JPS63163582A - Pattern recognizing device - Google Patents

Abstract

PURPOSE:To efficiently form a recognition dictionary good in a recognition performance without deteriorating the recognition ratio of an updated dictionary by controlling the learning of the recognition dictionary according to a specific learning parameter. CONSTITUTION:Plural feature vectors (input pattern) obtained in a feature extracting part 2 are sequentially stored in a storing part 3, a similarity arithmetic part 4 obtains the similarity to a category registered in the feature vector and the recognition dictionary memory 5 of an input sound and the similarity is stored in a similarity storing part 6. A learning parameter setting part 7 sets the learning parameter from the ratio of the number of patterns in which the difference between the similarity of the respective categories and the similarity of other category is above a certain threshold. A learning part 8 inputs the input sound pattern stored in the storing part 3, learns a dictionary pattern stored in the recognition dictionary memory 5 under the control of the learning parameter setting part 7 for respective categories to be recognized and the recognition dictionary is updated.

Description

Detailed Description of the Invention [Object of the Invention] (Field of Application in Industry-1-) The present invention relates to a pattern recognition device that effectively learns a recognition dictionary used for pattern recognition.

(Prior art) In pattern recognition in voice recognition devices such as voice word processors and in old-fashioned character recognition processing, in order to improve recognition performance (recognition rate), a large number of recognition target patterns are collected in advance and these patterns are Currently, recognition dictionaries are being trained using ``recognition dictionaries'' to improve recognition dictionaries.

Conventionally, when learning a recognition dictionary, many learning patterns are collected, the collected learning patterns are averaged for each recognition target category, and the averaged recognition target patterns are used as mark patterns (recognition dictionary). was registered in the recognition dictionary memory. Then, during subsequent pattern recognition processing, the standard patterns and human cover patterns of each category registered in the recognition dictionary memory as described above are used.
For example, human cover patterns have been recognized by matching using the DP matching method.

However, the standard patterns (recognition dictionaries) obtained in this way cannot be said to reflect the statistical distribution of the recognition target pattern, and are extremely inadequate for application to speech recognition for unspecified speakers, for example. .

On the other hand, as a recognition method that takes into account the statistical distribution of the recognition target,
For example, the composite similarity method and the subspace method are known. However, in these recognition methods, it is necessary to create a recognition dictionary in advance by considering the systematic distribution, so it is necessary to collect a huge amount of learning patterns and to It is necessary to repeatedly train the recognition dictionary using . For this purpose, pattern collection for learning the recognition dictionary,
The drawback is that the learning process requires a large amount of time.

Therefore, a technique has been proposed in which a recognition dictionary is created, and then the recognition dictionary is used to recognize and process human cover turns, while at the same time collecting human cover turns as learning patterns and changing the recognition dictionary through sequential learning.

However, in this method, if the learning parameters used as weights in the case of sequential learning are large, the change effect becomes too large, and the recognition rate of the updated dictionary may actually decrease. On the other hand, if a learning parameter is set to a value that is small enough not to deteriorate the recognition rate, there is a problem in that the number of updates required until convergence increases.

(Problem to be solved by the invention) As described above, in conventional pattern recognition devices of this type, if the learning parameter is too large, the recognition rate of the updated recognition dictionary will actually decrease, and the learning parameter will be too small for the correspondence. If it is too high, there is a problem in that the number of learning operations required until convergence increases.

The present invention was made in consideration of these circumstances, and its purpose is to reduce the number of updates of the recognition dictionary until convergence without lowering the recognition rate of the updated dictionary. An object of the present invention is to provide a pattern recognition device that can efficiently create a recognition dictionary with good recognition performance.

[B of the invention] (Means for solving the problem) The present invention compares feature parameters obtained by analyzing human cover patterns with a recognition dictionary using, for example, a composite similarity method, and generates an input pattern. In a pattern recognition device that recognizes and processes, the similarity (distance) of each recognition target category to two input patterns and the similarity (distance) of other categories
Find the number of patterns whose difference is 1- or less than a certain threshold, and the number of patterns whose difference is less than or equal to the same threshold, find the learning parameters necessary for learning the recognition dictionary according to these information, and calculate the learning parameters necessary for learning the recognition dictionary. The present invention is characterized in that the recognition dictionary is learned and updated according to learning parameters.

