 1, a table lookup may be sufficient instead of this calculation.

The degree of membership is defined by (Equation 4), and u _{t, h (t, 1)} + ... + u
_{t, h (t, K)} = 1, ut _{, h (t, K + 1)} = ... = ut _{, h (t, M)}
When = 0, the order of the magnitude of u _tm and the order of the _smallest d (y ₁ , μ _m ) are the same, so first calculate d (y _t , μ _m ) for all clusters , The calculation of the upper K u _tm is d
This means that it is sufficient to perform the processing for the lower K clusters of (y _t , μ _m ), and the amount of calculation can be reduced. That is,
The degree of membership at 1 ≦ k ≦ K is When Given by At this time, the fraction calculation of the denominator of (Equation 76) and the calculation of (Equation 77) are both K times. M = 256, K = 3
If で 6, this calculation amount is 1/40 to 1/80.

The degree of membership is defined by (Equation 4), and u _{t, h (t, K + 1)} = ... =
ut _{, h (t, M)} = _ut0 , ut _{, h (t, 1)} + ... + ut _{, h (t, M)} =
When 1 is set, the degree of membership at 1 ≦ k ≦ K is When (Equation 78) requires a fraction calculation of the denominator M times, but the order of the magnitude of u _tm and the order of the small d (y _t , μ _m ) are the same. D for all clusters
(Y _t , μ _m ) and ( _equation 79) i) u _tm is calculated by d
(Y _t, μ _m) may be carried out with respect to a cluster of lower K of.

Alternatively, for simplicity, the following is one method. For example, d _t0 = ｛d (y _t , μ _{h (K + 1)} ) +.
+ D (y _t , μ _{h (M)} )｝ / (M−K) or d _t0 = {d
(Y _t , μ _{h (K + 1)} + d (y _t , μ _{h (K)} )｝ / 2 and d (y _t ,
_{μ h (K + 1))} = ... = d (y t, and μ _M)) = d _t0, the (number 78) Is approximated by

Alternatively, the membership degree calculation means calculates the membership degree from the distance between the observation vector to be calculated and the representative vector of each cluster, and arranges them in order of the smallest as the first place having the smallest distance. For clusters below the K + 1 rank, a fixed value of 1 / K or less is set in advance. For clusters of K in ascending order of distance, the sum of the degree of membership is calculated from the individual distances and the above-mentioned constant value to become 1 from the individual distances. Is what you do.

In the case of the synergistic DP matching, since the standard pattern is a membership vector sequence, when registering the membership of the top N clusters of the membership also in the standard pattern,
The exact same method as performed for _utm can be used. That is, b ^(r) _{j, g (r, j, 1)} + ... + b ^(r)
_{j, g (r, j, N)} = 1, b ^(r) _{j, g (r, j, N + 1)} = ... = b ^(r)
_{When j, g (r, j, M)} = 0, K → N, h (h, k) → g
(R, j, n), u _{t, h (t, k)} → b ^(r) _{i, g (r, j, n)}
6) b ^(r) _jm can be obtained according to ⁽ _Equation 77). Similarly, b ^(r) _{j, g (r, j, N + 1)} = ... b ^(r) _{j, g (r, j, M)} = b
^(r) _j0 , b ^(r) _{j, g (r, j, 1)} + ... + b ^{(r) When} _{j, g (r, j, M)} = 1, ⁽ _{Equation 78} ^{) (} _Equation 79) B ^(r) according to ^(Equation 80 ⁾
_jm can be obtained.

