JP2007065491A - Pattern model generating device, pattern model evaluating device, and pattern recognizing device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a pattern model generating device capable of easily generating a new pattern model improving recognition probability of a signal causing recognition probability of a pattern model to decrease, a pattern model evaluating device suitably used to evaluate the recognition probability of the signal causing the recognition probability of the pattern model to decrease, and a pattern recognizing device which performs pattern recognition by using the pattern model improving the recognition probability. <P>SOLUTION: A data space consisting of a high-order signal pattern model including high-dimensional elements and a high-order noise pattern model is mapped to a data space consisting of low-order signal vectors and low-order noise vectors while the position relation among the high-order pattern models is approximated, and the low-order vector space is sectioned and visualized; and the pattern model for pattern recognition is generated by using signal data and noise data corresponding to low-order vectors at positions at a distance longer than a specified distance from the center. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a pattern model used for pattern recognition of a noise mixed signal in which various signals and various noises are mixed, and in particular, the recognition probability for a noise mixed signal that causes a reduction in the recognition probability of the pattern model. Suitable for evaluating the recognition probability for noise mixed signals that cause a reduction in recognition probability due to a pattern model, pattern model generation program and pattern model generation method capable of easily generating a new pattern model to be improved The present invention relates to a simple pattern model evaluation apparatus, a pattern model evaluation program and a pattern model evaluation method, and a pattern recognition apparatus, a pattern recognition program and a pattern recognition method for performing pattern recognition using the novel pattern model.

  In general, pattern recognition is performed by comparing a feature analysis unit that converts a signal to be recognized into a sequence of feature parameters, and comparing the feature parameter sequence obtained by the feature analysis unit with information about feature parameters prepared in advance. It consists of two parts: a feature matching unit that uses a matching result with a high degree of similarity as a recognition result. Hereinafter, explanation will be given by taking voice recognition as an example. Speech recognition consists of an acoustic analysis unit that converts a speech sample uttered by a speaker into a sequence of feature parameters, and a vocabulary word in which the feature parameter sequence obtained by the acoustic analysis unit is stored in a storage device such as a memory or a hard disk in advance. It is composed of two parts: a voice collating unit that collates with information on the feature parameter of the voice and uses the voice with the highest similarity as a recognition result.

Cepstrum analysis and linear prediction analysis are known as acoustic analysis methods for converting a speech sample into a series of feature parameters. It is stated.
In speech recognition, a technique for recognizing unspecified speaker speech is generally referred to as unspecified speaker speech recognition. In unspecified speaker voice recognition, information related to the characteristic parameters of vocabulary words is stored in the storage device in advance, so that the user does not need to register a word that the user wants to recognize as in the case of specific speaker voice recognition. .

  In addition, the Hidden Markov Model (HMM) method is generally used as a method for creating information on feature parameters of vocabulary words and for collating the information with a sequence of feature parameters converted from the input speech. It has been. In the method based on the HMM, speech units such as syllables, semi-syllables, phonemes, biphones, and triphones are modeled by the HMM. These models are generally called acoustic models.

The method for creating an acoustic model is described in detail in “Chapter 6: Theory and Implementation of Hidden Markov Models” of Non-Patent Document 1 above.
Further, the Viterbi algorithm described in “Chapter 6: Theory and Implementation of Hidden Markov Models” of Non-Patent Document 1 enables a person skilled in the art to easily configure an unspecified speaker speech recognition apparatus.

  Conventionally, there have been many trial and error methods for evaluating the performance of speech recognition. For example, the collected speech data group is appropriately divided into a learning data group and an evaluation data group, and an HMM acoustic model is generated from the divided learning data group. Then, the recognition performance of the acoustic model for the divided evaluation data group is evaluated. In this recognition performance evaluation, the average value, maximum value, minimum value, etc. of each recognition result are calculated. Further, by performing the division process a plurality of times so that the combination of the learning data group and the evaluation data group changes, and by averaging the performance evaluation results for the plurality of division results, a more objective evaluation However, the performance evaluation of speech recognition is empirical and not systematic.

  In addition, voice recognition is expected to be put to practical use in in-vehicle devices such as car navigation systems that require hands-free operation and robots that have voice interactive functions as a technology that provides voice input interfaces. At present, it cannot be said that 100% recognition probability is realized. The current situation is that the recognition probability often falls below 50% depending on the voice quality of the user and how to speak. In the past, the recognition performance was often evaluated using the average value of the recognition probability as a scale. However, when evaluating the performance as an input interface, the minimum value of the recognition probability is evaluated, and the value is, for example, 90%. As described above, it is necessary to guarantee performance in a form that can realize recognition performance of 90% or more for all users.

In addition to guaranteeing the minimum performance, it is desired to generate a pattern model that can exhibit sufficient recognition performance for data that causes a reduction in recognition probability.
For example, as a technique for generating a pattern model from speech data mixed with noise, there are a speech model noise adaptation system described in Patent Literature 1 and Patent Literature 2, and an acoustic model creation method described in Patent Literature 3.

  In the prior art described in Patent Document 1, noise is first clustered by SNR (signal-to-noise ratio), then various types of noise are clustered in a tree structure for each SNR, and noise is predetermined for each cluster. Creates a speech HMM (pattern model) from noise-mixed speech data mixed with SNR, selects a speech HMM that matches the SNR and noise type from the set, and uses it for speech recognition, making it robust against noise A simple voice recognition system.

In addition, the prior art described in Patent Document 2, which is an improved technique of Patent Document 1, is the likelihood that the same applicant as Patent Document 1 gives the selected speech HMM in addition to the prior art described in Patent Document 1. A speech recognition system that is robust against noise is constructed by linearly transforming a speech HMM that is selected so that becomes larger.
In the prior art described in Patent Document 3, first, noise is clustered by a statistical method, voice and noise are mixed with a predetermined SNR for each noise cluster, and noise removal processing is performed. By creating a speech HMM (pattern model) of noise-mixed speech every time, a speech recognition system that is robust against noise is constructed.

JP 2004-117624 A JP 2004-279466 A JP 2004-206063 A Lawrence Rabiner and Biing-Hwang Juang, "Fundamentals of Speech Recognition," Prentice Hall Signal Processing Series, 1993

  However, in the prior arts of Patent Document 1 to Patent Document 3 described above, since an audio HMM of noise mixed speech is created for each noise cluster, an optimal speech HMM is selected from the set of speech HMMs of noise mixed speech. When the process of selecting is required and the size of the noise cluster becomes large, a great deal of time and effort is required for the process of creating a set of voice HMMs of a plurality of noise-mixed voices and the process of selecting an optimal voice HMM of noise-mixed voices. There was a problem that it took.

  In addition, until now, no systematic method has been proposed for deriving combinations of learning data groups and evaluation data groups that cause the recognition probability to decrease, so the minimum performance of pattern model recognition performance is guaranteed. It was difficult to do. None of the prior arts in Patent Document 1 to Patent Document 3 has disclosed a systematic method for deriving a combination of a learning data group and an evaluation data group that causes a reduction in recognition probability.

  Therefore, the present invention has been made paying attention to such an unsolved problem of the conventional technology, and is a novel technique for improving the recognition probability for a noise mixed signal that causes a reduction in the recognition probability of the pattern model. Pattern model generation device, pattern model generation program and pattern model generation method, and pattern model suitable for evaluating recognition probability for noise mixed signals that cause reduction of recognition probability by pattern model An object is to provide an evaluation apparatus, a pattern model evaluation program, a pattern model evaluation method, a pattern recognition apparatus, a pattern recognition program, and a pattern recognition method for performing pattern recognition using the novel pattern model.

In order to achieve the above object, a pattern model generation apparatus according to claim 1 according to the present invention comprises:
A pattern model generation device that generates a pattern model used for pattern recognition of recognition target data to be recognized,
Signal data storage means for storing a plurality of noise-unmixed pre-acquired signal data related to a plurality of objects;
As an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on the signal data, the positional relationship between the higher-order signal pattern models in the data space Low-order signal vector space storage means for storing a low-order signal vector space that is mapped to a data space composed of low-order signal vectors of three dimensions or less in an approximate state,
Noise data storage means for storing a plurality of types of signal-unmixed noise data;
As an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four-dimensional or higher-dimensional elements generated based on the noise data, the positional relationship between the higher-order noise pattern models in the data space Low-order noise vector space storage means for storing a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state where
Among the plurality of low-order signal vectors constituting the low-order signal vector space, a low-order signal vector for generating a pattern model from among low-order signal vectors located at a predetermined distance or more from the center of gravity of all the low-order signal vectors. Low-order signal vector selection means for selecting,
Of the plurality of low-order noise vectors constituting the low-order noise vector space, a low-order noise vector for generating a pattern model from among the low-order noise vectors located at a predetermined distance or more from the center of gravity of all the low-order noise vectors. Low-order noise vector selection means for selecting
Based on the signal data corresponding to the low-order signal vector selected by the low-order signal vector selection means and the noise data corresponding to the low-order noise vector selected by the low-order noise vector selection means, the noise is added to each signal data. Noise mixed signal data generating means for generating noise mixed signal data obtained by mixing data;
Pattern model generation means for generating a pattern model obtained by learning a probability model using the noise mixed signal data as learning data.

  With such a configuration, it is possible to store a plurality of noise-unmixed signal data related to a plurality of objects acquired in advance by the signal data storage means, and the low-order signal vector space storage means As an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on signal data, the positional relationship between each higher-order signal pattern model in the data space is It is possible to store a low-order signal vector space that is mapped to a data space composed of low-order signal vectors of three dimensions or less in an approximate state, and a plurality of types of unmixed signals can be stored by the noise data storage means. It is possible to store noise data and generated based on the noise data by the low-order noise vector space storage means As an alternative to a data space composed of a plurality of higher-order noise pattern models composed of higher-dimensional elements than the three-dimensional elements, the positional relationship between the higher-order noise pattern models in the data space is approximated with three or less dimensions. It is possible to store a low-order noise vector space mapped to a data space composed of low-order noise vectors.

  Further, the low-order signal vector selecting means selects a low-order signal vector located at a position more than a predetermined distance from the center of gravity of all the low-order signal vectors among a plurality of low-order signal vectors constituting the low-order signal vector space. It is possible to select a low-order signal vector for generating a pattern model from among the plurality of low-order noise vectors constituting the low-order noise vector space by the low-order noise vector selection means. It is possible to select a low-order noise vector for generating a pattern model from among low-order noise vectors located at a predetermined distance or more from the center of gravity of the low-order signal vector. Signal data corresponding to the low-order signal vector selected by the means and the low-order noise vector selection means. It is possible to generate noise mixed signal data in which the noise data is mixed with each signal data based on the noise data corresponding to the low-order noise vector, and the noise mixed signal data is generated by pattern model generation means. It is possible to generate a pattern model obtained by learning a probability model as learning data.

  In the low-order signal vector space and the low-order noise vector space, the density of each vector tends to decrease as the distance from the center of gravity increases, and the density is low (outside and near the distribution). The noise mixed signal data generated from the signal data and noise data corresponding to the vector in FIG. 5 tends to generate a low recognition probability for a pattern model for an unspecified object generated using all data in each space, for example. is there. Therefore, noise mixed signal data is generated using data at a position away from the center of gravity by a predetermined distance or more (position range where the existence density is low), and a pattern model is generated using the generated noise mixed signal data as learning data. Therefore, it is possible to generate a pattern model that can improve the recognition probability for signal data that reduces the recognition probability of the pattern model.

In addition, the pattern model generated using data at a position more than a predetermined distance from the center of gravity is relatively good recognition for noise mixed signal data generated by data within a range less than the predetermined distance from the center of gravity. Probability tends to be obtained, so by generating a pattern model using data that is at a predetermined distance or more from the center of gravity, a pattern that can obtain a good recognition probability uniformly for all unspecified target signal data A model can be generated.
Here, the pattern model is, for example, a model of a signal pattern that performs matching on audio data when the predetermined data is audio data, and is a statistical model such as an HMM or a neural network, or a feature parameter. It is expressed using the time series.

  The signal data includes, for example, acoustic data such as human voice, cry data of wildlife such as wild birds, insects, moths, moths, animals, image data, infrared sensor data, acceleration sensor data, azimuth sensor data , Pressure sensor data, vibration sensor data such as piezoelectric elements and vibrometers and all other sensor data, physical data on the charging status of batteries such as lithium ion secondary batteries and fuel cells, electrocardiogram, electromyogram, blood pressure, Biological signal data such as weight, microarray data for genetic analysis, weather data such as temperature, humidity, and pressure, environmental data such as oxygen concentration and nitrogen oxide concentration, time series data such as economic trend data such as stock prices and prices, etc. There is.

The noise data includes noise data for sounds such as house life noise, factory noise, traffic noise, noise data for electrical / electronic circuit noise, noise data for vibration noise, and the like.
In addition, the pattern model is a model including a four-dimensional or higher-dimensional element. For example, in pattern recognition such as speech recognition, a high recognition performance is obtained unless at least a four-dimensional or higher feature parameter is used. This is because it has not been obtained, and in speech recognition, a feature parameter of three dimensions or less capable of realizing practically effective recognition performance has not been found at present.

  In addition, the signal data related to a plurality of objects includes, for example, data itself that can be measured from a plurality of objects, feature amounts extracted from the data, pattern models generated based on the feature amounts, and a text file that describes the contents thereof. This refers to the set. For example, it is a set of speech data uttered by a plurality of speakers, feature amounts extracted from the speech data, a pattern model generated based on the feature amounts, and a text file describing the utterance contents.

  The distance between each higher-order signal pattern model or the distance between each higher-order noise pattern model indicates the similarity between pattern models generated from signal data or noise data of a specific target. For example, Euclid There is a Mahalanobis generalized distance in which the distance or the distance for measuring the similarity is the inner product of the two vectors, and the angle formed by the two vectors is evaluated as the similarity. In the present invention, the distance includes, in addition, a Bhattacharyya distance, a square Euclidean distance, a cosine distance, a Pearson correlation, a Chebyshev, a city block distance (or Manhattan distance), a Minkowski sum, a Calbach information amount, Bayesian distance, Chernov distance, Euclidean distance based on the normal vector of the normal distribution of the pattern model configured by the HMM, and the average of the normal distribution of the pattern model normalized by the standard deviation of the normal distribution of the pattern model configured by the HMM There are a Euclidean distance based on a vector, a butteraria distance based on a normal distribution of a pattern model constituted by an HMM, and the like. That is, although it is referred to as a distance, anything that shows a degree of similarity may be used.

In addition, the process of mapping a four-dimensional or higher-dimensional data space to a three-dimensional or lower-dimensional data space in a state where the distance relationship between the pattern models is approximated is, for example, a small distance between pattern models. All pattern models are mapped to a low-dimensional space (for example, two-dimensional or three-dimensional space) so that the two pattern models are located close to each other and the two pattern models having a large distance between the pattern models are located far from each other. It becomes processing.
For example, when the Euclidean distance is used as the distance between pattern models, the pattern models with a close Euclidean distance in the mapped low-dimensional space mean that the pattern models are more similar than those with a long distance. It is done.