(Operation) According to the present invention, the similarity (distance) of each recognition target category to the human cover turn and the similarity (distance) of other categories
The recognition rate for each category can be determined from the number of patterns whose difference is above a certain threshold and the number of patterns whose difference is below the same threshold. Weighting coefficients in pattern update processing are controlled. In other words, if the change effect is too large and the recognition rate decreases, learning will be restricted, and if the recognition rate improves, learning will be performed more quickly by giving a larger market. Therefore, according to the present invention, the recognition performance of the recognition dictionary can be improved with a small number of learning times.

(Example) Hereinafter, an example of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of a speech recognition device according to this embodiment. Although this device performs pattern recognition of input speech, it can be similarly applied to pattern recognition such as handwritten character recognition, for example.

The audio input unit 1 includes a microphone and an amplifier, and is a part that converts input audio into an electrical signal.

The audio signal input through the audio input unit l is given to a feature extraction unit 2 consisting of, for example, an 8-channel filter bank, where it is frequency-decomposed and each frequency component is divided into eight points along the time axis. Sample 134 (8
X8) is converted into a human voice pattern of dimensional feature vectors. Note that feature extraction is performed using the filter analysis described above.
Analysis using fast Fourier transform or cepstral analysis is also an nf function.

11 obtained for one recognition object in this way
The two feature vectors (human cover turns) are sequentially stored in the storage unit 3 and subjected to similarity calculation processing by the similarity calculation unit 4.

The similarity calculation unit 4 calculates the feature vector f of the input voice,
Similarity S with the categories registered in the recognition dictionary memory 5
'if] is obtained by the following 1-lack similarity calculation. The nth eigenvalue and eigenvector of the covariance matrix obtained from many sample patterns are shown in advance, and these are stored in the recognition dictionary memory 5. Note that instead of using such similarity S, it is also possible to use Maharabinos' general distance, Euclidean distance, probability, likelihood, or the like.

The similarity S of the input pattern for each category obtained in this manner is stored in the similarity storage unit 6 in association with the input pattern.

The similarity S between the input pattern stored in the similarity storage unit 6 and the dictionary pattern of each category is guided to the learning parameter setting unit 7. Here, the ratio of the number of patterns for which the difference between the similarity of each category and the similarity of other categories exceeds a certain threshold to the total number of patterns (hereinafter referred to as "high similarity difference rate D") ) is required. In other words, for one recognition target category! The summer number feature vector is stored in storage unit 3.
Since the similarities S are obtained as many as the number of human cover turns with one recognition target category, and these similarities S are obtained as many as the number of categories, from these similarities S - 1-The high similarity difference rate mentioned above is determined. This high similarity difference rate indicates the degree to which a certain category can be distinguished from other categories, and the threshold for determining the high similarity difference is set higher than the threshold in recognition processing. Then, the learning parameter setting unit 7 sets the learning parameter W from this high similarity difference rate.

On the other hand, the learning section 8 inputs the human voice patterns stored in the storage section 3, and under the control of the learning parameter setting section 7, the learning section 8 inputs the dictionary patterns stored in the recognition dictionary memory 5 for each recognition target category. It's something to learn. The learning of dictionary patterns in the learning section 8 is performed, for example, by repeatedly updating the covariance matrix and performing KL expansion on the covariance matrix, as shown in FIG.

That is, in step a), the covariance matrix K of the recognition dictionary of the category to be learned is manually updated, basically by performing the operation of updating the covariance matrix as follows.
Update its covariance matrix K. However, ■ is a learning pattern, and is given as a 64-dimensional vector, for example.

Moreover, K is a covariance matrix after learning. And W is positive or negative (7) K Wotoru weighting coefficient (value of learning parameter)
If it is positive, it acts to increase the similarity of the characteristic kernel to the human cover pattern in the covariance matrix, and if it is negative, it acts to decrease the similarity.

This 10,000 update of the covariance matrix is repeated multiple times using sample patterns of the number of stages for each recognition target category according to the learning parameter W set by the learning parameter setting section 7 (step C). . Then, the covariance matrix obtained as a learning result is subjected to KL expansion to obtain its eigenvalues and eigenvectors (step d).

These eigenvalues and eigenvectors are registered in the recognition dictionary memory 5 as recognition words 8 (standard patterns) for each recognition target category (step e), and the recognition dictionary is updated.

In FIG. 1, the control section 9 controls the above-described pattern recognition processing and recognition dictionary learning processing, respectively.

In this device configured in this way, there are basically 4
Learning of the recognition dictionary (standard pattern) for each recognition target category is performed as described above. However, the KL expansion into a covariance matrix is usually performed using a calculation method such as the Jacobi method, the Householder-Sturm method, or the power method, and the calculation process requires a large amount of time.

