JP2020056918A - Sound data learning system, sound data learning method and sound data learning device - Google Patents

Abstract

To assist in compatibly reducing a processing load in execution of learning processing and generating a learnt model efficiently with high precision according to its use.SOLUTION: The sound data learning system extracts J feature quantities of a sound signal for learning and generates N feature quantity reduction blocks having parameters for defining the presence or absence of use. The sound data learning system executes: preparation processing to prepare data sets of k kinds of different patterns using a sound signal for learning; learning processing, for each of the N feature quantity reduction blocks, using the corresponding feature quantity reduction block and some of the data sets with respect to respective different data structures; and feature quantity reduction processing to repeat evaluation processing associated with use of each of J feature quantities targeting the remaining data sets. The sound data learning system selects, based upon a feature quantity reduction processing result of each of the N feature quantity reduction blocks having k kinds of evaluation results, one of the feature quantity reduction blocks as reduction information.SELECTED DRAWING: Figure 4

Description

  The present disclosure relates to a sound data learning system, a sound data learning method, and a sound data learning device that execute processing related to learning of input sound data.

  Patent Literature 1 discloses an apparatus for detecting an abnormal sound of a machine using learning data given in advance for normal operation of a mechanical sound. Specifically, this device separates an input frequency-domain signal into two or more types of signals having different sound properties from each other, and extracts an acoustic feature amount for each of the two or more types of signals. Further, the apparatus calculates the degree of abnormality of each of the two or more signals by using the extracted acoustic feature amounts and the respective models of the normal state of the two or more signals learned in advance. It is determined whether the signal in the frequency domain is abnormal based on the integrated abnormality degree obtained by integrating the abnormality degrees.

  In recent years, for the purpose of marketing, tagging video data, and the like, effective computing (Affective Computing) technology that analyzes (identifies) human emotions and emotions using a model based on machine learning has attracted attention. . For example, the most actively researched emotion analysis technology is analysis based on images, and analysis of emotions from facial expressions has been studied to the point where a certain degree of accuracy can be obtained. On the other hand, human emotions are included not only in facial expressions but also in uttered voices. For this reason, it is difficult to obtain the face image of the other party by telephone reception at a call center or the like, but it is difficult for the operator to recognize the emotion of the other party via the image. Therefore, it is considered important to read emotions from the uttered voice of the other party. However, unlike facial images, technology for analyzing emotions from uttered voices has not yet matured and research is being actively conducted.

JP 2017-90606 A

  In order to improve the accuracy of analysis results of human emotions and emotions via machine learning, it is important to generate a more suitable learned model by machine learning. However, in order to generate such a trained model in the related art including Patent Document 1 described above, a large amount of input data is required as learning data, and a large amount of dimensional features are required for machine learning. Had to be dealt with. Therefore, a processing load is imposed on an apparatus that performs machine learning, and it is difficult to efficiently and accurately generate a learned model according to a use.

  The present disclosure has been devised in view of the above-described conventional circumstances, and has a sound that supports both reduction of the processing load imposed at the time of executing a learning process and efficient and highly accurate generation of a learned model according to the application. It is an object to provide a data learning system, a sound data learning method, and a sound data learning device.

  The present disclosure relates to a microphone that collects a learning sound signal, an extraction unit that extracts J (J: a positive integer) feature amounts of the learning sound signal, and use of each of the J feature amounts. A generator (N: N: an integer equal to or greater than 2) configured by J parameters defining presence / absence of the feature amount reduction block, and k (k : Two or more specified values) A preparation process for preparing data sets of different types of patterns, and for each of the N feature amount reduction blocks, the feature amounts corresponding to k different data sets of the different patterns A learning process using the reduced block and a partial data set of the data of the learning sound signal, and using a learning result of the learning process, the remaining data of the learning sound signal. A feature amount reduction unit that performs a feature amount reduction process that repeats an evaluation process for use of each of the J feature amounts for the data set, and the N feature amounts each having k types of evaluation results. A learning control unit that selects one of the feature amount reduction blocks as J pieces of feature amount reduction information based on a feature amount reduction processing result for each reduction block; .

  Further, the present disclosure relates to a sound data learning method in a sound data learning system having a microphone, wherein the microphone collects a learning sound signal, and the learning sound signal J (J: positive integer). ) Number of feature amounts, and N (N: an integer equal to or greater than 2) feature amount reduction blocks composed of J parameters defining whether or not each of the J feature amounts is used. Generating each, using the data of the learning sound signal, preparing a data set of k (k: a specified value of 2 or more) different patterns, and preparing for each of the N feature amount reduction blocks In the k data sets of the different patterns, a part of the data of the corresponding feature amount reduction block and the data of the learning sound signal is used. A learning process using the data set, and an evaluation process for using each of the J feature amounts for the remaining data set of the data of the learning sound signal using a learning result of the learning process. Performing a feature amount reduction process that repeats the above. Based on the feature amount reduction process results for each of the N feature amount reduction blocks, each of which has k types of evaluation results, Selecting J pieces of feature amount reduction information as sound information learning methods.