  As a known method for converting a higher-order pattern model into a lower-order vector, the Sammon method (JW Sammon, “A nonlinear mapping for data structure analysis,” IEEE Trans. Computers, vol. C-18). , no.5, pp. 401-409, May 1969.), discriminant analysis (RA Fisher, "The use of multiple measurements in taxonomic Problems," Ann. Eugenics, vol. 7, no. Part II, pp. 179 -188,1936.), Aladjam method (M.A1adjem, "Multiclass discriminant mappings," Signa1 Process., Vol.35, pp.1-18,1994.), Neural network method (J.Mao et a1 ., "Artificial neural networks for feature extraction and mu1tivariate data projection," IEEE Trans.Neura1 Networks, vol.6, no.2, pp.296-317,1995.), Graph-based method (Y.Mori et al., "Comparison of 1ow-dimensional mapping techniques based on discriminatory information," Proc. 2nd International ICSC Symposium on Advances in Intelligent Data Analysis (AIDA'2001), CD-ROM Paper-no. 1724-166, Bangor, United Kingdom, 2001.), projection tracking method (JHFreidman et al., "A projection pursuit algorithm for exp1oratory data ana1ysis," IEEE Trans.Comput., Vol.C-23, no.9 , pp. 881-889, 1974), SOM method (see T. Kohonen, “Self-Organizing Maps,” Springer Series in Information Sciences, vol. 30, Berlin, 1995.) and the like.

Furthermore, the invention according to claim 2 is the pattern model generation device according to claim 1,
The high-order signal pattern model, the high-order noise pattern model, and the pattern model are configured by an HMM (Hidden Markov Model).
With such a configuration, each pattern model can be configured by an HMM (Hidden Markov Model), so that an appropriate pattern model can be generated for signal data with a temporal concept. It is.

  Here, the HMM is a pattern model of a time series signal, and an unsteady time series signal can be modeled by transitioning between a plurality of stationary signal sources. For example, the length of time of speech changes depending on the speaking speed, and the utterance content shows a characteristic shape on the frequency (referred to as spectrum envelope), but the shape depends on the person who speaks, the environment, the content, etc. Fluctuation occurs. HMM is a statistical model that can absorb such fluctuations. The HMM may be defined in any unit (for example, a word or a phoneme), and each HMM (here, “each” is, for example, a word includes a plurality of words, and a phoneme includes a plurality of phonemes. Is composed of a plurality of states, and each state is made up of statistically learned state transition probabilities and output probabilities (probability distributions such as normal distribution and mixed normal distribution). The transition probability absorbs the fluctuation of time expansion and contraction of the voice, and the output probability absorbs the fluctuation of the spectrum.

Furthermore, the invention according to claim 3 is the pattern model generation device according to claim 1 or 2,
The low-order signal vector space and the low-order noise vector space are two-dimensional or three-dimensional data spaces,
A plurality of low-order signal vectors constituting the low-order signal vector space are centered on the centroid of all low-order signal vectors, and the distance between the centroid and the low-order signal vector at the position farthest from the centroid is the radius. The outer circle or outer sphere is divided into n outer circles or inner spheres (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer circle or outer sphere. A plurality of concentric circles consisting of circles and inner circles or outer spheres and inner spheres, or annular or spherical regions formed between curved lines or curved surfaces of concentric spheres are divided into a plurality by radially extending lines or surfaces. Low-order signal vector space partitioning means for partitioning;
A plurality of low-order noise vectors constituting the low-order noise vector space are centered on the centroid of the low-order noise vector, and the distance between the centroid and the low-order noise vector at a position farthest from the centroid is a radius. The outer circle or outer sphere is divided into n outer circles or inner spheres (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer circle or outer sphere. A plurality of concentric circles consisting of circles and inner circles or outer spheres and inner spheres, or annular or spherical regions formed between curved lines or curved surfaces of concentric spheres are divided into a plurality by radially extending lines or surfaces. Low-order noise vector space classification means for classifying;
Low-order signal vector space display means for displaying each low-order signal vector constituting the low-order signal vector space together with the classification result;
Low-order noise vector space display means for displaying each low-order noise vector constituting the low-order noise vector space together with the classification result, and
The low-order signal vector selection means includes a plurality of concentric circles or concentric spheres composed of the outer circle or outer sphere and an inner circle or inner sphere having a radius equal to or greater than the predetermined distance in the displayed lower-order signal vector space. A low-order signal vector for generating a pattern model is selected from low-order signal vectors included in an annular or spherical region formed between each curve or between curved surfaces.
The low-order noise vector selection means includes a plurality of concentric circles or concentric spheres composed of the outer circle or outer sphere and an inner circle or inner sphere having a radius equal to or greater than the predetermined distance in the displayed lower-order noise vector space. A low-order noise vector for generating a pattern model is selected from low-order noise vectors included in an annular or spherical area formed between each curve or between curved surfaces. .

  With such a configuration, when the low-order signal vector space and the low-order noise vector space are two-dimensional data spaces, a plurality of low-order signal vector spaces are configured by the low-order signal vector space dividing means. A low-order signal vector with one outer circle centered on the center of gravity of all the low-order signal vectors and the radius between the center of gravity and the low-order signal vector located farthest from the center of gravity, and the center of gravity And an n-shaped inner circle (n is an integer of 1 or more) having a smaller radius than the outer circle, and an annular region formed between a plurality of concentric circles composed of the outer circle and the inner circle. It is possible to divide into a plurality by lines extending in the radial direction.

  Further, by the low-order noise vector space segmentation means, a plurality of low-order noise vectors constituting the low-order noise vector space are centered on the centroid of the low-order noise vector and located at a position farthest from the centroid. Partitioning into one outer circle whose radius is the distance from the low-order noise vector and n inner circles (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer circle; An annular region formed between the curved lines of a plurality of concentric circles composed of the outer circle and the inner circle can be divided into a plurality by a line extending in the radial direction.

Each low-order signal vector constituting the low-order signal vector space can be displayed together with the classification result by the low-order signal vector space display means, and the low-order noise vector space display means can display the low-order signal vector space. Each low-order noise vector constituting the next-order noise vector space can be displayed together with the classification result.
Further, the low-order signal vector selection means is a ring formed between the respective curves of a plurality of concentric circles each including the outer circle and an inner circle having a radius equal to or greater than the predetermined distance in the displayed low-order signal vector space. A low-order signal vector for pattern model generation can be selected from the low-order signal vectors included in the region of, and the low-order noise vector selection means is configured to select the low-order noise vector space in the displayed low-order noise vector space. A low-order noise vector for generating a pattern model is selected from low-order noise vectors included in an annular region formed between curves of a plurality of concentric circles each having an outer circle and an inner circle having a radius equal to or greater than the predetermined distance. It is possible to select.

  On the other hand, when the low-order signal vector space and the low-order noise vector space are three-dimensional data spaces, a plurality of low-order signal vectors constituting the low-order signal vector space are obtained by low-order signal vector space partitioning means. One outer sphere centered at the center of gravity of all the low-order signal vectors and having a radius as a distance between the center of gravity and the low-order signal vector at a position farthest from the center of gravity, and from the outer sphere centered on the center of gravity and Is divided into n inner spheres (n is an integer of 1 or more) having a small radius, and a spherical region formed between the curved surfaces of a plurality of concentric spheres composed of the outer sphere and the inner sphere, It is possible to divide into a plurality by the surface extending in the radial direction.

  Further, by the low-order noise vector space segmentation means, a plurality of low-order noise vectors constituting the low-order noise vector space are centered on the centroid of the low-order noise vector and located at a position farthest from the centroid. Partitioning into one outer sphere whose radius is a distance from a low-order noise vector and n inner spheres (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer sphere; A spherical region formed between the curved surfaces of a plurality of concentric spheres composed of the outer sphere and the inner sphere can be divided into a plurality by a surface extending in the radial direction.

Each low-order signal vector constituting the low-order signal vector space can be displayed together with the classification result by the low-order signal vector space display means, and the low-order noise vector space display means can display the low-order signal vector space. Each low-order noise vector constituting the next-order noise vector space can be displayed together with the classification result.
The low-order signal vector selection means is formed between the curved surfaces of a plurality of concentric spheres composed of the outer sphere and an inner sphere having a radius equal to or greater than the predetermined distance in the displayed low-order signal vector space. It is possible to select a low-order signal vector for generating a pattern model from low-order signal vectors included in a spherical region, and the low-order noise vector selection means is arranged in the displayed low-order noise vector space. A low-order noise vector for generating a pattern model from among low-order noise vectors included in a spherical region formed between curved surfaces of a plurality of concentric spheres each having an outer sphere and an inner sphere having a radius equal to or greater than the predetermined distance It is possible to select the next noise vector.

  Accordingly, an annular shape (or between curved surfaces) formed between a plurality of concentric circles (or concentric spheres) composed of an outer circle (or outer sphere) and n inner circles (or inner spheres). Since the (spherical) area is divided into a plurality of lines (or faces) extending in the radial direction and the low-order vector space is displayed in this way, the distribution of each pattern model is visually simplified. From the annular (or spherical) region formed by concentric circles (or concentric spheres) having a radius equal to or greater than a predetermined distance in the displayed low-order vector space. It is possible to easily select the next vector (for example, a low-order vector for each region divided into a plurality of lines (or surfaces) extending in the radial direction). Furthermore, by devising how to select this region, for example, by selecting low-order vectors according to the content of pattern recognition and the recognition target, the pattern model and recognition performance specialized for a specific recognition target can be improved. It becomes possible to easily generate an increasing pattern model.

Furthermore, the invention according to claim 4 is the pattern model generation device according to any one of claims 1 to 3,
A high-order signal pattern model generating means for extracting a feature quantity from the signal data and generating the high-order signal pattern model based on the extracted feature quantity;
A high-order noise pattern model generating means for extracting a feature value from the noise data and generating the high-order noise pattern model based on the extracted feature value;
As an alternative to a data space composed of a plurality of higher-order signal pattern models, data composed of lower-order signal vectors of three dimensions or less in a state in which the positional relationship between the higher-order signal pattern models in the data space is approximated Low-order signal vector space generating means for generating the low-order signal vector space formed by mapping to space;
As an alternative to the data space composed of the plurality of high-order noise pattern models, data composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the high-order noise pattern models in the data space is approximated Low-order noise vector space generation means for generating the low-order noise vector space mapped to the space.

  With such a configuration, it is possible to extract a feature quantity from the signal data by a high-order signal pattern model generation unit, and to generate the high-order signal pattern model based on the extracted feature quantity. It is possible to extract a feature amount from the noise data by the next noise pattern model generation means and generate the higher order noise pattern model based on the extracted feature amount.

Further, as a substitute for the data space composed of the plurality of higher-order signal pattern models by the low-order signal vector space generating means, the three-dimensional state is obtained by approximating the positional relationship between the higher-order signal pattern models in the data space. It is possible to generate the low-order signal vector space mapped to the data space composed of the following low-order signal vectors, and from the plurality of high-order noise pattern models by the low-order noise vector space generation means As an alternative to the configured data space, the lower order obtained by mapping to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the higher-order noise pattern models in the data space is approximated It is possible to generate a noise vector space.
Accordingly, each high-order pattern model and each low-order vector space can be generated. For example, it is easy to generate each high-order pattern model for added data, and to change or update each data space. Can be done.

On the other hand, in order to achieve the above object, the pattern model evaluation apparatus according to claim 5 is:
A pattern model evaluation device for evaluating the recognition performance of a pattern model used for pattern recognition of recognition target data to be recognized,
Signal data storage means for storing a plurality of noise-unmixed pre-acquired signal data related to the plurality of objects;
As an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on the signal data, the positional relationship between the higher-order signal pattern models in the data space Low-order signal vector space storage means for storing a low-order signal vector space that is mapped to a data space composed of low-order signal vectors of three dimensions or less in an approximate state,
Noise data storage means for storing a plurality of types of signal-unmixed noise data;
As an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four-dimensional or higher-dimensional elements generated based on the noise data, the positional relationship between the higher-order noise pattern models in the data space Low-order noise vector space storage means for storing a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state where
Among the plurality of low-order signal vectors constituting the low-order signal vector space, the low-order signal vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order signal vectors or that are uniformly centered on the center of gravity. Evaluation signal vector selection means for selecting a low-order signal vector for evaluation from the inside;
Among the plurality of low-order noise vectors constituting the low-order noise vector space, the low-order noise vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order noise vectors or that are uniformly centered on the center of gravity. Evaluation noise vector selection means for selecting a low-order noise vector for evaluation from the inside;
Based on the signal data corresponding to the low-order signal vector selected by the evaluation signal vector selection means and the noise data corresponding to the low-order noise vector selected by the evaluation noise vector selection means, each signal data Mixed signal data generation means for evaluation for generating mixed signal data for evaluation mixed with data; and
Performance evaluation means for evaluating the performance of the pattern model based on the mixed signal data for evaluation generated by the mixed signal data generation means for evaluation and the pattern model;
And an evaluation result output means for outputting an evaluation result of the performance evaluation means.

  With such a configuration, it is possible to store a plurality of noise-unmixed signal data related to a plurality of objects acquired in advance by the signal data storage means, and the low-order signal vector space storage means As an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on signal data, the positional relationship between each higher-order signal pattern model in the data space is It is possible to store a low-order signal vector space that is mapped to a data space composed of low-order signal vectors of three dimensions or less in an approximate state, and a plurality of types of unmixed signals can be stored by the noise data storage means. It is possible to store noise data and generated based on the noise data by the low-order noise vector space storage means As an alternative to a data space composed of a plurality of higher-order noise pattern models composed of higher-dimensional elements than the three-dimensional elements, the positional relationship between the higher-order noise pattern models in the data space is approximated with three or less dimensions. It is possible to store a low-order noise vector space mapped to a data space composed of low-order noise vectors.

  Further, the evaluation signal vector selection means is located at a position more than a predetermined distance from the center of gravity of all the low-order signal vectors among the plurality of low-order signal vectors constituting the low-order signal vector space, or centered on the center of gravity. It is possible to select a low-order signal vector for evaluation from uniform low-order signal vectors, and a plurality of low-order noise vectors constituting the low-order noise vector space are selected by the evaluation noise vector selection means. Among them, it is possible to select a low-order noise vector for evaluation from low-order noise vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order noise vectors or that are uniformly centered on the center of gravity. is there.

  Further, the signal data corresponding to the low-order signal vector selected by the evaluation signal vector selection means by the evaluation mixed signal data generation means and the noise corresponding to the low-order noise vector selected by the evaluation noise vector selection means Based on the data, it is possible to generate mixed signal data for evaluation in which each noise data is mixed with each signal data, and the mixed signal for evaluation generated in the mixed signal data generation unit for evaluation generated by the performance evaluation unit The performance of the pattern model can be evaluated based on the data and the pattern model, and the evaluation result of the performance evaluation unit can be output by the evaluation result output unit.