Next, the fourth invention of the present application will be described. This case is effective in the case of the HMM. The idea is that _utm , b
_The K when estimating _im and the K when recognizing
The difference is that there is no problem in theory. In particular, it is often desired that the amount of calculation be as small as possible in the case of recognition, regardless of whether a model is created. The least computational complexity is the discrete HM
M, which in FVQ / HMM, K =
This corresponds to the case where the calculation is performed as 1. Therefore, the model can be created by the FVQ / HMM method, and the model can be recognized by the discrete HMM method. As described above, the meaning of using the FVQ type is not to reduce the quantization distortion due to vector quantization by complementation, but to alleviate the insufficient number of learning samples when learning the parameters of the HMM, and to reduce the parameter. The effect of increasing the estimation accuracy of is larger. Therefore, performing model creation with the FVQ type and performing recognition with the discrete type at the time of recognition is slightly lower in performance than performing recognition with the FVQ type, but it is better than performing model creation and recognition with the discrete type. It can be experimentally confirmed that the recognition rate is improved particularly where the coat book size is large.

In the matching based on the linear expansion and contraction of the time axis, the input pattern and the standard pattern can be compared based on the comparison between the membership vectors. Also in this case, when the number of frames of the standard pattern is linearly expanded and contracted so as to match the number of frames of the input pattern, the definition of the similarity of Expression 7 can be used, and the number of frames of the input pattern is adjusted to the number of frames of the standard pattern. In the case of expanding and contracting linearly as much as possible, the definition of the similarity of (Expression 43) can be used.

Industrial Applicability According to the first invention, Kullbach-Lejbler Divergen
It is possible to provide an HMM device capable of performing word spotting with a small amount of calculation and with high accuracy by using a distance measure of ce.

According to the second invention, in the conventional DP matching based on vector quantization, a feature vector belongs to only one cluster. However, according to the present invention, a feature vector is assigned to a plurality of clusters by a degree of membership for each cluster. Or a cluster corresponding to the posterior probability of each cluster with respect to its feature vector, and based on the degree of belonging, the similarity between frames was defined by a probabilistic distance scale. This makes it possible to realize a pattern comparison apparatus that is resistant to spectrum fluctuations due to factors and requires only a small increase in the amount of calculation as compared with the conventional example.

According to the third aspect, instead of storing the occurrence probabilities of all clusters for each state of the HMM corresponding to each recognition unit,
By storing up to the N-th place in the order of the probability, and storing one common value assuming that the others have the same probability, it is possible to greatly reduce the required storage amount.

According to the fourth aspect of the present invention, the model is created by a synergistic FVQ / HM
Since the recognition is performed as M and the recognition is performed as the discrete HMM, an estimation error due to the shortage of the number of learning samples at the time of model creation is reduced, and a device that requires a small amount of calculation at the time of recognition can be realized.

──────────────────────────────────────────────────続 き Continued on the front page (58) Fields surveyed (Int. Cl. ⁶ , DB name) G10L 3/00 531 G10L 3/00 515 G10L 3/00 541 JICST file (JOIS)

Claims (29)

