JPH04305700A - Pattern recognition device - Google Patents

Abstract

PURPOSE:To enable the estimation of parameters for learning using a recognition dictionary and likelihood arithmetic individually with learning data of the recognition dictionary and parameter estimation data for likelihood arithmetic which are prepared separately. CONSTITUTION:A feature vector is extracted from a feature parameter obtained by analyzing input data and collated with the recognition dictionary 4 to find similarity and likelihood corresponding to the similarity is calculated and used to find the recognition result as to the input data. The learning using the recognition dictionary 4 is performed by using the feature vector extracted from recognition dictionary learning data 7 in a category to be recognized and the recognition result of the feature vector, and the parameter for likelihood arithmetic is estimated from the similarity found by matching the feature vector extracted from the parameter estimation data 8 for likelihood arithmetic and the recognition dictionary 4 after the learning and used for a next recognition process.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition device for recognizing characters, sounds, and the like.

0002

2. Description of the Related Art Conventionally, the following types of pattern recognition devices for recognizing characters, voices, etc. are known. For example, a pattern recognition device that recognizes speech patterns extracts a speech pattern from input speech data, compares the recognition target category of this pattern with the contents of a recognition dictionary, calculates the degree of similarity, and selects the highest Categories indicating the degree of similarity are output as recognition results.

[0003] In this case, the degree of similarity determined by matching the pattern with the recognition dictionary is a measure of how similar the pattern subjected to the similarity calculation is to the contents of the recognition dictionary. In other words, the similarity here only indicates which category the pattern is most similar to among the recognition target categories, and it is used to identify whether the extracted pattern belongs to a certain category or not. It was considered difficult to do so.

[0004] Conventionally, therefore, as a means for identifying categories, a predetermined threshold is set for the degree of similarity.
If the similarity obtained by matching the pattern with the recognition dictionary described above is smaller than the threshold, it is rejected.
A method is being considered in which a process is executed and if the value is larger than a threshold value, it is identified as the current category.

[0005] However, when recognizing a pattern of speech data with superimposed noise, for example, the threshold value set for the degree of similarity is set low in anticipation of a decrease in the degree of similarity overall. On the other hand, when pattern recognition is performed on noise-free voice data, there is a tendency to depend on the recognition conditions, such as setting a higher value, and it has been difficult to uniquely define the recognition conditions.

[0006] Therefore, in any case, using this kind of similarity is only a way of knowing how similar the extraction pattern and the recognition dictionary are, and by taking the similarity value, you can determine the probability. It cannot be said that it is an effective method in terms of whether accurate results can be obtained.

[0007] Conventionally, therefore, a method of converting the similarity value into a likelihood has been considered. In other words, this method makes it possible to express certainty by converting the similarity value into a likelihood and treating the similarity as a probability event. This allows it to be treated as a clear relationship. (Reference: Teru
Hiko Ukita et al. ”A Speak
er Independent Recognition
n Algorithm for Connected
Word Using Word Boundary
Hypothesizer”, Proceedin
gs of ICASSP86, pp-1077~1
080) However, even with this method, for example,
The parameters for likelihood calculation obtained from the dataset used for learning the recognition dictionary are significantly different from the parameters for likelihood calculation obtained from other datasets. Even if we estimate the parameters for the recognition dictionary using the same data set, we cannot expect any improvement in recognition performance. It was difficult to expect recognition performance to improve because there were no guidelines for determining whether the recognition was acceptable or not.

[0008]

[Problems to be Solved by the Invention] In this way, in the conventional pattern recognition device, when converting the similarity value obtained by matching the recognition dictionary to the input pattern, it is difficult to learn the recognition dictionary and calculate the likelihood. The problem is that it is difficult to expect pattern recognition with excellent recognition performance because there is no clear guideline on how to estimate the parameters for the pattern recognition.

The present invention has been made in view of the above circumstances, and provides a pattern recognition device that can efficiently perform learning of a recognition dictionary and estimate parameters for likelihood calculation, and is expected to improve recognition performance. The purpose is to provide.