  In the low-order signal vector space and the low-order noise vector space, the density of each vector tends to decrease as the distance from the center of gravity increases, and the density is low (outside and near the distribution). The noise mixed signal data generated from the signal data and noise data corresponding to the vector in FIG. 5 tends to generate a low recognition probability for a pattern model for an unspecified object generated using all data in each space, for example. is there. Accordingly, data at a position that is a predetermined distance or more away from the center of gravity is selected as evaluation data, noise mixed signal data is generated using the selected data, and the pattern model is evaluated using the noise mixed signal data. Thus, it becomes possible to accurately and systematically evaluate the recognition probability for data that causes a low recognition probability.

Furthermore, the invention according to claim 6 is the pattern model evaluation apparatus according to claim 5,
The low-order signal vector space and the low-order noise vector space are two-dimensional or three-dimensional data spaces,
A plurality of low-order signal vectors constituting the low-order signal vector space are centered on the centroid of all low-order signal vectors, and the distance between the centroid and the low-order signal vector at the position farthest from the centroid is the radius. The outer circle or outer sphere is divided into n outer circles or inner spheres (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer circle or outer sphere. A plurality of concentric circles consisting of circles and inner circles or outer spheres and inner spheres, or annular or spherical regions formed between curved lines or curved surfaces of concentric spheres are divided into a plurality by radially extending lines or surfaces. Low-order signal vector space partitioning means for partitioning;
A plurality of low-order noise vectors constituting the low-order noise vector space are centered on the centroid of the low-order noise vector, and the distance between the centroid and the low-order noise vector at a position farthest from the centroid is a radius. The outer circle or outer sphere is divided into n outer circles or inner spheres (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer circle or outer sphere. A plurality of concentric circles consisting of circles and inner circles or outer spheres and inner spheres, or annular or spherical regions formed between curved lines or curved surfaces of concentric spheres are divided into a plurality by radially extending lines or surfaces. Low-order noise vector space classification means for classifying;
Low-order signal vector space display means for displaying each low-order signal vector constituting the low-order signal vector space together with the classification result;
Low-order noise vector space display means for displaying each low-order noise vector constituting the low-order noise vector space together with the classification result, and
The evaluation signal vector selection means and the evaluation noise vector selection means are lines extending in the radial direction in the annular region or the spherical region of the displayed low-order signal vector space and low-order noise vector space. Alternatively, a low-order signal vector and a low-order noise vector for evaluation are selected from each region divided into a plurality of areas according to a plane,
The evaluation mixed signal data generation means includes each signal based on signal data and noise data respectively corresponding to a low-order signal vector and a low-order noise vector selected by the evaluation signal vector selection means and the evaluation noise vector selection means. It is characterized in that mixed signal data for evaluation in which each noise data is mixed with data is generated.

  With such a configuration, when the low-order signal vector space and the low-order noise vector space are two-dimensional data spaces, a plurality of low-order signal vector spaces are configured by the low-order signal vector space dividing means. A low-order signal vector with one outer circle centered on the center of gravity of all the low-order signal vectors and the radius between the center of gravity and the low-order signal vector located farthest from the center of gravity, and the center of gravity And an n-shaped inner circle (n is an integer of 1 or more) having a smaller radius than the outer circle, and an annular region formed between a plurality of concentric circles composed of the outer circle and the inner circle. It is possible to divide into a plurality by lines extending in the radial direction.

  Further, by the low-order noise vector space segmentation means, a plurality of low-order noise vectors constituting the low-order noise vector space are centered on the centroid of the low-order noise vector and located at a position farthest from the centroid. Partitioning into one outer circle whose radius is the distance from the low-order noise vector and n inner circles (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer circle; An annular region formed between the curved lines of a plurality of concentric circles composed of the outer circle and the inner circle can be divided into a plurality by a line extending in the radial direction.

Each low-order signal vector constituting the low-order signal vector space can be displayed together with the classification result by the low-order signal vector space display means, and the low-order noise vector space display means can display the low-order signal vector space. Each low-order noise vector constituting the next-order noise vector space can be displayed together with the classification result.
The evaluation signal vector selection means and the evaluation noise vector selection means are divided into a plurality of lines by the lines extending in the radial direction in the annular region of the displayed low-order signal vector space and the low-order noise vector space. It is possible to select an evaluation low-order signal vector and a low-order noise vector from each of the selected regions, and the evaluation mixed signal data generation means applies the selected low-order signal vector and low-order noise vector to the selected low-order signal vector and low-order noise vector. Based on the corresponding signal data and noise data, it is possible to generate mixed signal data for evaluation in which each signal data is mixed with each noise data.

  On the other hand, when the low-order signal vector space and the low-order noise vector space are three-dimensional data spaces, a plurality of low-order signal vectors constituting the low-order signal vector space are obtained by low-order signal vector space partitioning means. One outer sphere centered at the center of gravity of all the low-order signal vectors and having a radius as a distance between the center of gravity and the low-order signal vector at a position farthest from the center of gravity, and from the outer sphere centered on the center of gravity and Is divided into n inner spheres (n is an integer of 1 or more) having a small radius, and a spherical region formed between the curved surfaces of a plurality of concentric spheres composed of the outer sphere and the inner sphere, It is possible to divide into a plurality by the surface extending in the radial direction.

  Further, by the low-order noise vector space segmentation means, a plurality of low-order noise vectors constituting the low-order noise vector space are centered on the centroid of the low-order noise vector and located at a position farthest from the centroid. Partitioning into one outer sphere whose radius is a distance from a low-order noise vector and n inner spheres (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer sphere; A spherical region formed between the curved surfaces of a plurality of concentric spheres composed of the outer sphere and the inner sphere can be divided into a plurality by a surface extending in the radial direction.

Each low-order signal vector constituting the low-order signal vector space can be displayed together with the classification result by the low-order signal vector space display means, and the low-order noise vector space display means can display the low-order signal vector space. Each low-order noise vector constituting the next-order noise vector space can be displayed together with the classification result.
Further, the evaluation signal vector selection means and the evaluation noise vector selection means are divided into a plurality by the surface extending in the radial direction in the spherical region of the displayed low-order signal vector space and the low-order noise vector space. It is possible to select a low-order signal vector and a low-order noise vector for evaluation from each of the divided regions, and the mixed signal data generation means for evaluation includes the selected low-order signal vector and low-order noise vector. It is possible to generate mixed signal data for evaluation in which each signal data is mixed with each noise data based on the signal data and the noise data respectively corresponding to.

  Therefore, for example, by setting the radius of the inner circle (or inner sphere) to be greater than or equal to a predetermined distance from the center of gravity, the curve between the outer circle (or outer sphere) and the inner circle (or inner sphere) inside it is one. Each of the low-order vectors included in the annular (or spherical) region formed between the curved surfaces (or between the curved surfaces) is a low-order vector that is located at a predetermined distance or more from the center of gravity. Since this annular (or spherical) region is divided by a line (or surface) extending in the radial direction, the low-order vectors in the annular (or spherical) region are further finely divided. Furthermore, since the low-order vector for evaluation is selected from the low-order vector space thus divided, among the data that causes a low recognition probability for the pattern model for the unspecified object, for example, It is possible to select data that is not used to generate the pattern model as data for evaluation, etc., and generate noise mixed signal data from the evaluation data selected in this way. By evaluating the pattern model, it becomes possible to more accurately and systematically evaluate the recognition probability for noise mixed signal data that causes a low recognition probability.

On the other hand, in order to achieve the above object, a pattern recognition apparatus according to claim 7 is:
The pattern model generation device according to any one of claims 1 to 4,
Data acquisition means for acquiring recognition target data of the recognition target;
Pattern recognition means for performing pattern recognition of the acquired recognition target data based on the recognition target data acquired in the data acquisition means and the pattern model generated in the pattern model generation device;
Recognition result output means for outputting the recognition result of the pattern recognition means.

With such a configuration, it is possible to acquire recognition target data to be recognized by the data acquisition unit, and the recognition target data acquired by the data acquisition unit by the pattern recognition unit and the pattern model generation device It is possible to perform pattern recognition of the acquired recognition target data based on the pattern model generated in step 1, and the recognition result output means can output the recognition result of the pattern recognition means.
Therefore, since the pattern recognition of the recognition target data is performed by the pattern model that can improve the recognition probability for the signal data and noise data that reduce the recognition probability of the pattern model, the recognition probability is uniformly good for each recognition target data. It is possible to perform pattern recognition.

The invention according to claim 8 is the pattern recognition apparatus according to claim 8,
A pattern model evaluation apparatus according to claim 5 or 6 is provided.
With such a configuration, when new signal data or noise data is added, the performance of the pattern model can be evaluated. Based on this evaluation result, if the recognition performance decreases. The pattern model can be recreated, and the pattern recognition performance can be maintained and improved relatively easily.

On the other hand, in order to achieve the above object, a pattern model generation program according to claim 9 is provided:
A pattern model generation program for generating a pattern model used for pattern recognition of recognition target data to be recognized,
The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Among them, low-order signal vector selection means for selecting a low-order signal vector for pattern model generation from low-order signal vectors located at a predetermined distance or more away from the center of gravity of all low-order signal vectors,
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, Low-order noise vector selection means for selecting a low-order noise vector for pattern model generation from low-order noise vectors located at a predetermined distance or more away from the center of gravity of all low-order noise vectors;
Based on the signal data corresponding to the low-order signal vector selected by the low-order signal vector selection means and the noise data corresponding to the low-order noise vector selected by the low-order noise vector selection means, the noise is added to each signal data. Noise mixed signal data generating means for generating noise mixed signal data obtained by mixing data;
It includes a program for causing a computer to realize a function as a pattern model generation means for generating a pattern model obtained by learning a probability model using the noise mixed signal data as learning data.
With such a configuration, when the program is read by the computer and the computer executes processing according to the read program, the same operations and effects as those of the pattern model generation apparatus according to claim 1 can be obtained.

In order to achieve the above object, the pattern model evaluation program according to claim 10 comprises:
A pattern model evaluation program for evaluating the recognition performance of a pattern model used for pattern recognition of recognition target data to be recognized,
The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Evaluation signal vector for selecting a low-order signal vector for evaluation from among low-order signal vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order signal vectors, or that are uniformly centered on the center of gravity. A selection means;
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, Evaluation noise vector selection means for selecting a low-order noise vector for evaluation from low-order noise vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order noise vectors, or are uniformly centered on the center of gravity; ,
Based on the signal data corresponding to the low-order signal vector selected by the evaluation signal vector selection means and the noise data corresponding to the low-order noise vector selected by the evaluation noise vector selection means, each signal data Mixed signal data generation means for evaluation for generating mixed signal data for evaluation mixed with data; and
Performance evaluation means for evaluating the performance of the pattern model based on the mixed signal data for evaluation generated by the mixed signal data generation means for evaluation and the pattern model;
It includes a program for causing a computer to realize a function as an evaluation result output means for outputting an evaluation result of the performance evaluation means.
With such a configuration, when the program is read by the computer and the computer executes processing according to the read program, the same operation and effect as those of the pattern model evaluation apparatus according to claim 5 can be obtained.

In order to achieve the above object, a pattern recognition program according to claim 11 comprises:
The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Among them, low-order signal vector selection means for selecting a low-order signal vector for pattern model generation from low-order signal vectors located at a predetermined distance or more away from the center of gravity of all low-order signal vectors,
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, Low-order noise vector selection means for selecting a low-order noise vector for pattern model generation from low-order noise vectors located at a predetermined distance or more away from the center of gravity of all low-order noise vectors;
Based on the signal data corresponding to the low-order signal vector selected by the low-order signal vector selection means and the noise data corresponding to the low-order noise vector selected by the low-order noise vector selection means, the noise is added to each signal data. Noise mixed signal data generating means for generating noise mixed signal data obtained by mixing data;
Pattern model generation means for generating a pattern model obtained by learning a probability model using the noise mixed signal data as learning data;
Data acquisition means for acquiring recognition target data of the recognition target;
Pattern recognition means for performing pattern recognition of the acquired recognition target data based on the recognition target data acquired in the data acquisition means and the pattern model generated in the pattern model generation device;
It includes a program for causing a computer to realize a function as a recognition result output means for outputting a recognition result of the pattern recognition means.
With such a configuration, when the program is read by the computer and the computer executes processing according to the read program, the same operation and effect as those of the pattern recognition apparatus according to claim 7 can be obtained.

In order to achieve the above object, the pattern model generation method according to claim 12 comprises:
A pattern model generation method for generating a pattern model used for pattern recognition of recognition target data to be recognized,
The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Among them, a low-order signal vector selection step for selecting a low-order signal vector for pattern model generation from among low-order signal vectors located at a predetermined distance or more away from the center of gravity of all the low-order signal vectors;
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, A low-order noise vector selection step for selecting a low-order noise vector for generating a pattern model from low-order noise vectors located at a predetermined distance or more away from the center of gravity of all low-order noise vectors;
Based on the signal data corresponding to the low-order signal vector selected in the low-order signal vector selection step and the noise data corresponding to the low-order noise vector selected in the low-order noise vector selection step, the noise is added to each signal data. Noise mixed signal data generation step for generating noise mixed signal data obtained by mixing data;
And a pattern model generation step of generating a pattern model obtained by learning a probability model using the noise mixed signal data as learning data.
Thus, the same operation and effect as those of the pattern model generation device according to claim 1 can be obtained.

In order to achieve the above object, the pattern model evaluation method according to claim 13 comprises:
A pattern model evaluation method for evaluating the recognition performance of a pattern model used for pattern recognition of recognition target data to be recognized,
The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Evaluation signal vector for selecting a low-order signal vector for evaluation from among low-order signal vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order signal vectors, or that are uniformly centered on the center of gravity. A selection step;
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, An evaluation noise vector selection step for selecting a low-order noise vector for evaluation from low-order noise vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order noise vectors or that are uniformly centered on the center of gravity; ,
Based on the signal data corresponding to the low-order signal vector selected in the evaluation signal vector selection step and the noise data corresponding to the low-order noise vector selected in the evaluation noise vector selection step, each signal data An evaluation mixed signal data generation step for generating mixed signal data for evaluation mixed with noise data;
A performance evaluation step for evaluating the performance of the pattern model based on the mixed signal data for evaluation generated in the mixed signal data generation step for evaluation and the pattern model;
An evaluation result output step of outputting an evaluation result of the performance evaluation step.
Thus, the same operation and effect as those of the pattern model evaluation apparatus according to claim 5 can be obtained.

In order to achieve the above object, the pattern recognition method according to claim 14 comprises:
The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Among them, a low-order signal vector selection step for selecting a low-order signal vector for pattern model generation from among low-order signal vectors located at a predetermined distance or more away from the center of gravity of all the low-order signal vectors;
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, A low-order noise vector selection step for selecting a low-order noise vector for generating a pattern model from low-order noise vectors located at a predetermined distance or more away from the center of gravity of all low-order noise vectors;
Based on the signal data corresponding to the low-order signal vector selected in the low-order signal vector selection step and the noise data corresponding to the low-order noise vector selected in the low-order noise vector selection step, the noise is added to each signal data. Noise mixed signal data generation step for generating noise mixed signal data obtained by mixing data;
A pattern model generation step of generating a pattern model obtained by learning a probability model using the noise mixed signal data as learning data;
A data acquisition step of acquiring recognition target data of the recognition target;
A pattern recognition step for performing pattern recognition of the acquired recognition target data based on the recognition target data acquired in the data acquisition step and the pattern model generated in the pattern model generation step;
A recognition result output step of outputting a recognition result of the pattern recognition step.
Thus, the same operation and effect as the pattern recognition device according to claim 7 can be obtained.

