JPH08297495A - Method and device for learning signal pattern recognization device - Google Patents

Abstract

PURPOSE: To provide a method and a device for learning a signal pattern recognization device which can accomplish recognization with an efficiency and a precision higher than those of conventional methods using a small number of samples for adaption. CONSTITUTION: A feature extracting section 20 has a learnable operative parameter and converts an inputted signal pattern into a specified feature signal pattern for recognizing signal pattern. And, a classifying section 3 outputs at least one class with higher tolerance as a result of recognization of signal pattern by performing classifying processing for classsifying the converted feature signal pattern into at least one class out of plural predetermined classes. Then, an adjusting section 21 adaptively renews the learnable operation parameter of the feature extracting section 20 without changing the classifying process of the classifying section 3 so that the feature signal pattern outputted from the feature extracting section 20 approaches to the class model of known signal patteren when inputting the signal pattern of known class attribute to the feature extracting section 20.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning method and a learning device for a signal pattern recognition device for recognizing various signal patterns, and in particular, it is capable of expressing numerical values such as voices, characters, images, and medical measurements. The present invention relates to a learning method and a learning device for a signal pattern recognition device for recognizing any physical or chemical signal pattern.

[0002]

2. Description of the Related Art A conventional signal pattern recognition device basically comprises a feature extraction section 2 and a classification section 3, as shown in FIG. The feature extraction unit 2 responds to the signal pattern stored in the signal data memory 1, converts the signal into a signal having a predetermined feature expression format, and further, the classification unit 3
Is a class which is a result of the classification process by setting the class boundary in the feature space formed by the features of the converted signal using the class model stored in advance in the class model memory 7, that is, performing the classification process. The recognition result data is stored in the pattern recognition result data memory 4.

The purpose of learning of such a signal pattern recognition device is to consider all operating conditions and environments that may occur in the future, such as a voice input signal pattern of an unknown speaker or a voice input signal pattern passing through an unknown telephone line. In order to recognize a character input signal pattern or the like by an unknown writer with high accuracy, the parameter format of the device is selected, and the state of the parameter is determined and set.

However, at the learning stage, it is impossible to know all the operating condition information that may occur in the future, that is, the signal pattern information as a result. In general, the accuracy which is the recognition performance of the signal pattern recognition device. In a new operating environment that was unknown during the learning phase. Clearly, it is desired that the signal pattern recognizer has somehow the ability to adapt to new operating conditions that were not known during the learning phase.

The conventional solution for solving the above-mentioned problem is to compare the operation parameters of the classification unit 3 of the signal pattern recognition apparatus that are learned in advance so that they can operate with a certain degree of recognition accuracy in the learning stage. It was updated to improve the recognition performance by using a small number of adaptive signal pattern samples.

However, updating the classifying unit 3 using only a small number of adaptation signal patterns is not sufficient to appropriately update the class boundaries of all classes handled by the recognition device. Because,
This is because, in principle, the information about the classification included in these minority signal patterns is limited to only the related classes, in other words, the classes to which those signal patterns belong, or the classes involved in the classification determination. This insufficiency becomes more serious as the number of classes handled by the signal pattern recognition device increases. In order to alleviate this insufficiency, several improvement methods have been proposed so far (for example, in Reference 1, “Y. Nakato et al.,“ A Study on Un
supervised Speaker Adaptation of Continuous Parame
ter HMM ", Acoustic Society Japan, SP90-67, pp.79-8
6, 1990 ”and reference 2“ C.-H. Lee et al., “Speaker A
daptation Based on MAP Estimation of HMM Parameter
s ", Proceedings of IEEE, ICASSP93, Vol. 2, pp.558-
561, April 1993 ". ). In order to reduce the locality in the sample feature space of updating that occurs due to the use of a small number of samples in the framework of updating the operating parameters of the classifying unit 3, some improvement method is related to distribution or the like in the sample feature space. It made assumptions and made additional updates based on those assumptions.

[0007]

However, in the case of voice signal pattern recognition, these assumptions include various factors, such as new operating conditions such as a new speaker and a new acoustic system related to a telephone line. There is no guarantee that it is optimal for the whole, and many adaptation samples are required to obtain excellent adaptation results and obtain high recognition accuracy after adaptation.
As a result, no signal pattern recognizer has been achieved that has the capacity to be adequately adapted using only a small number of samples.

The present invention solves the above problems and uses a small number of adaptive samples to achieve a higher efficiency and higher recognition accuracy than the conventional example, and a learning method for a signal pattern recognition apparatus. And to provide a learning device.

[0009]

A learning method for a signal pattern recognition device according to a first aspect of the present invention has an operation parameter that can be learned, and an input signal pattern is recognized as a signal pattern recognition device. For extracting a feature signal pattern for converting the feature signal pattern converted by the feature extracting means into at least one of a plurality of predetermined classes. By doing so, a learning method for a signal pattern recognition device, comprising: a classifying unit that outputs at least one class with higher likelihood as a recognition result of a signal pattern, wherein When input to the feature extraction means, the feature signal pattern output from the feature extraction means, so as to approach the similar model of the known signal pattern, Without changing the classification process of the serial classification unit, and updates the learning possible operating parameters of the feature extraction means adaptively.

A learning method of the signal pattern recognition device according to a second aspect is the learning method of the signal pattern recognition device according to the first aspect, wherein the feature signal pattern output from the feature extracting means and the known signal are used. In order to minimize the loss corresponding to the similarity between the pattern and the model,
It is characterized by adaptively updating the learnable operation parameters of the feature extraction means.

According to a third aspect of the present invention, there is provided a learning device for a signal pattern recognition device, which has an operation parameter that can be learned and converts an input signal pattern into a predetermined characteristic signal pattern for recognition of the signal pattern. At least one of a plurality of predetermined classes of the characteristic extracting means for converting and the characteristic signal pattern converted by the characteristic extracting means.
A learning device for a signal pattern recognition device, comprising: a classifying unit that outputs at least one class having a higher likelihood as a recognition result of a signal pattern by executing a classification process for classifying into one. , When the signal pattern whose class attribute is known is input to the feature extraction means, the classification of the classification means is performed so that the feature signal pattern output from the feature extraction means approaches the class model of the known signal pattern. The present invention is characterized by comprising adjusting means for adaptively updating the learnable operation parameter of the feature extracting means without changing the processing.

A learning device for a signal pattern recognition device according to a fourth aspect is the learning device for a signal pattern recognition device according to the third aspect, wherein the adjusting means is a feature signal pattern output from the feature extracting means. , The learning operation parameters of the feature extraction means are adaptively updated so that the loss corresponding to the similarity with the known model of the signal pattern is minimized.