  In addition, the present disclosure relates to an extraction unit that is connected to a microphone that collects a learning sound signal and extracts J (J: a positive integer) feature amounts of the learning sound signal; A generation unit configured to generate N (N: an integer equal to or greater than 2) number of feature amount reduction blocks, each of which includes J parameters defining whether or not each is used, and using the data of the learning sound signal. , K (k: 2 or more specified values) of different types of pattern data sets, and for each of the N feature amount reduction blocks, k sets of the different pattern data sets are applied. A learning process using the feature amount reduction block and a partial data set of the data of the learning sound signal, and a learning result of the learning sound signal using a learning result of the learning process. And a feature amount reduction unit that performs a feature amount reduction process that repeats an evaluation process for each of the J feature amounts for the remaining data sets of the data, and k types of evaluation results. A learning control unit that selects one of the feature amount reduction blocks as J feature amount reduction information based on the feature amount reduction processing result for each of the N feature amount reduction blocks. A data learning device is provided.

  It should be noted that any combination of the above-described components, and a representation of the present disclosure converted between a method, an apparatus, a system, a recording medium, a computer program, a learned model (learning model), and the like are also included in the embodiments of the present disclosure. It is valid.

  According to the present disclosure, it is possible to support both the reduction of the processing load imposed at the time of executing the learning process and the efficient and high-accuracy generation of the learned model according to the application.

FIG. 2 is a block diagram showing a configuration example of a sound data learning system according to Embodiment 1. Explanatory drawing of an example of learning processing and evaluation processing by a classifier of a learning processing unit using cross-validation Explanatory conceptual diagram of feature amount reduction processing using a genetic algorithm (GA) Flow chart for explaining an example of an overall operation procedure of the sound data learning process. 4 is a flowchart illustrating an example of an operation procedure of a learning process using the reduced feature amount. Flowchart for explaining an example of an operation procedure of sound data analysis processing using a learned model A graph in which input sound data is associated with an example of an analysis result of the sound data FIG. 4 is a block diagram showing a configuration example of a sound data learning system according to a modification of the first embodiment.

  Hereinafter, an embodiment that specifically discloses the configuration and operation of a sound data learning system, a sound data learning method, and a sound data learning device according to the present disclosure will be described in detail with reference to the drawings as appropriate. However, an unnecessary detailed description may be omitted. For example, a detailed description of a well-known item or a redundant description of substantially the same configuration may be omitted. This is to prevent the following description from being unnecessarily redundant and to facilitate understanding by those skilled in the art. The accompanying drawings and the following description are provided to enable those skilled in the art to fully understand the present disclosure, and are not intended to limit the claimed subject matter.

(Embodiment 1)
FIG. 1 is a block diagram illustrating a configuration example of a sound data learning system 200 according to Embodiment 1. The sound data learning system 200 according to the first embodiment collects sounds or voices (hereinafter, collectively referred to as “sounds”) emitted by a target person using the microphone 110, and outputs the collected sound data via the audio interface 120. The information is input to the information processing device 140. The sound data learning system 200 uses, for example, machine learning using input sound data in order to generate a learned model capable of analyzing emotions or emotions (hereinafter, collectively referred to as “emotions”) included in sounds emitted by a target person. Alternatively, the information processing device 140 executes a learning process such as deep learning (see below).

  As shown in FIG. 1, the sound data learning system 200 has a configuration including a microphone 110, an audio interface 120, and an information processing device 140. In FIG. 1, the microphone 110 is abbreviated as “MIC”, and the interface is abbreviated as “I / F”. Although only one microphone 110 is shown in FIG. 1, a plurality of microphones 110 may be connected to the audio interface 120. One microphone 110 and one audio interface 120 may be provided in association with each other. In this case, the same number of audio interfaces 120 as the number of microphones 110 are provided.

  In the sound data learning system 200, the configuration of the audio interface 120 may be included so as to be taken into the information processing device 140. In this case, the information processing device 140 and the microphone 110 are directly connected. Good.

  The microphone 110 is configured to include a sound collection device that inputs a sound wave constituting a sound emitted by a target person, generates an audio signal of an electric signal, and outputs the audio signal.

  The audio interface 120 is an interface device for inputting sound data that converts an audio signal generated by the microphone 110 into digital data that can be subjected to various kinds of signal processing. The audio interface 120 has an input unit 121, an AD converter 122, a buffer 123, and a communication unit 124. In FIG. 1, the AD converter 122 is abbreviated as “ADC”.

  The input unit 121 has an input terminal for inputting an audio signal.

  The AD converter 122 converts an analog audio signal into digital sound data using a predetermined quantization bit and a sampling frequency. The sampling frequency of the AD converter 122 is, for example, 48 kHz.

  The buffer 123 has a storage capacity (memory) capable of temporarily storing sound data, and buffers sound data for a predetermined time. The buffer capacity of the buffer 123 is, for example, about 40 msec. By using a relatively small buffer capacity as described above, it is possible to reduce the delay of the learning process and the like in the sound data learning system 200.

  The communication unit 124 is configured using a communication circuit having a communication interface such as a USB (Universal Serial Bus), and transmits and receives data to and from an external device such as the information processing device 140. The communication unit 124 transmits the sound data input from the buffer 123 to the information processing device 140.

  The information processing device 140 is configured by, for example, a PC (Personal Computer) having a processor and a memory, and executes a sound data learning process or a sound data detection process. Hereinafter, the input data provided for the learning process in the sound data learning system 200 (that is, the sound data input for learning) is referred to as “learning sound data”. The input data (that is, the sound data input for detection described above) to be used for the detection processing (that is, the analysis processing of human emotion or emotion) in the sound data learning system 200 is referred to as “detection sound data”. The information processing device 140 may be configured by a device capable of transmitting and receiving data to and from the audio interface 120 such as a tablet terminal or a smartphone, instead of the PC. The information processing device 140 includes a communication unit 141, a processing unit 142, a storage unit 143, an operation input unit 144, and a display unit 145.