As described above, according to the pattern model generation device according to claim 1 or 2 of the present invention, it is possible to generate a pattern model that can improve the recognition probability for signal data that reduces the recognition probability of the pattern model. The effect of being able to be obtained.
According to the pattern model generation device of claim 3, in addition to the effect of claim 1 or claim 2, the distribution of each pattern model can be easily grasped visually, and the method of selecting a region For example, by selecting data according to the content of pattern recognition and the recognition target, a pattern model specialized for a specific recognition target or a pattern model with higher recognition performance may be generated. The effect that it can be obtained.

According to the pattern model generation device of claim 4, in addition to the effect of any one of claims 1 to 3, each higher-order pattern model generation process for added data, and each data space The effect that the change process, the update process, and the like can be easily performed is obtained.
Moreover, according to the pattern model evaluation apparatus of Claim 5 or Claim 6, the effect that the recognition probability with respect to the data which produces a low recognition probability can be evaluated correctly and systematically is acquired.

Moreover, according to the pattern recognition apparatus of Claim 7 or Claim 8, the effect that the pattern recognition with which the recognition probability becomes uniform uniformly with respect to each signal data can be performed is acquired.
According to the pattern model generation program of the ninth aspect, it is possible to generate a pattern model that can improve the recognition probability with respect to signal data that reduces the recognition probability of the pattern model.

In addition, according to the pattern model evaluation program of the tenth aspect, it is possible to accurately and systematically evaluate the recognition probability for data that causes a low recognition probability.
Further, according to the pattern recognition program of the eleventh aspect, it is possible to obtain an effect that it is possible to perform pattern recognition with a uniform recognition probability for each signal data.

According to the pattern model generation method of the twelfth aspect of the present invention, it is possible to generate a pattern model that can improve the recognition probability for signal data that reduces the recognition probability of the pattern model.
According to the pattern model evaluation method of the thirteenth aspect, it is possible to accurately and systematically evaluate the recognition probability for data that causes a low recognition probability.
In addition, according to the pattern recognition method of the fourteenth aspect, an effect is obtained that it is possible to perform pattern recognition with a uniform recognition probability for each signal data.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. 1 to 11 show a pattern model generation apparatus, a pattern model generation program and a pattern model generation method, a pattern model evaluation apparatus, a pattern model evaluation program and a pattern model evaluation method, a pattern recognition apparatus, a pattern recognition program, and It is a figure which shows embodiment of the pattern recognition method.

First, the functional configuration of the pattern recognition apparatus according to the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing a functional configuration of a pattern recognition apparatus according to the present invention.
The pattern recognition apparatus 100 includes a data storage unit 10 that stores signal data and multiple types of noise data related to a plurality of objects, a low-order vector space storage unit 12 that stores a low-order signal vector space and a low-order noise vector space, A high-order pattern model generation unit 14 that generates a high-order signal pattern model based on the signal data and generates a high-order noise pattern model based on the noise data, and a high-order signal pattern model and a high-order noise pattern model that are stored. A pattern model storage unit 16; and a low-order vector space generation unit 18 that generates a low-order vector space by mapping a data space constituted by a high-order pattern model to a data space constituted by low-order vectors. It has a configuration.

  The data storage unit 10 includes signal data related to a plurality of objects such as voice signal data uttered by a plurality of speakers, sensor output signal data output from sensors such as a plurality of infrared sensors, and a plurality of wildlife cry signal data. And noise data such as household noise, factory noise, and traffic noise, noise data such as electric / electronic circuit noise, vibration noise, and the like are stored in a predetermined area of the storage device 70 to be described later. Furthermore, in this embodiment, there is a function of storing signal data related to a plurality of objects in groups based on a plurality of specific conditions. For example, a large number of unspecified audio data obtained from multiple speakers can be converted into “speaker type” such as speaker name, male / female gender, child / adult / elder age, etc. Stored in groups based on specific conditions such as “utterance vocabulary” of sentences, words, etc., “speech style” such as utterance speed, speech volume, and dialect features. Further, each signal data is clean data with no noise mixing, and each noise data is clean noise data with no signal other than noise components.

  The low-order vector space storage unit 12 stores a low-order signal vector space data and low-order noise vector space data, which will be described later, sent from an external device via a network or the like, or sent via a storage medium. 70, and the low-order signal vector space data and the low-order noise vector space data generated by the low-order vector space generation unit 18 are stored in a predetermined region of the storage device 70 described later. is doing. Here, for low-order vector space data sent from an external device or the like, signal data and noise data related to a plurality of objects used at the time of generation are also acquired.

  The high-order pattern model generation unit 14 first selects signal data or noise data for each group stored in a predetermined area of the storage device 70 described later by the data storage unit 10 according to a selection instruction from the user. The extracted signal data or noise data has a function of extracting a four-dimensional or higher-dimensional feature quantity (also referred to as a feature parameter) by analysis processing such as cepstrum analysis or linear prediction analysis. Further, HMM learning is performed using the extracted high-dimensional feature value as learning data using a known EM algorithm or the like, and a pattern model including high-dimensional elements composed of the HMM after the learning is generated. It has a function to do. Here, for example, if the pattern model is signal data related to multiple objects, it is generated for each individual or individual constituting each group, and if it is noise data, it is generated for each type of noise constituting each group. Is done.

  The high-order pattern model storage unit 16 includes a high-order signal pattern model and a high-order noise pattern model that are sent from an external device via a network or the like or a storage medium. The high-order signal pattern model and high-order noise pattern model generated by the high-order pattern model generation unit 14 are stored in a predetermined area of the storage device 70 to be described later. Here, for high-order pattern models sent from an external device or the like, signal data and noise data related to a plurality of objects used at the time of generation are also acquired.

  The low-order vector space generation unit 18 uses a well-known Sammon method as an alternative to a data space composed of a plurality of high-order signal pattern models, in an approximate state of the distance relationship between the pattern models in the data space. It has a function of generating a low-order signal vector space by mapping to a data space composed of signal vectors composed of low-dimensional elements of three dimensions or less. Further, similarly, using the well-known Sammon method, as an alternative to the data space constituted by a plurality of higher-order noise pattern models, in a state where the distance relation between each pattern model in the data space is approximated, the three-dimensional or less It has a function to generate a low-order noise vector space by mapping to a data space composed of signal vectors composed of low-dimensional elements.

  Here, the Sammon method is the difference between the sum of mutual distances of vector information (higher order pattern model) in a high-dimensional space and the sum of mutual Euclidean distances of mapping position coordinates (lower-order vectors) in a low-dimensional space. This is a technique for optimizing the mapping position coordinates in the low-dimensional space by the steepest descent method so as to be minimized. That is, the high-order signal pattern model and the high-order noise pattern model are converted into, for example, a three-dimensional or two-dimensional low-order vector in a state in which the distance relation between the pattern models is approximated, and coordinates in a low-dimensional space. It will be mapped to a point. In the present embodiment, the generated low-order signal vector space and low-order noise vector space can be displayed by the information display unit 36 described later.

  The pattern recognition apparatus 100 further includes a low-order vector space partitioning unit 20 that partitions a low-order vector space composed of a plurality of low-order vectors into a plurality of regions, and a pattern model generation from the divided low-order vector spaces. A learning data selection unit 22 that acquires data corresponding to the selected low-order vector, and noise mixed signal data based on the signal data and noise data acquired by the learning data selection unit 22 A noise mixture signal data generation unit 26 that generates a pattern model generation unit 28 that generates a pattern model for pattern recognition based on the generated noise mixture signal data, and a pattern model storage unit that stores the generated pattern model 28.

  The low-order vector space partitioning unit 20 has a two-dimensional (or three-dimensional) low-order signal vector space composed of a plurality of low-order signal vectors, centered on the center of gravity of all the low-order signal vectors, and the center and the center thereof. An outer circle (or outer sphere) whose radius is the distance between the lowest-order vectors at the farthest positions, and n inner circles having a radius shorter than that of the outer circle (or outer sphere) centered on the center of gravity. It is divided by a circle (or inner sphere), and between each curve (or between curved surfaces) of a concentric circle (or concentric sphere) composed of the outer circle (or outer sphere) and n inner circles (or inner spheres). It has a function of dividing the formed annular (or spherical) region into a plurality of regions by lines (or surfaces) extending in the radial direction. Further, similarly to the low-order signal vector space, it has a function of dividing a low-order noise vector space composed of a plurality of low-order noise vectors into a plurality of regions.

  The learning data selection unit 22 assumes that the low-order signal vector space and the low-order noise vector space are two-dimensional (or three-dimensional) when there is a selection instruction from the user via the input device 74 described later. An area corresponding to a selection instruction from an annular (or spherical) area formed between curves (or between curved surfaces) of an outer circle (or outer sphere) and an inner circle (or inner sphere) inside the outer circle (or outer sphere). And the data corresponding to the low-order signal vector or low-order noise vector belonging to the selected region is acquired from the stored contents of the data storage unit 10. In addition, when there is no selection instruction, an outer circle (or outer sphere) and an inner circle (or inner sphere) on the inner side of the outer circle (or outer sphere) according to a preset selection method (for example, selection of all areas) A data storage unit selects a predetermined region in an annular (or spherical) region formed between curves (or between curved surfaces), and stores data corresponding to a low-order signal vector or a low-order noise vector belonging to the selected region 10 to obtain the function. Note that the low-order vector itself may be directly selected instead of the region.

Here, the description in the parentheses is a description showing a shape when the low-order vector space is a three-dimensional space. The same applies to the following description of the low-order vector space.
The noise mixed signal data generation unit 24 generates noise mixed signal data by mixing noise data with a predetermined SNR with respect to each signal data based on the signal data and noise data acquired by the learning data selection unit 22. It has a function to do.

The pattern model generation unit 26 uses the noise mixed signal data generated by the noise mixed signal data generation unit 24 as learning data, and learns an HMM in an initial state or a learned HMM to generate a pattern model for pattern recognition. have. In the present embodiment, as a learning method, the same method as that for the high-order signal pattern model and the high-order noise pattern model is used.
The pattern model storage unit 28 stores the pattern model generated by the pattern model generation unit 26 in a predetermined area of a storage device 70 to be described later, or is sent from an external device via a network or the like, or via a storage medium It has a function of storing the transmitted pattern model in a predetermined area of the storage device 70 to be described later.

  The pattern recognition apparatus 100 further performs evaluation from a low-order vector space divided into a plurality of regions in the low-order vector space partitioning unit 20 corresponding to a pattern model specified by the user via the input device 74 described later. An evaluation data selection unit 30 that selects a region and acquires data corresponding to a low-order vector belonging to the selected region from the data storage unit 10, and an evaluation that generates mixed signal data for evaluation from the acquired evaluation data Pattern signal evaluation unit 34 for evaluating the recognition performance of the pattern model based on the generated mixed signal data for evaluation 32, the generated mixed signal data for evaluation, and the pattern model; A configuration including a low-order vector space and an information display unit 36 for displaying a recognition result of a pattern recognition processing unit 40 to be described later You have me.

  The evaluation data selection unit 30 selects a low-order signal vector and a low-order noise vector for evaluation from the low-order signal vector space after segmentation and the low-order noise vector space after segmentation corresponding to the pattern model specified by the user It has a function to do. In this selection process, when there is a selection instruction from the user, a low-order vector belonging to the region according to the selection instruction is selected, and when there is no selection instruction, a low-order vector for evaluation according to a preset selection method is selected. A signal vector and a low-order noise vector are selected. Further, when a low-order signal vector and a low-order noise vector for evaluation are selected, signal data and noise data corresponding to the selected low-order signal vector and low-order noise vector are stored from the stored contents of the data storage unit 10. It has a function to acquire.

The evaluation mixed signal data generation unit 32 mixes noise data with a predetermined SNR to each signal data based on the evaluation signal data and noise data acquired by the evaluation data selection unit 30, and the evaluation mixed signal data. It has the function to generate.
The pattern model evaluation unit 34 extracts a feature amount from each evaluation mixed signal data generated in the evaluation mixed signal data generation unit 32, and based on the extracted feature amount and the pattern model specified by the user, Performs pattern recognition processing of the mixed signal data for evaluation. For example, a feature amount is extracted by performing cepstrum analysis, MFCC analysis, etc. for each frame (for example, frame length 20 ms, frame shift 10 ms) for the mixed signal data for evaluation, and the extracted feature amount is designated. The likelihood for each pattern model is calculated based on the pattern model, and pattern recognition is performed based on the calculated likelihood. For pattern recognition, for example, a label corresponding to the pattern model of maximum likelihood is selected as a recognition result. That is, if the signal data is speech data, the vocabulary corresponding to the pattern model with the highest likelihood is selected as the recognition result from the vocabulary corresponding to each pattern model for speech recognition. Further, the pattern model evaluation unit 34 has a function of evaluating the recognition performance of the pattern model designated by the user based on the recognition result of the pattern recognition processing.

  The information display unit 36 displays the low-order vector space after the division in the low-order vector space division unit 20 as a coordinate point on the two-dimensional space if it is two-dimensional, and as a coordinate point on the three-dimensional space if it is three-dimensional. It has a function of displaying together with a dividing line for dividing the area. Furthermore, it also has a function of displaying the evaluation result of the pattern model evaluation unit 34 and displaying the recognition result of the pattern recognition processing unit 40. Here, as the evaluation result, the recognition probability for each mixed signal data for evaluation is displayed, or a superior or inferior comment for the recognition probability is displayed. As the recognition result, for example, in the case of speech recognition using a speech pattern model, the vocabulary of the recognition result is displayed.

The pattern recognition apparatus 100 further includes a data acquisition unit 38 that acquires signal data to be recognized, and a pattern recognition processing unit 40 that recognizes the acquired signal data.
The data acquisition unit 38 has a function of acquiring signal data to be recognized. For example, when acquiring audio data, the analog data input from the microphone is input via a microphone and an A / D converter. Audio data converted into digital data by an A / D converter is acquired. Further, for example, when data to be recognized is acquired via the Internet or the like, the data is acquired via a network card or the like. For example, when acquiring signal data of a sensor detection result of an infrared sensor or the like to be recognized, a sensor output signal converted into digital data by an A / D converter is acquired.

  The pattern recognition processing unit 40 performs cepstrum analysis or MFCC for each frame (for example, a frame length of 20 ms and a frame shift of 10 ms) on the signal data acquired by the data acquisition unit 38 in the same manner as the pattern model evaluation unit 34. Analyzes and the like extract feature quantities, calculate likelihoods for each pattern model based on the extracted feature quantities and the pattern model corresponding to the signal data, and perform pattern recognition based on the calculated likelihoods It has a function. This recognition result is displayed by the information display unit 36.

