JPH064094A - HMM creation device, HMM storage device, likelihood calculation device, and recognition device - Google Patents

Abstract

(57) [Summary] [Purpose] An HMM that is accurate and requires a small amount of calculation.
Providing a recognition device using [Constitution] Continuous probability distribution H for creating continuous probability distribution HMM
MM creating means 104, fuzzy clustering means 106 for performing fuzzy clustering of learning vectors,
An occurrence degree calculation means 108 for calculating the occurrence degree of each cluster in each state of the HMM from the probability density function in each state of the continuous HMM is provided, and the occurrence degree is defined as the occurrence degree in each state of each cluster label and the discrete probability distribution H.
Create an MM.

Description

Detailed Description of the Invention

[0001]

[Industrial application] A new HMM (Hidden Markov model) applicable to pattern recognition such as voice recognition.
Model)) HMM creation device, HMM storage device, likelihood calculation device, and recognition device.

[0002]

2. Description of the Related Art An HMM is applicable to a general field of time-series signal processing, but for convenience of explanation, a voice recognition will be described below as an example.

First, a voice recognition device using an HMM will be described.

FIG. 5 is a block diagram of a voice recognition device using an HMM. The voice analysis unit 201 uses a well-known method such as a filter bank, Fourier transform, or LPC analysis for the input voice signal to set a fixed time interval (called a frame), for example, 1
Convert to a feature vector every 0 msec. Therefore, the input speech signal is a sequence of feature vectors Y = (y (1), y (2), ..., Y
(T)). T is the number of frames. The codebook 202 holds the labeled representative vector. The vector quantization unit 203 replaces each vector of the vector series Y with a label corresponding to the representative vector registered in the codebook 202 that is closest to the vector. The HMM creating unit 204
The HMM corresponding to each word which is a recognition vocabulary is created from the training data. That is, in order to create the HMM corresponding to the word v, first, the structure of the HMM (the number of states and the transition rules allowed between those states) is appropriately determined, and then the word v is repeated many times as described above. From the label sequences obtained by uttering, the state transition probabilities in the model and the label occurrence probabilities associated with the state transitions are calculated so that the occurrence probabilities of the label sequences are as high as possible. The HMM storage unit 205 stores the H thus obtained.
The MM is stored for each word. Likelihood calculator 20
6 calculates the likelihood of the label sequence of each model stored in the HMM storage unit 205 with respect to the label sequence of the unknown input speech to be recognized. The comparison determination unit 207 determines a word corresponding to the model that gives the maximum value of the likelihood of each model obtained by the likelihood calculation unit 206 as a recognition result.

The recognition by the HMM is specifically performed as follows. That is, the label sequence obtained for the unknown input is O = (o (1), o (2), ..., O (T)), the model corresponding to the word w is λ ^w , and the model λ ^w Let X = (x (1), x (2), ..., x (T)) be an arbitrary state sequence of length T generated by
Then, the likelihood of λ ^{w for} the label sequence O is [exact solution]

[0006]

[Equation 1]

[Approximate Solution]

[0008]

[Equation 2]

Or, taking the logarithm

[0010]

[Equation 3]

Is defined by Where P (x, y | λ ^w )
Is the joint probability of x and y in the model λ ^w .

Therefore, if, for example, (Equation 1) is used,

[0013]

[Equation 4]

Then, w ^ is the recognition result. The same applies when using (Equation 2) and (Equation 3).

P (O, X | λ) is obtained as follows.

Now, for the state q _i (i = 1 to I) of the HMMλ, for each state q _i , the occurrence probability b _i (o) of the label o and the state q _i (i = 1 to I) When the transition probability a _ij to q _j (j = 1 to I + 1) is given, the state sequence X = (x
(1), x (2), ..., x (T + 1)) and label series O = (o
(1), o (2), ..., o (T))

[0017]

[Equation 5]

It can be defined as Where π _{x (1)} is the state x (1)
Is the initial probability of. It is also assumed that x (T + 1) = I + 1 is the final state and no label is generated.

In this example, the input feature vector y is converted into a label, but there is also a method of using the feature vector y as it is in place of the label occurrence probability in each state and providing the probability density function of the feature vector y in each state. . At this time, the occurrence probability b _i (o) in the state q _i of the label o in (Equation 5) B _i when instead would use a probability density b _i (y) of the feature vector y in (hereinafter, the probability z is generated in b _i (z) is z state i when the label, z is a vector of (z) means the probability density of z). At this time, the above (Formula 1), (Formula 2), and (Formula 3) are as follows. [Exact solution]

[0020]

[Equation 6]

[Approximate Solution]

[0022]

[Equation 7]

Alternatively, the following equation can be obtained by taking the logarithm.

