JP3000642B2 - Pattern recognition method and standard pattern learning method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention relates to a pattern recognition method for recognizing a pattern represented as a time series of a feature vector such as an audio signal, and a standard pattern thereof is automatically generated from learning data. The present invention relates to a standard pattern learning method configured in a symmetric manner.

[Conventional technology]

As a pattern recognition method based on prediction of a time-series pattern, a “neural prediction model” is known. This method is described in Japanese Patent Application No. 1-344214 (hereinafter abbreviated as Reference 1) and the Transactions of the Acoustical Society of Japan, October 1989.
This is described in detail on pages 175 to 176 (hereinafter abbreviated as reference 2). In the neural prediction model, the feature vector at the time i of the input pattern and the state j as the local distance between the feature vector at the time i of the input pattern and the j-th state of the standard pattern model composed of the finite state transition network. The distance between the prediction vector and the feature vector at time i calculated from a plurality of feature vectors of the input pattern before time i-1 by the predictor (multi-layer perceptron) attached to the input pattern is used.

As a learning method for automatically determining a model parameter that defines a neural prediction model by learning, a method in which an initial value of a model parameter is appropriately given by a random number or the like and is corrected to an optimal value by an iterative algorithm is disclosed in the above-mentioned document. Is disclosed.

[Problems to be solved by the invention]

As described above, in the neural prediction model, when predicting the feature vector of the input pattern at time i, the predictor determines the prediction vector using only the feature vector of the input pattern before time i-1. For this reason, prediction is appropriately performed particularly when a feature vector at a certain time has a strong correlation with a feature vector before that time.

However, in the opposite case, that is, when a feature vector at a certain time has a strong correlation with a feature vector after that time, the prediction accuracy may not be improved.
For example, it is considered that the correlation between the rising part of the plosive and the transient part of the subsequent vowel after the plosion is stronger than the correlation with the silent part before the plosion.

Further, in the learning method of the model parameters, an appropriate method for setting the initial value is not known, and the initialization using a random number has been used so far. However, since the iterative learning method given in the above-mentioned documents 1 and 2 is an algorithm that converges to the minimum point of the learning evaluation function, it is desirable that the initial value of the parameter be given by a random number. There is a case where the learning ends with the convergence to the minimum point.

An object of the present invention is to predict a feature vector at a certain point in time, and incorporate not only a correlation with a feature vector before that point in time, but also a correlation with a feature vector after that point in time, thereby achieving a more accurate time series. It is to provide a pattern recognition method based on pattern prediction.

It is another object of the present invention to provide a standard pattern learning method for setting an initial value that enables better estimation of a parameter when a parameter of a standard pattern model is determined by iterative learning.

[Means for solving the problem]

A first invention is a pattern recognition method for recognizing an input pattern represented as a time series of a feature vector using a standard pattern model composed of a finite state transition network, wherein each state of the finite state transition network is input. Time i of the pattern
A predictor for calculating a prediction vector for the feature vector at time i from a plurality of feature vectors after +1; a local distance between the feature vector at time i of the input pattern and the j-th state of the finite state transition network d (i,
j) a distance D (a _i , A) between a feature vector a _i at time i of the input pattern and a prediction vector A _i (j) for the feature vector at time i by the predictor of state j.
_i (j)).

A second invention is a pattern recognition method for recognizing an input pattern represented as a time series of a feature vector using a standard pattern model composed of a finite state transition network, wherein each state of the finite state transition network is input. Time i of the pattern
Time i of a plurality of feature vectors and an input pattern before −1
A predictor that calculates a prediction vector for a feature vector at time i from a plurality of feature vectors after +1; and a localizer between the feature vector at time i of the input pattern and the j-th state of the finite state transition network. The distance d (i,
j) a distance D (a _i , A) between a feature vector a _i at time i of the input pattern and a prediction vector A _i (j) for the feature vector at time i by the predictor of state j.
_i (j)).