0010

[Means for Solving the Problems] The pattern recognition device of the present invention extracts feature vectors of a certain dimension from feature parameters obtained by analyzing input data, compares the extracted feature vectors with a recognition dictionary, and calculates the degree of similarity. , the likelihood corresponding to the similarity is calculated, and the calculated likelihood is used to obtain the recognition result for the input data, which is extracted from the recognition dictionary training data belonging to the recognition target category. The above recognition dictionary is trained using the feature vector and the recognition result of this feature vector, and the likelihood is calculated from the similarity obtained by comparing the feature vector extracted from the parameter estimation data for likelihood calculation with the learned recognition dictionary. The parameters for the degree calculation are estimated and used for the next recognition process.

[0011] Furthermore, the pattern recognition device of the present invention extracts a feature vector of a certain dimension from the feature parameters obtained by analyzing input data, compares this feature vector with a recognition dictionary to obtain a degree of similarity, and also calculates a degree of similarity corresponding to the degree of similarity. The system calculates the likelihood for the above input data and uses this likelihood to obtain the recognition result for the input data. The data set selection means is equipped with a data set selection means capable of selecting a data set to be provided, and executes recognition dictionary learning using the recognition result of the feature vector from the data set selected by the data set selection means for recognition dictionary learning. The parameters for likelihood calculation are estimated by comparing the feature vectors from the data set selected for estimating the parameters for likelihood calculation by the selection means with the recognition dictionary after learning, and the parameters for the next recognition dictionary are estimated using the previous Data sets including the data set selected for estimating parameters for likelihood calculation are selected, and data sets not used for recognition dictionary learning are selected for estimating parameters for likelihood calculation.

0012

As a result, according to the present invention, learning of the recognition dictionary and parameter estimation for likelihood calculation can be performed efficiently, and pattern recognition with excellent recognition performance can be realized.

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows a schematic configuration when the present invention is applied to a speech recognition device. In this case, reference numeral 1 denotes an audio input section, and the audio input section 1 converts an audio signal input through a microphone (not shown) into a digital signal and outputs the digital signal.

In this case, the audio input section 1 includes a low-pass filter (LPF) for removing high-frequency noise components, and a low-pass filter (LPF) for removing high-frequency noise components.
It is constituted by an A/D converter that converts an input audio signal taken in via F into a digital signal at a sampling frequency of 12 kHz and a quantization bit number of 12 bits, for example. Note that the digitization process for the input audio signal here may be performed to obtain a digital signal with a sampling frequency of 8 kHz and a quantization bit count of 16 bits, for example, depending on the specifications required for the input audio. It can be determined as appropriate depending on the recognition performance, etc. In addition, it is possible to input recognition dictionary learning data 7, which is data belonging to the recognition target category, and likelihood calculation parameter estimation data 8, which is used to estimate the likelihood calculation parameters, to the voice input unit 1. ing.

The output converted into a digital signal by the audio input section 1 is sent to the audio analysis section 2. This voice analysis section 2 is for analyzing the voice data given from the voice input section 1, and basically performs FFT analysis,
A method such as LPC analysis is used, for example, 8 msec
The feature parameters are calculated for each time.

The time series of feature parameters determined by the speech analysis section 2 is sent to the recognition processing section 3, where it is subjected to recognition processing. This recognition processing unit 3 includes a voice feature vector extraction unit 31 that extracts a voice feature vector using the feature parameters obtained by the voice analysis unit 2, a voice feature vector extracted by this voice feature vector extraction unit 31, and a recognition target. It is composed of a similarity calculation section 32 that performs pattern matching with the category recognition dictionary 4 and calculates the similarity, and a likelihood calculation section 33 that calculates the likelihood from the similarity value obtained by the similarity calculation section 32. There is.

In this case, the speech feature vector extracting unit 31 detects the beginning and end of the input speech using the speech feature parameters, and resamples and extracts the feature parameters of the speech section sandwiched between the start and end points. A voice feature vector of constant dimensions, such as 16 dimensions in the frequency direction and 16 dimensions in the time axis direction, as shown in the figure, is obtained. The speech feature vector obtained by the speech feature vector extraction section 31 is sent to the similarity calculation section 32 and the recognition dictionary learning section 6. Recognition dictionary learning section 6 here
In this method, a recognition dictionary is trained using a voice feature vector and recognition results from a recognition result output unit 5, which will be described later. The similarity calculation section 32 performs pattern matching between the speech feature vector provided by the speech feature vector extraction section 31 and the recognition target category registered in the recognition dictionary 4. The pattern matching in the similarity calculation unit 32 here uses a technique such as a composite similarity calculation, for example. Further, the matching result from the similarity calculation section 32 is sent to the likelihood calculation parameter estimating section 9, and is used for estimating the likelihood calculation parameter 10. The likelihood calculation section 33 converts the similarity value obtained by the similarity calculation section 32 into a likelihood using the likelihood calculation parameter 10.