[0024]

[Equation 8]

As described above, no matter which method is used, the final recognition result is that the HMMλ ^w for each word w is prepared for w = 1 to W, and the input speech signal Y is obtained.

[0026]

[Equation 9]

Or

[0028]

[Equation 10]

Is the recognition result of Y. Of course, Y here
Is the input label sequence according to the above method,
A feature vector series or the like.

[0030]

In such a conventional example, one that converts an input feature vector into a label is called a discrete probability distribution HMM, and one that uses the input feature vector as it is is called a continuous probability distribution HMM. At this time,
The characteristics of both of these are as follows.

In the discrete probability distribution HMM, the occurrence degree b _i (C _m ) of each label in each state in the calculation of the likelihood of the model with respect to the input label sequence is stored from the storage device stored in advance in association with the label. Although there is an advantage that the amount of calculation is very small because it can be executed by reading out, there is a problem that recognition accuracy deteriorates due to an error associated with quantization. To avoid this, it is necessary to increase the number of labels (the number of clusters), but as the number of labels increases, the number of learning patterns necessary for learning the model becomes enormous. Here, when the number of learning patterns is insufficient, the b _i (C _m ) may frequently become 0, and correct estimation cannot be performed. For example, the following occurs.

The codebook is created by converting the uttered voices of a large number of speakers into a feature vector sequence for all the words to be recognized, clustering the set of feature vectors, and labeling each cluster. . Each cluster has a representative vector of that cluster, called the centroid, which is usually the expected value of the vector classified into each cluster. The codebook stores these centroids in a form searchable by the label.

Now, consider a case where, for example, there is a word "Osaka" in the recognition vocabulary, and a model corresponding to this is made. A voice sample corresponding to the word "Osaka" uttered by a large number of speakers is converted into a feature vector sequence, each feature vector is compared with the centroid, and a label corresponding to the nearest centroid is a quantum of the feature vector. Will be In this way, the "Osaka"
Each voice sample for is converted into a label sequence. A model corresponding to the word "Osaka" is created by estimating the parameters of the HMM from the obtained label series so that the likelihood for the label series is maximized. The well-known Baum-Welch method or the like can be used for this estimation.

In this case, among the labels in the codebook, some may not be included in the learning label series corresponding to the word "Osaka". The occurrence probability of the label not included is estimated to be "0" in the learning process. Therefore, it is sufficient that there is a label that is not included in the label sequence used to create the model of "Osaka" in the label sequence in which the word "Osaka" converted at the time of recognition is converted. It is possible. In this case, the probability that the “Osaka” label sequence uttered at the time of recognition will occur from the “Osaka” model will be “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 speech sample used for learning the model, and if you look at the stage of the vector, it is sufficient to say “Osaka”. It may be recognized that It is quite possible that a label will be converted to a completely different label with a slight difference, even though they are similar at the vector level because they originally speak the same word. It can be easily imagined that such a situation adversely affects the recognition accuracy. Such problems occur more frequently as the number of clusters increases and the number of training data decreases.

In order to eliminate this problem, it is necessary to take measures such as smoothing and complementing labels that do not appear (are not included) in the training set. Starting from the idea of reducing the number of parameters using the concept called “conclusion”, when the 0 probability is estimated, it is replaced with a small amount instead of 0, or the boundary of the cluster such as fuzzy vector quantization. Although various methods for smoothing and complementing, such as a method for blurring, have been proposed, none of these methods fundamentally solve the above problem. Also, there are factors that must be determined empirically depending on the case, and there is no theoretical indicator that determines those factors.