A third invention is a pattern recognition method for recognizing an input pattern represented as a time series of a feature vector using a standard pattern model composed of a finite state transition network, wherein each state of the finite state transition network is input. Time i of the pattern
A forward predictor that calculates a forward prediction vector for the feature vector at time i from a plurality of feature vectors before -1;
A backward predictor for calculating a backward predictive vector for the feature vector at time i from a plurality of feature vectors of the input pattern after time i + 1, wherein the forward predictor for the feature vector a _i at time i of the input pattern and the state j Predicted distance d _F between the predicted vector A _i ^F (j) and the feature vector at time i according to
(A _i , A _i ^F (j)) between the feature vector a _i at time i of the input pattern and the backward predicted vector A _i ^B (j) for the feature vector at time i by the backward predictor in state j. D (d _F (a _i , A _i ^F (j)), d _B (a _i , A _i ^B (j) calculated from the backward predicted distance d _B (a _i , A _i ^B (j)) ))),
Local distance d (i, j) between the feature vector at time i of the input pattern and the j-th state of the finite state transition network
It is characterized by using as.

According to a fourth aspect, when a multilayer perceptron is used as a predictor associated with each state of the finite state transition network in the pattern recognition method according to the first or second or third aspect, its parameters are automatically determined from learning data. A standard pattern learning method, wherein a representative vector calculated from learning data is used as an initial value of a threshold value of a unit of an output layer of the multilayer perceptron associated with the state j.

[Action]

In the pattern recognition method and the standard pattern learning method of the present invention, the standard pattern model of each recognition target category is represented by a finite state transition network having a start state and an end state,
Each state of the finite state transition network has its own predictor. Each predictor receives as input a fixed-length feature vector sequence cut out from the feature vector time series of the input pattern after time i + 1, and outputs a prediction vector for a feature vector to appear at time i. The distance between this prediction vector and the feature vector actually input at time i is defined as a prediction error. The distance between the input pattern and the finite state transition network, which is the standard pattern, is, for example, the possible state transition when the start state and the end state are connected and the state transition is performed in synchronization with the time of the input pattern time series. Then, the transition that minimizes the cumulative value of the prediction error along the transition is determined by the dynamic programming, and the obtained cumulative value of the prediction error along the optimal transition is defined as the distance. The parameters of the predictor constituting the standard pattern are determined by learning using the steepest descent method using the prediction error as an evaluation function.

Hereinafter, the pattern recognition and standard pattern learning method of the present invention will be described in more detail. In the description, the case of recognizing a voice pattern will be described as an example, but the present invention can be similarly applied to other time-series patterns by reading the voice pattern portion into a pattern vector sequence.

Predictor according to a first aspect of the invention the time i + 1 feature vector series of subsequent input speech a _{i + 1,} a _{i + 2,} to predict a feature vector a _i should emerge from ... to time i in the time axis backward. The prediction vector A _i (j) by the predictor associated with the j-th state is represented by the following equation.

A _i (j) = F (W _j , a _{i + 1} , a _{i + 2} ,..., A _{i + τ} ) (1) where F (·) is associated with state j characterized by parameter W _j Is a nonlinear vector value function that gives the input / output relationship of the predicted predictor. Here, W _j is represented as a plurality of parameters. τ is the number of input speech feature vectors used for prediction. When such a predictor that performs backward prediction on the time axis is used, the accuracy of prediction when a feature vector at a certain time point has a strong correlation with a feature vector after that time point is improved, and thus recognition performance is improved. It is expected. For example, it is considered that the present method functions particularly effectively because the rising part of the plosive sound of the voice has a stronger correlation with the transient part of the subsequent vowel after the rupture than with the silent part before the rupture.

When a multilayer perceptron is used as such a predictor, a specific expression of Expression (1) is given as follows.