According to a fifth aspect of the present invention, there is provided a learning method for a signal pattern recognizing apparatus, which has a learnable operation parameter and converts an input signal pattern into a predetermined characteristic signal pattern for recognizing the signal pattern. At least one of a plurality of predetermined classes of the characteristic extracting means for converting and the characteristic signal pattern converted by the characteristic extracting means.
A learning method for a signal pattern recognition device, comprising: a classification unit that outputs at least one class having a higher likelihood as a recognition result of a signal pattern by executing a classification process for classifying into one. , When a signal pattern whose class attribute is unknown is input to the feature extracting means, the unknown signal pattern is the maximum likelihood having the highest likelihood among the plurality of classes which are the recognition results output from the classifying means. It is characterized in that the learnable operation parameter of the feature extracting means is adaptively updated without changing the classification process of the classifying means so as to approach the class model.

According to a sixth aspect of the present invention, there is provided a learning method for a signal pattern recognition apparatus according to the fifth aspect, wherein the characteristic signal pattern output from the characteristic extraction means is the same as the unknown characteristic pattern. However, it is characterized in that the learnable operation parameter of the feature extracting means is adaptively updated so that the loss corresponding to the similarity with the model of the maximum likelihood signal pattern is minimized.

According to a seventh aspect of the present invention, there is provided a learning device for a signal pattern recognizing device, which has a learnable operation parameter and converts an input signal pattern into a predetermined characteristic signal pattern for recognizing the signal pattern. At least one of a plurality of predetermined classes of the characteristic extracting means for converting and the characteristic signal pattern converted by the characteristic extracting means.
A learning device for a signal pattern recognition device, comprising: a classifying unit that outputs at least one class having a higher likelihood as a recognition result of a signal pattern by executing a classification process for classifying into one. , When a signal pattern whose class attribute is unknown is input to the feature extracting means, the unknown signal pattern is the maximum likelihood having the highest likelihood among the plurality of classes which are the recognition results output from the classifying means. It is characterized by comprising adjusting means for adaptively updating the learnable operation parameter of the feature extracting means without changing the classification process of the classifying means so as to approach the class model.

A learning device for a signal pattern recognition device according to an eighth aspect is the learning device for a signal pattern recognition device according to the seventh aspect, wherein the adjusting means is a feature signal pattern output from the feature extracting means. , A feature that the learned parameter of the feature extracting means is adaptively updated so that the loss corresponding to the similarity with the unknown but maximum likelihood signal pattern type model is minimized. And

[0017]

In the learning method for the signal pattern recognition device according to claim 1 configured as described above, the feature extraction means has a learnable operation parameter, and the input signal pattern is a signal. For converting into a predetermined characteristic signal pattern for pattern recognition, the classification means classifies the characteristic signal pattern converted by the characteristic extraction means into at least one of a plurality of predetermined classes. By executing the classification processing of, at least one class with higher likelihood is output as the recognition result of the signal pattern. Here, when a signal pattern having a known class attribute is input to the feature extracting means, the feature signal pattern output from the feature extracting means approximates the class model of the known signal pattern to the classifying means. The operation parameters that can be learned by the feature extraction means are adaptively updated without changing the classification process of.

Further, in the learning method of the signal pattern recognition device according to claim 2, in the learning method of the signal pattern recognition device according to claim 1, preferably, the feature signal pattern output from the feature extraction means, The learnable operation parameter of the feature extracting means is adaptively updated so that the loss corresponding to the similarity with the known model of the signal pattern is minimized.

In the learning device of the signal pattern recognition device according to the third aspect of the present invention, the feature extracting means comprises:
Having an operation parameter that can be learned, the input signal pattern is converted into a predetermined feature signal pattern for recognition of the signal pattern, the classification means, the feature signal pattern converted by the feature extraction means, By performing a classification process for classifying into at least one of a plurality of predetermined classes, at least one class with higher likelihood is output as a recognition result of the signal pattern. Here, the adjusting means, when a signal pattern whose class attribute is known is input to the feature extracting means, makes the feature signal pattern output from the feature extracting means close to the class model of the known signal pattern. Further, the learnable operation parameter of the feature extracting means is adaptively updated without changing the classification process of the classifying means.

Further, in the learning device of the signal pattern recognition device according to claim 4, in the learning device of the signal pattern recognition device according to claim 3, the adjusting means is preferably output from the feature extracting means. The learnable operation parameter of the feature extracting means is adaptively updated so that the loss corresponding to the similarity between the feature signal pattern and the known model of the signal pattern is minimized.

In the learning method of the signal pattern recognition device according to claim 5 of the present invention, the feature extracting means is
Having an operation parameter that can be learned, the input signal pattern is converted into a predetermined feature signal pattern for recognition of the signal pattern, the classification means, the feature signal pattern converted by the feature extraction means, By performing a classification process for classifying into at least one of a plurality of predetermined classes, at least one class with higher likelihood is output as a recognition result of the signal pattern. Here, when a signal pattern whose class attribute is unknown is input to the feature extracting means, the unknown signal pattern has the highest likelihood among the plurality of classes which are the recognition results output from the classifying means. To approach the maximum likelihood class model,
The learnable operation parameter of the feature extraction means is adaptively updated without changing the classification process of the classification means.

Further, in the learning method of the signal pattern recognition device according to claim 6, in the learning method of the signal pattern recognition device according to claim 5, it is preferable that the feature signal pattern output from the feature extraction means is , The learning parameters of the feature extracting means are adaptively updated so that the loss corresponding to the similarity with the unknown but maximum likelihood signal pattern type model is minimized.

In the learning device of the signal pattern recognition device according to the seventh aspect of the present invention, the feature extracting means comprises:
Having an operation parameter that can be learned, the input signal pattern is converted into a predetermined feature signal pattern for recognition of the signal pattern, the classification means, the feature signal pattern converted by the feature extraction means, By performing a classification process for classifying into at least one of a plurality of predetermined classes, at least one class with higher likelihood is output as a recognition result of the signal pattern. Here, the adjusting means, when inputting a signal pattern whose class attribute is unknown to the feature extracting means, selects the unknown signal pattern among the plurality of classes which are the recognition results output from the classifying means. The learnable operation parameter of the feature extraction means is adaptively updated without changing the classification processing of the classification means so as to approach the maximum likelihood class model having a high likelihood.

