JPH08254991A - Pattern recognition device - Google Patents

Abstract

PURPOSE: To provide a pattern recognition device capable of obtaining threshold values rationally and improving a recognition rate. CONSTITUTION: This device is provided with a threshold value setting part 6 setting threshold values by calculating threshold values for every pattern from the statistical quantity of the mean value and the standard deviation of the adaptabilities of patterns at the time of recognizing patterns previously registered in a storage part 5 in first order correctly and a recognition discriminating means 8 outputting a signal that an errorneous recognition is present when the adaptabilities of the patterns shown in the first order as the recognition result by a recognition means 7 are less than the threshold values is provided in a pattern matching part 9. Since the signal that the errorneous recognition is present is outputted when the adaptabilities of patterns shown in the first order as the result of the recognition, patterns in which probabilities of errorneous recognitions are high arc actively eliminated and, thus, the reliability of a pattern recognition is largely improved and, moreover, a recognition rate is improved because other recognition means and a human can perform the backup of the pattern recognition device.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition apparatus such as voice recognition and image recognition in information processing technology, and more particularly to a technology for preventing erroneous recognition thereof.

[0002]

2. Description of the Related Art Conventionally, as an effective method for preventing erroneous recognition not only by a voice recognition device but also by a pattern recognition device, a first method for increasing recognition accuracy peculiar to the recognition device and a pattern to be recognized are limited. And there is a second method of thresholding.

[0003]

However, the first method is difficult to realize because it requires a huge amount of calculation, and thus it is difficult to reliably prevent erroneous recognition, and the second method is difficult. However, there is a problem that the recognition rate cannot be improved because there is no rational standard regarding how to set the threshold.

The present invention solves the above problems,
An object of the present invention is to provide a pattern recognition device that can reasonably obtain a threshold value and improve the recognition rate.

[0005]

In order to solve the above problems, a pattern recognition apparatus of the present invention includes a characteristic parameter analyzing section for extracting characteristic parameters from a signal for pattern recognition, and a result of analysis by the characteristic parameter analyzing section. A pattern matching unit for recognizing the pattern by obtaining the matching degree between the registered pattern and the pre-registered pattern, and an average value of the matching degrees of the patterns when the registered pattern is correctly recognized in the first order in advance. A threshold value setting unit for calculating and setting a threshold value for each pattern from the statistical amount of standard deviation is provided, and the pattern matching unit recognizes the conformity of the pattern shown as the first rank as a result of the recognition. It is characterized in that a means for outputting a signal with erroneous recognition is added when it is less than the threshold value.

[0006]

With the above structure, the threshold value for each pattern is obtained from the statistical value of the average value and the standard deviation when the registered pattern is correctly recognized in the first rank, and the pattern shown as the first rank as a result of recognition. When the degree of conformity of is less than the threshold value, a signal with erroneous recognition is output.

[0007]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a voice recognition device which is an example of the pattern recognition device of the present invention.

As shown in the figure, the audio signal input from the audio input unit 1 such as a microphone is input to the preprocessing unit 2. The pre-processing unit 2 is composed of a noise filter and a level detection circuit that detects a word section when the voice signal is equal to or higher than a certain threshold, and is used for word section detection and recognition of pre-emphasis in addition to noise removal. The preceding preparation process is performed and output to the acoustic analysis unit 3. The acoustic analysis unit 3 includes a band-pass filter group, an A / D converter, and a microcomputer, performs spectrum analysis, extracts characteristic parameters, and sends the characteristic parameters to the recognition unit 4 including a microcomputer.

The recognition unit 4 includes a plurality of patterns (eg, male, female,
The storage unit 5 in which babies' utterances and patterns having different utterance lengths) are registered, and from the statistical value of the average value of the fitness and the standard deviation when recognizing the registered pattern in the first order in advance, Threshold setting section 6 for calculating and setting thresholds
(Details will be described later), a recognition unit 7 for obtaining the matching degree of two patterns, and a recognition determination unit 8 described later,
The pattern matching unit 9 recognizes a word, outputs the recognition result, and outputs a message to that effect when there is a possibility of erroneous recognition.

A threshold setting method in the threshold setting unit 6 will be described with reference to the flowchart of FIG. First, when a threshold value setting command signal is input from the outside (step-1), the No1 (n = 1) word is sequentially (step-2), and the registered patterns stored in the storage unit 5 are stored for each same word. An instruction for measuring the conformity of the (sample) is output to the recognition means 7 (step-3). The recognizing means 7 outputs the pattern matching degree to the threshold value setting unit 6 although the pattern is correctly recognized in the first rank.