Here, U ₀ (j), U ₁ (j),..., U _τ (j) are coupling coefficient matrices between the units of the perceptron, θ ₀ (j), θ ₁
(J) is the threshold vector of the unit of the output layer and the intermediate layer, f
(•) represents a vector obtained by applying a sigmoid function to each component of the argument vector. In this case, the parameter W _{j in} equation (1) is W _j = {U ₀ (j), U ₁ (j),..., U _τ (j), θ ₀ (j), θ ₁ (j)} (3) It is. Although an example in which a three-layer perceptron is used is shown here, the same applies to other cases such as four layers. Regarding multilayer perceptrons, see the publication "PDP
Model ”(Sangyo Tosho, 1989).

The specific expression of the local distance d (i, j) between the feature vector a _i at time i of the input voice and the j-th state of the standard pattern model (finite state transition network) is the above-described prediction vector Using A _i (j), for example, is given as follows:

d (i, j) = D (a _i , A _i (j)) (4) = {a _i −A _i (j)} ² (5) Here, when the square distance is used as the distance between the vectors However, the following discussion holds in the same way even when a distance between other vectors (a Mahalanobis distance or the like) that is generally used is used.

The discussion so far relates to the predictor according to the first invention, but the predictor according to the second invention replaces the expression (1) representing the input / output of the predictor in the above discussion with the following expression. Given by

A _i (j) = F (W _j , a _i-1 , a _i-2 , ..., a _{i + τ1} , a _{i + 1} , a _{i + 2} , ..., a
_{i + τ2} ) (6) Here, τ ₁ and τ ₂ represent the number of feature vectors of the input speech before and after time i used for prediction. Also in this case, similarly to the case of the first invention, the nonlinear vector value function F (·)
Can be configured using a multilayer perceptron. According to this method, it is possible to improve the prediction accuracy of a pattern having a strong correlation in both the forward and backward directions on the time axis.

Next, the method of the third invention will be described. In this case, two types of predictors, a forward predictor and a backward predictor, are attached to the j-th state of the finite state transition network. The prediction vector given by each predictor is represented by the following equation.

^{_{A i F (j) = F}} (W i F, a i-1, a i-2, ..., a i + τ1) (7) A i B (j) = F (W j B, a i + 1, a _{i + 2} , ..., a _{i + τ2} ) (8) Forward prediction vector A _i ^F (j) obtained from the forward predictor
, A forward prediction distance d _F (a _i , A _i ^F (j)) is defined as follows:

d _F (a _i , A _i ^F (j)) = {a _i −A _i ^F (j)} ² (9) Also regarding the backward predicted distance d _B (a _i , A _i ^B (j)) The same is true.

d _B (a _i , A _i ^B (j)) = {a _i −A _i ^B (j)} ² (10) Again, the square distance is used as an example of the distance, but other distances are used. , The following discussion can be similarly performed.

The local distance d (i, j) between the feature vector of the input voice at time i and the j-th state of the standard pattern model (finite state transition network) is given by the following equation.

d (i, j) = D (d F (a i, A i F (j)), d B (a i, A i B (j))) (11) where the function D (· defining the local distance For example, a method in which the smaller of the forward predicted distance and the backward predicted distance is selected as the local distance can be considered. According to this method, the more accurate one of the forward prediction and the backward prediction can be automatically selected and used, so which time series to be handled has a stronger correlation in the forward or backward direction of the time axis. Without knowing in advance, highly accurate prediction can be realized.

As described above, when the local distance is given by any of the first, second, and third inventions, the overall distance (cumulative distance) between the input pattern and the finite state transition network, which is the standard pattern, is calculated. Can be defined. Known definitions of this distance include, for example, a definition in which the state transition is deterministic (DP matching) and a stochastic definition (hidden Markov model). “DP matching” and “Hidden Markov model” are described in detail in the publication “Speech Recognition by Stochastic Model” (edited by the Institute of Electronics, Information and Communication Engineers, written by Seiichi Nakagawa, 1988, hereinafter abbreviated as Ref. 3). For example
When the definition by DP matching was adopted, the cumulative distance between the input pattern and the standard pattern (finite state transition network) linked the start state and the end state, and performed the state transition in synchronization with the time of the input pattern. Of the possible state transitions,
The transition that minimizes the cumulative value of the prediction error along the transition is determined by dynamic programming, and is given by the obtained cumulative value of the prediction error along the optimal transition. The specific algorithm is given in detail in References 1 and 2.