Further, in the learning device of the signal pattern recognition device according to claim 8, in the learning device of the signal pattern recognition device according to claim 7, the adjusting means is preferably output from the feature extracting means. In order to minimize the loss corresponding to the similarity between the characteristic signal pattern and the unknown but maximum likelihood signal pattern type model,
The learnable operation parameter of the feature extraction means is adaptively updated.

[0025]

Embodiments of the present invention will be described below with reference to the drawings. <First Embodiment> FIG. 1 shows a signal pattern recognition apparatus according to the first embodiment of the present invention. The signal pattern recognition device according to the first embodiment has a learnable operation parameter,
A feature extraction unit 20 that converts an input signal pattern into a predetermined feature signal pattern for signal pattern recognition.
And a classification process for classifying the feature signal pattern converted by the feature extraction unit 20 into at least one of a plurality of classes corresponding to a plurality of class models determined in advance and stored in the class model memory 22. By executing
The classification unit 3 of the conventional example that outputs at least one class with higher likelihood as a recognition result of a signal pattern, and when a signal pattern with a known class attribute is input to the feature extraction unit 20, the class model memory 22 Without changing the classification process of the classification unit 3 so that the characteristic signal pattern output from the characteristic extraction unit 20 can be brought closer to the similar model of the known signal pattern by referring to the similar model stored in A feature extraction pattern adjusting unit 21 that adaptively updates the learnable operation parameter of the feature extracting unit 20 is featured. Here, the feature extraction pattern adjustment unit 21
Preferably, the feature extraction unit 20 can perform learning so that the loss corresponding to the similarity between the feature signal pattern output from the feature extraction unit 20 and the known model of the signal pattern is minimized. Adaptively update operating parameters.

In the first embodiment, the signal shown in FIG. 1 is used for the purpose of recognizing a signal pattern belonging to any one of M classes (C _i ; i = 1, ..., M). Consider a pattern recognizer. Here, the signal pattern is
It may be a fixed dimension vector or a variable length signal pattern (sequence of the vectors). This signal pattern recognition device includes a feature extraction unit 20 and a classification unit 3. Let Φ be the learned parameter set of the feature extraction unit 20, and let Λ be the learned parameter set of the classification unit 3. The feature extraction unit 20 converts the signal pattern x input from the signal data memory 1 into the signal pattern recognition device into a feature signal pattern Y and outputs the feature signal pattern Y to the classification unit 3. That is, the process of the feature extraction unit 20 is expressed by the following mathematical expression 1.

[0027]

## EQU1 ## Y = f (x; Φ)

Here, f (·; Φ) is a function representing a feature extraction function, and the function is determined by the operation parameter set Φ. The classifying unit 3 stores the class model ω _i (Λ) in advance in the class model memory 22 for each class C _i as a function of the motion parameter set Λ in the feature space formed by the converted features, and uses this. Then, the discriminant function g representing the degree of membership of the input signal pattern x belonging to each class C _i
The function value is calculated using _i (Y; Λ) or g _i (f (x; Φ); Λ). That is, the classification unit 3 outputs the class that is the classification result using the classification determination rule represented by the following Expression 2.

[0029]

## EQU2 ## C (x) = C _j , where j = argmin _i g _i (Y; Λ)

Here, C (•) is a classification processing operation,
argmin _i (·) is a function indicating the value of the variable i when the function value becomes the minimum. The use of this classification decision rule is
This means that the discriminant function g _i (Y; Λ) is defined such that its value decreases as the input feature signal pattern Y and the corresponding model ω _i (Λ) become more similar. The states or set values of the learnable operating parameter sets Λ and Φ have been learned in advance in the learning stage by using the learning sample set appropriately, that is, at least to achieve high accuracy with respect to the learning sample set. Suppose.

In the above situation, the adaptive learning method according to the present invention is implemented as follows. For the sake of clarity, the proposed method is based on the adaptive feature extraction-based pattern recognition method.
recognition method) (hereinafter referred to as the AFE method). The AFE method can also be implemented by two different methods, based on the presence or absence of information about the class to which the adaptation sample belongs. If the class to which the adaptive sample belongs is known, AFE can be implemented in a so-called supervised learning format. The SAFE method (Supervised Adaptive)
Feature Extraction-based pattern recognition metho
d). Further, when the class to which the adaptive sample belongs is unknown, the AFE method is executed in the unsupervised learning format. This is the USAFE method (UnSupervised Adaptive Fea
ture Extraction-based pattern recognition method)
Call. The USAFE method will be described in the second embodiment.

In the first embodiment using the SAFE method, first, the processing for adaptive learning is executed in the feature extraction parameter adjusting unit 21, and as shown in FIG. 1, in a new operating environment in which the class to which it belongs is known. Signal data x ₁ (εC _k ) which is an adaptation sample is input from the signal data memory 1 to the feature extraction unit 20. Feature extraction parameter adjustment unit 2
1 is a model of the class C _k (model of the characteristic signal pattern) in which the characteristic signal pattern Y ₁ converted from the signal data x ₁ which is the adaptation sample by the characteristic extraction unit 20 is stored in the type model memory 22. The operation parameter Φ that determines the feature extraction function f (x; Φ) of the feature extraction unit 20 is updated so as to approach ω _k (Λ). Specifically, when an input signal pattern with a known class attribute is input to the feature extraction unit 20,
The similarity between the model ω _k (Λ) of the class C _k and the characteristic signal pattern Y ₁ of the input signal pattern whose class attribute is known is
A loss l _k (Y ₁ ; Λ) represented by the following expression 3 is newly defined, and the operating parameter Φ is updated so that this loss becomes as small as possible.

[0033]

## EQU00003 ## l _k (Y ₁ ; Λ) = l _k (f (x ₁ ; Φ); Λ)

In order to reduce the loss, it is natural to think of this similarity as a kind of distance between the model ω _k (Λ) of class C _k and the characteristic signal pattern Y ₁ . Various existing methods such as the gradient method can be used to minimize the loss for such one adaptive sample. However, considering that the input signal pattern x has a variable length such as a voice signal, it is particularly preferable to refer to, for example, Document 3 “S. Katagiri et al.,“ A Generalized Probabilis ”.
tic Descent Method ", Acoustic Society Japan Procee
ding of Fall Conference, pp.141-142, September 1990 "and Reference 4" B.-H. Juang et al., "Discriminative Lear
ning for Minimum Error Classification ", IEEE Trans
action on Signal Processing, Vol. 40, No. 12, pp.3
043-3054, December 1992 ”, the well-known generalized probabilistic descent method.
It is effective to use an update rule based on the Descent method (hereinafter referred to as the GPD method).

