JP2507308B2 - Pattern recognition learning method - Google Patents

Description

TECHNICAL FIELD OF THE INVENTION The present invention relates to a highly practical pattern recognition learning method capable of accurately recognizing an input voice, for example.

[Technical background of the invention and its problems]

Recently, pattern recognition processing technology for voice has developed,
Widely applied to product management in factories and various telephone services. In addition, application to speech word processors is being promoted, and demands for expanding recognition targets and improving recognition performance are increasing.

By the way, in a specific speaker recognition device in which the speakers to be recognized are limited, by creating a standard pattern (speech recognition dictionary) using the voice uttered by the specific speaker, large vocabulary recognition and continuous recognition are relatively easy. It is possible to improve the recognition performance for spoken voice. That is, since the speaker to be recognized is specified, it is possible to easily improve the performance of the voice recognition dictionary for the speaker and improve the recognition performance. Specifically, for example, when recognizing an input voice by DP matching or a simple similarity method, the average of voice patterns input for learning the recognition dictionary is used as a standard pattern to improve the performance of the recognition dictionary. It can be planned.
In an extreme case, the standard pattern can be created from only one learning pattern, and the input voice can be recognized using this standard pattern.

However, if a voice pattern that is slightly deviated from the above learning pattern due to pattern deformation is input, the similarity value with, for example, the standard pattern decreases, so the recognition rate is not so improved. By the way, 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 with respect to the deviation of the input voice pattern from the average pattern. Therefore, there is a problem in that the recognition rate cannot be expected to improve to a certain degree or more, no matter how many learning patterns are increased.

On the other hand, a recognition method using a covariance matrix such as a composite similarity method or a quadratic recognition function has been proposed as a recognition method for effectively recognizing an input voice pattern having a deviation from the average pattern. According to these recognition methods, for example, the covariance matrix reflects the distribution of deviations with respect to the average pattern. Therefore, by increasing the learning patterns, the recognition rate is much higher than that of the simple similarity method described above. Can be achieved.

However, on the other hand, in order to obtain a covariance matrix that correctly reflects the deviation distribution of the recognition target pattern and create a high-performance recognition dictionary, collect a huge amount (number) of learning patterns,
I needed to learn these.

[Object of the Invention]

The present invention has been made in consideration of such circumstances.
The purpose of this is pattern recognition with high versatility, which enables highly accurate recognition processing of the input pattern according to the degree of collection of the learning pattern or the property of the feature vector extracted from the input pattern. To provide a learning method.

[Outline of Invention]

INDUSTRIAL APPLICABILITY The present invention is used for creating a recognition dictionary in a pattern recognition device for recognizing the input pattern by collating a feature vector of the input pattern obtained by analyzing the input pattern with a previously registered recognition dictionary. According to the number of learning patterns or the number of dimensions of the feature vector extracted from the input pattern used for the recognition process, the number of axes of the recognition dictionary used for matching with the feature vector is controlled to improve the recognition performance. The recognition process can be efficiently performed while maintaining a sufficiently high level.

[Effects of the Invention] Thus, according to the present invention, according to the number of inputs of the learning pattern used to create the recognition dictionary or the number of dimensions of the feature vector extracted from the input pattern, the matching with the feature vector is provided. The number of axes of the recognized dictionary is controlled. Therefore, for example, the input pattern can be recognized with high accuracy by an appropriate recognition processing procedure according to the degree of learning of the recognition dictionary or according to the quality of the input pattern (the number of dimensions of the feature vector). Therefore, it is possible to improve the recognition rate as well as the efficiency of the recognition processing by appropriate recognition processing according to the degree of the learning without making much effort to learn the recognition dictionary. . Further, 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, so that the versatility of the recognition device can be improved, etc. The practically great effect can be obtained.

Example of Invention

An embodiment of the method of the present invention will be described below with reference to the drawings.

Although the input voice uttered word by word is described as a recognition target pattern here, the same can be applied to a pattern recognition device that recognizes characters, figures, and the like. Further, the unit to be recognized may be determined according to its specifications, for example, a phoneme, a syllable, and a continuously uttered voice.