  The communication unit 141 is configured using a communication circuit having a communication interface such as a USB (Universal Serial Bus), and transmits and receives data to and from an external device such as the audio interface 120. The communication unit 141 inputs sound data sent from the audio interface 120 and sends the sound data to the processing unit 142.

  The processing unit 142 is configured by a processor such as a CPU (Central Processing Unit), a DSP (Digital Signal Processor) or an FPGA (Field Programmable Gate Array). The processing unit 142 executes various processes according to a predetermined program and data stored in the storage unit 143 such as a memory, for example, and performs learning sound data learning processing or detection sound data detection processing described below. The processing unit 142 uses the RAM and ROM (see below) of the storage unit 143 during operation, and temporarily stores data or information generated or obtained by the processing unit 142 in the above-described RAM.

  The processing unit 142 functions as a functional configuration by cooperating with a RAM and a ROM in the storage unit 143, and a control unit 151 that comprehensively controls execution of various processes and a learning processing unit 152 that performs a learning process. , A detection processing unit 153 that performs a detection process, and a determination processing unit 154 that performs a determination process. Details of each processing in the control unit 151, the learning processing unit 152, the detection processing unit 153, and the determination processing unit 154 will be described later.

  The storage unit 143 includes a semiconductor memory such as a random access memory (RAM) and a read only memory (ROM), and a storage device such as a solid state drive (SSD) or a hard disk drive (HDD). The ROM of the storage unit 143 stores a program and data for executing the functions of the learning process, the detection process, and the determination process, and various setting data referred to at the time of executing the learning process, the detection process, and the determination process. You. The RAM, SSD, or HHD of the storage unit 143 stores sound data to be analyzed and various data generated at the time of learning processing, detection processing, or determination processing of the sound data.

  The operation input unit 144 has input devices such as a keyboard, a mouse, a touchpad, and a touch panel. The operation input unit 144 receives a user's input operation when using the sound data learning system 200 and outputs the operation to the processing unit 142.

  The display unit 145 has a display device such as a liquid crystal display (LCD) or an organic EL (Electroluminescence) display. The display unit 145 displays a display screen (not shown) when the processing unit 142 executes processing such as learning processing, detection processing, and determination processing.

  The control unit 151 functions as a controller that controls the overall operation of the information processing device 140, performs control processing for controlling the operation of each unit in the processing unit 142 of the information processing device 140, It performs data input / output processing, data calculation (calculation) processing, and data storage processing. When acquiring the learning sound data transmitted from the communication unit 141 based on a user operation using the operation input unit 144, the control unit 151 passes the learning sound data to the learning processing unit 152 and instructs the learning processing unit 152 to execute the learning process. I do. Further, when acquiring the analysis sound data transmitted from the communication unit 141 based on the user operation using the operation input unit 144, the control unit 151 passes the analysis sound data to the detection processing unit 153 and performs the detection processing (that is, Analysis process).

  The learning processing unit 152 as an example of the extraction unit, according to an instruction from the control unit 151, sets a predetermined J dimension (J: positive) of the learning sound data (one mode of the learning sound signal) passed from the control unit 151. An integer, for example, 988 as an example of about 1000. The same applies hereinafter.) The learning processing unit 152, which is an example of a learning unit, performs a learning process (see below) by using feature amount data of the extracted learning sound data (one mode of the learning sound signal), thereby performing a detection process. A learned model for analyzing a human emotion included in the sound data (one mode of the detection sound signal) is generated.

  In the first embodiment, the information processing device 140 achieves both reduction of the load at the time of executing the learning process and generation of an efficient and highly accurate learned model. Therefore, the learning processing unit 152 does not perform the learning process using all the high-dimensional (eg, 988-dimensional) feature amounts as in the conventional machine learning, but reduces a part of the high-dimensional feature amounts. Execute the learning process above.

  The feature amount is characteristic data included in the learning sound data or the detection sound data, and is, for example, “loudness”, “intensity”, or “MFCC (Mel-Frequency Cepstrum Coefficients)”. , “ZCR (Zero Crossing Ratio)”, and “F0 (Fundamental Frequency). Loudness is the intensity of sound perceived by human hearing, and is one of the sensations. MFCC is a characteristic quantity indicating vocal tract characteristics, and ZCR indicates that the signal level of the learning sound data or the detection sound data is 0 (zero) within a certain period of time. ), And this value tends to increase in the voice section.F0 is the fundamental frequency of the voice.Five (five-dimensional) feature values are illustrated here for reference. Type of quantity But it is not limited to these.

  The outline of the learning processing in the learning processing unit 152 is as follows. Details will be described in detail with reference to FIGS.

(1) The learning processing unit 152 uses a genetic algorithm (GA) in order to reduce the number of dimensions of the feature amount, and uses a DNN (Deep) in an identifier (Identifier, that is, an identification algorithm) in the learning process. Neural Network).
(2) When the genetic algorithm is executed, the learning processing unit 152 generates a plurality of individuals (see below) having genes (see below) in the current generation, and calculates an evaluation value of each individual by DNN. .
(3) Based on the evaluation value of each individual in the current generation calculated by the DNN, the learning processing unit 152 selects a plurality (for example, two )select. The learning processing unit 152 generates a next-generation individual by combining the plurality of selected individuals to generate a new individual.