(57) [Claims] 1. A system to be analyzed has a plurality of states, a feature vector space is clustered, and a representative book of each cluster is stored in a form retrievable by its label. And a cluster occurrence probability storage means for storing the occurrence probability of each label (ie, the occurrence probability of each cluster) in (i), and the degree of belonging of the observation vector to each cluster (ie, the observation vector of each cluster) using the codebook. Posterior probability for
And a product sum of a logarithmic value of the occurrence probability of each cluster stored in the calculated cluster occurrence probability storage unit or the equivalent thereof. Vector generation degree calculation means for calculating the degree of occurrence of the observation vector in each state of the system, wherein the cluster generation probability storage means comprises a cluster in which the generation probability of the cluster is N + 1 or less. Is a common value that is not zero, and is calculated so that the sum of cluster probabilities becomes 1. 2. A method according to claim 1, wherein said membership degree calculating means calculates the membership degree for clusters having a membership degree equal to or lower than the K + 1th rank to be zero and the sum of the membership degrees to be one. Signal analyzer. 3. A system to be analyzed has a plurality of states, a feature vector space is clustered, and a representative book of each cluster is stored in a form retrievable by its label. And a cluster occurrence probability storage means for storing the occurrence probability of each label (ie, the occurrence probability of each cluster) in (i), and the degree of belonging of the observation vector to each cluster (ie, the observation vector of each cluster) using the codebook. Posterior probability for
And a product sum of the logarithmic value of the calculated degree of membership of the observation vector to each cluster and the occurrence probability of each cluster stored in the cluster occurrence probability storage means, or equivalent thereto. An observation vector occurrence degree calculating means for calculating an observation vector occurrence degree in each state of the system by calculating an appropriate amount of the observation vector. 4. A cluster occurrence probability storage means in each state stores a cluster calculated so that the sum of the clusters having a predetermined probability N up to the highest N is equal to 1; 4. The signal analysis device according to claim 3, wherein the occurrence probability is zero. 5. The degree of membership calculation means sets the degree of membership of each observation vector to each cluster to a common non-zero value if the degree of membership is less than or equal to the (K + 1) th rank, and sets the sum of the degrees of membership to one. The signal analyzer according to claim 3, wherein the signal is calculated. 6. The signal analyzing apparatus according to claim 1, wherein each state is a state of a hidden Markov model. 7. A codebook in which a feature vector space is clustered and a representative vector of each cluster is stored in a form that can be searched by its label, and a probability of occurrence of each label in each state (ie, a probability of occurrence of each cluster) ) Is stored, and the codebook is used to calculate the degree of membership of the observed vector to each of the clusters (that is, the posterior probability of each cluster to the observed vector). Calculate the product sum of the calculated degree of belonging of the observation vector to each cluster and the logarithmic value of the occurrence probability of each cluster stored in the cluster occurrence probability storage means or an equivalent amount, and calculate the observation vector An observation vector occurrence degree calculating means for calculating the occurrence degree in each state of the system, The estimation of the raster occurrence probability is calculated using the observation vector occurrence degree calculation means. At the time of recognition, the degree of belonging of the observation vector is set to 1, the maximum degree of belonging is set to 1, and all other degrees of belonging are set to 0. Signal analysis device characterized in that the calculation is performed as follows. 8. A cluster storage means into which a feature vector is to be classified, and for vectors x and y to be compared, the degree of belonging of each vector to each cluster or the posterior probability of each cluster to each vector ( A membership degree calculating means for calculating the membership degree vector having the membership degree of each vector as an element, and a distance or similarity between the membership degree vectors. And a similarity calculating means for calculating the degree.