FIG. 1 is a schematic configuration diagram of a voice recognition device to which the embodiment system is applied.

The input voice input through the voice input unit 1 including a microphone or the like is given to the feature extraction unit 2, and the feature vector of the input voice pattern is extracted. The feature extraction unit 2 converts the input voice into, for example, a cutoff frequency of 5.6K.
After a low pass filter of 12 Hz, the sampling frequency of 12 KHz
It is quantized into a bit digital signal. The digitized input voice data is analyzed through a band pass filter group of, for example, 8 channels, further smoothed and logarithmically converted, and then, for example, as a time series of the characteristic parameters of the input voice every 10 msec. It is outputting.

Instead of this digital filter analysis processing, linear prediction analysis (LPC analysis) or the like may be performed to obtain the time series of the characteristic parameters.

From the time series of the characteristic parameters analyzed in this way, the characteristic vector of the input speech pattern is obtained by, for example, detecting a voice section and cutting out the feature parameter based on the detection result of the voice section. Specifically, the characteristic parameters (8 points in the frequency axis direction) of each division point (8 points in the time axis direction) obtained by dividing the voice section into seven equal parts are extracted from the time series of the characteristic parameters, and the extracted input voice ( For example, it is output as a feature vector of a word voice pattern. Therefore, in this case, the feature vector of the input voice pattern is extracted as a 64-dimensional vector.

The feature vector of the input voice pattern thus obtained by the feature extraction unit 2 is given to the recognition unit 3 at the time of the input voice recognition processing, and given to the learning unit 4 at the time of learning. In the word recognition device for a specific speaker, voice input of a voice pattern for learning is performed prior to utterance of voice used for recognition processing. Then, under the control of the control unit 5, the learning unit 4 performs a process of creating a recognition dictionary (standard pattern) according to the learning pattern. The recognition dictionary (standard pattern) created by the learning unit 4 is stored in the standard pattern dictionary memory 6 and is used for the recognition process of the input voice pattern in the recognition unit 3.

Reference numeral 7 in the figure is a display unit for displaying the recognition result of the input voice pattern by the recognition unit 3.

The control unit 5 controls the learning method of the recognition dictionary in the learning unit 4 based on the number of input voice patterns used by the learning unit 4 to create the recognition dictionary. In addition, the control unit 5
Is a table preset using 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 by the learning unit 4 to create the recognition dictionary. By referring to this, the number of axes of the recognition dictionary used for the recognition process of the input voice pattern in the recognition unit 3 is controlled.

The recognition unit (standard pattern) is created in the learning unit 4 as the recognition dictionary φ1 of the first axis when only one learning voice pattern for a certain recognition target category is input, for example. Then, when the second and third learning patterns are input, the Schmitt orthogonalization is applied to the first axis recognition dictionary φ1 to obtain a second axis recognition dictionary φ2 and a third axis recognition dictionary. φ3. Similarly, every time a learning pattern is input, the Schmitt orthogonalization is performed to sequentially create the recognition dictionary φm for the Mth axis.

The recognition dictionary is created by the orthogonalization of the Schmidt. For example, when the Mth new learning pattern fm is input, the standard pattern φm of the Mth axis is calculated. Is calculated as

In this way, the standard patterns from the first axis to the M-th axis, which are obtained according to the number of input voice patterns provided for the learning process, are used as the recognition dictionary in the standard pattern memory 6
Stored in.

In addition, in the word voice recognition device used by the unspecified speaker,
Since the variation of the input voice pattern is larger than that in the case of using a specific speaker, for example, the feature vector is changed in the time axis direction.
Data of 16 points and 16 points in the frequency axis direction are extracted as a 256-dimensional vector.

Then, when the number of learning patterns for creating the recognition dictionary is 1 to 10, the covariance matrix of these learning patterns is calculated, and the eigenvalue and eigenvector are obtained by expanding the covariance matrix by KL, A preset number of axes, for example, standard patterns (recognition dictionary) φ1, φ2, ~ φ4 from the first axis to the fourth axis
And If the number of learning patterns for creating the recognition dictionary further increases and the number becomes 11 to 30, the covariance matrix of these learning patterns is calculated in the same manner including the previously input learning patterns, and the KL expansion is performed. By processing, a new preset number of axes, for example, standard patterns (recognition dictionaries) φ1, φ2, ~ φ6 from the first axis to the sixth axis, is obtained, and the previously obtained recognition dictionary is updated. To do.