On the other hand, in the continuous probability distribution HMM, the distribution shape is given in advance in the form of a function such as a normal distribution and the parameters defining this function are estimated from the learning data. Therefore, the number of parameters to be estimated is small, the parameters can be accurately estimated with less learning patterns than the discrete type, there is no need to consider smoothing and complementation, and generally a higher recognition rate than the discrete type can be obtained. It is reported that

The number of parameters in the 4-state 3-loop HMM as shown in FIG. 6 is compared between the discrete type and the continuous type, for example, as follows. In the case of the discrete type, if the type of label used is 256, the label occurrence probability is 2
56 × 3 = 768, the transition probability is 6, and a total of 874 are required per model. In the case of the continuous type, if a 10-dimensional normal distribution is used, the average vector is 10 × 3 = 30, the covariance matrix is 55 × 3 = 165 (symmetric matrix), and the transition probability is 6; The parameter value of the continuous type is 1/4 or less of that of the discrete type.

However, the continuous type is excellent in recognition accuracy, but there is a problem that the amount of calculation is much larger than that of the discrete type. That is, assuming that the input feature vector y (t) has a normal distribution of the mean vector μ _i and the variance-covariance matrix Σ _{i in} the state i, the calculation of the occurrence probability (density) of y (t) in the state i (y (t) −μ _i ) ^T Σ _i ⁻¹ (y (t) −μ _i ), which is necessary. For example, in a 10-dimensional continuous HMM, this calculation alone requires 110 multiplications. And
This is (number of states x number of input frames) times this for one model. Therefore, if the above model is assumed when the number of input frames is 50, (y (t) −μ _i ) ^T Σ _i ⁻¹ (y (t) −μ _i ) of one model is required. The number of multiplications in the calculation is 110 × 3 × 50 = 16500
When the number of words is 500, this is further multiplied by 500. That is, in that case, only the multiplication of this part is 82
It requires 50,000 times.

In the case of the discrete type, once the vector quantization calculation is completed, it is only necessary to read the occurrence probability of the label from the storage device according to the label as described above. Also, y
The calculation required for vector quantization of (t) is the calculation of the distance or the similarity between 256 representative vectors and y (t) in the above example. Now the distance (Euclidean distance)
^{If 2} , then the calculation required to label y (t) is ² times 10 subtractions, 10 multiplications and 10 additions.
56 times. Therefore, in 50 frames, if only multiplication is considered, 10 × 256 × 50 = 1280 thousand times. If vector quantization is performed by a method called a binary search, 256 is 2log ₂ 256 =
The number is changed to 16, which means 10 × 16 × 50 = 8000 times.

As described above, when the discrete type is used, the amount of calculation is remarkably reduced, and when the number of recognized words is increased in the continuous type, the amount of calculation is also increased in proportion thereto. This calculation is necessary only when the signal is vector-quantized once, and this calculation amount does not change even if the number of recognized words increases.

In short, the discrete type has a small calculation amount but has a problem in recognition accuracy, and the continuous type has a good recognition accuracy but has a problem in the calculation amount.

In consideration of such problems of the conventional HMM, the present invention provides an HMM creating device, an HMM storage device, a likelihood calculating device, and a recognizing device which have high recognition accuracy and can reduce the amount of calculation. With the goal.

[0043]

According to the present invention, a continuous probability density distribution HMM creating means, a clustering means for fuzzy clustering a training vector set and giving a label to each cluster, and a cluster according to each cluster in each state of the HMM and a label. Is provided with a training vector included in each cluster and a label occurrence degree calculation means for calculating from the probability density function in each state of the continuous probability density distribution HMM, and the label occurrence degree of the label occurrence degree in each state of the HMM is generated. A discrete probability distribution HMM is created by obtaining it as the output of the degree calculating means.

[0044]

In the present invention, the probability density function in each state of the HMM is obtained by the continuous probability distribution HMM creating means, the training vector set is clustered by the fuzzy clustering means, each cluster is labeled, and the degree of label occurrence is calculated. A discrete probability distribution HMM is created by calculating the degree of occurrence of each cluster and hence the label in each state of the HMM from the training vector included in each cluster and the probability density function in each state of the continuous probability density distribution HMM by means. To do.

[0045]

Embodiments of the present invention will be described below with reference to the drawings.

Here, the definitions of symbols to be used hereinafter will be collectively described. For the sake of simplicity, the states q _i , q _j, etc. will be simply referred to as i, j, etc. unless misunderstanding occurs. In addition, it is assumed that the model learning is performed for one word, and when it is necessary to distinguish between the models, the number corresponding to the model is added to the right shoulder of the parameter as a subscript, and this is usually used. I will omit it. The definition is as follows.