An example of the SAFE method using the GPD method is a voice recognition device as shown in FIG. 3, for example. This voice recognition device, as shown in FIG.
2, an A / D converter 42, and a voice recognition signal pattern recognition device 43 having the same configuration as the device shown in FIG. The voiced voice is input to the microphone 41,
After being converted into an audio signal, it is input to the A / D converter 42. The A / D converter 42 A / D-converts the input voice signal at a predetermined sampling frequency to convert it into a digital voice signal and inputs it as an input signal pattern to the voice recognition signal pattern recognition device 43, and outputs the signal pattern. The voice recognition process is performed by performing the recognition process, and the voice recognition result data is output from the signal pattern recognition device 43.

The case of using the apparatus of FIG. 1 in the example of FIG. 3 will now be described. The signal input from the A / D converter 42 to the feature extraction unit 20 is a variable-length speech waveform signal x ₁ given for adaptive learning. The data of the signal x ₁ is input to the feature extraction unit 20, and in response to this, the feature extraction unit 20 executes the feature extraction processing to move the fixed-length time window on this signal x _1. However, the sequence of fixed-dimensional acoustic feature vectors, which is obtained by the following equation 4, is output as the feature signal pattern Y ₁ .

[0037]

[Formula 4] Y ₁ = {ν ₁ , ..., ν _τ , ..., ν _{T (1)} }

Here, T (1) is the length of the vector sequence Y ₁ of the feature signal pattern, and ν _τ is the τ th acoustic feature vector, and this vector is an application example of many speech recognition devices. In this case, a feature vector related to the short-time power spectrum such as cepstrum is used. Also,
It is assumed that the model ω _k (Λ) of the class C _k is represented as the reference signal pattern R in the characteristic signal pattern space by the following Expression 5.

[0039]

## EQU5 ## R = {ξ ₁ , ..., ξ _τ , ..., ξ _{T (r)} }

Here, T (r) is the length of the vector series R which is the reference signal pattern. In the GPD method used for learning the signal parameter recognition device, a misclassification measure is used in addition to the loss function reflecting the classification result and the discrimination function similar to the above. In the SAFE method of the first embodiment, these are all defined as the same function. C above
Any reasonable distance or a function thereof can be used for the similarity of _k , but in particular, a nonlinear time based on dynamic programming between the vector sequence R of the reference signal pattern and the characteristic signal pattern Y _1. Using the cumulative squared Euclidean distance D _θ obtained by accumulating the squared Euclidean distance between the feature vectors of R and Y _{1 associated} with the route θ within the expansion / contraction matching range used for axis expansion / contraction, It can be given by the following equation 6.

[0041]

(Equation 6)

Where d _k (f (x ₁ ; Φ); Λ) is the misclassification measure and ζ is a positive constant. Θ is the number of possible stretch matching paths. Further, here, although the learnable operation parameter set Λ is included in each function, it is used as a constant that is not updated. The adjustable variable in Equation 6 above is Φ. This is because the SAFE method fixes the state or set value of the operation parameter of the classification unit 3 and allows only the feature extraction unit 20 to learn. GPD with learning as a loss variable for this input signal x ₁ and, as a result, Φ that reduces the same discriminant function
For example, reference 5 “T. Komori et al.,“ FD Trai
ning of Dynamic Programming-Based Speech Recognize
rs ", Jounal of Acoustic Society Japan (E), Vol. 1
3, No. 6, pp.341-349, November 1992 ”can be implemented by using the update rule of the GPD method. This completes the learning process of the SAFE method for one adaptation signal pattern sample.

Then, S for this one adaptation sample
The learning process of the AFE method is performed using the adaptive signal pattern sample {x
₂ , x ₃ , ...} and repeat. This repetition can be similar to the repetition of the GPD method. In this way, the learning process of the SAFE method performed by continuously and adaptively giving a new adaptation sample is completed.

As demonstrated by the GPD method and the learning vector quantization which is one of the practical learning methods, this SAFE learning is performed by the recognizer as the learning progresses while replacing the adaptive sample. Input signal pattern samples under new operating conditions
The feature extraction parameter adjusting unit 21 is provided so that the feature signal pattern can be converted into a feature signal pattern that is similar to the feature signal pattern that is actually approximated by the model, which is used when learning a type model that is built in the recognition device in advance. The operation parameter Φ in the feature extraction unit 20 is gradually updated by.

In the first embodiment, in order to end the repetition of the above-mentioned update processing, it is preferable to use one end condition of the following examples. (I) When the value of loss for a certain adaptation pattern falls below a predetermined threshold value, the updating process is terminated. (II) If the recognition result for a certain adaptation pattern is correct, the updating process is terminated. (III) When the tabulation or average of the loss values for a plurality of adaptation patterns falls below a predetermined threshold value, the updating process is terminated. (IV) When the average recognition result for a plurality of adaptation patterns exceeds a predetermined threshold value, the updating process is terminated. (V) Using the pattern set independent of the adaptation pattern prepared for evaluation, replacing the adaptation pattern in each of the above (I) to (IV) with the evaluation pattern obtained from this new set. Then, each condition is verified. When any one of the above conditions (I) to (IV) is satisfied, the updating process is terminated. (VI) If the above conditions (I) to (V) are selectively or comprehensively combined and the combined conditions are simultaneously satisfied,
The update process ends.

In the embodiment using the GPD method,
It is possible to use various route usage methods including the case of using only the best route in the sense that the cumulative distance by controlling the parameter ζ is small, similarly to the GPD method.

Furthermore, in the above example, one reference pattern for each class is used, but formalization for a recognition device using a plurality of reference patterns for each class may be performed in exactly the same way as the GPD method. As the class model, any rational model such as a hidden Markov model or an artificial neural network (neural network) can be used.
As the feature extraction function, any rational function can be used according to the type of pattern to be recognized.

Further, an example of the update rule based on the GPD method will be described. Basically, the update rule depends on the configuration of the signal pattern recognition device used, in particular, the type of device operation parameter. Here, in particular, a distance classifier using a plurality of reference vectors is used for the classifying unit 3, and the feature extracting unit 20 performs feature extraction within a relatively short time window.
Using a filterbank calculation based on the Fourier transform,
A voice recognition device for recognizing M classes will be described below. The input audio signal x is converted into a feature vector series X of a power spectrum represented by the following Expression 7 by performing calculation processing such as clipping with a short time window and frequency conversion based on the Freier conversion. In the above description, an index such as {x ₁ , x ₂ , ...} Is added as an input signal for description, but in the description of this detailed example, the index of the signal is omitted for simplicity.