Here, in order to obtain the posterior probability as the likelihood, when we have a similarity value Si for the category Ci, we will obtain the probability P (Ci | Si) of actually belonging to Ci. Since the probability P(Ci | Si) cannot be determined directly, from Bayes' theorem,

[0020]

[Math 1]

It can be found by:

[0022] Note that P(Si|Ci) has a category of C
The probability that the similarity degree Si is taken when i is equal, P(Ci), is the probability that the category Ci appears, and can be assumed to be equal. This makes it possible to derive P(Ci|Si) from P(Si|Ci). (Reference: Institute of Electronics, Information and Communication Engineers Pattern Recognition and Understanding Study Group Material PRU87-18 (1)
987)) In FIG. 1, P(Si|Ci) extracts speech feature parameters by inputting the speech data 8 for likelihood parameter estimation into the speech input section 1, and extracts speech feature parameters through speech analysis in the speech analysis section 2. After extracting the feature vector, it can be determined from the distribution of similarity values calculated by the similarity calculation unit 32. In reality, by using the similarity between the recognition dictionary 4 of the category to which the input speech belongs and the speech feature vector, finding the distribution function that the similarity distribution follows, and finding some parameters that determine the distribution function, It can be determined for each category.

[0023] Also,

[0024]

[Math 2]

[0025] Also, the similarity distribution can be obtained using the similarity between the recognition dictionary of the full-power category to be recognized and the speech feature vector in the same manner as P(Si|Ci). And these P(Si|Ci)
By obtaining P(Ci|Si) from , it becomes possible to convert the similarity into likelihood (posterior probability).

Then, the result obtained by converting the similarity into a likelihood in the likelihood calculation section 33 is sent to the recognition result output section 5, where the recognition result is output. Further, the recognition result from the recognition result output section 5 is sent to the recognition dictionary learning section 6.

Here, the recognition processing unit 3 has been described with the voice feature vector extraction unit 31 and the similarity calculation unit 32 separately, but these are combined into one and the recognition processing unit 3 continuously processes the input voice without detecting the beginning and end of the input voice. It is also possible to determine the similarity of each category using a word spotting method using pattern matching.

Next, the operation of the embodiment configured as described above will be explained.

First, to explain the case of matching continuous speech patterns as shown in FIG. Temporarily set each analysis frame whose characteristic parameters have been determined from the series as the terminal point E,
Based on these terminal points E, a starting end candidate section S consisting of a plurality of starting points satisfying a certain voice section condition is provisionally set. Then, the series of feature parameters of the hypothetical voice section indicated by these start and end points is resampled in the time axis direction, and
A feature vector of a predetermined number of dimensions that differs between voice sections is obtained using the terminal point E as a reference. Then, set the terminal point E like this
A plurality of feature vectors with a predetermined number of dimensions obtained based on the above are sequentially and continuously extracted while shifting the terminal point E in the time axis direction, and the degree of similarity between each feature vector and the recognition dictionary 4 is determined. A speech feature vector indicating the maximum similarity and start/end information are obtained for each category from the similarity series obtained for the feature vectors. By doing so, errors in detecting the start and end points can be reduced.

Next, the operation of the apparatus of this embodiment will be explained with reference to FIG.

First, the recognition dictionary learning data 7 is analyzed,
An initial recognition dictionary 4 is created based on the voice feature vector extracted after detecting the beginning and end. (Step A1
).

Next, the degree of similarity is determined based on the speech feature vector extracted from the parameter estimation data 8 for likelihood calculation,
The likelihood calculation parameter estimation unit 9 estimates the likelihood calculation parameters from the similarity distribution, and creates the initial likelihood calculation parameters 10 (step A2).

Then, using the initial recognition dictionary 4 created in step 1, the degree of similarity with the speech feature vector extracted from the recognition dictionary training data 7 is determined (step A3), and the initial likelihood calculation parameter 10 is calculated. , to convert the similarity into likelihood (step A4). Then, the recognition result output unit 5 executes recognition processing based on the likelihood obtained in step A4 (step A5).