  The learning processing unit 152 performs the series of processes (1) to (3) as a process in one generation, and repeats the series of processes described above for the M-th generation in the same manner. Can be obtained. M is, for example, a specified integer value of 2 or more.

  Here, the gene possessed by the individual is a parameter that defines whether or not the feature amount of the learning sound data or the detection sound data used by the DNN is used, and is specifically “1” or “0”. “1” is a parameter instructing the use of the feature amount. “0” is a parameter instructing non-use of the feature amount.

  Here, the algorithm or method for realizing the learning processing executed in the learning processing unit 152 may be performed using, for example, one or more statistical classification techniques. Examples of the statistical classification technology include linear classifiers (Linear Classifiers), support vector machines (Support Vector Machines), quadratic classifiers (Quadratic Classifiers), kernel density estimation (Kernel Estimation), decision trees (Decision Trees), Artificial Neural Networks, Bayesian Techniques and / or Networks, Hidden Markov Models, Binary Classifiers, Multi-Class Classifiers ) Clustering (Clustering Technique), Random Forest (Random Forest Technique), Logistic Regression (Logistic Regression Technique), Linear Regression (Linear Regression Technique), Gradient Boosting Technique, and the like. However, the statistical classification technique used is not limited to these. Further, the learning process may be performed by the processing unit 142 in the information processing device 140, or for example, a server device 340 communicably connected to the information processing device 140 using a network (see FIG. 8). It may be done in.

  The detection processing unit 153, which is an example of the detection unit, converts the sound data for detection (an example of a sound signal) passed from the control unit 151 into a learned state generated by the learning processing unit 152 in accordance with an instruction from the control unit 151. By inputting the data to the model, a process of analyzing the emotion of a person included in the detection sound data (one mode of the detection sound signal) (that is, a detection process) is executed.

  The determination processing unit 154 outputs the detection result of the detection processing of the detection processing unit 153 as the determination result of the emotion of the person included in the detection sound data. The output destination of the determination result by the determination processing unit 154 may be the storage unit 143 or the display unit 145, or may be an external device (not shown) communicably connected to the information processing device 140. Further, the processing of the determination processing unit 154 may be executed by the detection processing unit 153, and in this case, the configuration of the determination processing unit 154 may be omitted.

  Next, details of the learning processing in the learning processing unit 152 will be described with reference to FIGS. FIG. 2 is an explanatory diagram illustrating an example of a learning process and an evaluation process performed by the discriminator of the learning processing unit 152 using cross validation. FIG. 3 is an explanatory conceptual diagram of a feature amount reduction process using a genetic algorithm (GA).

  In order to make the description of FIGS. 2 and 3 easy to understand, the learning sound data VC1 is, for example, five data blocks BLK1, BLK2, BLK3, and BLK4 as sound data of voices uttered for testing by five persons. , BLK5. For example, the data block BLK1 may be the first person, the data block BLK2 may be the second person, the data block BLK3 may be the third person, the data block BLK4 may be the fourth person, and the data block BLK5 may be the fifth person's voice. Alternatively, the data blocks BLK1, BLK2, BLK3, BLK4, BLK5 may be data obtained by simply dividing the entirety of the learning sound data VC1 into five equal parts. The learning sound data VC1 illustrated in FIG. 2 may be replaced with a predetermined J-dimensional (eg, 988-dimensional) feature amount of the learning sound data VC1. J is a predetermined positive integer. In this case, the data blocks BLK1 to BLK5 are feature amount data blocks obtained by dividing the 988-dimensional feature amount into five equal parts.

  As shown in FIG. 2, upon receiving the learning sound data VC1 from the control unit 151, the learning processing unit 152 as an example of the feature amount reducing unit uses the cross validation to perform the learning sound data VC1 corresponding to the sample data. Is divided into k (k: a specified value of 2 or more) data blocks BLK1, BLK2, BLK3, BLK4, and BLK5. Here, for example, k = 5.

  Further, as the learning processing unit 152 as an example of the feature amount reducing unit divides the learning sound data VC1 into k data blocks BLK1 to BLK5, the learning processing unit 152 includes k types of patterns (in other words, the same number of divisions). Data sets having different roles are prepared and generated (preparation processing). The cross-validation means that, in statistics, sample data (for example, sound data for learning VC1) is divided into a predetermined number, a part of which is first analyzed (for example, learning processing), and the remaining part is analyzed (for example, learning processing). (For example, evaluation processing using learning results) to verify or confirm the validity of the analysis itself.

  Specifically, the learning processing unit 152 includes a data set of learning sound data of a first pattern, a data set of learning sound data of a second pattern, a data set of learning sound data of a third pattern, and a data set of learning sound data of a third pattern. And a data set of the learning sound data of the fifth pattern. The data set of the learning sound data of each of the first to fifth patterns has a different data block as a target of the learning processing and the evaluation processing in the learning processing unit 152.

  In the data set of the learning sound data of the first pattern, the data set EST1 corresponding to the data block BLK1 is used for the evaluation processing, and the data set LRN1 corresponding to the data blocks BLK2 to BLK5 is used for the learning processing.

  In the data set of the learning sound data of the second pattern, the data set EST2 corresponding to the data block BLK2 is used for the evaluation processing, and the data set LRN2 corresponding to the data blocks BLK1, BLK3 to BLK5 is used for the learning processing.

