JPS62114082A - Pattern recognition learning system - Google Patents

Abstract

PURPOSE:To execute a recognition processing efficiently by controlling the number of axes of a recognition dictionary in accordance with the number of dimensions of a feature which is extracted from an input pattern offered for the recognition processing. CONSTITUTION:An input voice which has been inputted through a voice input part 1 is given to a feature extracting part 1, and a feature vector of its input voice pattern is extracted. The feature vector of the input voice pattern which as been derived by the feature extracting part 2 is given to a recognizing part 3, and a learning part 4, at the time of a recognition processing of the input voice, and at the time of learning, respectively. In the learning part 4, a generation processing of a recognition dictionary (standard pattern) conforming to a learning pattern is executed, and it is stored in a standard pattern dictionary memory 6. A control part 5 controls a learning method of the recognition dictionary in the learning part 4 from the number of the input voice pattern which have been used for generating the recognition dictionary in the learning part 4, and also controls the number of axes of the recognition dictionary used for the recognition processing of the input voice pattern in the recognizing part 3, in accordance with the number of dimensions of the feature vector of the input voice.

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention provides a highly practical pattern recognition learning method that can, for example, accurately recognize input speech [Technical background of the invention and its problems] Recently, Pattern recognition processing technology for voice has been developed and is widely applied to product management in factories and various telephone services. Applications to voice word processors are also progressing, and demands for expanding recognition targets and improving recognition performance are becoming stronger.

By the way, with a specific speaker recognition device that restricts the speakers to be recognized, it is relatively easy to recognize large phrases by creating a standard pattern (speech recognition dictionary) using the speech uttered by the specific speaker. It is possible to improve recognition performance for continuously uttered voices.

In other words, since the speaker to be recognized is specified,
It becomes possible to easily improve the performance of the speech recognition dictionary for the speaker and improve its recognition performance. Specifically, when recognizing input speech using DP matching or simple similarity method, for example, the performance of the recognition dictionary can be improved by using the average of the input speech patterns for learning the recognition dictionary as the standard pattern. It is possible. In extreme cases, 1
A standard pattern can be created even from just one learning pattern, and input speech can be recognized using this standard pattern.

However, if a speech pattern slightly deviated from the learning pattern is input due to pattern deformation, the similarity Hlf[ with the standard pattern, for example, will decrease, so that the recognition rate cannot be expected to improve much. Incidentally, when the number of learning patterns is increased, the standard pattern obtained as the average pattern is improved, and the recognition rate can be improved.

However, a problem still remains regarding the deviation of the input speech pattern from the above-mentioned average pattern. Therefore, there is a problem in that no matter how many learning patterns are increased, the recognition rate cannot be expected to improve beyond the level that can be achieved.

On the other hand 1. Recognition methods using covariance matrices, such as a composite similarity method and a quadratic discriminant function, have been proposed as recognition methods for effectively recognizing input speech patterns that deviate from the average pattern. According to these recognition methods, for example, the covariance matrix reflects the distribution of deviations from the average pattern, so by increasing the number of learning patterns, it is possible to achieve a much higher level of zero recognition than the simple similarity reduction method described above. rate can be achieved.

However, on the other hand, in order to obtain a covariance matrix that accurately reflects the distribution of deviations in the recognition target pattern and to create a high-performance recognition dictionary, it is necessary to collect and learn a huge amount (number) of learning patterns. was there.

[Purpose of the invention]

The present invention has been made in consideration of such circumstances, and its purpose is to adjust the input pattern according to the degree of collection of learning patterns or according to the nature of the feature vector extracted from the input pattern. The purpose of this invention is to provide a highly versatile pattern recognition learning method that can recognize and process patterns with high precision.

[Summary of the Invention] The present invention provides a pattern recognition device that recognizes an input pattern by comparing a feature vector of the input pattern obtained by analyzing the input pattern with a recognition dictionary registered in advance, comprising: a recognition dictionary. The number of axes of the recognition dictionary used for matching with the feature vector is controlled according to the number of Gakuga patterns used to create it or the number of dimensions of the feature vector extracted from the input pattern used for recognition processing. In this way, recognition processing can be performed efficiently while maintaining sufficiently high recognition performance.

〔Effect of the invention〕

Thus, according to the present invention, the number of input learning patterns used to create the recognition dictionary or the number of dimensions of the feature vector I-L extracted from the input pattern,
The number of axes of the recognition dictionary used for comparison with the .