Next, in step A6, it is determined whether the data is complete. Here, the operations of steps A3 to A5 are repeatedly executed until the determination is YES due to data completion.

When it is determined in step A6 that the data has ended, the recognition dictionary learning unit 6 updates the recognition dictionary 4 using the voice feature vector used in the similarity calculation and the recognition result of the recognition result output unit 5 (step A6). A7).

Here, the learning associated with updating the recognition dictionary 4 is performed by updating the covariance matrix and performing principal component analysis as described below. Specifically, the covariance matrix is updated using equation (1).

[0037]

[Math 3]

Note that the update coefficient α is changed according to the recognition result.

Then, a plurality of eigenvalues and eigenvectors are obtained from the thus updated covariance matrix by K-L expansion, which is one type of principal component analysis, and these are used as a recognition dictionary for composite similarity. (Reference: Proceedings of the Acoustical Society of Japan, pp. 109-110 (October 1986)) Next, the speech feature vector extracted from the parameter estimation data 8 for likelihood calculation using the updated recognition dictionary 4. (Step A8). Next, in step A9, it is determined whether the data has ended. Here, YE due to the end of data
The operation of step A8 is repeatedly executed until it is determined that S is determined.

When it is determined in step A9 that the data has ended, the parameters for likelihood calculation are re-estimated using the similarity value obtained in step A8 and used for the next likelihood calculation (step A10). ).

Thereafter, in step A11, it is determined whether learning of the recognition dictionary is completed. Here, the operations from step A3 onwards are repeatedly executed until the determination is YES due to the completion of learning, and when the determination is YES, the process is terminated.

Therefore, in this way, the recognition dictionary 4 is trained using the feature vector extracted from the recognition dictionary training data 7 and the recognition result of the feature vector, and the recognition dictionary 4 is trained using the feature vector extracted from the recognition dictionary training data 7 and the recognition result of the feature vector is used. The parameters for the above-mentioned likelihood calculation are estimated from the similarity obtained by comparing the learned feature vector with the recognition dictionary 4 after learning, and are used for the next recognition process. It is possible to enrich the recognition dictionary, estimate the optimal likelihood calculation parameters for the recognition dictionary, and expect an improvement in recognition performance. Furthermore, in this embodiment, by having the recognition dictionary learning data 7 and the likelihood calculation parameter estimation data 8 separately, it is possible to expect a highly reliable recognition dictionary 4 to be learned.

Here, if the recognition dictionary learning data 7 and the likelihood calculation parameter estimation data 8 are the same,
As shown in FIG. 5(a), the distribution of similarity obtained by this becomes higher as the learning of the recognition dictionary 4 progresses, and There is a large deviation from the similarity distribution shown by the data. As a result, recognition performance cannot be improved by estimating the likelihood calculation parameters using the similarity distribution obtained from the speech data for recognition dictionary training, and by using this likelihood calculation parameter to recognize unknown input speech. It's something you can't expect. From this result as well, it can be understood that the reliability of the recognition dictionary is increased by separating the recognition dictionary learning data 7 and the likelihood calculation parameter estimation data 8.

Next, FIG. 6 shows another embodiment of the present invention.

In this case, FIG. 6 shows a data set section in which various data are selected by the data set selection section 11 in place of the recognition dictionary learning data 7 and the likelihood calculation parameter estimation data 8 shown in FIG. 12 is used. The other parts are the same as those in FIG. 1, and the same parts are given the same reference numerals and the explanation thereof will be omitted.

Here, the data set selection unit 11 selects data from the data set unit 12 to be used in the above-described recognition dictionary learning process and likelihood calculation parameter estimation. The data set section 12 is configured as follows.

FIG. 7 is a schematic diagram for explaining the data set section 12. As shown in FIG. In this case, the data set section 12
For each recognition target category, there are multiple data sets 121 storing, for example, 50 data (in the illustrated example, n of #1 to #n).
) have. The plurality of data sets 121 of the data set unit 12 are used as recognition dictionary learning data 122 when learning a recognition dictionary by the operation of the data selection unit 11, and as data for estimating parameters for likelihood calculation when estimating parameters for likelihood calculation. Each is selected as data 123 and sent to the audio input section 1.