[0049]

X = {χ (1), ..., χ (t), ..., χ (T ₀ )}

Where χ (t) is the feature vector of the F-dimensional power spectrum at the t-th time window position,
T ₀ is the sequence length. In the first embodiment using the AFE method, this sequence is an input signal pattern input to the adaptive feature extraction unit 20. The feature extraction unit 20
A function calculation processor of the feature extraction function f (·; Φ) that executes filtering determined by the adaptive operation parameter Φ is provided, and the feature vector series X is represented by the following equations 8 and 9. It is converted into an S-dimensional feature vector Y filtered by the filter bank and output for each window position using the extraction function f (·; Φ).

[0051]

Ν (t) [s] = f (χ (t); Φ) [s]

Equation 9] Y = {ν (1), ..., ν (t), ..., ν (T 0)}

Here, ν (t) is a feature vector converted from χ (t). Therefore, the feature extraction function f (χ
(T); Φ) transforms the F-dimensional feature vector χ (t) into an S-dimensional feature vector ν (t), where the function f (χ
(T); Φ) [s] represents the s-th dimension component of the output of the function f (χ (t); Φ). Similarly below, [s] represents the s-th component of the corresponding vector. The feature vector Y is an output of the adaptable feature extraction unit 20 and is an input to the classification unit 3.

In the learning process by the AFE method based on the GPD method, the feature extraction parameter adjustment unit 21 updates this operation parameter Φ, and as a result, the feature extraction function f (·) of the feature extraction unit 20 is changed to a new operation environment. To adapt. The classifying unit 3 has B _i reference signal patterns for classifying the classes C _i as shown in the following Expression 10.

[0054]

Λ = {{R _i ¹ , ..., R _i ^b , ..., R _i ^Bi }} _{i = 1} ^{i = M}

The structure of this reference signal pattern is basically the same as the above-mentioned feature vector sequence. In this example, the discriminant function of the class C _i is represented by the following Expression 11.

[0056]

[Equation 11]

Here, D (X, R _i ^b ; Φ) is given by
It is represented by.

[0058]

(Equation 12)

D (X, R _i ^b ; Φ) represented by the equation 12
Is the vector sequence Y of the signal pattern converted by the operation parameter Φ and the vector sequence length T of the reference pattern.
_i ^b is a generalization of the cumulative distance between the vector sequences R _i ^b having, it is decomposed as sequentially follows. Note that both ζ and ξ are predetermined positive constants. Also, Θ _i ^b is the number of possible matching paths allowed for matching with R _i ^b . further,
D _θ (X, R _i ^b ; Φ) is defined on the θ-th matching path allowed in the nonlinear time axis expansion / contraction method based on dynamic programming between the feature vector Y and the reference pattern vector series R _i ^b. The route distance is defined by the following equation 13.

[0060]

(Equation 13)

[0061] _δ iθt ^b in the number 13, the t th feature vector r _i ^b of the vector sequences R _i ^b of the reference pattern
(T) and the θ _t ^ib- th feature vector ν of the feature vector series Y that is associated with it via the θ-th matching path.
It is a local distance to (θ _t ^ib ), and is defined as the following Expression 14.

[0062]

[Equation 14]

Here, w _it ^b is a weighting coefficient associated with r _i ^b (t).

Next, the feature extraction unit 20 based on the GPD method.
The update rule regarding the learnable operating parameter Φ of is described. This update rule is expressed by the following Expression 15.

[0065]

[Formula 15] Φ _{τ + 1} = Φ _τ + δΦ _τ

Here, τ is a time index indicating the number of repetitions in the repeated update of GPD. For the input signal data x belonging to the class C _k , which is given for adaptive updating,
The update amount δΦ _τ is represented by the following Expression 16 according to the stochastic descent theorem. Note that ∇ _phi below represents the partial derivative with respect to [Phi.

[0067]

[Number 16] _{δΦ τ = -εU∇ φ g k (} X; Λ, Φ)

[0068] vector _∇ φ to determine the amount of update _δΦ τ
g _k (X; Λ, Φ) is expressed by the following equations 17 to 19 stepwise.
Given as shown in.

[0069]

[Equation 17]

(Equation 18)

[Formula 19]

Here, ε is a small positive number that decreases monotonically, and U is a positive definite matrix.

Further, a case where the above update rule is simplified will be described. By bringing the above positive constants ζ and ξ close to infinity, the discriminant function g _i (X; Λ, Φ) can be approximated in a simple form as in the following Expression 20.

[0072]

(Equation 20)

Here, b'represents the reference vector of the best minimum generalized cumulative distance in the class, and θ'represents the path of the best minimum cumulative distance in each non-linear time-axis stretch matching. In this case, the vector _{∇ φ g k (X, Λ} ; Φ) for determining the update amount of the number 16 is expressed by the following equation 21.

[0074]

[Equation 21]

As described above, according to the embodiment of the present invention, the classification unit 3 sets the class boundary by updating the operation parameter of the feature extraction unit 20 (that is, the classification judgment is executed). In order to adapt the feature signal pattern space itself, compared with the conventional method in which the class boundary is individually updated in the classifying unit 3 while fixing the operation parameter of the feature extracting unit 20 and applying heuristic or empirical knowledge, It is possible to perform the adaptation process to the new operating condition with higher efficiency, and it is possible to achieve high recognition accuracy even under the new operating condition by using only a small number of adaptation samples as compared with the conventional example.

<Second Embodiment> FIG. 2 shows a second embodiment of the present invention.
FIG. 3 is a block diagram of a signal pattern recognition device according to the embodiment.
In the second embodiment, the above USFE method is used. The signal pattern recognition apparatus according to the second embodiment has a feature extraction unit 30 that has an operation parameter that can be learned and that converts an input signal pattern into a predetermined feature signal pattern for signal pattern recognition. Performs a classification process for classifying the characteristic signal pattern converted by the characteristic extraction unit 30 into at least one of a plurality of classes corresponding to a plurality of class models stored in the class model memory 22 in advance. By doing so, at least one class having a higher likelihood is provided as the recognition result of the signal pattern, and the classification unit 3 of the conventional example is provided, and the class model memory 22 in which a plurality of class models are stored in advance is referred to, When a signal pattern whose class attribute is unknown is input to the feature extraction unit 30, the unknown signal pattern is selected from among the plurality of classes which are the recognition results output from the classification unit 3. Without changing the classification process of the classification unit 3, the feature extraction unit 30 does not change so as to approach the maximum likelihood class model having a high likelihood (the plurality of class models are stored in the class model memory 22). A feature extraction parameter adjusting unit 31 that adaptively updates the learnable operation parameter is further provided. Here, in particular, the feature extraction parameter adjustment unit 31 preferably has a loss corresponding to the similarity between the feature signal pattern output from the feature extraction unit 30 and the unknown model of the signal pattern. The feature is to adaptively update the learnable operation parameter of the feature extraction unit 30 so as to minimize the value. Hereinafter, differences from the first embodiment will be described.