Next, a standard pattern learning method for automatically determining parameters characterizing a standard pattern model (finite state transition network) from learning data will be described. As the learning method, the methods given in References 1 and 2 can be applied as they are to the learning of the standard pattern model of the present invention. In this method, the initial values of the model parameters are appropriately given by random numbers, etc.
In this method, the cumulative value of the prediction error with respect to the learning data is used as an evaluation function to correct the value to an optimum value by an iterative correction algorithm. However, since the learning method based on this iterative correction is an algorithm that converges to a minimum point of the learning evaluation function, if the initial values of the parameters are given by random numbers, the learning method remains converged to an undesirable minimum point of the evaluation function. There is a case where learning ends. The standard pattern learning method according to the present invention provides a parameter initial value setting method that solves this problem when a multilayer perceptron is used as a predictor and enables better estimation of model parameters than random numbers.

In this case, the parameters to be estimated are the parameters of the equation (2). Here, the initial values of these parameters are given as follows.

Here, _j is a representative vector calculated from learning data by a method described later. Also the symbol Represents that an initial value is given by a random number whose absolute value is sufficiently smaller than _j . By performing such initialization, Expression (2) can be approximated as follows at the beginning of the iterative correction learning.

A _i (j) to _j (18) By this approximation, the prediction vector becomes a constant value regardless of the feature vector of the input speech. At this time, the local distance in Expression (4) is as follows.

d (i, j) = {a _i − _j } ² (19) This is between the feature vector at time i of the input pattern and the feature vector at time j of the standard pattern used in speech recognition by ordinary DP matching. Can be considered as a distance. Therefore, a feature vector (representative vector) of a standard pattern created from learning data can be used as _j as an initial value of θ ₀ (j). Specific _j
An example of the setting method is shown below. M-th of recognition target category s
The feature vector of the j-th frame of the learning data (m = 1,..., M _s ) is b _j (s, m), (j = 1,..., J). Here, it is assumed that the difference between the lengths of the individual occurrences of the training data has been normalized using DP matching or the like. At this time, the initial value θ ₀ (j) of the predictor of the state j of the standard pattern model of the category s is set as follows.

When the initial values are set in this way, the predictor outputs a representative standard pattern regardless of the feature vector of the input pattern at the beginning of learning, so that the zeroth approximation has started from normal pattern matching. Compared to initializing the model with random numbers, the possibility of convergence to an undesirable minimum is greatly reduced.

〔Example〕

1 to 5 show a flowchart of recognition by the pattern recognition method of the present invention. The input pattern feature vector time series a ₁ ,..., A I of length _I , and the parameters of the standard pattern model are externally given. Suppose This flowchart embodies the recognition method described in the section of operation, and the notation of variables and the like follows the one given there. However, the subscript s of the variable represents the category (s = 1,..., S) of the recognition target. Hereinafter, description will be given along the flow.

In step 101 of FIG. 1, variables are initialized. The details are shown in FIG. Steps 201 to 2 in FIG.
Initial setting of the counter is performed in 03. In step 204, the storage area for the local distance d _s (i, j) and the accumulated prediction error g _s (i, j) is initialized. In steps 205 to 210, the counter is incremented and the condition is determined, and all s, i, j
Are performed in step 204. Step 2
Numeral 12 sets the value at the start point of the cumulative prediction error of each category s.

Returning to FIG. 1, in steps 102 to 104, the counter is initialized. In step 105, the local distance is calculated. This is a portion corresponding to the first, second, and third inventions. In the case of the first invention, the local distance is calculated in steps 301 to 302 in FIG. The calculation in step 301 corresponds to the equation (2) described in the section of the operation, and the predictor is realized by a perceptron having a three-layer structure. Step 302 is a case where the square distance of Expression (4) is used as the local distance.