Here, the data selection unit 11 selects the data set 121 at the same time as the recognition dictionary learning data 122.
and likelihood calculation parameter estimation data 123 so as not to overlap. In addition, dataset 1 selected as data 123 for estimating parameters for likelihood calculation
21 is selected as the recognition dictionary learning data 122 during the next recognition dictionary learning.

Next, FIG. 8 shows the procedure for selecting the data set 121 in the data set unit 12 by the data selection unit 11. In this case, it is assumed that the data set unit 12 has five (=N) data sets 121, and each data set 121 stores, for example, 50 data for each recognition target category.

First, n is set to 1 (step B1), and n=
1, the data set 121 of #1 is selected as the recognition dictionary learning data 122 (step B2), and the voice input unit 1
sent to. As a result, data set 12 of #1
1, the recognition dictionary 4 is trained according to the flowchart of FIG. 4 described above (step B3). In this case, since the #1 data set 121 is selected, the number of data of each recognition target word used for learning of the recognition dictionary 4 is 50. Note that it is assumed that the initial dictionary and the parameters for likelihood calculation are determined in advance from another data set.

Next, the value obtained by adding 1 to n is set to j (step B4), and when j=2, the data set 121 of #2 is selected as the data 123 for estimating the parameter for likelihood calculation (
Step B5), the audio is sent to the audio input section 1. As a result, the process of estimating the likelihood calculation parameters according to the flowchart of FIG. 4 described above is executed using the #2 data set 121 (step B6).

Next, 1 is added to n to make n=2 (step B7), and in the next step B8, it is compared with the number N (=5) of data sets 121. Here, n is smaller than N
Since it is judged as O, the process returns to step B2, and this time,
When n=2, data sets 121 of #1 and #2 are selected as recognition dictionary learning data 122 (step B2).
, learning of the recognition dictionary 4 is executed using these #1 and #2 data sets 121 (step B3). In this case, since the #1 and #2 data sets 121 are selected, the number of data used for learning the recognition dictionary 4 is:
From 50 pieces to 100 pieces.

Next, the value obtained by adding 1 to n is set to j (step B4), and when j=3, the data set 121 of #3 is selected as the parameter estimation data 123 for likelihood calculation (
Step B5), using this #3 data set 121, a process of estimating the parameters for likelihood calculation is executed (step B6).

Next, 1 is added to n to make n=3 (step B7), and in the next step B8, data set 1 is
Compare with the number N (=5) of 21.

[0055] Thereafter, in step B8, if n is smaller than N, then N
If it is determined to be O, the above-described process is repeatedly executed. On the other hand, if n becomes larger than N (=5) and it is determined to be YES, the process is terminated.

In this case, in the third recognition dictionary learning process, data sets #1 to #3 are selected as the recognition dictionary learning data 122, and these data sets 121 to #3 are used to perform recognition. Learning of dictionary 4 is executed. In this case, the number of data used for learning the recognition dictionary 4 is 1
There will be 50 pieces. Further, the #4 data set 121 is selected as the parameter estimation data 123 for likelihood calculation.

In addition, in the fourth recognition dictionary learning process, the data sets #1 to #4 are selected as the recognition dictionary learning data 122, and these data sets 121 to #4 are used to create the recognition dictionary. 4 learning is executed. In this case, the number of data used for learning the recognition dictionary 4 is 200.
Become an individual. In addition, data 1 for estimating parameters for likelihood calculation
The data set 121 of #5 is selected as the data set 23.

Therefore, in this way, the data sets 121 for recognition dictionary training and the data sets 121 for likelihood calculation parameter estimation are separately selected from among the plurality of data sets 121, and the data set 121 for the recognition dictionary training is selected separately. set 121
The recognition dictionary 4 is trained using the recognition results of the feature vectors from this data set 121, and the feature vectors from the data set 121 selected for likelihood calculation parameter estimation are compared with the learned recognition dictionary. The parameters for the likelihood calculation are estimated, and furthermore, for the next learning of the recognition dictionary, the dataset 121 including the previous data set 121 for estimating the parameters for the likelihood calculation is selected, and the parameter estimation for the likelihood calculation is performed. In this case, the dataset 121 that has not been selected for recognition dictionary training is selected from among the remaining datasets 121, so that the learning process of the recognition dictionary 4 can be performed in detail, and the dataset 121 used for similarity calculation can be performed in detail. The recognition dictionary can be made even more complete, and the process of estimating the likelihood calculation parameters for the recognition dictionary can also be made more complete, making it possible to expect further improvement in recognition performance.