The similarity calculating means calculates the distance or the similarity.
Let a = (a ₁ , ..., a _M ) and b
= (B ₁ , ..., b _M ), the distance or similarity is calculated as one of the following or an equivalent amount thereof, and the feature vectors x and y are calculated using the distance or similarity. A signal analysis device characterized in that the distance or similarity is determined. 9. A cluster storage unit into which a feature vector is to be classified, a degree of belonging of each vector of the vector series to each cluster, and a degree of membership vector having the degree of belonging of each vector to each cluster as an element. , A standard pattern storage means that simultaneously represents the recognition units to be collated with a membership degree vector sequence, and an input pattern comprising a membership degree vector sequence obtained as an output of the membership degree calculation means. Matching means for matching with the pattern, and as a result of the matching,
9. The signal analyzer according to claim 8, wherein a similarity or a distance between the input pattern and the standard pattern is calculated. 10. The time axis of either the input pattern or the standard pattern, or both time axes, are linearly or non-linearly expanded and contracted, and the time axes of both patterns are adjusted to obtain the distance or the distance between the corresponding membership vectors. A similarity calculating means for calculating the similarity; and a cumulative similarity for accumulating the distance or the similarity along the time axis of one of the input pattern and the standard pattern, or both of them. 10. The signal analyzing apparatus according to claim 9, further comprising a calculating unit, wherein the accumulated value is used as a distance or a similarity between the input pattern and the standard pattern. 11. A similarity calculating means for calculating the degree of belonging or similarity between the degree of belonging vectors, each of the degree of belonging vectors constituting the input pattern, and the degree of belonging constituting a standard pattern to be matched with the input pattern. With each of the vectors, the similarity between the associated degree of association vectors of both patterns, the time value of either the input pattern or the standard pattern, or the cumulative value accumulated along both time axes is the minimum or 10. The signal analyzing apparatus according to claim 9, further comprising: a dynamic planning unit that optimally corresponds to the maximum value and calculates the maximum value. 12. The membership vector corresponding to the frame t of the input pattern _{a t = (a t1, ...} , a tM), the membership vector corresponding to the frame j of the reference pattern b _j = (b _j1 , ...,
b _jM ), the k-th (t, j) coordinate on the matching path is x (k) = (t (k), j (k)), and the weighting factor at x (k) is w (x (k) ) _Is the similarity between at _(k) and bj _(k) , and the vector sequence at ₍₁₎ , ...,
a _{t (K)} and b _{j (1),} ..., a cumulative similarity along the path of the b _{j (K)} Where 1 ≦ n ≦ k−1, t
When (k) -t (kn) = 1, w (x (kn +
12. The signal analyzer according to claim 11, wherein 1))... + W (x (k)) = 1. 13. The matching path is t (k) -t (k-
13. The signal analyzer according to claim 12, wherein 1) = 1, w (x (k)) = 1. 14. The matching path is represented by x (k) = (t,
j), for k-1 ≧ n ≧ 1, (1) x (k−1) =
(T−1, j−n) or x (k−1) = (t−1, j),
(2) x (k-1) = (t-1, j-1) or x (k-
1) = (t−1, j) and m = 2,.
m) = (t−1, j−m), (3) For m = 1,..., n−1, x (km) = (t, j−m), x (k− n) = (t
−1, j−n), (4) For m = 1,..., N−1, x (k
-M) = (t, j-m), x (kn) = (t-1, j-
n), (5) x (k-1) = (t-1, j-1) or x
For (k-1) = (t-1, j) and m = 2, ..., n, x
(Km) = (t−1, j−m), and w (x (k)) = 1 for the path (1) and w (x (k)) for the path (2). x (k)) = 1, w (x (km + 1)) =