I = 1, 2, ..., I + 1: i-th state [a _ij ]: transition matrix a _ij : transition probability from state i to state j r: training pattern number for the model to be created ( r =
1, ..., R) y ^(r) (t): observation vector in the t-th frame of the training pattern r o ^(r) (t): observation label b _i (y ^{(r (r)} in the t-th frame of the training pattern ^{r )} (t)): probability density b _i (o ^(r) (t)) of observation vector y ^(r) (t) of frame t of training pattern r in observation state o of frame t of training pattern r ^(r) (t) occurrence degree in state i (probability, probability density, etc.) y ^(r) = (y ^(r) (1), y ^(r) (2), ..., y ^(r) (T ^(r) )): vector sequence of training pattern r (where r = 1, 2, ...,
R) O ^(r) = (o ^(r) (1), o ^(r) (2), ..., o ^(r) (T ^(r) )): r-th label sequence for word w ( However, r = 1, 2, ...
.., R) X ^(r) = (x ^(r) (1), x ^(r) (2), ..., x ^(r) (T ^(r) ), x ^(r)
(T ^(r) +1)): State sequence corresponding to X ^(r) or O ^(r) x ^(r) (t): State at frame t of the r-th training pattern for word w T ^(r) : words several frames of training patterns of the r th for _{w μ i: b i (y} ) mean vector sigma _i of: b _i (y) of the covariance matrix xi] _i: defining a probability distribution of the observed vectors in state i Set of parameters (ξ _i = {μ _i , Σ _i }) λ _i = [ξ _i , {a _ij } _{j = 1, ..., I + 1} ]: Set of parameters of state i λ = {λ _i }: Set of all parameters (a model having λ as a parameter is also referred to as model λ) P (Y | λ): Probability that observation vector series Y occurs from model λ Density P (O | λ): Probability that the observed label sequence O occurs from the model λ π _i : Probability that the state i occurs at t = 1 First, a method for learning the continuous probability distribution HMM will be described.

The problem is that the likelihood function P (Y ⁽¹⁾ , Y ⁽²⁾ , ^... , Y ^(R) | λ) is applied to the prepared training patterns of r = ^{1 to R.}
Is to estimate the parameter λ that maximizes.

If Y ^(r) are independent of each other,

[0050]

[Equation 11]

Is given by Where the following auxiliary function Q
Define (λ, λ ').

[0052]

[Equation 12]

At this time, the following can be said. Q (λ, λ ')
If ≧ Q (λ, λ), P (Y ⁽¹⁾ , ..., Y ^(R) | λ ') ≧ P (Y ⁽¹⁾ ,
, Y ^(R) | λ), and the equal sign holds when λ ′ = λ. Therefore,

[0054]

[Equation 13]

If it is possible to obtain, by repeatedly applying (Equation 13) with λ ^* → λ, λ becomes P
(Y ⁽¹⁾ , ..., Y ^(R) | λ) stop point, that is, P (Y ⁽¹⁾ , ..., Y
The maximum value of ^(R) | λ) or the point that gives the saddle point is converged, and the change rate of P (Y ⁽¹⁾ , ..., Y ^(R) | λ) becomes equal to or less than a predetermined threshold value. A local optimum solution can be obtained by repeating the operation.

Next, a method of estimating parameters using Q (λ, λ ') will be described.

By modifying (Equation 12), the following equation is obtained.

[0058]

[Equation 14]

From the above description, if Q (λ, λ ') is regarded as a function of λ', and λ'where Q (λ, λ ')> Q (λ, λ) is found,
It becomes an updated version of λ, and P (Y ⁽¹⁾ , ..., Y ^(R)
| λ) is a constant value for λ ', so remove this

[0060]

[Formula 41]

Then, it is similar to finding λ ′ such that Q ′ (λ, λ ′)> Q ′ (λ, λ). However, here

[0062]

[Equation 15]

It is said that.

(Equation 14) becomes as follows.