In the case of the second invention, the local distance is calculated in steps 401 and 402 in FIG. The calculation in step 401 is the realization of the predictor of equation (6) using a three-layer perceptron. U ₀ ^(s) , U _k ^(s) , V _k ^(s) , θ ₀ ^(s) (j), θ ₁ ^(s) (J)
Are parameters characterizing the three-layer perceptron.
The function f (•) is a sigmoid functio
n). In step 402, the local distance is calculated as 2 in equation (4).
This is the case where the riding distance is used.

In the case of the third invention, the local distance is calculated in steps 501 to 507 in FIG. The calculation in step 501 is the realization of the forward predictor of equation (7) with a three-layer perceptron.
The calculation in step 502 is realized by realizing the backward predictor of equation (8) with a three-layer perceptron. Step 503 is the case where the squared distance of equation (9) is used as the predicted forward distance, and step 504 is the case where the squared distance of equation (10) is used as the predicted backward distance. In step 505, the smaller of the forward predicted distance and the backward predicted distance is selected, and in steps 506 and 507, the value is set as the local distance.

Returning to FIG. In steps 106 to 115, using the local distance given in step 105, the distance (cumulative prediction error) between the input pattern and the standard pattern model is calculated based on the dynamic programming given in documents 1 to 3. I'm calculating. Steps 106 to 108 perform the recurrence formula calculation of the dynamic programming. In steps 109 to 114, the increment of the counter and the condition judgment are performed, and the calculation is performed for all frames, categories, and states. In step 115, a category having the smallest cumulative prediction error at the terminal point is selected as a recognition result.

The standard pattern learning method of the present invention provides a parameter initial value setting method in the standard pattern learning method for automatically determining parameters characterizing a standard pattern model (finite state transition network) from learning data. 6 and 7 show the flow of processing of the initial value setting method given to the equations (12) to (21) described in the section of operation. In step 603 of FIG. 6, a predictor (multi-layer perceptron) calculates the average value of the feature vectors of a plurality of learning data according to equation (21).
(The threshold value vector of the unit of the output layer) is set as the initial value. In FIG. 7, a threshold vector (step 703) and an inter-unit coupling coefficient matrix (step 704) of each state of each category are initialized with random numbers according to equation (12). In steps 703 and 704, the RND (ma
g) represents a uniform random number whose absolute value is less than mag. As described in the operation section, a value sufficiently smaller than the absolute value of the initial value of θ ₀ ^s (j) is used as mag.

〔The invention's effect〕

As described above, according to the present invention, when predicting a feature vector at a certain time, not only a correlation with a feature vector before that time but also a correlation with a feature vector after that time are incorporated. To provide a pattern recognition method based on highly accurate time-series pattern prediction and a standard pattern learning method for setting initial values that enable better estimation of parameters when repeatedly determining parameters of a standard pattern model. Can be.

[Brief description of the drawings]

FIG. 1 is a diagram showing a flowchart when DP matching is employed as a definition of a cumulative prediction error in the pattern recognition method of the present invention, and FIG. 2 is a diagram showing a flowchart showing an initialization algorithm in the flowchart of FIG. FIG. 3 is a diagram showing a flowchart of calculating a local distance when the predictor of the first invention is used, and FIG. 4 is a flowchart of calculating a local distance when using the predictor of the second invention. FIG. 5 is a diagram showing a flowchart for calculating a local distance when the predictor of the third invention is used. FIGS. 6 and 7 show initial values of model parameters by the standard pattern learning method of the present invention. FIG. 4 is a diagram illustrating a flowchart for setting a value.