  In the data set of the learning sound data of the third pattern, the data set EST3 corresponding to the data block BLK3 is used for the evaluation processing, and the data set LRN3 corresponding to the data blocks BLK1, BLK2, BLK4, and BLK5 is used for the learning processing. You.

  In the data set of the learning sound data of the fourth pattern, the data set EST4 corresponding to the data block BLK4 is used for the evaluation processing, and the data set LRN4 corresponding to the data blocks BLK1 to BLK3 and BLK5 is used for the learning processing.

  In the data set of the learning sound data of the fifth pattern, the data set EST5 corresponding to the data block BLK5 is used for the evaluation processing, and the data set LRN5 corresponding to the data blocks BLK1 to BLK4 is used for the learning processing.

  As illustrated in FIG. 3, the learning processing unit 152 as an example of the generation unit employs an evolutionary algorithm (for example, a genetic algorithm) in order to reduce a feature amount used for the learning process of the learning sound data VC1. Generating a first generation of a plurality (for example, N: an integer of 2 or more) different individuals 1, individuals 2,..., Individuals N (one mode of the feature amount reduction block) in the genetic algorithm (for example, step St5 in FIG. 4) reference).

  Evolutionary Algorithm (EA) refers to a field of evolutionary computation, which is a part of Artificial Intelligent (AI), and is a general term for a population-based metaheuristic optimization algorithm. is there. For example, an evolutionary algorithm is a computational algorithm inspired by an evolutionary mechanism such as reproduction, mutation, genetic recombination, natural selection, and survival of the fittest. The evolutionary algorithm is not limited to the genetic algorithm described above, and genetic programming or evolutionary programming may be employed instead of the genetic algorithm. Genetic algorithms use a sequence of numbers (eg, binary numbers) to search for a solution to the problem, and generate individuals one after another in addition to selection and mutation. Genetic programming is basically the same as genetic algorithm, but the solution is expressed in the form of a tree structure, and expresses mathematical formulas and program codes. Evolutionary programming uses a stochastic function that represents the superiority of the solution in the population as the fitness function of the solution.

  Each individual has a bit size consisting of the same number of bit strings as a predetermined number of feature quantities (eg, 988 dimensions), and is configured as a number sequence in which, for example, binary parameters “0” and “1” are randomly arranged. Is done. As described above, “1” is a parameter instructing the use of the feature amount. “0” is a parameter instructing non-use of the feature amount. For example, the feature amount of the learning sound data VC1 is “loudness, intensity, MFCC, ZCR, F0,...”, And the individual is composed of a bit string of “1, 0, 1, 1, 0,. In this case, at least the feature amount of “loudness, MFCC, ZCR” is used, and at least the feature amount of “intensity, F0” is not used. It is assumed that the arrangement of the feature amounts (that is, the arrangement of the above-described “loudness, intensity, MFCC, ZCR, F0,...”) Is predetermined.

  The learning processing unit 152, which is an example of the feature amount reducing unit, first obtains a data set LRN1 (that is, the data blocks BLK2 to BLN2) used from the first generation individual 1 for the learning process in the learning sound data of the first pattern. The bit string of the ordinal part corresponding to BLK5) is extracted. The learning processing unit 152 as an example of the feature amount reducing unit corresponds to each parameter (that is, “0”, “1”) constituting the extracted bit string and the data blocks BLK2 to BLK5 used in the learning processing. By performing an AND operation with each of the feature amounts, the feature amounts used in the learning processing are reduced. Then, the learning processing unit 152 as an example of the feature amount reducing unit executes the learning process of the classifier (that is, the algorithm for identification) in the learning processing unit 152 using the reduced feature amount. After the learning processing, the learning processing unit 152, which is an example of the feature amount reducing unit, uses the classifier after the learning processing to use the data set EST1 used for the evaluation processing among the learning sound data of the first pattern described above. The evaluation process (that is, the data block BLK1) is executed, and a score (for example, “0.6”, see FIG. 3) as an evaluation result (one mode of the feature amount reduction process result) is output.

  Next, the learning processing unit 152, which is an example of the feature amount reducing unit, obtains a data set LRN2 (that is, a data block BLK1, BLK3 to BLK5) are extracted. The learning processing unit 152, which is an example of the feature amount reduction unit, assigns the parameters (that is, “0” and “1”) constituting the extracted bit string and the data blocks BLK1, BLK3 to BLK5 used in the learning process. By performing an AND operation with each of the corresponding feature amounts, the feature amounts used in the learning processing are reduced. Then, the learning processing unit 152 as an example of the feature amount reducing unit executes the learning process of the classifier (that is, the algorithm for identification) in the learning processing unit 152 using the reduced feature amount. After the learning process, the learning processing unit 152, which is an example of the feature amount reducing unit, uses the classifier after the learning process to use the data set EST2 used for the evaluation process among the learning sound data of the second pattern described above. (Ie, the data block BLK2) is evaluated, and a score (for example, “0.7”, see FIG. 3) as an evaluation result is output. The learning processing unit 152 repeats the same processing for the learning sound data of the third pattern and the fourth pattern.