[0065]

[Equation 16]

[0066] Re-estimate of when maximized for the first term π _i 'π _i π _i ^* is

[0067]

[Equation 17]

[0068] Re-estimate of when maximized for a _ij 'from the second term on the right side a _ij a _ij ^* is

[0069]

[Equation 18]

[0070] From the third term on the right side μ _i ', Σ _i' be maximized for, mu _i, re-estimated value of each ^{_{Σ i μ i *, Σ i}} * is

[0071]

[Formula 19]

[0072]

[Equation 20]

Here, ξ ^(r) _ij (t) is calculated as follows. That is,

[0074]

[Equation 21]

In summary,

[0076]

[Equation 22]

It is

At this time

[0079]

[Equation 23]

[0080]

[Equation 24]

The following recurrence formula holds. Therefore, α
^(r) ₁ (1) = 1 and an appropriate initial value is given to the parameter λ, and α ^(r) _j (t) is calculated according to ⁽ Equation 23) for t = 1 to T ^(r) +1, j = 1 to I + 1. To β ^(r) _{I + 1} (T ^(r) +1)
(15) can be calculated by sequentially calculating β ^(r) _i (t) according to (Equation 24) for t = T ^(r) +1 to 1 and i = I to 1 with = 1.

The actual calculation procedure for parameter estimation is as follows.

(1) L ₁ = ∞ (2) For i, j = 1 to I λ _i = {(a _ij ) _{j = 1, ..., I + 1} , μ
Give an appropriate initial value to _i , Σ _i }.

(3) r = 1 to R, t = 2 to T ^(r) , i =
For 1 to I + 1, α ^(r) _i (t) is calculated according to (Equation 23) with λ = {λ _i }.

(4) r = 1 to R, t = 2 to T ^(r) , i =
^Let β ^(r) _i (t) and ξ ^(r) _ij (t) for 1 to I + 1 be λ = {λ
_i } is calculated according to (Equation 24) and (Equation 22).

(5) For r = 1 to R, i, j = 1 to I + 1, the numerator of (Equation 18) (Equation 19) (Equation 20): a _{ij, num} (r), μ _{i, num} (r ), Σ _{i, num} (r) and denominator: Den _i (r) = a _{ij, denom} (r) = μ _{i, denom} (r) = Σ
Calculates _{i, denom} (r).

[0087] _{(6) a ij, μ i} , re-estimated value of Σ _i a _ij ^*, μ _i
^* , Σ _i ^* is calculated according to the following (number).

[0088]

[Equation 25]

(7) For i, j = 1 to I + 1, a _ij =
The re-estimated parameter set λ = {λ _i } is obtained by performing the substitution a _ij ^* , μ _i = μ _i ^* , Σ _i = Σ _i ^* .

(8) r = 1 to R, t = 2 to T ^(r) , i =
Parameter set λ obtained in step (7) for 1 to I + 1
Against

[0091]

[Equation 26]

Calculate

(9) | L ₁ −L ₂ | / L ₁ > ε, then L ₂ =
Leave L ₁ and go to step (4), otherwise end.

Ε in the step (9) is an appropriately small positive number that determines the width of convergence, and its value is selected as a practical value depending on the situation.

As described above, the continuous probability distribution HMM is obtained. In the present invention, the discrete probability distribution HMM is obtained by fuzzy clustering based on this, and the procedure is as follows.

(1) Fuzzy clustering of learning vectors is performed to calculate M clusters. Cluster name is C
_{1, C 2, ···, C} m, ···, and C _M.

(2) The HMM using the continuous HMM
The degree of occurrence of C _m (m = 1, ..., M) in each state is calculated.

Here, various methods of defining the degree of occurrence of each label can be considered. That is, (a) the weighted average of the probability density of the training vector by the degree of membership of the training vector with respect to C _m , (b) the probability density of the occurrence of a centroid of C _{m in} the state i, and (a) and (b) It can be considered that the clusters are normalized so that the sum of the occurrence degrees in each state of the cluster becomes 1. Here, as an example of the present invention,
The case where the normalization is not performed by the methods (a) and (b) will be described as an example. B _i (y) used in the following equation is the continuous HM
It is obtained from the estimated parameters of M. In this case, the occurrence degree b ^w of the cluster C _m in the state i of the word ^w
_im is given by the following equation.

[0099]

[Equation 27]

However,

[0101]

[Equation 28]

Where u ^{v (r)} _m (t) is the vector y
^The degree of membership of ^{v (r)} (t) in the cluster C _m . y ^{v (r)} (t)
Is the training vector of the r-th frame of the word v. Further, b ^w _i (y ^{v (r)} (t)) is the probability density of the vector y ^{v (r)} (t) in the state i of the continuous probability distribution model corresponding to the word w.

As the clustering method in the step (1), a well-known fuzzy clustering method can be used. As the data to be clustered, the HMM is used.
It is possible to use the entire set of feature vectors forming the pattern corresponding to the word speech of w = 1 to W used for learning.