Continuation of the front page (56) References JP-A-4-324500 (JP, A) Proceedings of the Acoustical Society of Japan Fall Meeting, 1990, 2-P-17, "Using a semisyllable neural prediction model Speech recognition ”, p. 163-164, (announced on September 20, 1990) Proceedings of the Acoustical Society of Japan Fall Meeting, 1990, I-8, 1-8-22, "Speech Recognition by Neural Network Predictive HMM", p. 43-44, (published September 19, 1990) Transactions of the Institute of Electronics, Information and Communication Engineers, Vol. J 73-D-II, No. 8, August 1990, “Unspecified speaker speech recognition using a neural prediction model”, p. 1315-1321, (published August 25, 1990) IEICE Technical Report [Voice], Vol. 91, No. 95, SP91-14, “Hierarchical neural network with time series processing function”, p. 63-70, (issued on June 20, 1991) IEICE Technical Report [Voice], Vol. 89, No. 90, SP89-23, “Speech Recognition Using Prediction Model by Neural Network”, p. 81-87, (published June 22, 1989) IEICE Technical Report [Voice], Vol. 89, No. 340, SP89-83, "Neural Network Driven HMM", p. 55-62, (issued December 14, 1989) Proceedings on 1990 IEEE International Conferencing on Acoustics, Speech and Signal Processing, Vol. 1, "S8.8 Speaker-Independence World Recognition usng a Neural Prediction Model", p. 441-444, (3-6, April 1990) (58) Fields investigated (Int. Cl. ⁷ , DB name) C10L 3/00 535 C10L 3/00 521 C10L 3/00 539 C10L 9/10 301 G06F 15 / 18 JICST file (JOIS) INSPEC

Claims (4)

(57) [Claims] 1. A pattern recognition method for recognizing an input pattern represented as a time series of feature vectors using a standard pattern model composed of a finite state transition network, wherein each state of the finite state transition network is an input pattern. Time i +
A predictor for calculating a prediction vector for a feature vector at time i from a plurality of feature vectors subsequent to 1; a local distance between the feature vector at time i of the input pattern and the j-th state of the finite state transition network d (i, j)
The feature vector a _i at time i of the input pattern
And a distance D (a _i , A) between a prediction vector A _i (j) for a feature vector at time i by the predictor in state j.
_i (j)) using a pattern recognition method. 2. A pattern recognition method for recognizing an input pattern represented as a time series of a feature vector using a standard pattern model comprising a finite state transition network, wherein each state of the finite state transition network is an input pattern. Time i-
Time i + of a plurality of feature vectors before 1 and the input pattern
A predictor for calculating a prediction vector for a feature vector at time i from a plurality of feature vectors subsequent to 1 and a local vector between the feature vector at time i of the input pattern and the j-th state of the finite state transition network. Distance d (i, j)
The feature vector a _i at time i of the input pattern
And a distance D (a _i , A) between a prediction vector A _i (j) for a feature vector at time i by the predictor in state j.
_i (j)) using a pattern recognition method. 3. A pattern recognition method for recognizing an input pattern represented as a time series of a feature vector using a standard pattern model composed of a finite state transition network, wherein each state of the finite state transition network is an input pattern. Time i-
1. A forward predictor that calculates a forward prediction vector for a feature vector at time i from a plurality of feature vectors before 1 and a backward predictor that calculates a backward prediction vector for a feature vector at time i from a plurality of feature vectors of the input pattern after time i + 1. A forward prediction distance d _F between a feature vector a _i at time i of the input pattern and a forward prediction vector A _i ^F (j) for the feature vector at time i by the forward predictor in state j. (A
_i , A _i ^F (j)) and the backward vector between the feature vector a _i at time i of the input pattern and the backward predicted vector A _i ^B (j) for the feature vector at time i by the backward predictor of state j. predicted distance d _B amount is calculated from the ^{_{(a i, a i B (}} j)) D (d F (a i, a i F (j)), d B (a i, a i B (j)) ) Is used as a local distance d (i, j) between the feature vector at time i of the input pattern and the j-th state of the finite state transition network. 4. When a multilayer perceptron is used as a predictor associated with each state of a finite state transition network in the pattern recognition method according to any one of claims 1 to 3, its parameters are automatically obtained from learning data. A standard pattern learning method to be determined, wherein a representative vector calculated from learning data is used as an initial value of a threshold value of a unit of an output layer of a multilayer perceptron associated with a state j.