  Furthermore, the pattern recognition apparatus 100 is provided with a computer system for realizing the control of the above-described units on software, and the hardware configuration thereof is a central processing unit responsible for various controls and arithmetic processing as shown in FIG. A PCI (Central Processing Unit) 60 that is a device, a RAM (Random Access Memory) 62 that constitutes a main storage device (Main Storage), and a ROM (Read Only Memory) 64 that is a read-only storage device. (Peripheral Component Interconnect) bus and ISA (Industrial Standard Architecture) bus etc. are connected by various internal and external buses 68, and this bus 68 is entered and exited. Via a force interface (I / F) 66, an external storage device (Secondary Storage) 70 such as an HDD (Hard Disk Drive), an output device 72 such as a printing means, a CRT, an LCD monitor, an operation panel, a mouse, a keyboard, An input device 74 such as a scanner and a network L for communicating with an external device (not shown) are connected.

  When the power is turned on, a system program such as BIOS stored in the ROM 64 or the like is stored in various dedicated computer programs stored in the ROM 64 in advance, or in a CD-ROM, DVD-ROM, flexible disk (FD), or the like. Various dedicated computer programs installed in the storage device 70 via the medium or the communication network L such as the Internet are similarly loaded into the RAM 62, and the CPU 60 performs various operations according to instructions described in the program loaded in the RAM 62. Each function of each means as described above can be realized on software by performing predetermined control and arithmetic processing by making full use of resources.

  Next, an example of a flow of pattern pattern generation processing for pattern recognition using the pattern recognition apparatus 100 configured as described above will be described with reference to the flowchart of FIG. Here, FIG. 3 is a flowchart showing an example of a pattern model generation process for pattern recognition. As shown in the flowchart of FIG. 3, the pattern recognition apparatus 100 first proceeds to step S <b> 100, and determines whether or not the low-order vector space partitioning unit 20 has received a pattern model generation instruction from the user via the input device 74. If it is determined and it is determined that there is a generation instruction (Yes), the process proceeds to step S102. If not (No), the determination process is repeated until there is a generation instruction.

When the process proceeds to step S102, the low-order vector space division unit 20 reads low-order signal vector space data corresponding to the generation instruction from the user from a predetermined area of the storage device 70, and the read data is stored in the RAM 62 as a predetermined data. By storing in the area, low-order signal vector space data is acquired, and the process proceeds to step S104.
In step S <b> 104, the low-order vector space division unit 20 reads low-order noise vector space data corresponding to the generation instruction from the user from a predetermined area of the storage device 70, and stores the read data in the predetermined area of the RAM 62. Thus, the low-order noise vector space data is acquired, and the process proceeds to step S106.

  In step S106, the low-order vector space partitioning unit 20 partitions the low-order signal vectors constituting the space into a plurality of regions based on the low-order signal vector space data acquired in step S102 and acquired in step S104. Based on the low-order noise vector space data, the low-order noise vector constituting the space is divided into a plurality of regions, and the process proceeds to step S108.

  In step S108, the low-order vector space dividing unit 20 determines whether or not the manual selection mode is selected. If it is determined that the manual selection mode is selected (Yes), the information display unit 36 is divided. The subsequent low-order vector space data is provided and the process proceeds to step S110. If not (No), the process proceeds to step S122. Here, the manual selection mode is a mode in which the user can select an arbitrary region via the input device 74 from specific regions in the low-order signal vector space and the low-order noise vector space after the division. Further, if the low-order signal vector space and the low-order noise vector space are two-dimensional (or three-dimensional), the specific region is the outer circle (or outer sphere) and the inner one of the inner circle as described above. An annular (or spherical) area formed between curves (or curved surfaces) with a circle (or inner sphere). This area is divided into multiple areas by radially extending lines (or faces). It is divided. There is an automatic selection mode corresponding to this manual selection mode, and when this mode is set, a predetermined area is automatically selected from the specific area according to a preset selection method.

When the process proceeds to step S110, in the information display unit 36, the low-order vector space and the low-order noise vector space based on the data of the low-order signal vector space and the low-order noise vector space are used to display each low-order vector and the division result as CRT or The display is output to the output device 72 constituted by an LCD monitor or the like, and the process proceeds to step S112.
In step S112, the learning data selection unit 22 determines whether or not a predetermined region has been selected by the user from the above-described specific regions of the divided low-order signal vector space and low-order noise vector space. If it is determined (Yes), the process proceeds to step S114. If not (No), the determination process is repeated until it is selected.

  When the process proceeds to step S114, the learning data selection unit 22 causes the signal data and noise data corresponding to the low-order signal vector and the low-order noise vector belonging to the area selected in step S112 to be stored in a predetermined area of the storage device 70. , And the read data is stored in a predetermined area of the RAM 62 to acquire the signal data and noise data, and the process proceeds to step S116.

  In step S116, the noise mixed signal data generation unit 24 generates noise mixed signal data by mixing noise with each signal data at a preset SNR based on the signal data and noise data acquired in step S114. The process proceeds to S118. Here, in the present embodiment, the SNR is configured such that the user can set an arbitrary value or select from several candidates.

In step S118, the pattern model generation unit 26 analyzes the noise mixed signal data generated in step S116, extracts high-dimensional (eg, 10 to 40-dimensional) feature quantities of four or more dimensions, and extracts the extracted high A pattern model for pattern recognition is generated by learning the HMM using the dimension feature quantity, and the process proceeds to step S120.
In step S120, the pattern model storage unit 28 stores the pattern model generated in step S118 in association with each recognition target in a predetermined area of the storage device 70, and the process proceeds to step S100.

  On the other hand, if the automatic selection mode is set instead of the manual selection mode in step S108, and the process proceeds to step S122, the learning data selection unit 22 determines whether a predetermined area is selected from the specific area according to a preset selection method. In addition to selecting an area, signal data and noise data corresponding to the low-order signal vector and the low-order noise vector belonging to the selected area are read from the predetermined area of the storage device 70, and the read data is stored in the predetermined area of the RAM 62. By storing the signal data and the noise data, the process proceeds to step S116.

Furthermore, an example of the flow of low-order vector space data generation processing in the pattern recognition apparatus 100 will be described based on the flowchart of FIG. Here, FIG. 4 is a flowchart showing an example of a low-order vector space data generation process.
As shown in the flowchart of FIG. 4, the pattern recognition apparatus 100 first proceeds to step S <b> 200, and determines whether there is an instruction to generate low-order vector space data in the high-order pattern model generation unit 14. When it determines (Yes), it transfers to step S202, and when that is not right (No), a determination process is repeated until there exists an instruction | indication.

When the process proceeds to step S202, the high-order pattern model generation unit 14 determines whether the low-order vector space data generation instruction is for the signal data, and if it is determined that the instruction is for the signal data (Yes) moves to step S204, otherwise (No) moves to step S216.
When the process proceeds to step S204, the high-order pattern model generation unit 14 reads the signal data corresponding to the generation instruction from the predetermined area of the storage device 70, and stores the read signal data in the predetermined area of the RAM 62. Thus, the signal data is acquired and the process proceeds to step S206.

In step S206, the high-order pattern model generation unit 14 analyzes the signal data acquired in step S204, extracts a high-dimensional feature quantity of four or more dimensions, and proceeds to step S208.
In step S208, the high-order pattern model generation unit 14 learns the HMM using the high-dimensional feature value extracted in step S206 to generate a high-order signal pattern model, and the process proceeds to step S210.

In step S210, the high-order pattern model storage unit 16 stores the high-order signal pattern model generated in step S208 in a predetermined area of the storage device 70, and the process proceeds to step S212.
In step S212, the low-order vector space generation unit 18 changes the data space composed of the high-order signal pattern model stored in step S210 to a data space composed of signal vectors of three dimensions or less using the Sammon method. To generate low-order signal vector space data, and the process proceeds to step S214.

In step S214, the low-order vector space storage unit 12 stores the low-order signal vector space data generated in step S212 in a predetermined area of the storage device 70, and the process proceeds to step S200.
On the other hand, when the process proceeds to step S216 instead of signal data in step S202, the high-order pattern model generation unit 14 reads out noise data corresponding to the generation instruction from a predetermined area of the storage device 70, and the read noise data. Is stored in a predetermined area of the RAM 62 to acquire the signal data, and the process proceeds to step S218.

In step S218, the high-order pattern model generation unit 14 analyzes the noise data acquired in step S216, extracts high-dimensional feature quantities of four or more dimensions, and proceeds to step S220.
In step S220, the high-order pattern model generation unit 14 learns the HMM using the high-dimensional feature value extracted in step S218, generates a high-order noise pattern model, and proceeds to step S222.
In step S222, the high-order pattern model storage unit 16 stores the high-order noise pattern model generated in step S220 in a predetermined area of the storage device 70, and the process proceeds to step S224.

In step S224, in the low-order vector space generation unit 18, the data space composed of the high-order noise pattern model stored in step S222 is composed of low-dimensional noise vectors of three dimensions or less using the Sammon method. By mapping to the data space, low-order noise vector space data is generated and the process proceeds to step S226.
In step S226, the low-order vector space storage unit 12 stores the low-order noise vector space data generated in step S224 in a predetermined area of the storage device 70, and the process proceeds to step S200.

Further, an example of the flow of the pattern recognition process for pattern recognition in the pattern recognition apparatus 100 will be described based on the flowchart of FIG. Here, FIG. 5 is a flowchart showing an example of pattern pattern evaluation processing for pattern recognition.
As shown in the flowchart of FIG. 5, the pattern recognition apparatus 100 first proceeds to step S <b> 300, and in the low-order vector space partition unit 20, determines whether or not there has been a pattern model evaluation instruction from the user via the input device 74. If it is determined and it is determined that there is an evaluation instruction (Yes), the process proceeds to step S302. If not (No), the determination process is repeated until there is an evaluation instruction.

When the process proceeds to step S302, the low-order vector space partition unit 20 reads low-order signal vector space data corresponding to the pattern model specified in the evaluation instruction from a predetermined area of the storage device 70, and the read data Is stored in a predetermined area of the RAM 62, low-order signal vector space data is acquired, and the process proceeds to step S304.
In step S304, the low-order vector space partitioning unit 20 reads low-order noise vector space data corresponding to the pattern model specified in the evaluation instruction from the predetermined area of the storage device 70, and the read data is stored in the predetermined area of the RAM 62. , The low-order noise vector space data is acquired, and the process proceeds to step S306.

  In step S306, the low-order vector space partitioning unit 20 partitions the low-order signal vectors constituting the space into a plurality of regions based on the low-order signal vector space data acquired in step S302 and acquired in step S304. Based on the low-order noise vector space data, the low-order noise vector constituting the space is divided into a plurality of regions, and the process proceeds to step S308.

In step S308, the low-order vector space segmentation unit 20 determines whether or not the manual selection mode is selected. If it is determined that the manual selection mode is selected (Yes), the information display unit 36 is segmented. The subsequent low-order vector space data is given, and the process proceeds to step S310. Otherwise (No), the process proceeds to step S322.
When the process proceeds to step S322, in the information display unit 36, the low-order vector space and the low-order noise vector space based on each data of the low-order signal vector space and the low-order noise vector space are displayed on the CRT or The display is output to the output device 72 constituted by an LCD monitor or the like, and the process proceeds to step S312.

In step S312, the evaluation data selection unit 30 determines whether or not a predetermined region has been selected by the user from the above-described specific regions of the low-order signal vector space and the low-order noise vector space after classification. If it is determined (Yes), the process proceeds to step S314. If not (No), the determination process is repeated until it is selected.
When the process proceeds to step S314, the evaluation data selection unit 30 stores the signal data and noise data corresponding to the low-order signal vector and the low-order noise vector belonging to the area selected in step S312 in the predetermined area of the storage device 70. , And the read data is stored in a predetermined area of the RAM 62 to acquire the signal data and noise data, and the process proceeds to step S116.

  In step S116, in the evaluation mixed signal data generation unit 32, based on the signal data and noise data acquired in step S114, noise is mixed in each signal data with a preset SNR to generate evaluation mixed signal data. Then, the process proceeds to step S318. Here, in the present embodiment, the SNR is configured such that the user can set an arbitrary value or select from several candidates.

In step S318, the pattern model evaluation unit 34 analyzes the evaluation mixed signal data generated in step S116 to extract high-dimensional feature values, and extracts the extracted feature values and the pattern model specified in the evaluation instruction. And pattern recognition is performed based on the recognition result, and the process proceeds to step S320.
In step S320, the information display unit 36 displays the evaluation result of step S318 on the monitor of the output device 72, and the process proceeds to step S300.

  On the other hand, if the automatic selection mode is set instead of the manual selection mode in step S308, and the process proceeds to step S322, the evaluation data selection unit 30 performs predetermined processing from the specific area described above according to a preset selection method. While selecting a low-order vector, reading out the signal data and noise data corresponding to the selected low-order vector from the storage device 70, and storing the read data in a predetermined area of the RAM 62, the signal data and the noise data And the process proceeds to step S316.

Furthermore, an example of the flow of pattern recognition processing in the pattern recognition apparatus 100 will be described based on the flowchart of FIG. Here, FIG. 6 is a flowchart showing an example of pattern recognition processing.
As shown in the flowchart of FIG. 6, the pattern recognition apparatus 100 first proceeds to step S <b> 400, and the data acquisition unit 38 determines whether or not there is a pattern recognition instruction from the user via the input device 74. If it is determined that there is (Yes), the process proceeds to step S402. If not (No), the determination process is repeated until there is a recognition instruction.

When the process proceeds to step S402, the signal data acquisition unit 38 determines whether or not the signal data to be recognized has been acquired. If it is determined that the signal data has been acquired (Yes), the A / D converter converts the signal data into digital data. After conversion, the process proceeds to step S404. If not (No), the determination process is repeated until acquisition or an instruction to end the recognition process is given.
When the process proceeds to step S404, the pattern recognition processing unit 40 analyzes the signal data acquired in step S402 to extract high-dimensional feature values, and the process proceeds to step S406.

In step S406, the pattern recognition processing unit 40 performs pattern recognition processing based on the high-dimensional feature amount extracted in step S404 and the pattern model specified in the recognition instruction, and the information display unit displays the recognition result information and the recognition. The result display instruction is given and the process proceeds to step S408. In step S408, the information display unit 36 includes a CRT or LCD monitor constituting the output device 72 in accordance with the information on the recognition result from the pattern recognition processing unit 40 and the display instruction. The recognition result is displayed on the screen, and the process proceeds to step S410.
In step S410, the data acquisition unit 38 determines whether or not there has been an instruction to end the recognition process from the user via the input device 74. If it is determined that there is (Yes), the process proceeds to step S400. In the case (No), the process proceeds to step S402.

Next, the operation of the present embodiment will be described with reference to FIGS.
Here, FIG. 7 is a diagram illustrating an example of a two-dimensional low-order signal vector space for audio signal data. FIG. 8 is a diagram illustrating an example of the low-order signal vector space of FIG. 7 after region segmentation. FIG. 9 is a diagram illustrating an example of a two-dimensional low-order noise vector space for noise data. FIG. 10 is a diagram illustrating an example of the low-order noise vector space of FIG. 9 after region segmentation. Moreover, FIG. 11 is a figure which shows the evaluation result of the pattern model produced | generated based on real data using the method of this invention.