Fuzzy clustering is performed as follows, for example. For simplification of notation, serial numbers are assigned to vectors forming all word sounds used for learning, and y ₁ , ..., Y _N are set, and the degree of membership (membership function) of y _{n to} the cluster C _m is determined. u _mn . In the normal clustering method, it is recognized whether a certain vector y _n belongs to the cluster C _m (u _mn = 1) or not (u _mn = 0), whereas in the fuzzy clustering method, y
It admits that _n belongs to several clusters to different degrees.

One specific method is to calculate the centroid (center vector, average vector) of each cluster C _m by y.
_0m (m = 1, ..., M) and the dissimilarity between y _n and the centroid y _0m is d _mn = d (y _n , y _0m ),

[0106]

[Equation 29]

One finds u _mn and y _0m that minimize. The objective function J _m is differentiated with respect to y _0m and u _mn , and the condition

[0108]

[Equation 30]

By using, the necessary conditions for locally minimizing J _m are shown as the following (Equation 31) and (Equation 32).

[0110]

[Equation 31]

[0111]

[Equation 32]

Here, F represents so-called fuzzyness, and 1 <F. When F → ∞, u _mn → 1 / M for m = 1, ..., M, and when F → 1,

[0113]

[Expression 33]

Therefore,

[0115]

[Equation 34]

[0116] That is, as F increases, y _n
The ambiguity of which cluster belongs to increases, and as F approaches 1, it approaches so-called hard clustering that uniquely determines the cluster to which y _n belongs.

The actual procedure of fuzzy clustering is as follows. (1) The training vector set is appropriately divided into C ₁ , ..., C _M to form initial clusters.

Initialize u _mn appropriately. (2) Obtain the average vector y _0m of each cluster according to ( _Equation 31). (3) When y _n ≠ y _0m , u _mn is updated using the result of step (2) according to ( _Equation 32). When y _n = y _0m , u _mn = 1 when m = n, and u _mn = 0 when m ≠ n. (4) If the convergence condition is satisfied, the process is terminated, and if not, the process returns to step (2).

[0119] As the convergence condition in step (4), in the iterative calculation, the amount of change between u _mn after being updated with the previous u _mn is updated when the e _mn, e _mn
It is conceivable that the case where the total sum of m and n of ² becomes equal to or less than a predetermined convergence judgment value ε is converged, and the case where it does not reach the convergence judgment value is not converged.

1 and 2 show an embodiment of the HMM creating apparatus of the present invention, the structure and operation of which will be described at the same time.

The feature extraction unit 101 converts the speech signals of the training words r = 1 to R ^w prepared for creating the model corresponding to the word w (= 1, ..., W) into the feature vector by the well-known method. series

[0122]

[Equation 35]

Convert to

The word pattern storage unit 102 is RAM, RO
M, various discs, storing R ^w learning words for creating the model λ ^w in the form of the feature vector series.

The buffer memory 103 takes out R ^w word patterns for w stored in the word pattern storage unit 102 and temporarily stores them.

The parameter estimation unit 104 uses the model λ
Perform steps (1) to (9) to create ^w, and use the word w
Estimate the model λ ^w corresponding to.

The first parameter storage unit 105 temporarily stores the re-estimated value of the parameter obtained in the step (6). The parameter estimation unit 104 re-estimates using the value of the parameter storage unit 105.

The fuzzy clustering unit 106 is stored in the word pattern storage unit 102.

[0129]

[Equation 36]

The feature vector set is fuzzy clustered into M clusters. T ^{w (r)} is the number of frames of the r-th training pattern of word w. The label of the m-th cluster is C _m and the centroid is y _0m .

The vector membership degree storage unit 107 stores the degree of membership U ^{w (r)} (t) = (u ^{w (r)} _{1 in} each cluster of y ^{w (r)} (t) obtained by the fuzzy clustering unit 106. (t), ..., u
^{w (r)} _M (t)) ^T is stored in a form that can be referred to by a combination of w, r, and t. Here, U ^{w (r)} (t) is referred to as a membership vector of y ^{w (r)} (t).

The label generation degree calculation unit 108 uses the probability density function of the model λ ^w stored in the parameter storage unit 105 to determine the degree of membership U ^{w (r)} (t) stored in the vector membership degree storage unit 107. And the vector y ^{w (r)} (t) stored in the word pattern storage unit 102 according to (Equation 27), the degree of occurrence of C _m in the HMM state i of the word w b ^w _im
To calculate.