0, w (x (km−1)) = 0 for the same (3),
13. The signal analysis according to claim 12, wherein w (x (k−n + 1)) = 1 and w (x (km−1)) = 1 / n for (4) and (5). apparatus. 15. Distance or similarity calculating means for calculating the similarity, the membership vector corresponding to the frame t of the input pattern _{a t = (a t1, ..} , a tM), the frame j of the reference pattern The corresponding membership degree vector is represented by b _j = (b _j1 , ..., b _jM ),
Let the k-th (t, j) coordinate on the matching path be x
When (k) = (t (k), j (k)) and the weighting coefficient at x (k) is w (x (k)), _Is the similarity between at _(K) and bj _(K) , and the vector sequence at ₍₁₎ , ...,
a _{t (K)} and the similarity along the path between b _{j (1)} , ..., b _{j (K)} And for 1 ≦ n ≦ k−1, j
When (k) -j (kn) = 1, w (x (kn +
12. The signal analyzing apparatus according to claim 11, wherein 1))... + W (x (k)) = 1. 16. The matching path is j (k) -j (k-
16. The signal analyzer according to claim 15, wherein 1) = 1 and w (x (k)) = 1. 17. The matching path is defined as x (k) = (t,
j), for k-1 ≧ n ≧ 1, (1) x (k−1) =
(T−n, j−1) or x (k−1) = (t, j−1),
(2) x (k-1) = (t-1, j-1) or x (k-
1) = (t, j−1), m = 2,..., N, x (k−
m) = (tm, j−1), (3) For m = 1,..., n−1, x (km) = (tm, j), x (k− n) = (t
−n, j−1), (4) For m = 1,..., N−1, x (k
-M) = (tm, j), x (kn) = (tm, j-
1), (5) x (k-1) = (t-1, j-1) or x
For (k-1) = (t, j-1) and m = 2, ..., n, x
(Km) = (tm, j-1), and w (x (k)) = 1 for the route (1) and w (x (k)) for the route (2). x (k)) = 1, w (x (km + 1)) =
0, w (x (km−1)) = 0 for the same (3),
16. The signal analysis according to claim 15, wherein w (x (k-n + 1)) = 1 and w (x (k-m + 1)) = 1 / n for (4) and (5). apparatus. 18. When the degree of belonging of the feature vector of the frame j of the standard pattern to the cluster m is b _jm , and the number of clusters is M, the magnitude is taken from b _{j1 ,.} N b _{j, g (j, 1)} , b _{j, g (j, 2)} , ..., b _{j, g (j, N)} (g (j,
n) is the label of the n-th largest cluster in the frame j of the standard pattern, N ≦ M) is the same value, and the rest is a fixed value b ₀ and b _{j, g (j, 1)} +... + b _{j, g (j , N)} + b ₀ (M−
N) = 1 or logarithmic values log b _{j, g (j, 1)} , log b _{j, g (j, 2)} , ..., log b
_{j, g (j, N)} , signal analysis apparatus according to claim 8, further comprising a standard pattern storage means for storing in the form of log b _0. 19. When the degree of belonging of the feature vector of the frame j of the standard pattern to the cluster m is b _jm and the number of clusters is M, the magnitude is taken from b _{j1 ,.} N b _{j, g (j, 1)} , b _{j, g (j, 2)} , ..., b _{j, g (j, N)} (g (j,
n) is the label of the n-th largest cluster in frame j of the standard pattern, N ≦ M) is b _{j, g (j, 1)} +.
b _{j, g (j, N)} = 1, the remainder is b
_9. The signal analyzing apparatus according to claim 8 _{, wherein j, g (j, N + 1)} =... = b _{j, g (j, M)} = 0 is stored. 20. When the degree of membership of the feature vector y _t of the frame t of the input pattern to the cluster m is u _tm , and the number of clusters is M, the degree of membership to which y _t is to be transformed is
K u's in order of magnitude from u _t1 , ..., u _{tM
t, h (t, 1)} , ut _{, h (t, 2)} , ..., ut _{, h (t, K)} (h (t, k)
Is the label of the k-th largest cluster in the frame t of the input pattern, K ≦ M) is the same value, and the rest is a constant value u ₀ and u _{t, h (t, 1)} +... + Ut _{, h (t, K) )} + U ₀ (M−K)
9. The signal analyzing apparatus according to claim 8, wherein a value calculated so that = 1 is set. 21. When the degree of membership of the feature vector y _t of the frame t of the input pattern to the cluster m is u _tm and the number of clusters is M, the degree of membership y _t is to be transformed is
K u's in order of magnitude from u _t1 , ..., u _{tM
t, h (t, 1)} , ut _{, h (t, 2)} , ..., ut _{, h (t, K)} (h (t, k)
Is the label of the k-th largest cluster in the frame t of the input pattern, K ≦ M) is u _{t, h (t, 1)} +.
_{t, h (t, K)} = 1, the value calculated so that the remainder is u
_9. The signal analyzing apparatus according to claim 8 _{, wherein t, h (t, K + 1)} =... = ut _{, h (t, M)} = 0. Similarity between the j-th frame of the t-th frame and the standard pattern of 22. Input pattern, b _j1, ..., b N pieces of b _j taken in order of size from among _{jM, g (j, 1)} , b