Hereinafter, it is assumed that the low-order signal vector space for the speech signal data and the low-order noise vector space for the noise data have already been generated. The operation will be described.
When the pattern recognition device 100 receives a pattern model generation instruction for the audio signal data from the user via the input device 74 (“Yes” in step S100), the pattern recognition device 100 outputs the audio data corresponding to the generation instruction. The low-order signal vector space data is read from the storage device 70, the read data is stored in a predetermined area of the RAM 62 (step S102), and the low-order noise vector space data of the noise data for the audio data is also read from the storage device 70. The read data is stored in a predetermined area of the RAM 62 (step S104). In the present embodiment, the low-order signal vector space of the audio data is a space composed of two-dimensional low-order signal vectors. When each low-order signal vector is visualized and displayed as coordinate points, for example, As shown in FIG. 7, each low-order signal vector is planarly displayed as a coordinate point in a two-dimensional space. Here, the low-order signal vector space shown in FIG. 7 is created as voice data using voice data in which 145 men actually uttered 175 Japanese words in a plurality of utterance styles. A high-order signal pattern model for 561 specific speakers is generated from the data, and a data space composed of the 561 high-order signal pattern models is mapped to a two-dimensional low-order signal vector space. It is.

  In the low-order vector space segmentation unit 20, the two-dimensional low-order signal vector space data is centered on the center of gravity of all the low-order signal vectors constituting the low-order signal vector space, for example, as shown in FIG. The outer circle whose radius is the distance from the center to the lowest order signal vector located farthest from the center is divided into one inner circle having a radius shorter than the radius of the outer circle, and An annular region formed between the curve with the inner circle is equally divided into eight regions by a line extending in the radial direction (step S106). FIG. 8 shows an example in which a low-order signal vector space created from a high-order signal pattern model for 100 specific speakers created from speech data in which 100 men uttered 1000 city names in Japan is segmented. It is.

  On the other hand, in the present embodiment, the low-order noise vector space of the noise data is a space composed of a two-dimensional low-order noise vector, like the low-order signal vector space of the voice data, and each low-order noise vector When this is visualized as a coordinate point and displayed, each low-order noise vector is planarly displayed as a coordinate point in a two-dimensional space, as shown in FIG. Here, the low-order noise vector space shown in FIG. 9 actually generates a high-order noise pattern model from 2000 kinds of noise data including the sound of hot water in a bath tub and the sound of hitting a desk with a pen. This is obtained by mapping a data space constituted by the generated higher-order noise pattern model to a two-dimensional lower-order noise vector space.

  In the low-order vector space segmentation unit 20, as shown in FIG. 10, the center of gravity of all the low-order noise vectors constituting the two-dimensional low-order noise vector space is the center, and the position farthest from this center. Is divided into an outer circle whose radius is the distance to the low-order noise vector in FIG. 1 and one inner circle having a radius shorter than the radius of the outer circle, and is further formed between the curves of the outer circle and the inner circle. The annular region is equally divided into eight regions by a line extending in the radial direction (step S106).

In both the low-order signal vector space and the low-order noise vector space, the distance from the center of gravity of the low-order vector that is farthest from the center of gravity is 1, and the radius of the inner circle is the distance from the center of gravity. The position was such that the ratio was 0.6.
Here, a case will be described in which the audio signal data and noise data for pattern model generation are acquired according to a preset selection method, assuming that the automatic selection mode is set.

  In this embodiment, a speaker corresponding to one lower-order signal vector for each man and woman located near the center of each region in the eight regions obtained by dividing the annular region is selected as a speaker for evaluation. It is set to be. The generation voice signal data is set to select voice signal data for speakers corresponding to all the low-order signal vectors belonging to the eight regions except for the 16 male and female evaluation speakers. . That is, a low-order signal vector is selected from an annular region in which the ratio of the distance from the center of gravity is in the range of 0.6 to 1.0. The reason for selecting the low-order signal vector from the annular region in this way is that, in speech recognition of an unspecified speaker, a pattern model generally corresponding to a low-order vector located far from the center of gravity has low recognition performance. This is because a pattern model corresponding to a low-order vector located near the center of gravity tends to have a higher recognition performance, so that a pattern model that is relatively far from the center of gravity and has a relatively low recognition performance is selected.

  On the other hand, in the present embodiment, as shown in FIG. 10, eight regions obtained by partitioning the annular region of the low-order noise vector space are divided into “TR_1, TR_2, It is divided into four regions of “TR_3, TR_4” and four regions of “EV_1_2, EV_2_3, EV_3_4, EV_4_1” which are evaluation regions. The noise data for pattern model generation is set to select noise data corresponding to all low-order noise vectors belonging to each zone of “TR_1, TR_2, TR_3, TR_4”. That is, similarly to the low-order signal vector, the low-order noise vector is selected from an annular region in which the ratio of the distance from the center of gravity is in the range of 0.6 to 1.0. The reason for selecting from the annular area is the same as the reason for selecting the audio signal data.

  Therefore, from the low-order signal vector space illustrated in FIG. 8, eight segmented regions in the annular region are selected, and a total of 16 men, eight men and eight women belonging to the selected eight segmented regions. The audio signal data for all the remaining speakers excluding the evaluation speaker is read from the storage device 70 and stored in a predetermined area of the RAM 62. Here, the sound data is clean sound data in which noise is not mixed as described above. Further, from the low-order noise vector space shown in FIG. 10, four regions “TR_1, TR_2, TR_3, TR_4” which are regions for pattern model generation in the eight divided regions of the annular region are selected. Noise data corresponding to all low-order noise vectors belonging to the four areas is read from the storage device 70 and stored in a predetermined area of the RAM 62 (step S122).

When the sound signal data and noise data for pattern model generation are selected and acquired in this way, the noise mixed signal data generation unit 24 connects the acquired noise data and acquires the connected noise. Noise mixed signal data is generated by mixing clean audio data with a predetermined SNR (for example, 20 dB) (step S116).
When the noise mixed signal data is generated, the pattern model generation unit 26 learns the HMM in the initial state prepared for speech recognition by using each noise mixed signal data as learning data by a known EM algorithm. A pattern model for speech recognition for a specific speaker is generated (step S118). In other words, since the pattern model for speech recognition is generated using the speech signal data and noise data corresponding to the pattern model with low recognition performance, a pattern model specialized for noise mixed signal data with low recognition performance is generated. Will be.

Furthermore, in the present embodiment, for the evaluation process described later, the following seven types of noise mixed signal data are generated, and each of the noise mixed signal data is used as learning data, and the generated pattern model for speech recognition is generated. Are re-learned by the EM algorithm to generate seven types of pattern models for evaluation.
(1) Noise mixed signal data obtained by concatenating noise data corresponding to all the low-order noise vectors belonging to the regions of TR_1 and TR_2, and mixing this with the selected speech signal data for pattern model generation at SNR 20 dB.

(2) Noise mixed signal data obtained by concatenating noise data corresponding to all the low-order noise vectors belonging to the regions of TR_2 and TR_3, and mixing this with the selected speech signal data for pattern model generation at SNR 20 dB.
(3) Noise mixed signal data obtained by concatenating noise data corresponding to all low-order noise vectors belonging to the regions of TR_3 and TR_4 and mixing this with the selected speech signal data for pattern model generation at SNR 20 dB.
(4) Noise mixed signal data obtained by concatenating noise data corresponding to all low-order noise vectors belonging to the regions of TR_4 and TR_1, and mixing this with the selected speech signal data for pattern model generation at SNR 20 dB.

(5) The noise data corresponding to all the low-order noise vectors belonging to the four regions “TR_1, TR_2, TR_3, TR_4”, excluding the four regions for evaluation, are concatenated, and the selected pattern is SNR 20 dB. Noise mixed signal data mixed with audio signal data for model generation.
(6) Noise mixed signal data obtained by concatenating noise data corresponding to all the low-order noise vectors in the low-order noise vector space and mixing the noise data with the selected sound signal data for pattern model generation at SNR 20 dB.
(7) Noise mixed signal data obtained by mixing exhibition noise, which has been conventionally used, with audio signal data for pattern model generation selected at SNR 20 dB.

Seven types of pattern models for speech recognition generated by re-learning the noise mixed signal data (1) to (7) as learning data are stored in a predetermined area of the storage device 70 by the pattern model storage unit 28. (Step S120).
Further, the actual operation of the evaluation process for the generated seven types of pattern models for speech recognition will be specifically described.

When receiving an evaluation instruction for the generated speech recognition pattern model from the user via the input device 74 (the branch of “Yes” in step S300), the pattern recognition device 100 uses this speech recognition device. Low-order signal vector space data and low-order noise vector space data corresponding to the pattern model are read from the storage device 70 and stored in a predetermined area of the RAM 62 (steps S302 and S304). Further, the low-order signal vector space data and the low-order noise vector space data are classified as shown in FIGS. 8 and 10 in the same manner as the pattern model generation process.
Again, the automatic selection mode is set, and the sound data and noise data for pattern model evaluation are selected according to a preset selection method.

  As described above, since it is set so that a total of 16 speakers, 8 males and 8 females, are selected as the evaluation speakers from the annular area illustrated in FIG. The unit 30 selects a low-order signal vector corresponding to the 16 men and women excluded at the time of generating the pattern model, reads out audio signal data corresponding to the selected low-order signal vector from the storage device 70, and stores a predetermined area of the RAM 62. To store.

  Furthermore, the lowest central noise vector is selected for each of the four regions “EV_1_2, EV_2_3, EV_3_4, EV_4_1”, which are the evaluation regions, and corresponds to the selected four low-order noise vectors. Noise data to be evaluated is selected as noise data for evaluation to be mixed with audio signal data for evaluation for the above seven types of pattern models for voice recognition. In the example of FIG. 10, low-order noise vectors are selected for four types of noises, that is, a bell sound of a bicycle, a ring tone of a mobile phone, a sound of cutting an adhesive tape, and a shutter sound of a camera, and the selected low-order noise vector is selected. Corresponding noise data is read from the storage device 70 and stored in a predetermined area of the RAM 62 (step S322).

  Then, in the mixed signal data generation unit 32 for evaluation, the four types of selected and acquired noise data are converted into the selected and acquired audio signal data (audio signal data with names of 1000 cities in Japan uttered by 16 men and women, respectively). ) Are separately mixed at SNR 20 dB, and four types of evaluation mixed signal data groups are generated (step S316). In this embodiment, the noise is mixed so as to be located in the voice section of the voice signal data with names of 1000 cities in Japan.

  When the four types of mixed signal data groups for evaluation are generated in this way, the pattern model evaluation unit 34 for each of the mixed signal data groups for evaluation with respect to the seven types of pattern models for speech recognition. Then, a speech recognition process is performed on each evaluation mixed signal data as a recognition target. Based on the speech recognition result for each pattern model, the recognition probability for each evaluation mixed signal data group is calculated, and the average value of the recognition probabilities for the four types of evaluation noise data is calculated for each pattern model. (Step S318). The recognition probability and the average recognition probability for each evaluation noise for each of the seven types of pattern models calculated in this way are displayed by the information display unit 36, resulting in the evaluation results shown in FIG. .

  In FIG. 11, "EV_" indicates the type of noise. If EV_1_2, the bell sound of the bicycle, EV_2_3 is the ringtone of the mobile phone, EV_3_4 is the sound of cutting the adhesive tape, and EV_4_1 is the camera's ring tone. The shutter sound is heard. In FIG. 11, “Base” indicates a pattern model corresponding to the above (7), and “TR_1_2, TR_2_3, TR_3_4, TR_4_1” indicate pattern models corresponding to the above (1) to (4), respectively. “Peri” indicates a pattern model corresponding to the above (5), and “All” indicates a pattern model corresponding to the above (6). In addition, “Ave.” in FIG. 11 indicates the average value of the recognition probabilities for the four types of evaluation mixed signal data for each pattern model.

  As shown in FIG. 11, the speech recognition probabilities of the pattern models corresponding to the four types of pattern model generation regions “TR_1_2, TR_2_3, TR_3_4, TR_4_1” are four types of “EV_1_2, EV_2_3, EV_3_4, EV_4_1”. It can be seen that the recognition probability is higher for the mixed signal data for evaluation with respect to the noise selected from the evaluation region than for the pattern model corresponding to “Base”. From this, it was generated from signal data and noise data selected from an annular region in which the ratio of the distance from the center of gravity is in the range of 0.6 to 1.0 (that is, a position relatively far from the center of gravity). The pattern model can be said to be more effective for speech recognition with respect to the mixed signal data for evaluation than the pattern model generated from the noise mixed signal data mixed with the conventional exhibition noise.

  Furthermore, as shown in FIG. 11, the pattern model corresponding to TR_1_2 has a higher recognition probability than the other pattern models for the mixed signal data for evaluation corresponding to EV_1_2, and similarly to TR_3_4 and TR_4_1, respectively. It can be seen that the recognition probability of the corresponding pattern model shows a higher recognition probability than the other pattern models for the speech recognition of the mixed signal data for evaluation corresponding to EV_3_4 and EV_4_1. Further, the recognition probability of the pattern model corresponding to TR_2_3 also shows a relatively high recognition probability of “87.9%” with respect to the speech recognition of the evaluation mixed signal data corresponding to EV_2_3. From this, when performing speech recognition on the evaluation mixed signal data corresponding to the evaluation area sandwiched between the two pattern model generation areas, it was generated using the noise data of these two areas. It can be said that it is effective to use a pattern model.

  Furthermore, as shown in FIG. 11, the pattern model corresponding to “Peri” is “TR_1_2, TR_2_3, TR_3_4” for speech recognition for the evaluation mixed signal data corresponding to “EV_1_2, EV_2_3, EV_3_4, EV_4_1”, respectively. , TR_4_1 ", it can be seen that the recognition probabilities are obtained by averaging the respective recognition probabilities of the pattern models respectively corresponding to TR_4_1". Also, as shown in FIG. 11, the average recognition probability of the pattern model corresponding to “Peri” (Ave. column in FIG. 11) is higher than the average recognition probability of the pattern model corresponding to “All”. I understand that. From this, the pattern model generated using the audio signal data and the noise data corresponding to the low-order vector selected from the annular region where the ratio of the distance from the center of gravity is in the range of 0.6 to 1.0 is It can be said that it is more effective for speech recognition with respect to the mixed signal data for evaluation than the pattern model using all noise data. It can also be seen that mixing all types of noise data into one pattern model does not necessarily give good results.

  Note that a pattern model generated using signal data and noise data corresponding to a low-order vector selected from an annular region in which the ratio of the distance from the center of gravity is in the range of 0.6 to 1.0 is Also relatively good for noise mixed signal data generated from speaker speech signal data and noise data corresponding to low-order vectors in the inner circle region where the distance ratio is less than 0.6 It has a recognition probability. Therefore, generally good recognition performance can be achieved by generating a pattern model using speech signal data and noise data corresponding to low-order vectors selected from an annular region that is more than a predetermined distance from the center of gravity. Can be generated. Hereinafter, the reason for this will be described with reference to FIGS.