Therefore, the input pattern can be recognized with high precision by using an appropriate recognition processing sequence depending on, for example, the degree of training of the recognition dictionary or the quality of the input pattern (number of dimensions of the feature vector). Therefore, without putting much effort into learning the recognition dictionary, it is possible to improve the recognition rate and recognition processing efficiency by performing appropriate recognition processing according to the degree of learning. . In addition, the number of axes of the recognition dictionary used for matching with the feature vector is set according to the number of dimensions of the feature vector extracted from the input pattern, making it possible to increase the versatility of the recognition device. A great practical effect can be achieved.

[Embodiments of the invention]

Hereinafter, one embodiment of the method of the present invention will be described with reference to the drawings.

Although input speech uttered in units of words will be described here as a pattern to be recognized, the present invention is similarly applicable to a pattern recognition device that recognizes characters, figures, and the like. Also, regarding the units to be recognized, for example, phonemes, syllables,
Furthermore, it may be determined according to the specifications, such as continuously uttered voice.

FIG. 1 is a schematic configuration diagram of a speech recognition device to which an embodiment method is applied.

An incoming voice input through a voice input section 1 consisting of a microphone or the like is given to a feature extraction section 2, where a feature vector of the input voice pattern is extracted. The feature extraction unit 2 converts the input audio into, for example, a cutoff frequency of 5.
．． After passing through a low-pass filter of 6 Kllz, it is quantized into a 12-pitch digital signal at a sampling frequency of 121 (llz).Then, this digitized input audio data is converted into, for example, an 8-channel band bus signal. After analysis processing through a filter ll'f, further smoothing and logarithmic transformation, the input speech is outputted as a time series of characteristic parameters of the input speech every 10 m5ec, for example.

Note that instead of this digital filter analysis processing, linear predictive analysis (LPG analysis) or the like may be performed to obtain the time series of the characteristic parameters.

From the time series of the feature parameters analyzed in this manner, the feature vector of the input speech pattern is determined by, for example, detecting the fIM section and cutting out the feature parameters based on the detection result of this speech section. Specifically, from the time series of the feature parameters, the feature parameters (8 points in the frequency axis direction) of each division point (8 points in the time axis direction) of dividing the speech interval into 7 equal parts are extracted, and these are extracted from the input speech ( For example, it is output as a feature vector of a word (speech) pattern.

Therefore, in this case, the feature patterns 1 to 1 of the input voice pattern are extracted as 64-dimensional vectors.

The feature vector of the input speech pattern thus obtained by the feature extraction section 2 is given to the recognition section 3 during input speech recognition processing, and is given to the learning section 4 during learning. Note that in a word recognition device for a specific speaker, a speech pattern for learning is uttered and inputted prior to uttering the speech to be subjected to recognition processing. Under the control of the control section 5, the learning section 4 creates a recognition dictionary (standard pattern) according to the learning pattern. The recognition dictionary (standard pattern) created by the learning section 4 is stored in the standard pattern 'fIIm memory 6, and is used for the recognition process of the input speech pattern in the recognition section 3.

Note that 7 in the figure is a display unit that displays the recognition result of the input speech pattern by the recognition unit 3.

The control unit 5 causes the learning unit 4 to receive one voice based on the number of human voice butters used to create the recognition dictionary in the learning unit 4.
It controls the learning method of the Noru recognition dictionary. Further, the control unit 5 uses at least one of the number of dimensions of the feature vector of the input speech extracted by the feature extraction unit 2, or the number of feature vectors used in the learning unit 4 to create a recognition dictionary, The number of axes of the recognition dictionary used in the recognition process of the input speech pattern in the recognition unit 3 is controlled.

The creation of the recognition dictionary (standard pattern) in the learning section 4 is performed by, for example, inputting only one learning speech pattern for the recognition target category consisting of the recognition wA of the first axis.
This is done as dictionary φ1. When the second and third learning patterns are input, they are made into Schmidt orthogonal to the recognition dictionary φ1 of the first axis, [dictionary φ2 of the second axis, recognition dictionary φ2 of the third axis] It is set to φ3. Similarly, each time a learning pattern is input, a recognition dictionary φm of the cylinder M axis is sequentially created by Schmidt's orthogonalization.

The creation of the Is dictionary using Schmidt's orthogonalization is performed, for example, by calculating the standard pattern φm of the M-th axis when the M-th new learning pattern fn+ is input.

The standard patterns from the first axis to the Mth axis obtained according to the number of input speech patterns subjected to the learning process in this way are stored in the standard pattern memory 6 as a recognition dictionary.

Note that in a word speech recognition device that is used by an unspecified speaker, the fluctuation of the input speech pattern is larger than in the case of a specific speaker, so for example, the feature vector is
Data of 6 points and 16 points in the frequency axis direction is extracted as a 256-dimensional vector.