Therefore, in this device, the above-mentioned updating process to the covariance matrix is efficiently executed according to the learning parameter W, and the covariance matrix K is expanded into KL, thereby effectively performing the learning process in a short time. ing.

FIG. 3 shows the flow of learning processing of the recognition dictionary.

When learning the recognition dictionary, first, the number of updates 1 is initialized to (1) (step A), and one of the human cover patterns (learning patterns) stored in the storage unit 3 is manually input to the similarity calculation unit 4. Then, similarity calculation is performed for each category (step C).

The obtained similarity S for each category is sequentially stored in the similarity storage unit 6. Next, based on this similarity S, the learning parameter setting section 7 performs the following processing. That is, for example, the difference between the degree of similarity Sx between a dictionary pattern of category X, which is a human cover turn, and the degree of similarity Sy between a dictionary pattern of category The ratio n between the number nl that is less than the set threshold value θ and the total number of patterns N (-nl + n2)
l/N (high similarity difference rate) DI is determined (step E).

This high similarity difference rate DI is determined for each input pattern and each category. Then, a learning parameter W for each category is set based on the calculated high similarity difference rate DI for each category (step E). This learning parameter W
indicates the value of a weighting coefficient used, for example, in the updating process of the covariance matrix, and is set to, for example, the initial value as shown in the table below.

Learning parameters (weighting coefficients) set in this way
According to W, the process of updating the covariance matrix of the recognition dictionary of the recognition target category is performed, for example, as follows (
Step F).

That is, if it is not certain that it is a recognition target, that is, if the difference between the similarity of the feature vector S regarding a certain recognition target category X and the similarity of another recognition target category Y is equal to or more than a predetermined threshold, The covariance matrix Ky of Y is
For example, it is updated with a negative weight (subtraction) as Ky −Ky −α■vL. At the same time, the covariance matrix Kx of category X, for example, KX −KX +α
Update with positive weight (addition) as VV[.

In addition, if a reliable recognition result is obtained, that is, if the difference between the similarity of the feature vector S for a certain recognition target category X and the similarity of another recognition target category Y is less than a predetermined threshold, , the covariance matrix Kx of the category X is
For example, update with positive weight (addition) as Kx −KX + (ZVV'. Note that the above α
indicates the value of the learning parameter.

Then, the covariance matrix updated in this way is KL
In step G), the 1iIi'Ik value of the covariance matrix obtained by this and the INI-r vector are used as a new recognition word N (standard pattern) in the recognition dictionary memory 5.
, and update its recognition dictionary (Step H).
.

Then, the control parameter i is incremented (step I), and the similarity of the previous learning pattern is calculated again using this updated new recognition dictionary (step J).

Then, calculate the high similarity difference rate DI again (step K
).

Thereafter, it is determined whether the learning of the recognition dictionary is sufficient (step L). This determination is made when the high similarity difference rate DI obtained by performing similarity calculation of the input pattern using the updated recognition dictionary is a predetermined high similarity difference rate Dt.
This is done by determining whether or not the threshold value (DI threshold value for determining the end of learning) is reached. If the predetermined high similarity difference rate Dt has not been reached, the current high similarity difference rate is compared with the high similarity difference rate D (1-1) when using the recognition dictionary before learning mentioned above. The rate DI is more positive than D (+-1).
It is determined whether or not there is a problem (step M).

Through such determination processing, if a predetermined high similarity difference rate D1 is obtained by learning two consecutive recognition dictionaries,
Assuming that the purpose of learning the recognition dictionary has been achieved, the learning process ends. Further, if a predetermined high similarity difference rate DI is not obtained, the above-described learning process is repeatedly executed.

At this time, if an improvement in the high similarity difference rate is confirmed by learning the recognition dictionary, the above-mentioned learning process of the recognition dictionary is repeatedly executed using the value of the learning parameter W described above as is, and the high similarity difference rate is If the rate decreases, the value of the learning parameter is set to, for example, 1/2 (step N),
The learning of the recognition dictionary described above is repeatedly executed. This iterative process is repeated until the predetermined high similarity difference rate Dt is obtained, thereby completing the learning of the recognition dictionary.

In addition, if a limit is set on the number of times of repeated learning of the recognition dictionary, that is, the number of times of updating the covariance matrix, and if a high similarity difference rate of 1 cannot be expected even after updating the covariance matrix a predetermined number of times, If it is determined that the learning pattern has no learning effect on the recognition dictionary, and the learning process is performed in the middle, the number of learning times can be further reduced.