As shown in Table 1, the threshold value setting unit 6 stores the input pattern conformity (step-4) and takes the statistics.

[0012]

[Table 1]

The statistical values are, for example, considered to be a normal distribution.
If it is done, the recognition means will take the correct first order.
Average μ obtained from statistics of conformity Y of turn X_xAnd standard deviation
Difference σ _xIs calculated (step-5), and α_xThe putter
As an appropriate constant depending on X, (μ_x−α_x・
σ_x) Is used as the threshold value of pattern X (step
Up-6).

That is, when the statistics of the goodness of fit Y of the correctly recognized ones are taken and it is assumed that they follow a normal distribution, (μ _x −α _x · σ _x ) <Y <(μ
95% of the data is in the range defined by ( _x + α _x · σ _x ). Even if the pattern is unknown in advance, what is recognized correctly should follow the normal distribution, and the goodness of fit in the case of misrecognition should be smaller even if it is recognized as the first rank. Introducing α _x to determine
If (μ _x −α _x · σ _x ) <Y, it is considered that the recognition is correct. That is, the threshold value is (μ _x −α _x · σ _x ).

Step-3 for all words (N)
~ 6 are executed (steps -7, 8). By setting the threshold value of the threshold value setting unit 6, the threshold value of each word (μ
_x −α _x · σ _x ) is set.

The recognition means 7 of the pattern matching section 9 is
When a certain pattern is presented by the acoustic analysis unit 3 in the state where the threshold value is obtained, the matching degree between the word pattern obtained as a result of the acoustic analysis and the registered pattern registered in the storage unit 5 , The pattern X shown as the first rank is output as a recognition result, and the matching degree Y
Is output to the recognition determination means 8.

The recognition determining means 8 determines that the input goodness of fit Y is
The threshold value (μ _x −α _x stored in the threshold value setting unit 6
When σ _x ) or more, the recognition result is good, and when it is less than the threshold value (μ _x −α _x · σ _x ), false recognition is output.

Thus, the reliability can be greatly improved by obtaining the threshold value from the statistics and positively outputting the possibility when there is a possibility of erroneous recognition.
Furthermore, another recognition means or a person can back up the voice recognition device, and the recognition rate can be improved.

In the recognition means 7 of the pattern matching section 9, there are two types of erroneous recognition: a. False recognition between registered patterns, b. The unregistered pattern may be erroneously recognized as a registered pattern, but the present invention is applicable to both.

In the present embodiment, the voice recognition device is described, but the invention can be similarly applied to the image recognition device.

[0021]

As described above, according to the present invention, the threshold value for each pattern is obtained from the statistical value of the average value and the standard deviation when the registered pattern is correctly recognized in the first order,
When the conformity of the pattern shown as the first rank as a result of recognition is less than the threshold value, a signal with false recognition is output, so that a high probability of false recognition can be positively excluded, and reliability can be improved. The pattern recognition device can be backed up by other recognition means or human by positively outputting the possibility of erroneous recognition, and the recognition rate can be improved. .

[Brief description of drawings]

FIG. 1 is a configuration diagram of a voice recognition device according to an embodiment of the present invention.

FIG. 2 is a flowchart of a threshold value setting method for the voice recognition device.

FIG. 3 is a distribution diagram of fitness of the same word in the same speech recognition apparatus.

[Explanation of symbols]

 1 voice input unit 2 pre-processing unit 3 acoustic analysis unit (feature parameter analysis unit) 4 recognition unit 5 storage unit 6 threshold value setting unit 7 recognition unit 8 recognition determination unit 9 pattern matching unit

Claims (1)

[Claims] 1. A feature parameter analysis unit for extracting a feature parameter from a signal for pattern recognition, a pattern obtained as a result of the analysis by the feature parameter analysis unit, and a degree of matching between a pattern registered in advance to obtain a pattern. A pattern matching unit for recognizing the pattern is provided, and a threshold value for each pattern is calculated and set from the average value of the goodness of fit of the pattern and the statistical amount of the standard deviation when the registered pattern is correctly recognized in the first order. A threshold value setting unit is provided, and the pattern matching unit is provided with a unit for outputting a signal indicating erroneous recognition when the conformity of the pattern shown as the first rank is less than the threshold value as a result of recognition. A pattern recognition device characterized in that