  Similarly, the learning processing unit 152, which is an example of the feature amount reduction unit, first obtains a data set LRN5 (that is, a data set LRN5 used for learning processing from among the learning sound data of the fifth pattern described above from the first generation individual 1). The bit string of the ordinal part corresponding to the data blocks BLK1 to BLK4) is extracted. The learning processing unit 152 as an example of the feature amount reduction unit corresponds to each parameter (that is, “0”, “1”) constituting the extracted bit string and the data blocks BLK1 to BLK4 used in the learning processing. By performing an AND operation with each of the feature amounts, the feature amounts used in the learning processing are reduced. Then, the learning processing unit 152 as an example of the feature amount reducing unit executes the learning process of the classifier (that is, the algorithm for identification) in the learning processing unit 152 using the reduced feature amount. After the learning process, the learning processing unit 152 as an example of the feature amount reducing unit uses the discriminator after the learning process, and uses the data set EST5 used for the evaluation process among the learning sound data of the fifth pattern described above. The evaluation process (that is, data block BLK5) is executed, and a score (for example, “0.6”, see FIG. 3) is output as the evaluation result.

  The learning processing unit 152 performs a total of five scores (for example, “0.6”, “0.7”, “0.7”, “0”) by learning processing and evaluation processing using the first generation individual 1. .5 "," 0.6 "), and outputs the average value (AVE)" 0.62 "as the final evaluation score of the first generation individual 1.

  In addition, the learning processing unit 152 similarly executes the learning process and the evaluation process using the individual 1 described above for each individual from the first generation individual 2 to the individual N. In the description of FIG. 2 and FIG. 3, the procedure of the learning process and the evaluation process by the learning processing unit 152 using each individual is the same, and thus the detailed description of the procedure is omitted.

  The learning processing unit 152 performs a total of five scores (for example, “0.4”, “0.4”, “0.5”) by performing learning processing and evaluation processing using the first-generation individual 2 (see above). , “0.8”, “0.4”), and outputs the average value (AVE) “0.50” as the final evaluation score of the first generation individual 2.

  Similarly, the learning processing unit 152 performs a total of five scores (for example, “0.6”, “0.4”, “0.4”) by performing learning processing and evaluation processing using the first-generation individual N (see above). 0.6 ”,“ 0.8 ”,“ 0.5 ”), and outputs the average value (AVE)“ 0.58 ”as the final evaluation score of the first generation individual N.

  Next, the learning processing unit 152 randomly selects two individuals from the first generation N individuals according to a probability distribution according to the final evaluation score of each of the first generation individuals 1 to N, By performing an AND operation on parameters (elements) constituting each individual, individuals of the next generation (for example, the second generation) are generated one after another, and N are generated similarly to the first generation. For example, as shown in FIG. 3, the learning processing unit 152 selects the first generation individual 1 and the individual 2 and generates the second generation individual 1. The learning processing unit 152 generates the second generation individual 2 by selecting the first generation individual 2 and the individual N. Similarly, the learning processing unit 152 selects the first generation individual 1 and the individual N, and generates the second generation individual N. Thus, the learning processing unit 152 generates N individuals of the second generation by the genetic algorithm.

  Thereafter, the learning processing unit 152 executes the same learning and evaluation processing as the learning and evaluation processing using each of the first generation N individuals for each of the N individuals of the second generation. I do.

  The learning processing unit 152 performs a total of five scores (for example, “0.7”, “0.7”, “0.6”, “0”) by performing learning processing and evaluation processing using the second-generation individual 1. .5 "," 0.7 "), and outputs the average value (AVE)" 0.64 "as the final evaluation score of the second generation individual 1.

  The learning processing unit 152 performs a total of five scores (for example, “0.6”, “0.4”, “0.6”) by performing learning processing and evaluation processing (see above) using the second-generation individual 2. , “0.6”, “0.5”), and outputs the average value (AVE) “0.54” as the final evaluation score of the second generation individual 2.

  Similarly, the learning processing unit 152 performs a total of five scores (eg, “0.7”, “0.5”, “0.5”) by performing learning processing and evaluation processing using the second-generation individual N (see above). 0.7 ”,“ 0.7 ”,“ 0.6 ”), and outputs the average value (AVE)“ 0.62 ”as the final evaluation score of the second generation individual N.

  In addition, the learning processing unit 152 randomly selects two individuals from the second generation and subsequent N individuals according to the probability distribution according to the final evaluation score of each individual 1 to individual N from the second generation. Then, by performing an AND operation on the parameters (elements) constituting each individual, individuals of the next generation (for example, the third generation and thereafter) are generated one after another, and N are generated similarly to the first generation and the second generation. As described above, the learning processing unit 152 generates N individuals of the Mth generation set as the predetermined upper limit in the genetic algorithm. M is a predetermined integer of 3 or more. Similarly, the learning processing unit 152 obtains the final evaluation score of each individual by executing learning processing and evaluation processing using each individual for each of the N individuals of the M-th generation. I do.

  The learning processing unit 152 as an example of the learning control unit selects an individual having the best evaluation among the final evaluation scores of the individuals of the first to Mth generations, and selects the individual (that is, the use of the feature amount). Is output as reduction information.

  Next, an overall operation procedure of the sound data learning process by the sound data learning system 200 according to the first embodiment will be described with reference to FIG. FIG. 4 is a flowchart illustrating an example of an overall operation procedure of the sound data learning process. The processing illustrated in FIG. 4 is mainly executed by the processing unit 142 of the information processing device 140.

  In FIG. 4, the information processing apparatus 140 inputs and acquires the learning sound data collected by the microphone 110 in the processing unit 142 via the audio interface 120 (St1). The information processing apparatus 140 uses the learning processing unit 152 to extract feature amount data having a known number of dimensions (for example, 988 dimensions) from the learning sound data acquired in step St1 (St2).