In addition, the characteristic nucleus of the recognition dictionary previously obtained is updated using the characteristic nucleus of the newly input learning pattern,
This may be KL expanded to update its recognition dictionary.

Similarly, when the number of learning patterns increases to 30 or more, the KL expansion of the covariance matrix causes the first axis to
Obtain the recognition dictionary (standard pattern) φ1, φ2, ~ φ10 up to 10 axes.

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

Then, in the recognition unit 3, the standard pattern φi (i = 1, 2,
~ M) is used to collate with the input voice pattern f to be subjected to the recognition processing by calculating the next composite similarity S.

However, ‖φi‖ is normalized to (1),
λi is a coefficient.

Here, the number M of axes of the recognition dictionary (standard pattern) used for this recognition processing is the number of learning patterns used for creating the recognition dictionary, or the dimension of the feature vector extracted by the feature extraction unit 2. According to the number, it is controlled under the control of the control unit 5.

For example, when the number of learning patterns used to create the recognition dictionary is 1 to 10, only the standard patterns up to the fourth axis are obtained according to the above-mentioned example. Is instructed to perform the standard pattern of the first to fourth axes, the feature vector of the input voice, and the similarity calculation. Further, when the number of learning patterns used for learning the recognition dictionary is 11 to 30, according to the above-described example, it is instructed to perform the composite similarity calculation using the recognition dictionary up to the sixth axis, and the learning pattern If the number is 31 or more,
It is instructed to perform composite similarity calculation using the recognition dictionary up to the 10th axis.

In this way, the control unit 5 determines the number of axes of the recognition dictionary to be used in the calculation of the composite similarity with the feature vector of the input voice according to the number of learning patterns used for learning (creating) the recognition dictionary. It controls and eliminates the waste of calculation of the recognition process.

The number of axes of the recognition dictionary used in the recognition process may be individually controlled for each recognition target category, or the number of axes of each recognition dictionary obtained for each of the plurality of recognition target categories may be controlled. It is also possible to control the overall control unitedly according to the one having the lowest number of axes.

That is, FIG. 5 shows an example of the number of learning patterns learned for creating the recognition dictionary for the 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 by the category.

Specifically, only the recognition dictionaries up to the fourth axis are required for the category "Akita", but the recognition dictionaries up to the tenth axis are required for the categories "Tokyo" and "Osaka".

However, the number of axes of the recognition dictionary used for the recognition process may be changed for each category, but since the control may be complicated, for example, the recognition dictionary with the smallest number of axes is selected. The composite similarity calculation for each recognition target category may be performed according to the heading and the number of axes thereof. By doing so, it is possible to unify the evaluation conditions of the similarity value for each recognition target category, so that it is possible to perform highly accurate recognition processing.

That is, in this example, the recognition process may be performed using only the recognition dictionary up to the fourth axis.

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

For example, as shown in FIG. 2, the segmentation processing unit 8 cuts out each monosyllabic of the input voice from the time series (for example, 16 channels) of the feature parameters of the input voice extracted by the feature extraction unit 2. . For the vowel component, the analysis output of the 16 channels is cut out as one feature pattern of the vowel part as a characteristic pattern for vowel recognition, and for the consonant component, the analysis output of the 16 channels is adjacent to each other.
As a characteristic parameter of 8 channels grouped by channel, 8 frames including a portion from a consonant to a vowel are cut out as a consonant pattern.

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

The control unit 5 independently controls the learning unit 4 to create a recognition dictionary for such vowel patterns and consonant patterns. For example, when creating a vowel pattern recognition dictionary, the number of recognition dictionary creation axes is controlled according to the number of learning patterns, as shown in FIG.
When creating a recognition dictionary for consonant patterns, the number of axes for creating the recognition dictionary is controlled according to the number of learning patterns, as shown in FIG. 4, for example. Such a recognition dictionary may be created by Schmitt orthogonalization as described above, but it is preferable to use the covariance matrix of the learning pattern as the KL expansion to obtain its eigenvalues and eigenvectors, and use this as the recognition dictionary. . When updating the recognition dictionary, for example, the covariance matrix for the unspecified speaker is updated by the learning pattern of the specified speaker, and the updated covariance matrix is expanded by KL to make the recognition dictionary the specified speaker. It should be adapted.