The second parameter storage unit 109 stores the word w =
A means for storing parameters corresponding to 1 to W,
The parameters corresponding to the respective words w = 1, ..., W are stored in the parameter storage units 1 ,. That is, the transition probability corresponding to each state of each word is read from the first parameter storage unit 105 and stored in a form that can be referred to by w, i, j.
Further, the degree of label occurrence in each state of each word is read from the label occurrence degree calculation unit 108, and w,
It is stored in a form that can be referred to by i and m.

FIGS. 3 and 4 show another embodiment, in which the vector membership degree storage unit 107 of FIGS. 1 and 2 is replaced with a centroid storage unit for storing a centroid, and the label generation degree calculation unit 1 is used.
In 08, b ^w _im is calculated as b ^w _im = b ^w _i (y _0m ).

As described above, the discrete probability distribution HMM is created.

As described above, the continuous probability density distribution HMM is first created, the set of vectors forming the pattern set used for learning is fuzzy clustered, and the occurrence degree b _im of the cluster m in the state i of the HMM is calculated as the continuous probability. It is obtained by using the probability density obtained as the distributed HMM and converted into the discrete probability distribution HMM.

Next, the structure and operation of an apparatus for recognizing an actual input voice using the above model will be described at the same time.

FIG. 7 is a block diagram of the recognition device.

The feature extraction unit 401 is
The configuration and function are exactly the same as those of the feature extraction unit 101 in FIG.

The codebook 403 stores the centroid of each cluster stored in the cluster vector storage unit of the HMM creating apparatus shown in FIGS. 1, 2, 3 and 4.

The fuzzy vector quantizer 402 outputs the feature vector y (t) output from the feature extractor 401 and the representative vector y _0m (m = 1, ..., M) of each cluster stored in the codebook. To fuzzy vector quantization of y (t). That is, from the above formula, y (t)
Degree of membership u _m (t) for cluster C _m (m = 1, ..., M)
To calculate. That is, y (t) is the membership vector (u ₁ (t) ,.
.., u _M (t)) converted to ^T.

The parameter storage unit 404 stores the parameters shown in FIGS.
The parameter storage unit 109 has exactly the same configuration and function as the parameter storage unit 109 of FIGS. 3 and 4, and the parameter storage unit w stores the model parameters corresponding to the word w (= 1, ..., W). Remembered

The likelihood calculating section 405 calculates the likelihood of each model for the membership vector sequence obtained at the output of the fuzzy vector quantizing section 402, using the contents of the parameter storage section 404. That is, the likelihood calculation unit w uses the contents of the parameter storage unit w. The likelihood is calculated by calculating the occurrence degree f ^w _i (y (t)) of y (t) in state i.

[0144]

[Equation 37]

Given that b _i (o (t)) in (Equation 5) is f ^w _i (y (t)) and a _ij is a ^w _ij , ( _Equation 1), (Equation 2), (Equation 2) 3) etc. are used. When calculating (Equation 1), for Y ^(r) in ^(Equation 26)

[0146]

[Equation 38]

The calculation of α _{I + 1} (T + 1) for the input pattern Y is performed in exactly the same manner as the calculation of. T is the number of Y frames.

When (Equation 2) and (Equation 3) are used, the likelihood can be obtained by the well-known Viterbi method. When using (Equation 3), it becomes as follows. (1) Initial value setting The initial probability of the state i of the word w is π ^w _i , i = 1, ..., I
About (expression 39)

[0149]

[Formula 39]

(2) Calculation of recurrence formula [Formula 40] is executed for t = 2, ..., T, j = 1 ,.

[0151]

[Formula 40]

(3)

[0153]

[Equation 42]

Φ _{I + 1} (T + 1) in step (3) is the likelihood of the model w (word w) with respect to Y.

The comparison / determination unit 406 compares and determines which of the likelihood calculation units 1, ..., W included in the likelihood calculation unit 405 has the maximum output, and determines the corresponding word as the recognition result. This is output, and the calculation corresponding to (Equation 4) is executed.

Although the present embodiment has been described as recognizing a word, the present invention may be replaced with a phoneme, a syllable, or the like, and may be applied to patterns other than voice. .