  Here, FIG. 12 is a diagram illustrating an example of an utterance style. FIG. 13 is a diagram showing an example of a visualized low-order vector space of a pattern model for speech recognition. FIG. 14 is a diagram illustrating an example in which the low-order vector space of FIG. 13 is partitioned. FIG. 15 is a diagram showing recognition probabilities for the pattern model generated based on the category contents of FIG. 12 to 15 are diagrams generated based on actual experiments.

  First, voice data obtained by uttering a plurality of word lists consisting of 175 words in a plurality of utterance styles as shown in FIG. 12 are recorded in 145 men, and an HMM is used using all of the recorded voice data. A pattern model for unspecified speaker speech recognition for men having high-dimensional elements was generated by learning. Next, the generated pattern model is used as an initial model, and for each combination of each speaker and the utterance style instructed at the time of recording, the initial model is learned from each speaker's voice data by connection learning. Generated. Hereinafter, the generated pattern model is referred to as a speaker / utterance style acoustic model.

  Then, the data space consisting of this speaker / utterance style acoustic model consisting of high-dimensional elements is mapped to the data space consisting of two-dimensional vectors using the Sammon method as described above, and each low-order vector is mapped. Displayed as coordinate points in a two-dimensional space. The display result is as shown in FIG. FIG. 13 also shows the voicing speed and voicing volume axes obtained visually. The low-order vector space visualized in this way is, similarly to the above, an outer circle whose radius is the distance to the low-order vector farthest from the center of gravity, and one inner circle having a smaller radius than this outer circle. And an annular region formed between the curved lines of concentric circles composed of an outer circle and an inner circle was divided into four zones by lines extending in the radial direction. The classification result is as shown in FIG. As shown in FIG. 14, the low-order vector in FIG. 13 is divided into five regions, zone 1 to zone 5.

  In this way, when the low-order vector space of the speaker / utterance style acoustic model is divided into five zones, each zone of zones 1 to 5, each of zones 2 to 5 and all of the zones is Five pattern models and zones corresponding to each of the zones 1 to 5 using the voice data (excluding evaluation data) used to create the speaker / utterance style acoustic model corresponding to the low-order vector belonging to the zone A total of seven pattern models were generated: pattern models corresponding to 2 to 5 and pattern models corresponding to all zones. Hereinafter, the generated pattern model is referred to as a zone acoustic model. The creation method of the zone acoustic model is the same as that of the speaker / utterance style acoustic model.

  When the zone acoustic model is generated, the audio data corresponding to the low-order vectors (1-A to 5-D in the figure) labeled as the evaluation data are used in the low-order vector space shown in FIG. The word speech recognition of the above seven zone acoustic models was performed, and the recognition probability was evaluated. Here, when selecting the evaluation data, those having different distances from the center of gravity of the low-order vector space were appropriately selected. The evaluation data used was a mixture of exhibition noise and SNR of 20 dB. The word speech recognition result is as shown in FIG.

When observing FIGS. 13 to 15 in detail, the following seven of A to G are understood.
A. Zone 1 is mostly occupied by plots of normal speech styles (（).
B. Zone 2 is dominated by a plot of the utterance style (Δ) of the loud voice.
C. In Zone 3, where the mora enhancement (x) plot is dominant, a significant groove is observed between other utterance modes.
D. Zone 4 is dominated by loud (□) and Lombard (■) plots, and is located on the opposite side of zone 2 across which zone 1 is dominant.
E. Zone 5 in which the fast (O) and high voice (●) plots are dominant is located on the opposite side of Zone 3 from which Zone 1 is dominant.
F. The recognition performance becomes lower as it is located in the periphery of the low-order vector space.
G. The zone acoustic models corresponding to the collective zones of zones 2 to 5 have almost the same recognition probability as the zone acoustic models corresponding to the collective zones of all zones.

  That is, the zone acoustic model corresponding to the divided low-order vector space and the collective zones of zones 2 to 5 is the A. ~ F. And G. above. Since each of the above-mentioned characteristics is provided, a pattern model is generated by using voice signal data and noise data corresponding to a low-order vector selected from an annular region within a predetermined distance from the center of gravity. It is possible to generate a pattern model that can obtain excellent recognition performance.

  For the generation method of the low-order vector space, the visualization method of the low-order vector space, etc., a paper published by the present inventors ("M. Shozakai and G. Nagano," Analysis of Speaking Styles by Two-Dimensional Visualization of Aggregate of Acoustic Models, “Proc. ICSLP, vol. 1, pp. 717-720, Jeju, Korea, Oct. 2004.”).

  As described above, the pattern recognition apparatus 100 according to the present invention has a low-dimensional data space composed of high-order signal pattern models including high-dimensional elements in a state where the distance relation between the high-order signal pattern models is approximated. Consists of a two-dimensional (or three-dimensional) low-order signal vector space mapped to a data space consisting of low-order signal vectors containing elements, and a high-order noise pattern model containing high-dimensional elements. A two-dimensional (or three-dimensional) low-order image that maps the data space to a data space consisting of low-order noise vectors containing low-dimensional elements in a state in which the distance relationship between each high-order noise pattern model is approximated. The noise vector space is an outer circle (or outer sphere) whose radius is the distance to the low-order vector farthest from the center of gravity, and smaller than this outer circle (or outer sphere). The inner circles (or inner spheres) of different radii are used to distinguish between the concentric circles composed of the outer and inner spheres (or between the concentric spheres composed of the outer and inner spheres). The formed annular (spherical) region can be divided into a plurality of lines (or surfaces) extending in the radial direction. Furthermore, the distance from the center of the low-order vector located farthest from the center is set to 1, and a pattern model is generated from an annular region whose ratio to this distance is in the range of 0.6 to 1.0. It is possible to select signal data and noise data and generate a pattern model for recognition using the selected signal data and noise data. Therefore, since the signal data and noise data located at a predetermined distance or more from the center have unique characteristics, a pattern in which a good recognition probability can be obtained evenly for normal data and such unique data. A model can be generated.

  Further, the pattern recognition apparatus 100 of the present invention uses the above-described two-dimensional (or three-dimensional) low-order signal vector space and low-order noise vector space to determine the distance to the low-order vector that is farthest from the center of gravity. The outer circle (or outer sphere) with a radius and the inner circle (or inner sphere) with a smaller radius than this outer circle (or outer sphere) are separated, and concentric circles composed of the outer circle and inner circle (or An annular (spherical) region formed between each curve (or between curved surfaces) of concentric spheres consisting of an outer sphere and an inner sphere) can be divided into a plurality of lines (or surfaces) extending radially. It is. Further, the distance from the center of the low-order vector located farthest from the center is set to 1, and the pattern model is evaluated from an annular region whose ratio to the distance is in the range of 0.6 to 1.0. It is possible to select signal data and noise data, and to evaluate a pattern model for recognition using mixed signal data for evaluation generated using the selected signal data and noise data. Therefore, signal data and noise data that are more than a predetermined distance away from the center have unique characteristics, and these unique data are a major factor that degrades pattern model recognition performance. It is possible to easily and systematically evaluate the minimum performance of the pattern model by evaluating the pattern model using the above data.

In the above embodiment, the data storage unit 10 corresponds to the signal data storage unit and the noise data storage unit according to claim 1 or 5, and the low-order vector space storage unit 12 has the configuration according to claim 1 or 5. This corresponds to low-order signal vector space storage means and low-order noise vector space storage means.
Moreover, in the said embodiment, the high-order pattern model production | generation part 14 respond | corresponds to the high-order signal pattern model production | generation means and high-order noise pattern model production | generation means of Claim 4, The low-order vector space production | generation part 18 is This corresponds to the low-order signal vector space generating means and the low-order noise vector space generating means.

  In the above-described embodiment, the low-order vector space partitioning unit 20 corresponds to the low-order signal vector space partitioning unit and the low-order noise vector space partitioning unit according to claim 3 or 6, and the learning data selection unit 22 Corresponds to the low-order signal vector selection means and the low-order noise vector selection means according to any one of claims 1, 3, 9, and 11, and the noise mixed signal data generation unit 24 includes: The pattern model generation unit 26 corresponds to the pattern model generation unit according to any one of claims 1, 9, and 11, and corresponds to the noise mixed signal data generation unit according to any one of 9 and 11. .

Moreover, in the said embodiment, the evaluation data selection part 30 respond | corresponds to the evaluation signal vector selection means and the evaluation noise vector selection means of any one of Claim 5, 6, and 10, and is mixed for evaluation. The signal data generation unit 32 corresponds to the evaluation mixed signal data generation unit according to any one of claims 5, 6, and 10, and the pattern model evaluation unit 34 includes the performance evaluation unit according to claim 5 or 10. Corresponding to
Moreover, in the said embodiment, the data acquisition part 38 respond | corresponds to the data acquisition means of Claim 7 or 11, and the pattern recognition process part 40 respond | corresponds to the pattern recognition means of Claim 7 or 11.

Moreover, in the said embodiment, the information display part 36 respond | corresponds to the evaluation result output means of Claim 5 or 10, and the recognition result output means of Claim 7 or 11.
In the above embodiment, an example in which a pattern model for speech recognition is generated using a part of signal data and a part of noise data in the annular portion in consideration of the evaluation process has been described. However, the pattern model may be generated using all audio signal data and all noise data corresponding to low-order vectors belonging to the annular region.

  In the above-described embodiment, a single device has a function of generating a pattern model, a function of evaluating a pattern model, and a function of performing pattern recognition. However, the present invention is not limited to this. The pattern model generation device may be configured with only the functions necessary for generating the pattern model, or the pattern model evaluation device may be configured with only the functions necessary for evaluating the pattern model. The pattern recognition device may be composed of the pattern model generated by the pattern model generation device and the functions necessary for pattern recognition. Various combinations of devices such as a device having a pattern model generation function and an evaluation function and a device having a pattern model evaluation function and a pattern recognition function may be configured by freely combining the functions.

  Further, in the above embodiment, for evaluation from a plurality of low-order signal vectors constituting the low-order signal vector space, a low-order signal vector located at a position more than a predetermined distance from the center of gravity of all the low-order signal vectors. When selecting a low-order signal vector and for evaluation from a plurality of low-order noise vectors constituting a low-order noise vector space, from a low-order noise vector located at a position more than a predetermined distance from the center of gravity of all the low-order noise vectors The case of selecting a low-order noise vector has been described. However, the present invention is not limited to this, and as shown in FIG. 16, when visualized in two dimensions (or three dimensions), the outer circle (or outer sphere) and a plurality of A plurality of annular (or spherical) regions formed between inner circles (or inner spheres) are divided into a plurality of regions by the radially extending lines (or surfaces). Low-order noise vector of the low-order signal vector or evaluation for evaluation from the vicinity area of the may be selected. In this case, it is possible not only to evaluate the minimum value of the recognition performance in the real environment but also to comprehensively evaluate the recognition performance including the maximum value of the recognition performance in the real environment. That is, the signal data corresponding to the low-order signal vector for evaluation located near the center of gravity of all the low-order signal vectors has a higher recognition probability and is located near the curve (or curved surface) of the outer circle (or outer sphere). The signal data corresponding to the low-order signal vector for evaluation to be obtained has a lower recognition performance. Furthermore, the noise data corresponding to the low-order noise vector for evaluation located near the centroid of all the low-order noise vectors has a higher recognition probability and is located near the curve (or curved surface) of the outer circle (or outer sphere). The lower the probability of recognition, the more noise data corresponding to the evaluation low-order noise vector.

  In the above embodiment, the low-order signal vector space generation unit and the low-order noise vector space generation unit require a large amount of processing and storage capacity. Are independently selected, and a low-order signal vector selected from a plurality of low-order signal vectors constituting a low-order signal vector space and a low-order signal selected from a plurality of low-order noise vectors constituting a low-order noise vector space The next noise vector is mixed at a predetermined SNR to generate noise mixed signal data, and the probability model can be learned using the noise mixed signal data as learning data. Therefore, even when the types of signal data increase and it is necessary to recreate a low-order signal vector space, the low-order noise vector space does not have to be recreated. On the other hand, it is needless to say that the low-order signal vector space need not be re-created even when the number of types of noise data increases and the low-order noise vector space needs to be re-created.

According to the above-described embodiment, when the size of the noise cluster increases, the present invention makes a great deal of processing for creating a set of noise mixed speech HMMs and selecting an optimal noise mixed speech HMM from the set of noise mixed speech HMMs. It was possible to solve the problem of the prior art of Patent Document 1 to Patent Document 3 that takes time and labor.
Furthermore, according to the above embodiment, the present invention discloses a systematic method for deriving a combination of a learning data group and an evaluation data group that cause a reduction in recognition probability, which could not be achieved by the conventional technique. It has come.