In this way, according to the present device, the weight (
The learning parameter) is set to a value that greatly improves the recognition rate, and the learning of the recognition dictionary is controlled according to this learning parameter, so the recognition rate of the updated dictionary does not decrease and the recognition performance is improved. A recognition dictionary can be created efficiently.

In addition, while controlling the learning parameters of the input pattern in this way, the recognition dictionary is trained with greater weighting in the direction of increasing the recognition rate, making it possible to perform the learning process quickly and efficiently. It becomes possible.

Note that the present invention is not limited to the embodiments described above. For example, in the above embodiment, to control whether learning is performed or not, the ratio of the total (high similarity However, instead of this high similarity difference rate, we calculated the number of person kataans for which the difference between the similarity of the human-powered category and the maximum similarity of other categories is less than a threshold of 1-. A ratio to the number of input patterns that are equal to or less than the same threshold value may be used. Even with such a value, the pattern recognition dictionary represents the degree of separation from other categories, so it can be used until a certain degree of separation is achieved.

In addition, for example, the number of times the same learning pattern is used repeatedly and the predetermined high similarity difference rate Dt for terminating the learning process can be modified in various ways depending on the recognition performance required of the device. be. Furthermore, the learning parameter setting algorithm can be modified in various ways. For example, learning parameters may be set by finding a fusion matrix between the recognition dictionary and the input pattern for each category to be recognized.

Although the speech pattern recognition process and the learning of the recognition dictionary have been explained here as an example, the present invention can be applied to pattern recognition processing apparatuses according to various recognition algorithms, such as handwritten character recognition.

In short, the present invention collects learning patterns and discards them according to the degree of similarity to human cover patterns and the learning processing status of the recognition dictionary, and can be implemented with various modifications without departing from the gist thereof. I can do it.

[Effects of the Invention] As explained above, according to the present invention, the learning parameters required when learning the recognition dictionary are set to values that improve the recognition rate, based on the similarity calculation result between the input pattern and the dictionary pattern. Since the recognition dictionary is trained after setting, the recognition rate does not decrease due to updating the recognition dictionary. Moreover, in the present invention, since a large learning parameter is set in the direction of improving the recognition rate, the number of updates until convergence can be reduced, and a recognition dictionary with good recognition performance can be quickly created. .

[Brief explanation of the drawing]

FIG. 1 is a schematic configuration diagram of a speech recognition device according to an embodiment of the present invention, FIG. 2 is a diagram showing the basic flow of recognition dictionary learning processing in the same device, and FIG. 3 is a diagram showing learning parameters in the same device. FIG. 3 is a diagram showing the flow of learning processing of a recognition dictionary using settings of and learning parameters thereof. DESCRIPTION OF SYMBOLS 1... Voice input unit, 2... Feature extraction unit, 3... Storage unit, 4... Similarity calculation unit, 5... Recognition dictionary memory, 6... Similarity storage unit, 7 ...Learning parameter setting section, 8.Learning section, 9.Control section. Applicant's agent Patent attorney Takehiko Suzue Figure 1 Figure 2 Figure 3

Claims (4)

[Claims] (1) means for analyzing an input pattern to obtain characteristic parameters of the input pattern; means for comparing the characteristic parameters with a recognition dictionary to determine the degree of similarity (distance) of the input pattern; A means for storing the degree of similarity (distance) stored in the means, and a degree of similarity (distance) stored in this means.
Find the number n1 of patterns for which the difference between the similarity (distance) of a certain category and the similarity (distance) of another category is greater than or equal to a predetermined threshold, and the number n2 of patterns for which the difference is less than or equal to the same threshold. A pattern recognition device comprising: means for determining learning parameters necessary for learning the recognition dictionary according to this information; and means for learning the recognition dictionary according to the learning parameters. (2) Learning of the recognition dictionary involves updating the covariance matrix and its KL
2. The pattern recognition device according to claim 1, wherein the learning parameter is used as a weighting coefficient in the updating process of the covariance matrix. (3) The learning parameters are determined by the similarity (distance) of each recognition target category and the similarity (distance) of other categories.
The value of the learning parameter of the input pattern for the variance matrix of the recognition target category is determined according to the number of patterns for which the difference is above a certain threshold value and the number of patterns for which the difference is below the same threshold value, and When the recognition rate decreases, the value of the learning parameter is set smaller than the previous set value, and when the recognition rate improves, the value of the learning parameter is maintained at the previous set value. The pattern recognition device according to item 1. (4) The means for determining the learning parameter is n1/(n1+n2) determined from the numbers n1 and n2 of the patterns.
2. The pattern recognition device according to claim 1, wherein the learning parameter is determined according to the value of .