In the USAFE method of the second embodiment, the model ω _m (of the class C _{m to} which the adaptive input sample signal pattern of unknown class attribute is judged by the signal pattern recognition device in the state before update) Λ) (the class model is stored in the class model memory 22) is used, and the same updating process as the SAFE method of the first embodiment is performed. If the classification judgment for the signal pattern of unknown class attribute for adaptation is correct, the USAFE method leads the same result as the SAFE method. However, the correctness of this judgment is not generally guaranteed. As the learning of the USAFE method progresses, the feature extraction unit 30 acquires the state of the signal pattern recognition device (specifically, the feature) from the unknown input signal pattern sample under the new operation condition immediately before the adaptive update accompanying each input. In the state immediately before the adaptive update associated with each input of the operation parameter Φ of the extraction unit 30), a feature signal pattern closer to the maximum likelihood type model obtained by executing the classification processing represented by the above-mentioned Equation 2 is obtained. So that it can be converted
The operation parameter Φ of the feature extraction unit 30 is gradually updated.

That is, the feature extraction pattern adjusting unit 31
Is the characteristic signal pattern Y ₁ converted from the unknown input signal data x ₁ is the characteristic signal of the model ω _m (Λ) of the maximum likelihood class C _m of the plurality of models output from the classification unit 3. The operation parameter Φ that determines the feature extraction function f (x; Φ) of the feature extraction unit 30 is updated so as to approach the pattern.
Specifically, the model ω _m (Λ) of the maximum likelihood class C _m and the characteristic signal pattern Y ₁ converted from the unknown input signal pattern
And the similarity l _m (Y ₁ ;
Λ) is newly defined, and the operating parameter Φ is updated so that this loss becomes as small as possible. The models of a plurality of classes are stored in the model memory 22.

As described above, according to the embodiment of the present invention, the classification unit 3 sets the class boundary (that is, executes the classification process) by updating the operation parameter of the feature extraction unit 30. In order to adapt the feature signal pattern space itself, compared to the conventional method in which the class boundary is individually updated in the classifying unit 3 while fixing the operation parameter of the feature extracting unit 30 and applying heuristic or empirical knowledge, It is possible to perform the adaptation process to the new operating condition with higher efficiency, and it is possible to achieve high recognition accuracy even under the new operating condition by using only a small number of adaptation samples as compared with the conventional example.

The SAFE method of the first embodiment and the second method
Also in the USAFE method of the above embodiment, the selection of the adaptive signal pattern may be performed randomly or may be learned based on scientific knowledge about the characteristics of the signal pattern.

In the above embodiment, the feature extraction unit 20,
30, a classification unit 3, a feature extraction parameter adjustment unit 21,
The numeral 31 is composed of, for example, a digital computer.

[0082]

As described above in detail, the learning method for the signal pattern recognition apparatus according to the first aspect of the present invention has the operation parameter that can be learned, and the input signal pattern Feature extraction means for converting into a predetermined feature signal pattern for recognition, and classification processing for classifying the feature signal pattern converted by the feature extraction means into at least one of a plurality of predetermined classes. By performing at least one of the more likely
A learning method for a signal pattern recognition device, comprising: a classifying unit that outputs two classes as a recognition result of a signal pattern, wherein when the signal pattern having a known class attribute is input to the feature extracting unit, The learnable operation parameter of the feature extracting means is adaptively changed without changing the classification process of the classifying means so that the feature signal pattern output from the extracting means approaches the similar model of the known signal pattern. Update. Therefore, by updating the operation parameter of the feature extracting means, the classifying means adapts the feature signal pattern space itself on which the classification processing is performed. Therefore, the operation parameter of the feature extracting means is fixed to be heuristic or empirical. Compared with the conventional method of individually updating the class boundaries in the classification means while using the knowledge, it is possible to perform the adaptation processing to the new operating condition with higher efficiency, and the number of adaptations is smaller than that of the conventional example. High recognition accuracy can be achieved even under new operating conditions by using only the specimen.

The learning method of the signal pattern recognition device according to claim 2 is the learning method of the signal pattern recognition device according to claim 1, wherein the characteristic signal pattern output from the characteristic extraction means and the known signal are used. In order to minimize the loss corresponding to the similarity between the pattern and the model,
The learnable operation parameter of the feature extraction means is adaptively updated. As a result, higher recognition accuracy can be achieved with a simpler configuration as compared with the conventional example.

According to a third aspect of the present invention, there is provided a learning device for a signal pattern recognition device, which has a learnable operation parameter and converts an input signal pattern into a predetermined characteristic signal pattern for recognition of the signal pattern. At least one of a plurality of predetermined classes of the characteristic extracting means for converting and the characteristic signal pattern converted by the characteristic extracting means.
A learning device for a signal pattern recognition device, comprising: a classifying unit that outputs at least one class having a higher likelihood as a recognition result of a signal pattern by executing a classification process for classifying into one. , When the signal pattern whose class attribute is known is input to the feature extraction means, the classification of the classification means is performed so that the feature signal pattern output from the feature extraction means approaches the class model of the known signal pattern. The present invention is characterized by comprising adjusting means for adaptively updating the learnable operation parameter of the feature extracting means without changing the processing. Therefore, by updating the operation parameter of the feature extracting means, the classifying means adapts the feature signal pattern space itself on which the classification processing is performed. Therefore, the operation parameter of the feature extracting means is fixed to be heuristic or empirical. Compared with the conventional method of individually updating the class boundaries in the classification means while using the knowledge, it is possible to perform the adaptation processing to the new operating condition with higher efficiency, and the number of adaptations is smaller than that of the conventional example. High recognition accuracy can be achieved even under new operating conditions by using only the specimen.

A learning device for a signal pattern recognition device according to a fourth aspect of the present invention is the learning device for a signal pattern recognition device according to the third aspect, wherein the adjusting means is a feature signal pattern output from the feature extracting means. , The learning operation parameters of the feature extraction means are adaptively updated so that the loss corresponding to the similarity with the known model of the signal pattern is minimized. As a result, higher recognition accuracy can be achieved with a simpler configuration as compared with the conventional example.