  The information processing device 140 determines, in the learning processing unit 152, which generation of the genetic algorithm has been reached, with reference to, for example, information stored in the RAM of the storage unit 143 (St3). If the information processing apparatus 140 determines that the current variable does not reach the Mth generation of the genetic algorithm (St3, NO), the information processing apparatus 140 learns the variable i = 1 if the current variable i is not “1”. The information set in the processing unit 152 and indicating that the current time is the first generation is temporarily stored in the RAM of the storage unit 143 in the processing unit 142 (St4).

  After step St4, the information processing device 140 performs a process for reducing the feature amount used for the learning process of the classifier (that is, the identification algorithm) in the learning processing unit 152 (that is, the feature amount reduction process) in the genetic algorithm. A plurality of individuals (for example, individuals 1 to N) of the i-th generation (for example, the first generation) are generated in the learning processing unit 152 (St5, see FIG. 3).

  The information processing apparatus 140 uses a cross-validation-based classifier (that is, a discrimination algorithm) in the learning processing unit 152 using each of a plurality of individuals of the i-th generation (for example, the first generation) generated in step St5. The learning process and the evaluation process are performed by the learning processing unit 152 (St6). Since the details of steps St5 and St6 have been described with reference to FIGS. 2 and 3, detailed description thereof will be omitted here.

  The information processing device 140 selects, in the learning processing unit 152, the individual who has obtained the best final evaluation score among the final evaluation scores of the individual 1 to individual N of the i-th generation (for example, the first generation) ( St7).

  The information processing device 140 causes the processing unit 142 to determine whether or not the final evaluation score of the individual selected in step St7 is equal to or greater than a predetermined value (St8). If the information processing device 140 determines that the final evaluation score of the individual selected in step St7 is equal to or more than the predetermined value (St8, YES), the information processing apparatus 140 determines the individual corresponding to the final evaluation score equal to or more than the predetermined value. The information is stored in the storage unit 143 as reduction information and output (St9).

  On the other hand, when it is determined in step St7 that the final evaluation score of the selected individual is not greater than or equal to the predetermined value (St8, NO), the processing of the information processing device 140 returns to step St3.

  On the other hand, when the information processing device 140 determines that the current time has reached the Mth generation of the genetic algorithm (St3, YES), the information processing device 140 has a plurality of individuals 1 to N in each of the first to Mth generations. In the learning processing unit 152, the individual that has obtained the best final evaluation score among the final evaluation scores obtained by the above is selected (St10). The information processing device 140 saves and outputs the individual selected in step St10 as the reduction information in the storage unit 143 (St9).

  Next, an operation procedure of a learning process using the reduced feature amount by the sound data learning system 200 according to the first embodiment will be described with reference to FIG. FIG. 5 is a flowchart illustrating an example of an operation procedure of a learning process using the reduced feature amount. The processing illustrated in FIG. 5 is mainly executed by the learning processing unit 152 of the information processing device 140 after all the processing illustrated in FIG.

  5, the information processing apparatus 140 inputs the learning sound data collected by the microphone 110 at a time different from the time of step St1 shown in FIG. Acquire (St11). The information processing device 140 executes a predetermined data extension process on the learning sound data acquired in step St1 (St12). The data extension process executed in step St12 is, for example, at least one of the following six types, but is not limited thereto.

  The learning processing unit 152 adds an echo component to the learning sound data as a first data extension process. Accordingly, the learning processing unit 152 can generate learning sound data that contributes to generation of a highly accurate learned model, as compared to a case where the first data expansion processing is not performed.

  The learning processing unit 152 adds a predetermined noise component (e.g., equivalent to an indoor environment) in addition to the first data expansion process as the second data expansion process. Thereby, the learning processing unit 152 can generate learning sound data that contributes to generation of a highly accurate learned model, as compared with the case where the first data expansion processing is performed.

  The learning processing unit 152 reduces the speech speed (that is, the speed of the voice of a person) to 90% in addition to the second data extension process as the third data extension process. Thereby, the learning processing unit 152 can generate learning sound data that contributes to generation of a highly accurate learned model, as compared with the case where the first or second data expansion processing is executed.

  The learning processing unit 152 increases the speech speed (that is, the speed of the voice of a person) to 110% in addition to the third data extension process as the fourth data extension process. Accordingly, the learning processing unit 152 can generate learning sound data that contributes to generation of a highly accurate learned model, as compared with the case where any one of the first to third data expansion processes is executed.

  The learning processing unit 152 adds white noise (for example, the SN ratio is 24 dB) in addition to the fourth data expansion process as the fifth data expansion process. Accordingly, the learning processing unit 152 can generate learning sound data that contributes to generation of a highly accurate learned model, as compared with a case where any one of the first to fourth data expansion processes is executed.

  The learning processing unit 152 adds white noise (for example, the SN ratio is 36 dB) in addition to the fifth data expansion process as the sixth data expansion process. Accordingly, the learning processing unit 152 can generate learning sound data that contributes to generation of a highly accurate learned model, as compared with a case where any one of the first to fifth data expansion processes is executed.

  The information processing device 140 reads out the reduction information (see Step St9 in FIG. 4) stored in the storage unit 143 and inputs the information to the learning processing unit 152 (St13).