And the number of learning patterns for creating a recognition 1liv document is 1~
In the case of 10 learning patterns, calculate the covariance matrix of these learning patterns, perform KL expansion of this covariance matrix to find its eigenvalues and eigenvectors, and then calculate the standard patterns from the 1st axis to the 4th axis ( recognition dictionary) φ1. φ2. ~φ4. Then, when the number of learning patterns for creating a recognition dictionary increases further and the number reaches 11 to 30, the covariance matrix of these learning patterns including the previously input learning patterns is calculated in the same way, and the KL expansion is performed. After processing, a new standard pattern (recognition dictionary) from the first axis to the sixth axis (recognition dictionary) φ1゜φ2. ~φ6 is determined, and the previously determined recognition dictionary is updated.

Note that the previously obtained characteristic kernel of the recognition dictionary is updated using the characteristic kernel of the newly input learning pattern, and this is expanded into KL to update the recognition dictionary.
good.

Similarly, when the number of learning patterns increases to 30 or more, the recognition dictionary (standard pattern) φ1. φ2.
~Find φ10.

In this way, recognition dictionaries with up to a predetermined number of axes are created according to the number of learning patterns used to create recognition dictionaries,
It is stored in the standard pattern dictionary memory 6.

In the recognition unit 3, the standard pattern φi (i=1
．． 2. ~m) is used to perform comparison with the input speech pattern f to be subjected to recognition processing by calculating the following composite similarity S.

S=2m ((λi //λ1)(φi, f)21”1 ÷11φi 112 II f ll2) Cape Σ((λ
i/λ1) (φi, f)21=1 ÷II f 112 ) However, 11φ+ ++ is normalized to (1), and λ1 is a coefficient.

Here, the number of axes M of the recognition dictionary (standard pattern) used in this recognition process is the number of learning patterns used to create the recognition dictionary, or the dimension of the feature vector extracted by the feature extraction unit 2. The controller 5 controls the controller 5 according to the number of controllers.

For example, the number of learning patterns used to create the recognition dictionary is 1.
~10, since only the standard patterns up to the 4th axis have been obtained, the control unit 5 inputs the standard patterns from the 1st to 4th axes to the recognition unit 3.
If the number of learning patterns used for learning the recognition dictionary is 11 to 30, the recognition dictionary up to the 6th axis If the number of learning patterns is 31 or more, it instructs J: to perform a composite similarity calculation using recognition dictionaries up to the 10th axis.

In this way, the control unit 5 controls the i! ! Learning dictionary (
The number of axes of the recognition dictionary used for the composite similarity calculation with the input speech feature patterns is controlled according to the number of learning patterns used in the process (creation), thereby reducing waste of recognition processing needles. I'm there.

The number of axes of the recognition dictionary used in recognition processing may be controlled individually for each recognition target category, or the number of axes of each recognition dictionary determined for each of multiple recognition target categories may be controlled individually. The overall control may be unified depending on the one with the lowest number of axes among them.

That is, Fig. 5 shows an example of the number of learning patterns learned to create a recognition dictionary for a word category and the number of axes of the recognition dictionary created by the learning. There is a difference in the number of axes of the recognition dictionary created depending on the category.

In terms of layout, only the recognition dictionary up to the 4th axis is required for the category "Akita"1, but "Tokyo""
For the category "Osaka", Is dictionaries up to axis 10 are required.

Although it is possible to change the number of axes of the recognition dictionary used for recognition processing for each category, there is a risk that the control will become unnecessarily complicated. Composite similarity calculations for each recognition target category may be performed in accordance with the heading and its number of axes. If J: is used, it is possible to unify the evaluation conditions of similarity values for each recognition target category, and therefore it is possible to perform recognition processing with a high accumulation of power.

In other words, in this example, the recognition process can be performed using the recognition dictionary up to the fourth axis.

By the way, when performing speech [, for example, the feature vector of the vowel component and the feature vector of the consonant component of the input speech pattern may be extracted separately and subjected to recognition processing respectively.

For example, as shown in FIG. 2, in the segmentation processing section 8,
Each monosyllable of the input speech is extracted from the time series (for example, 16 channels) of the feature parameters of the input speech extracted by the feature extraction unit 2. For the vowel component, one frame of the vowel part was extracted from the analysis output of the 16 channels as a feature pattern for vowel recognition, and for the consonant component, the analysis output of the 16 channels was summarized into two adjacent channels. As a characteristic parameter of the channel, eight frames including the transition part from a consonant to a vowel are extracted as a consonant pattern.

In this case, the number of dimensions of the vowel pattern feature vector used for recognition processing is 16 dimensions, and the consonant pattern feature vector is a 64-dimensional vector.