It is a block diagram which shows the function structure of the pattern recognition apparatus which concerns on this invention. It is a figure which shows the hardware constitutions of the pattern recognition apparatus. It is a flowchart which shows an example of the production | generation process of the pattern model for pattern recognition. It is a flowchart which shows an example of the production | generation process of low-order vector space data. It is a flowchart which shows an example of the evaluation process of the pattern model for pattern recognition. It is a flowchart which shows an example of a pattern recognition process. It is a figure which shows an example of the two-dimensional low-order signal vector space with respect to audio | voice signal data. It is a figure which shows an example of the low-order signal vector space of FIG. 7 after area | region division. It is a figure which shows an example of the two-dimensional low-order noise vector space with respect to noise data. It is a figure which shows an example of the low-order noise vector space of FIG. 9 after area | region division. It is a figure which shows the evaluation result of the pattern model produced | generated based on actual data using the method of this invention. It is a figure which shows an example of an utterance style. It is a figure which shows an example of the low-order vector space visualized of the pattern model for speech recognition. It is a figure which shows an example which divided the low-order vector space of FIG. It is a figure which shows the recognition probability with respect to the pattern model produced | generated based on the division content of FIG. In two-dimensional visualization, it is a figure which shows the vertex of each area | region where the some cyclic | annular area | region formed between an outer circle and a some inner circle was divided into plurality by the line extended in a radial direction.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Pattern recognition apparatus 10 Data storage part 12 Low-order vector space storage part 14 High-order pattern model generation part 16 High-order pattern model storage part 18 Low-order vector space generation part 20 Low-order vector space division part 22 Data selection part 24 for learning Noise mixed signal data generation unit 26 Pattern model generation unit 28 Pattern model storage unit 30 Evaluation data selection unit 32 Evaluation mixed signal data generation unit 34 Pattern model evaluation unit 36 Information display unit 38 Data acquisition unit 40 Pattern recognition processing unit

Claims (14)

A pattern model generation device that generates a pattern model used for pattern recognition of recognition target data to be recognized,
Signal data storage means for storing a plurality of noise-unmixed pre-acquired signal data related to a plurality of objects;
As an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on the signal data, the positional relationship between the higher-order signal pattern models in the data space Low-order signal vector space storage means for storing a low-order signal vector space that is mapped to a data space composed of low-order signal vectors of three dimensions or less in an approximate state,
Noise data storage means for storing a plurality of types of signal-unmixed noise data;
As an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four-dimensional or higher-dimensional elements generated based on the noise data, the positional relationship between the higher-order noise pattern models in the data space Low-order noise vector space storage means for storing a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state where
Among the plurality of low-order signal vectors constituting the low-order signal vector space, a low-order signal vector for generating a pattern model from among low-order signal vectors located at a predetermined distance or more from the center of gravity of all the low-order signal vectors. Low-order signal vector selection means for selecting,
Of the plurality of low-order noise vectors constituting the low-order noise vector space, a low-order noise vector for generating a pattern model from among the low-order noise vectors located at a predetermined distance or more from the center of gravity of all the low-order noise vectors. Low-order noise vector selection means for selecting
Based on the signal data corresponding to the low-order signal vector selected by the low-order signal vector selection means and the noise data corresponding to the low-order noise vector selected by the low-order noise vector selection means, the noise is added to each signal data. Noise mixed signal data generating means for generating noise mixed signal data obtained by mixing data;
A pattern model generation device, comprising: a pattern model generation unit configured to generate a pattern model obtained by learning a probability model using the noise mixed signal data as learning data.   2. The pattern model generation apparatus according to claim 1, wherein the high-order signal pattern model, the high-order noise pattern model, and the pattern model are configured by an HMM (Hidden Markov Model). The low-order signal vector space and the low-order noise vector space are two-dimensional or three-dimensional data spaces,
A plurality of low-order signal vectors constituting the low-order signal vector space are centered on the centroid of all low-order signal vectors, and the distance between the centroid and the low-order signal vector at the position farthest from the centroid is the radius. The outer circle or outer sphere is divided into n outer circles or inner spheres (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer circle or outer sphere. A plurality of concentric circles consisting of circles and inner circles or outer spheres and inner spheres, or annular or spherical regions formed between curved lines or curved surfaces of concentric spheres are divided into a plurality by radially extending lines or surfaces. Low-order signal vector space partitioning means for partitioning;
A plurality of low-order noise vectors constituting the low-order noise vector space are centered on the centroid of the low-order noise vector, and the distance between the centroid and the low-order noise vector at a position farthest from the centroid is a radius. The outer circle or outer sphere is divided into n outer circles or inner spheres (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer circle or outer sphere. A plurality of concentric circles consisting of circles and inner circles or outer spheres and inner spheres, or annular or spherical regions formed between curved lines or curved surfaces of concentric spheres are divided into a plurality by radially extending lines or surfaces. Low-order noise vector space classification means for classifying;
Low-order signal vector space display means for displaying each low-order signal vector constituting the low-order signal vector space together with the classification result;
Low-order noise vector space display means for displaying each low-order noise vector constituting the low-order noise vector space together with the classification result, and
The low-order signal vector selection means includes a plurality of concentric circles or concentric spheres composed of the outer circle or outer sphere and an inner circle or inner sphere having a radius equal to or greater than the predetermined distance in the displayed lower-order signal vector space. A low-order signal vector for generating a pattern model is selected from low-order signal vectors included in an annular or spherical region formed between each curve or between curved surfaces.
The low-order noise vector selection means includes a plurality of concentric circles or concentric spheres composed of the outer circle or outer sphere and an inner circle or inner sphere having a radius equal to or greater than the predetermined distance in the displayed lower-order noise vector space. A low-order noise vector for generating a pattern model is selected from low-order noise vectors included in an annular or spherical area formed between each curve or between curved surfaces. The pattern model generation apparatus according to claim 1 or 2. A high-order signal pattern model generating means for extracting a feature quantity from the signal data and generating the high-order signal pattern model based on the extracted feature quantity;
A high-order noise pattern model generating means for extracting a feature value from the noise data and generating the high-order noise pattern model based on the extracted feature value;
As an alternative to a data space composed of a plurality of higher-order signal pattern models, data composed of lower-order signal vectors of three dimensions or less in a state in which the positional relationship between the higher-order signal pattern models in the data space is approximated Low-order signal vector space generating means for generating the low-order signal vector space formed by mapping to space;
As an alternative to the data space composed of the plurality of high-order noise pattern models, data composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the high-order noise pattern models in the data space is approximated 4. The pattern model generation device according to claim 1, further comprising: a low-order noise vector space generation unit configured to generate the low-order noise vector space mapped to a space. 5. . A pattern model evaluation device for evaluating the recognition performance of a pattern model used for pattern recognition of recognition target data to be recognized,
Signal data storage means for storing a plurality of noise-unmixed pre-acquired signal data related to the plurality of objects;
As an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on the signal data, the positional relationship between the higher-order signal pattern models in the data space Low-order signal vector space storage means for storing a low-order signal vector space that is mapped to a data space composed of low-order signal vectors of three dimensions or less in an approximate state,
Noise data storage means for storing a plurality of types of signal-unmixed noise data;
As an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four-dimensional or higher-dimensional elements generated based on the noise data, the positional relationship between the higher-order noise pattern models in the data space Low-order noise vector space storage means for storing a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state where
Among the plurality of low-order signal vectors constituting the low-order signal vector space, the low-order signal vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order signal vectors or that are uniformly centered on the center of gravity. Evaluation signal vector selection means for selecting a low-order signal vector for evaluation from the inside;
Among the plurality of low-order noise vectors constituting the low-order noise vector space, the low-order noise vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order noise vectors or that are uniformly centered on the center of gravity. Evaluation noise vector selection means for selecting a low-order noise vector for evaluation from the inside;
Based on the signal data corresponding to the low-order signal vector selected by the evaluation signal vector selection means and the noise data corresponding to the low-order noise vector selected by the evaluation noise vector selection means, each signal data Mixed signal data generation means for evaluation for generating mixed signal data for evaluation mixed with data; and
Performance evaluation means for evaluating the performance of the pattern model based on the mixed signal data for evaluation generated by the mixed signal data generation means for evaluation and the pattern model;
An evaluation result output means for outputting an evaluation result of the performance evaluation means. The low-order signal vector space and the low-order noise vector space are two-dimensional or three-dimensional data spaces,
A plurality of low-order signal vectors constituting the low-order signal vector space are centered on the centroid of all low-order signal vectors, and the distance between the centroid and the low-order signal vector at the position farthest from the centroid is the radius. The outer circle or outer sphere is divided into n outer circles or inner spheres (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer circle or outer sphere. A plurality of concentric circles consisting of circles and inner circles or outer spheres and inner spheres, or annular or spherical regions formed between curved lines or curved surfaces of concentric spheres are divided into a plurality by radially extending lines or surfaces. Low-order signal vector space partitioning means for partitioning;
A plurality of low-order noise vectors constituting the low-order noise vector space are centered on the centroid of the low-order noise vector, and the distance between the centroid and the low-order noise vector at a position farthest from the centroid is a radius. The outer circle or outer sphere is divided into n outer circles or inner spheres (n is an integer of 1 or more) centered on the center of gravity and having a smaller radius than the outer circle or outer sphere. A plurality of concentric circles consisting of circles and inner circles or outer spheres and inner spheres, or annular or spherical regions formed between curved lines or curved surfaces of concentric spheres are divided into a plurality by radially extending lines or surfaces. Low-order noise vector space classification means for classifying;
Low-order signal vector space display means for displaying each low-order signal vector constituting the low-order signal vector space together with the classification result;
Low-order noise vector space display means for displaying each low-order noise vector constituting the low-order noise vector space together with the classification result, and
The evaluation signal vector selection means and the evaluation noise vector selection means are lines extending in the radial direction in the annular region or the spherical region of the displayed low-order signal vector space and low-order noise vector space. Alternatively, a low-order signal vector and a low-order noise vector for evaluation are selected from each region divided into a plurality of areas according to a plane,
The evaluation mixed signal data generation means includes each signal based on signal data and noise data respectively corresponding to a low-order signal vector and a low-order noise vector selected by the evaluation signal vector selection means and the evaluation noise vector selection means. 6. The pattern model evaluation apparatus according to claim 5, wherein mixed signal data for evaluation obtained by mixing each noise data with the data is generated. The pattern model generation device according to any one of claims 1 to 4,
Data acquisition means for acquiring recognition target data of the recognition target;
Pattern recognition means for performing pattern recognition of the acquired recognition target data based on the recognition target data acquired in the data acquisition means and the pattern model generated in the pattern model generation device;
And a recognition result output means for outputting a recognition result of the pattern recognition means.   The pattern recognition apparatus according to claim 7, comprising the pattern model evaluation apparatus according to claim 5. A pattern model generation program for generating a pattern model used for pattern recognition of recognition target data to be recognized,
The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Among them, low-order signal vector selection means for selecting a low-order signal vector for pattern model generation from low-order signal vectors located at a predetermined distance or more away from the center of gravity of all low-order signal vectors,
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, Low-order noise vector selection means for selecting a low-order noise vector for pattern model generation from low-order noise vectors located at a predetermined distance or more away from the center of gravity of all low-order noise vectors;
Based on the signal data corresponding to the low-order signal vector selected by the low-order signal vector selection means and the noise data corresponding to the low-order noise vector selected by the low-order noise vector selection means, the noise is added to each signal data. Noise mixed signal data generating means for generating noise mixed signal data obtained by mixing data;
A pattern model generation program comprising a program for causing a computer to realize a function as a pattern model generation unit that generates a pattern model obtained by learning a probability model using the noise mixed signal data as learning data. A pattern model evaluation program for evaluating the recognition performance of a pattern model used for pattern recognition of recognition target data to be recognized,
The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Evaluation signal vector for selecting a low-order signal vector for evaluation from among low-order signal vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order signal vectors, or that are uniformly centered on the center of gravity. A selection means;
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, Evaluation noise vector selection means for selecting a low-order noise vector for evaluation from low-order noise vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order noise vectors, or are uniformly centered on the center of gravity; ,
Based on the signal data corresponding to the low-order signal vector selected by the evaluation signal vector selection means and the noise data corresponding to the low-order noise vector selected by the evaluation noise vector selection means, each signal data Mixed signal data generation means for evaluation for generating mixed signal data for evaluation mixed with data; and
Performance evaluation means for evaluating the performance of the pattern model based on the mixed signal data for evaluation generated by the mixed signal data generation means for evaluation and the pattern model;
A pattern model evaluation program comprising a program for causing a computer to realize a function as an evaluation result output means for outputting an evaluation result of the performance evaluation means. The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Among them, low-order signal vector selection means for selecting a low-order signal vector for pattern model generation from low-order signal vectors located at a predetermined distance or more away from the center of gravity of all low-order signal vectors,
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, Low-order noise vector selection means for selecting a low-order noise vector for pattern model generation from low-order noise vectors located at a predetermined distance or more away from the center of gravity of all low-order noise vectors;
Based on the signal data corresponding to the low-order signal vector selected by the low-order signal vector selection means and the noise data corresponding to the low-order noise vector selected by the low-order noise vector selection means, the noise is added to each signal data. Noise mixed signal data generating means for generating noise mixed signal data obtained by mixing data;
Pattern model generation means for generating a pattern model obtained by learning a probability model using the noise mixed signal data as learning data;
Data acquisition means for acquiring recognition target data of the recognition target;
Pattern recognition means for performing pattern recognition of the acquired recognition target data based on the recognition target data acquired in the data acquisition means and the pattern model generated in the pattern model generation device;
A pattern recognition program comprising a program for causing a computer to realize a function as a recognition result output means for outputting a recognition result of the pattern recognition means. A pattern model generation method for generating a pattern model used for pattern recognition of recognition target data to be recognized,
The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Among them, a low-order signal vector selection step for selecting a low-order signal vector for pattern model generation from among low-order signal vectors located at a predetermined distance or more away from the center of gravity of all the low-order signal vectors;
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, A low-order noise vector selection step for selecting a low-order noise vector for generating a pattern model from low-order noise vectors located at a predetermined distance or more away from the center of gravity of all low-order noise vectors;
Based on the signal data corresponding to the low-order signal vector selected in the low-order signal vector selection step and the noise data corresponding to the low-order noise vector selected in the low-order noise vector selection step, the noise is added to each signal data. Noise mixed signal data generation step for generating noise mixed signal data obtained by mixing data;
And a pattern model generation step of generating a pattern model obtained by learning a probability model using the noise mixed signal data as learning data. A pattern model evaluation method for evaluating the recognition performance of a pattern model used for pattern recognition of recognition target data to be recognized,
The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Evaluation signal vector for selecting a low-order signal vector for evaluation from among low-order signal vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order signal vectors, or that are uniformly centered on the center of gravity. A selection step;
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, An evaluation noise vector selection step for selecting a low-order noise vector for evaluation from low-order noise vectors that are located at a predetermined distance or more away from the center of gravity of all the low-order noise vectors or that are uniformly centered on the center of gravity; ,
Based on the signal data corresponding to the low-order signal vector selected in the evaluation signal vector selection step and the noise data corresponding to the low-order noise vector selected in the evaluation noise vector selection step, each signal data An evaluation mixed signal data generation step for generating mixed signal data for evaluation mixed with noise data;
A performance evaluation step for evaluating the performance of the pattern model based on the mixed signal data for evaluation generated in the mixed signal data generation step for evaluation and the pattern model;
An evaluation result output step of outputting an evaluation result of the performance evaluation step. The data space generated as an alternative to a data space composed of a plurality of higher-order signal pattern models composed of four-dimensional or higher-dimensional elements generated based on a plurality of noise-unmixed signal data related to a plurality of objects A plurality of low-order signal vectors constituting a low-order signal vector space mapped to a data space composed of low-order signal vectors of three dimensions or less in a state in which the positional relationship between the high-order signal pattern models in FIG. Among them, a low-order signal vector selection step for selecting a low-order signal vector for pattern model generation from among low-order signal vectors located at a predetermined distance or more away from the center of gravity of all the low-order signal vectors;
Each height in the data space generated as an alternative to a data space composed of a plurality of higher-order noise pattern models composed of four or more high-dimensional elements generated based on a plurality of types of signal-unmixed noise data Among a plurality of low-order noise vectors constituting a low-order noise vector space mapped to a data space composed of low-order noise vectors of three dimensions or less in a state in which the positional relationship between the second-order noise pattern models is approximated, A low-order noise vector selection step for selecting a low-order noise vector for generating a pattern model from low-order noise vectors located at a predetermined distance or more away from the center of gravity of all low-order noise vectors;
Based on the signal data corresponding to the low-order signal vector selected in the low-order signal vector selection step and the noise data corresponding to the low-order noise vector selected in the low-order noise vector selection step, the noise is added to each signal data. Noise mixed signal data generation step for generating noise mixed signal data obtained by mixing data;
A pattern model generation step of generating a pattern model obtained by learning a probability model using the noise mixed signal data as learning data;
A data acquisition step of acquiring recognition target data of the recognition target;
A pattern recognition step for performing pattern recognition of the acquired recognition target data based on the recognition target data acquired in the data acquisition step and the pattern model generated in the pattern model generation step;
And a recognition result output step of outputting a recognition result of the pattern recognition step.

Images