Note that the present invention is not limited to the embodiments described above; for example, in the embodiments described above, voice data is used as input data, and an example of recognition and learning processing for this voice data is described. However, it is also possible to perform recognition and learning processing on character data instead of voice data. There are no particular limitations on the number of dimensions of the feature vectors used in the recognition processing and learning processing in this case, or on the pattern matching method.

The key point of the present invention is that the data used for learning the recognition dictionary and the data for estimating the likelihood calculation parameters are separately provided, and the recognition dictionary and likelihood calculation parameters generated from each data are used to calculate the input data. It performs recognition processing and learning processing, and can be implemented with various modifications without departing from the gist.

[0061]

[Effects of the Invention] According to the pattern recognition device of the present invention, learning of a recognition dictionary and parameter estimation for likelihood calculation can be performed efficiently with high reliability, realizing pattern recognition with excellent recognition performance. can do.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a schematic configuration of an embodiment of the present invention.

FIG. 2 is a diagram showing an example of audio feature vectors in the embodiment of FIG. 1;

FIG. 3 is a diagram for explaining a case of consecutive voice pattern matching in the embodiment of FIG. 1;

FIG. 4 is a flowchart for explaining the operation of the embodiment of FIG. 1;

FIG. 5 is a diagram for explaining a state when recognition dictionary learning data and likelihood calculation parameter estimation data are the same in the embodiment of FIG. 1;

FIG. 6 is a block diagram showing a schematic configuration of another embodiment of the present invention.

FIG. 7 is a schematic diagram for explaining a data set section used in another embodiment of FIG. 6;

FIG. 8 is a flowchart for explaining the procedure of the data set selection process in the data set unit by the data selection unit in another embodiment of FIG. 6;

[Explanation of symbols]

1...Speech input section, 2...Speech analysis section, 3...Recognition processing section, 3
1... Audio feature vector extraction unit, 32... Similarity calculation unit, 3
3... Likelihood calculation unit, 4... Recognition dictionary, 5... Recognition result output unit,
6... Recognition dictionary learning section, 7... Recognition dictionary learning data, 8...
Likelihood calculation parameter estimation data, 9... Likelihood calculation parameter estimation unit, 10... Likelihood calculation parameters, 11...
Data set selection section, 12...Data set section, 121...
data set.

Claims (3)

[Claims] Claim 1: Extract a feature vector of a certain dimension from the feature parameters obtained by analyzing input data, compare this feature vector with a recognition dictionary to determine the degree of similarity, calculate the likelihood corresponding to the degree of similarity, and calculate the degree of similarity. In a pattern recognition device that uses likelihood to obtain a recognition result for the input data, the recognition dictionary is trained using the feature vector extracted from the recognition dictionary training data and the recognition result of the data. The parameter for the likelihood calculation is estimated from the similarity obtained by comparing the feature vector extracted from the parameter estimation data for the likelihood calculation with the recognition dictionary after learning, and used for the next recognition process. pattern recognition device. [Claim 2] A feature vector of a certain dimension is extracted from feature parameters obtained by analyzing input data, and this feature vector is compared with a recognition dictionary to obtain a degree of similarity, and a likelihood corresponding to the degree of similarity is calculated. In a pattern recognition device that uses likelihood to obtain recognition results for the above input data, it is now possible to select a dataset for recognition dictionary learning processing and parameter estimation for likelihood calculation from multiple datasets prepared in advance. The data set selection means executes recognition dictionary learning using the feature vectors from the data set selected for recognition dictionary learning by the data set selection means and the recognition results of the data. The parameters for the likelihood calculation are estimated by comparing the feature vector from the dataset selected for estimating the parameters for the likelihood calculation with the recognition dictionary after learning, and the current likelihood calculation is used for the next learning of the recognition dictionary. A pattern recognition apparatus characterized in that a data set is selected including a data set selected for parameter estimation for a likelihood calculation, and a data set not used for recognition dictionary learning is selected for a parameter estimation for a likelihood calculation. 3. A pattern recognition apparatus according to claim 1, wherein learning of a recognition dictionary and estimation of parameters for likelihood calculation are repeatedly executed in the pattern recognition apparatus according to claim 1 or 2.