When the input voice is recognized, the dimension of the vector is different depending on whether the feature pattern is the vowel feature vector or the consonant feature vector, and therefore the recognition unit 3 provides the recognition process. The number of axes of the recognition dictionary used for calculating the composite similarity with the feature vector is controlled according to the dimension number of the feature vector.

As a result, the recognition process using the recognition dictionary having the optimum number of axes according to the feature vector of the input pattern provided for the recognition process can be performed efficiently and accurately.

Even in the 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 depends on the number of learning patterns used to create the recognition dictionary. May be controlled. That is, according to the information of both the number of dimensions of the feature vector and the number of learning patterns used to create the recognition dictionary,
You may make it control the number of axes of the recognition dictionary used for a recognition process.

As described above, according to the present device, according to the number of learning patterns used for creating the recognition dictionary, or the number of dimensions of the feature vector of the input pattern provided for the recognition process,
The number of axes of the recognition dictionary used for matching with the feature vector of the input pattern is controlled, and the recognition processing is performed with the optimum number of axes. Therefore, it is possible to perform effective recognition processing by using a recognition dictionary having an appropriate number of axes without imposing a burden on the user of the recognition device. Also, regardless of the number of learning patterns used to create the recognition dictionary, appropriate recognition processing is performed according to the number of times of learning of the recognition dictionary.

Therefore, without spending effort to collect learning data,
In addition, practically great effects such as simple and effective pattern recognition can be achieved according to the number of dimensions of the feature vector of the input pattern to be recognized.

The present invention is not limited to the above embodiments. For example, the selection setting standard of the number of axes of the recognition dictionary may be determined according to the learning method of the recognition dictionary. In addition, here, the standard pattern was created using only the learning pattern of the user, but a standard pattern for the unspecified speaker is prepared in advance, and this is used as the standard pattern of the first axis, and the learning pattern is used. The standard patterns of the second and subsequent axes may be sequentially created according to the input of. Further, the already created standard pattern may be updated to be adapted to the specific speaker. Further, the learning of the recognition dictionary may be further developed to create a high-performance recognition dictionary for unspecified speakers. Furthermore, although voice recognition has been described as an example in the embodiment, the present invention can be similarly applied to character and graphic recognition processing. In short, the present invention can be variously modified without departing from the scope of the invention. You can

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram of a voice recognition device to which an embodiment system of the present invention is applied, FIG. 2 is a diagram showing a schematic configuration of another embodiment device, and FIGS. 3 and 4 are respectively recognition dictionary. FIG. 5 is a diagram showing the relationship between the number of learning patterns used for creation and the number of axes of the created recognition dictionary, and FIG. 5 shows an example of the number of learning patterns of the recognition dictionary and the number of axes of the recognition dictionary that differ depending on the recognition target category. It is a figure. 1 ... voice input section, 2 ... feature extraction section, 3 ... recognition section,
4 ... Learning section, 5 ... Selection section, 6 ... Standard pattern dictionary memory, 6 ... Display section, 8 ... Segmentation processing section.

Claims (2)

(57) [Claims] 1. When an input pattern is analyzed to extract a feature vector of a feature of a predetermined dimension of the input pattern, and the feature vector is collated with a recognition dictionary registered in advance, the input pattern is recognized, The number of axes of the recognition dictionary used for matching with the feature vector is controlled according to the dimension number of the feature vector used for matching with the recognition dictionary or the number of learning patterns used for creating the recognition dictionary. A pattern recognition learning method characterized by: 2. The input pattern recognition process is performed by the composite similarity method, and the recognition dictionary is created by KL expansion of the characteristic kernel or Schmidt orthogonalization process. The pattern recognition learning method described in the first item of the range.