The control unit 5 independently controls the learning unit 4 to create recognition dictionaries for such vowel patterns and consonant patterns. For example, when creating a recognition dictionary for vowel patterns, the number of axes for creating the recognition dictionary is controlled according to the number of learning patterns, as shown in Figure 3, and when creating a recognition dictionary for consonant patterns, for example, As shown in FIG. 4, the number of axes for creating the recognition dictionary is controlled according to the number of learning patterns. Although such a recognition dictionary may be created by Schmidt's orthogonalization as described above, it is preferable to perform a KL expansion on the covariance matrix of the learning pattern to obtain its eigenvalues and eigenvectors, and use this as the recognition dictionary. .

In addition, when updating the ms dictionary, for example, the covariance matrix for unspecified speakers is updated using the learning pattern of a specific speaker, and the updated covariance matrix is subjected to KL expansion. I Just try to adapt the dictionary to the specific speaker.

When input speech is recognized, the number of dimensions of the vector is different depending on whether the feature pattern is a vowel feature vector or a consonant feature vector. The number of axes of the recognition dictionary used for composite similarity with the feature vector is controlled according to the number of dimensions of the feature vector provided to the feature vector.

As a result, the recognition process using the recognition dictionary with the optimal number of axes is performed efficiently and accurately according to the feature vector of the input pattern used for the recognition process.

In addition, even in pattern recognition processing by selecting the number of axes of the recognition dictionary according to the number of dimensions of the feature vector, the number of axes of the recognition dictionary used for the recognition processing is determined according to the number of learning patterns used to create the recognition dictionary. The number may be controlled. In other words, according to the information on both the number of dimensions of the feature vector and the number of learning patterns used to create the ill dictionary,
The number of axes of the recognition dictionary used for recognition processing may be controlled.

As explained above, according to the present device, the feature vector of the input pattern and the number of dimensions of the feature vector of the input pattern to be subjected to recognition processing are The number of axes of the recognition dictionary used for verification is controlled, and recognition processing is performed using the optimal number of axes. Therefore, it is possible to perform effective recognition processing using a recognition dictionary with an appropriate number of axes without placing a burden on the user of the recognition device. Furthermore, regardless of the number of learning patterns used to create a recognition dictionary, appropriate recognition processing is performed according to the number of times the recognition dictionary is trained.

Therefore, it is possible to easily and effectively perform pattern recognition without spending effort on collecting learning data, and according to the number of dimensions of the feature vector of the input pattern to be recognized. Great effects can be achieved.

Note that the present invention is not limited to the embodiments described above. For example, the selection setting criteria for the number of axes of the recognition dictionary may be determined depending on the learning method of the recognition dictionary. In addition, here, the standard pattern was created using only the user's learning pattern, but a standard pattern for unspecified speakers is prepared in advance, and this is used as the standard pattern for the first axis, and the learning pattern The standard patterns for the second and subsequent axes may be created sequentially according to the input. Alternatively, a standard pattern that has already been created may be updated and adapted for a specific speaker.

Further, learning of the recognition dictionary may be further developed to create a high-performance recognition dictionary for unspecified speakers. Further, although the embodiments have been described using voice recognition as an example, the present invention can be similarly applied to character and figure recognition processing, and in short, the present invention can be implemented with various modifications without departing from the gist thereof. I can do it.

[Brief explanation of drawings]

FIG. 1 is a schematic configuration diagram of a speech recognition device to which one embodiment of the present invention is applied, FIG. 2 is a diagram that does not show the schematic configuration of another embodiment of the device, and FIGS. 3 and 4 are recognition dictionaries, respectively. FIG. 3 is a diagram showing an example of the number of learning patterns in the learning pattern textbook used for creating the learning pattern book and the number of axes of the recognition dictionary. DESCRIPTION OF SYMBOLS 1... Voice input part, 2... Feature extraction part, 3... Recognition part, 4... Learning part, 5... Selection part, 6... Standard pattern dictionary memory, 6... Display section, 8... Segmentation processing section.

Claims (2)

[Claims] (1) Analyzing an input pattern to extract a feature vector of a predetermined number of dimensions of the input pattern, and comparing this feature vector with a recognition dictionary registered in advance to recognize the input pattern, the recognition dictionary The number of axes of the recognition dictionary used for comparison with the feature vector is controlled according to the number of dimensions of the feature vector used for comparison with the feature vector or the number of learning patterns used to create the recognition dictionary. A pattern recognition learning method featuring: (2) The input pattern recognition process is performed by the subspace method or the composite similarity method, and the recognition dictionary is created by averaging the learning patterns, KL expansion of the characteristic kernel, or Schmidt orthogonalization. The pattern recognition learning method according to claim 1, which is carried out by.