Further, in the present embodiment, the distribution of the feature vector is described as following a single normal distribution in each state, but in the present invention, a more precise label generation degree is obtained by using a so-called mixed distribution. Of course, it is possible.

The present invention is applicable not only to the voice recognition device but also to other time series signal processing fields.

Each means of the present invention may be realized by software using a computer, or may be realized by using a dedicated hardware circuit having each of these functions.

[0160]

As is apparent from the above description,
The present invention provides an HMM creating means for creating a continuous probability density distribution HMM, a clustering means for fuzzy clustering a training vector set to give a label to each cluster, and a degree of occurrence of each label according to each cluster in each state of the HMM. Is provided with a training vector belonging to each cluster and a label generation degree calculating means for calculating from the probability density function in each state of the continuous probability density distribution HMM. Estimate error due to bias is eliminated and discrete HMM
It is possible to realize a model that takes advantage of the small calculation amount of.

Further, by using fuzzy clustering as the clustering, the number of clusters can be reduced and the estimation accuracy can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing a part of an embodiment of an HMM creating apparatus according to the present invention.

FIG. 2 is a block diagram showing the remaining part of the embodiment of the HMM creating apparatus according to the present invention.

FIG. 3 is a block diagram showing a part of an embodiment of an HMM creating apparatus according to the present invention.

FIG. 4 is a block diagram showing the remaining part of the embodiment of the HMM creating apparatus according to the present invention.

FIG. 5 is a block diagram illustrating a conventional example of a voice recognition device using an HMM.

FIG. 6 is an HMM configuration diagram showing a configuration of a continuous probability distribution HMM.

FIG. 7 is a block diagram showing an embodiment of a voice recognition device using an HMM constructed according to the present invention.

[Explanation of symbols]

 101 ... Feature extraction unit 102 ... Word pattern storage unit 103 ... Buffer memory 104 ... Parameter estimation unit 105 ... Parameter storage unit 106 ... Fuzzy clustering unit 107 ...・ ・ ・ Vector belonging degree storage unit 108 ・ ・ ・ Label generation degree calculation unit 109 ・ ・ ・ ・ Parameter storage unit

Claims (7)

[Claims] 1. An HMM for creating a continuous probability density distribution HMM.
Creating means, fuzzy clustering of the training vector set to give a label to each cluster, and the degree of occurrence of each label according to each cluster in each state of the HMM, the training vectors belonging to each cluster And the continuous probability density distribution HMM
And a label generation degree calculating means for calculating from a probability density function in each state of HMM creation apparatus. 2. A state transition probability storing means for storing the state transition probability obtained by the HMM creating apparatus according to claim 1, and a label occurrence degree storing means for storing the occurrence degree of each label in each state. HMM characterized by
Storage device. 3. A fuzzy vector quantizing means for calculating the degree of membership of each vector of a feature vector sequence forming an input pattern, which belongs to a cluster according to claim 1, and stored in an HMM storage device according to claim 2. From the state transition probability and the degree of label occurrence in each state, the HMM described by the parameters stored in the HMM storage device,
A likelihood calculating device, comprising: a likelihood calculating unit that calculates a likelihood for the input pattern. 4. The likelihood calculation device according to claim 3 is provided for each recognition unit, the likelihood of each recognition unit model for an input signal is calculated, and the input signal is recognized as the recognition signal from the value of the likelihood. A recognition device characterized by determining which of the units. 5. The label generation degree calculation means calculates each of the training vectors from the probability density function of the state i of the continuous probability density distribution HMM when the cluster is C _m (m = 1, ..., M). obtains the probability density, calculates the weighted mean by membership for C _m of the training vectors in said probability density, according to claim 1, characterized in that the occurrence rate b _im of C _m a weighted average該荷in the state i HMM creation device. 6. The label generation degree calculating means represents C _m from the probability density function of the state i of the continuous probability density distribution HMM when the cluster is C _m (m = 1, ..., M). The HMM creating apparatus according to claim 1, wherein the probability density of the vector is calculated, and the probability density is defined as the degree of occurrence of C _m in state i, b _im . 7. A label occurrence rate calculating means is further from the _{b im, b im '= b} im / (b i1 + ··· + b iM) comprises generating the degree normalizing means for calculating, the normalized occurrence 7. The HMM creating apparatus according to claim 5, wherein the degree b _im ′ is the degree of occurrence of C _m in state i.