_{j, g (j, 2)} , ..., b _{j, g (j, N)} (where g (j, n) is the label of the n-th largest cluster in frame j of the standard pattern, N ≦ M) b _{j, g (j, 1)} + ... + b _{j, g (j, N)} + b ₀
And (M-N) = 1 values b ₀ the calculated so that, u _tm or u _t1, which is calculated for all clusters, ..., in correspondence with the order of magnitude from the u _tM u _{t, h (t, 1)} + ... + u
K u calculated so that _{t, h (t, K)} = 1
_{t, h (t, 1)} , ut _{, h (t, 2)} , .., ut _{, h (t, K)} (where h (t, k) is the k-th largest cluster in frame t of the input pattern) Label, K ≦ M), 9. The signal analyzing apparatus according to claim 8, wherein: 23. The similarity between the t-th frame of the input pattern and the j-th frame of the standard pattern corresponds to b _{j, g (} b _{j1 ,. j, 1)} + ... + b
N b's calculated so that _{j, g (j, N)} = 1
_{j, g (j, 1)} , b _{j, g (j, 2)} , ..., b _{j, g (j, N)} (g (j, n)
Is the label of the n-th largest cluster in the frame j of the standard pattern, N ≦ M) and u _tm calculated for all clusters or u _{t1 ,.} K u _{t, h (t, 1)} , u _{t, h (t, 2)} , ..., u
_{t, h (t, K)} (h (t, k) is the label of the k-th largest cluster in the frame t of the input pattern, K ≦ M)
to _{u t, h (t, 1} ) + ... + u t, h (t, K) + u 0 (M-K) = 1 values u ₀ calculated so that, 9. The signal analyzing apparatus according to claim 8, wherein: 24. The similarity between the t-th frame of the input pattern and the j-th frame of the standard pattern is determined from b _jm or b _j1 ,..., B _jM calculated for all clusters. _{, G (j, 1)} + ... + bj _{, g (j, N)} =
N b _{j, g (j, 1)} , b calculated to be 1
_{j, g (j, 2)} , ..., b _{j, g (j, N)} (where g (j, n) is the label of the n-th largest cluster in frame j of the standard pattern, N ≦ M) K u _{t, h (t, 1)} , u _{t, h (t, 2)} , ..., ut _{, h ( )} taken in order of magnitude from u _t1 , ..., _{ut M t, K)}
(H (t, k) is the label of the k-th largest cluster in frame t of the input pattern, K ≦ M) and u
_t, with respect to _{h (t, 1) + ...} + u t, h (t, K) + u 0 (M-K) = 1 values u ₀ calculated so that, 9. The signal analyzing apparatus according to claim 8, wherein: 25. The similarity between the t-th frame of the input pattern and the j-th frame of the standard pattern is calculated from b _jm or b _j1 ,..., B _jM calculated for all clusters. N b _{j, g (j, 1)} , b _{j, g (j, 2)} , ..., b
_{j, g (j, N)} (g (j, n) is the label of the n-th largest cluster in frame j of the standard pattern, N ≦ M)
b _{j, g (j, 1)} +... + b _{j, g (j, N)} + b ₀ (M−N) = 1 b ₀ and u _t1 , ..., _ut M K, u _{t, h (t, 1)} , calculated so that u _{t, h (t, 1)} +... + Ut _{, h (t, K)} = 1 in the order of magnitude from
u _{t, h (t, 2)} , ..., ut _{, h (t, K) where} h (t, k) is the label of the k-th largest cluster in frame t of the input pattern, K ≦ M On the other hand, The signal analysis device according to claim 1, wherein 26. The degree of membership is calculated from the distance between the vector for which the degree of membership is to be calculated and the representative vector of each cluster. 19. The method according to claim 5, wherein the distances are used as they are up to N, and the degree of membership is calculated using a common value for clusters whose rank is equal to or smaller than K + 1 or N + 1. The signal analyzer according to the above. 27. The method according to claim 27, wherein the common value is K + 1 or N +
27. The signal analyzing apparatus according to claim 26, wherein clusters having a rank of 1 or less are distances and averages for each cluster. 28. A method according to claim 28, wherein the common value is K + 1 or N +
27. The signal analysis device according to claim 26, wherein the cluster having the rank of 1 or less is an average of a minimum distance and a maximum distance. 29. The degree of membership is determined from the distance between the observed vector for which the degree of membership is to be calculated and the representative vector of each cluster is small, and is determined in advance for clusters K + 1 or N + 1 or less by 1 / K or 1 / N. The following constant values are calculated, and the clusters of K or N are calculated in order from the smallest distance so that the sum of the degree of belonging becomes 1 from the respective distances and the constant value, or 18. The signal analyzer according to 18.