According to a fifth aspect of the present invention, there is provided a learning method for a signal pattern recognizing device, which has a learnable operation parameter and converts an input signal pattern into a predetermined characteristic signal pattern for recognizing the signal pattern. At least one of a plurality of predetermined classes of the characteristic extracting means for converting and the characteristic signal pattern converted by the characteristic extracting means.
A learning method for a signal pattern recognition device, comprising: a classification unit that outputs at least one class having a higher likelihood as a recognition result of a signal pattern by executing a classification process for classifying into one. , When a signal pattern whose class attribute is unknown is input to the feature extracting means, the unknown signal pattern is the maximum likelihood having the highest likelihood among the plurality of classes which are the recognition results output from the classifying means. It is characterized in that the learnable operation parameter of the feature extracting means is adaptively updated without changing the classification process of the classifying means so as to approach the class model. Therefore, by updating the operation parameter of the feature extracting means, the classifying means adapts the feature signal pattern space itself on which the classification processing is performed. Therefore, the operation parameter of the feature extracting means is fixed to be heuristic or empirical. Compared with the conventional method of individually updating the class boundaries in the classification means while using the knowledge, it is possible to perform adaptation processing to new operating conditions with higher efficiency,
High recognition accuracy can be achieved even under new operating conditions by using only a small number of adaptation samples as compared with the conventional example.

The learning method of the signal pattern recognition device according to a sixth aspect is the learning method of the signal pattern recognition device according to the fifth aspect, wherein the feature signal pattern output from the feature extraction means is the unknown. Is adaptively updated so as to minimize the loss corresponding to the similarity with the model of the maximum likelihood signal pattern. As a result, higher recognition accuracy can be achieved with a simpler configuration as compared with the conventional example.

According to a seventh aspect of the present invention, there is provided a learning device for a signal pattern recognizing device, which has a learnable operation parameter and converts an input signal pattern into a predetermined characteristic signal pattern for recognizing the signal pattern. At least one of a plurality of predetermined classes of the characteristic extracting means for converting and the characteristic signal pattern converted by the characteristic extracting means.
A learning device for a signal pattern recognition device, comprising: a classifying unit that outputs at least one class having a higher likelihood as a recognition result of a signal pattern by executing a classification process for classifying into one. , When a signal pattern whose class attribute is unknown is input to the feature extracting means, the unknown signal pattern is the maximum likelihood having the highest likelihood among the plurality of classes which are the recognition results output from the classifying means. It is characterized by comprising adjusting means for adaptively updating the learnable operation parameter of the feature extracting means without changing the classification process of the classifying means so as to approach the class model. Therefore, by updating the operation parameter of the feature extracting means, the classifying means adapts the feature signal pattern space itself on which the classification processing is performed. Therefore, the operation parameter of the feature extracting means is fixed to be heuristic or empirical. Compared with the conventional method of individually updating the class boundaries in the classification means while using the knowledge, it is possible to perform the adaptation processing to the new operating condition with higher efficiency, and the number of adaptations is smaller than that of the conventional example. High recognition accuracy can be achieved even under new operating conditions by using only the specimen.

A learning device for a signal pattern recognition device according to an eighth aspect is the learning device for a signal pattern recognition device according to the seventh aspect, wherein the adjusting means is a feature signal pattern output from the feature extracting means. , A feature that the learned parameter of the feature extracting means is adaptively updated so that the loss corresponding to the similarity with the unknown but maximum likelihood signal pattern type model is minimized. And As a result, higher recognition accuracy can be achieved with a simpler configuration as compared with the conventional example.

[Brief description of drawings]

FIG. 1 is a block diagram of a signal pattern recognition device according to a first embodiment of the present invention.

FIG. 2 is a block diagram of a signal pattern recognition device according to a second embodiment of the present invention.

FIG. 3 is a block diagram showing an application example using the signal pattern recognition device of the first and second embodiments according to the present invention.

FIG. 4 is a block diagram of a conventional signal pattern recognition device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Signal data memory, 3 ... Classification part, 5, 6 ... Signal pattern recognition result data memory, 20, 30 ... Feature extraction part, 21, 31 ... Feature extraction parameter adjustment part, 7, 22 ... Model memory, 41 ... Microphone, 42 ... A / D converter, 43 ... Speech recognition signal pattern recognition device.

[Procedure amendment]

[Submission date] September 14, 1995

[Procedure Amendment 1]

[Document name to be amended] Statement

[Name of item to be corrected] 0025

[Correction method] Change

[Correction content]

[0025]

Embodiments of the present invention will be described below with reference to the drawings. <First Embodiment> FIG. 1 shows a signal pattern recognition apparatus according to the first embodiment of the present invention. The signal pattern recognition device according to the first embodiment has a learnable operation parameter,
A feature extraction unit 20 that converts an input signal pattern into a predetermined feature signal pattern for signal pattern recognition.
And a classification process for classifying the feature signal pattern converted by the feature extraction unit 20 into at least one of a plurality of classes corresponding to a plurality of class models determined in advance and stored in the class model memory 22. By executing
The classification unit 3 of the conventional example that outputs at least one class with higher likelihood as a recognition result of a signal pattern, and when a signal pattern with a known class attribute is input to the feature extraction unit 20, the class model memory 22 Without changing the classification process of the classification unit 3 so that the characteristic signal pattern output from the characteristic extraction unit 20 can be brought closer to the similar model of the known signal pattern by referring to the similar model stored in A feature extraction parameter adjusting unit 21 that adaptively updates the learnable operation parameter of the feature extracting unit 20 is featured. Here, the feature extraction parameter adjustment unit 21
Is preferably learned by the feature extraction unit 20 so that the loss corresponding to the similarity between the feature signal pattern output from the feature extraction unit 20 and the known model of the signal pattern is minimized. Adaptively update possible operating parameters.

[Procedure Amendment 2]

[Document name to be amended] Statement

[Correction target item name] 0040

[Correction method] Change

[Correction content]

Here, T (r) is the length of the vector series R which is the reference signal pattern. In the GPD method used for learning the signal pattern recognition device, a misclassification measure is used in addition to the loss function reflecting the classification result and the discrimination function similar to the above. In the SAFE method of the first embodiment, these are all defined as the same function. C _k above
Any reasonable distance or a function thereof can be used for the similarity of, but in particular, between the vector sequence R of the reference signal pattern and the characteristic signal pattern Y ₁ , a nonlinear time axis based on dynamic programming is used. Using the cumulative squared Euclidean distance D _θ obtained by accumulating the squared Euclidean distances between the feature vectors of R and Y _{1 associated} with the route θ within the expansion / contraction matching range used for expansion / contraction, Can be given as in Equation 6.

[Procedure 3]

[Document name to be amended] Statement

[Correction target item name] 0076

[Correction method] Change

[Correction content]

<Second Embodiment> FIG. 2 shows a second embodiment of the present invention.
FIG. 3 is a block diagram of a signal pattern recognition device according to the embodiment.
In the second embodiment, the USAFE method is used. The signal pattern recognition apparatus according to the second embodiment has a feature extraction unit 30 that has an operation parameter that can be learned and that converts an input signal pattern into a predetermined feature signal pattern for signal pattern recognition. Performs a classification process for classifying the characteristic signal pattern converted by the characteristic extraction unit 30 into at least one of a plurality of classes corresponding to a plurality of class models stored in the class model memory 22 in advance. By doing so, at least one class having a higher likelihood is provided as the recognition result of the signal pattern, and the classification unit 3 of the conventional example is provided, and the class model memory 22 in which a plurality of class models are stored in advance is referred to When a signal pattern whose class attribute is unknown is input to the feature extraction unit 30, the unknown signal pattern is selected from among the plurality of classes which are the recognition results output from the classification unit 3. Without changing the classification process of the classification unit 3, the feature extraction unit 30 does not change so as to approach the maximum likelihood class model having a high likelihood (the plurality of class models are stored in the class model memory 22). A feature extraction parameter adjusting unit 31 that adaptively updates the learnable operation parameter is further provided. Here, in particular, the feature extraction parameter adjustment unit 31 preferably corresponds to the similarity between the feature signal pattern output from the feature extraction unit 30 and the unknown but maximum likelihood signal pattern type model. The feature is that the learnable operation parameter of the feature extraction unit 30 is adaptively updated so as to minimize the loss. Hereinafter, differences from the first embodiment will be described.

[Procedure amendment 4]

[Document name to be amended] Statement

[Correction target item name] 0078

[Correction method] Change

[Correction content]

That is, the feature extraction parameter adjusting unit 31
Is the characteristic signal pattern Y ₁ converted from the unknown input signal data x ₁ is the characteristic signal of the model ω _m (Λ) of the maximum likelihood class C _m of the plurality of models output from the classification unit 3. The operation parameter Φ that determines the feature extraction function f (x; Φ) of the feature extraction unit 30 is updated so as to approach the pattern.
Specifically, the model ω _m (Λ) of the maximum likelihood class C _m and the characteristic signal pattern Y ₁ converted from the unknown input signal pattern
And the similarity l _m (Y ₁ ;
Λ) is newly defined, and the operating parameter Φ is updated so that this loss becomes as small as possible. The models of a plurality of classes are stored in the model memory 22.

Claims (8)

[Claims] 1. A feature extraction unit that has an operation parameter that can be learned and that transforms an input signal pattern into a predetermined feature signal pattern for recognition of the signal pattern, and a feature transformed by the feature extraction unit. A classifying unit that outputs at least one class having a higher likelihood as a recognition result of the signal pattern by performing a classification process for classifying the signal pattern into at least one of a plurality of predetermined classes. A learning method for a signal pattern recognition device comprising: a characteristic signal pattern output from the feature extracting means when a signal pattern having a known attribute is input to the feature extracting means. Of the feature pattern extracting means without changing the classification process of the classifying means so as to approximate the model of the signal pattern of Learning signal pattern recognition device and updates the data adaptively. 2. The feature extraction means is capable of learning so that the loss corresponding to the similarity between the feature signal pattern output from the feature extraction means and the known model of the signal pattern is minimized. The method for learning a signal pattern recognition device according to claim 1, wherein various operating parameters are adaptively updated. 3. A feature extraction unit that has an operation parameter that can be learned and that converts an input signal pattern into a predetermined feature signal pattern for recognition of the signal pattern, and a feature converted by the feature extraction unit. A classifying unit that outputs at least one class having a higher likelihood as a recognition result of the signal pattern by performing a classification process for classifying the signal pattern into at least one of a plurality of predetermined classes. A learning device for a signal pattern recognition device comprising: a characteristic signal pattern output from the feature extracting means when a signal pattern having a known attribute is input to the feature extracting means. Of the feature pattern extracting means without changing the classification process of the classifying means so as to approximate the model of the signal pattern of Learning device of the signal pattern recognition apparatus characterized by comprising an adjusting means for updating the data adaptively. 4. The adjusting means is configured to minimize the loss corresponding to the similarity between the characteristic signal pattern output from the characteristic extracting means and the similar model of the known signal pattern. The learning device for a signal pattern recognition device according to claim 3, wherein the learnable operation parameter of the extraction means is adaptively updated. 5. A feature extraction unit that has an operation parameter that can be learned and that converts an input signal pattern into a predetermined feature signal pattern for recognizing the signal pattern, and a feature converted by the feature extraction unit. A classifying unit that outputs at least one class having a higher likelihood as a recognition result of the signal pattern by performing a classification process for classifying the signal pattern into at least one of a plurality of predetermined classes. A learning method for a signal pattern recognition device comprising: a recognition pattern output from the classification means when the unknown signal pattern is input to the feature extraction means. Without changing the classification processing of the above classification means, the above-mentioned special processing is performed so as to approach the maximum likelihood class model having the highest likelihood among the resulting plural classes. Learning signal pattern recognition apparatus characterized by updating the learning possible operating parameters of the extraction means adaptively. 6. The feature so that the loss corresponding to the similarity between the feature signal pattern output from the feature extracting means and the unknown but maximum likelihood signal pattern type model is minimized. 6. The learning method for a signal pattern recognition device according to claim 5, wherein the learnable operation parameter of the extraction means is adaptively updated. 7. A feature extraction unit having a learnable operation parameter and converting an input signal pattern into a predetermined feature signal pattern for recognition of the signal pattern, and a feature converted by the feature extraction unit. A classifying unit that outputs at least one class having a higher likelihood as a recognition result of the signal pattern by performing a classification process for classifying the signal pattern into at least one of a plurality of predetermined classes. A learning device for a signal pattern recognition device comprising: a recognition pattern output from the classification means when the unknown signal pattern is input to the feature extraction means. Without changing the classification processing of the above classification means, the above-mentioned special processing is performed so as to approach the maximum likelihood class model having the highest likelihood among the resulting plural classes. Learning device of the signal pattern recognition apparatus characterized by a learnable operating parameters of the extraction means provided with adjustment means for adaptively updated. 8. The adjusting means minimizes the loss corresponding to the similarity between the characteristic signal pattern output from the characteristic extracting means and the unknown but maximum likelihood signal pattern type model. 8. The learning device for a signal pattern recognition device according to claim 7, wherein the learnable operation parameter of the feature extracting means is adaptively updated.