  The information processing apparatus 140 uses the learning sound data subjected to the data expansion processing in step St12 and the reduction information input in step St13 to convert the predetermined feature amount array (see above) into the reduction information (individual). The learning processing unit 152 extracts only the item of the feature amount corresponding to the parameter “1”. The information processing device 140 causes the learning processing unit 152 to acquire feature amount data corresponding to the extracted item (St14).

  The information processing device 140 uses the data of the reduced feature amount acquired in Step St14 to execute the learning processing of the discriminator (that is, the algorithm for identification) in the learning processing unit 152 in the learning processing unit 152 (St15). . The information processing apparatus 140 generates or updates a learned model, which is a classifier (that is, an algorithm for identification), as a result of the learning processing in step St15, and stores the model in the storage unit 143 (St16).

  Next, an operation procedure of sound data analysis processing by the sound data learning system 200 according to the first embodiment will be described with reference to FIGS. FIG. 6 is a flowchart illustrating an example of an operation procedure of sound data analysis processing using a learned model. FIG. 7 is a graph in which input sound data is associated with an example of an analysis result of the sound data. The processing illustrated in FIG. 6 is mainly executed by the detection processing unit 153 of the information processing device 140.

  6, the information processing apparatus 140 causes the detection processing unit 153 to read the learned model stored in the storage unit 143 (St21). The information processing device 140 inputs and acquires the detection sound data collected by the microphone 110 through the audio interface 120 in the processing unit 142 (St22).

  The information processing device 140 uses the learned model read out in Step St21 to execute analysis processing (identification processing) of the detection sound data acquired in Step St22 in the detection processing unit 153 (St23). The detection processing unit 153 determines, for example, the emotion of the person who uttered the voice corresponding to the detection sound data by analyzing the detection sound data (identification processing).

  The information processing apparatus 140 outputs the identification result (see FIG. 7) of the analysis processing (identification processing) in Step St23 to the display unit 145 (St24). The horizontal axis in FIG. 7 is time. The vertical axis of the upper graph in FIG. 7 indicates the magnitude (for example, sound pressure) of the sound data, and the vertical axis of the lower graph of FIG. 7 indicates a result obtained by normalizing the identification result. The upper graph in FIG. 7 is a voice waveform of the input detection sound data. For example, the subject had an utterance of “joy” while uttering the voice. The lower graph in FIG. 7 shows the results of the analysis processing (identification processing) performed by the detection processing unit 153. For example, “joy” (abbreviation for happy: happy), “normal” (abbreviation for nor: normal), “ It is a line graph which shows the degree of three kinds of emotions of "anger" (ang: abbreviation of angel).

  For example, as shown in FIG. 7, an erroneous detection of “anger” occurs instantaneously in the latter half of the utterance break, but since the identification result for the entire uttered voice is generally “joy”, It can be seen that highly accurate analysis results have been obtained. Note that the erroneously detected portion is a portion where the utterance is interrupted due to breathing or the like, and it can be considered that the cause of the erroneous detection is that there are fewer features included in the section where the utterance is clearly uttered.

  As described above, the sound data learning system 200 according to the first embodiment includes the microphone 110 that collects the learning sound signal and the extraction unit that extracts J (J: a positive integer) feature amounts of the learning sound signal. Having. The sound data learning system 200 includes N (N: an integer equal to or greater than 2) feature amount reduction blocks (for example, a genetic algorithm) configured by J parameters that specify whether or not each of the J feature amounts is used. Each generation unit has a generation unit that generates a gene (for example, an individual having a parameter of “0” or “1”) in each generation. The sound data learning system 200 prepares k (k: 2 or more specified values) types of different data sets of learning sound signal data, and performs N feature amount reduction blocks (for example, individuals) for each of N pieces of feature amount reduction blocks. For each different data set, a learning process using the corresponding feature amount reduction block and a partial data set of the learning sound signal data, and a learning sound signal data using the learning result of the learning process. A feature amount reduction unit that executes a feature amount reduction process of repeating the evaluation process regarding the use of each of the J feature amounts for the remaining data sets. The sound data learning system 200 reduces one of the feature amount reduction blocks to J feature amounts based on the feature amount reduction processing result for each of the N feature amount reduction blocks, each of which has k types of evaluation results. It has a learning control unit for selecting as information.

  Thereby, the information processing device 140 (an embodiment of the sound data learning device) of the sound data learning system 200 reduces the feature amount of the learning sound data when performing the learning process of the discriminator for the learning sound data. The processing load on the learning process can be reduced in that it is possible. In addition, the information processing apparatus 140 does not use all the high-dimensional (eg, 988-dimensional) feature amounts in the related art, but uses only the reduced feature amounts according to a specific use such as analyzing a human emotion. Can be used effectively and efficiently, and a highly accurate learned model can be generated. Therefore, according to the sound data learning system 200 or the information processing apparatus 140, it is possible to support both the reduction of the processing load imposed at the time of executing the learning process and the efficient and high-precision generation of the learned model according to the application. Becomes

  Further, the learning processing unit 152 of the information processing device 140 stores the selected feature amount reduction block in the storage unit 143 (an example of a memory) as reduction information. Thereby, the sound data learning system 200 or the information processing device 140 can store the feature amount reduction block selected by the feature amount reduction process, and thus the learning process for generating the learned model requires the learning sound data. Since the feature amount can be carefully selected, an efficient learned model can be generated.

  Further, the learning processing unit 152 of the information processing apparatus 140 determines that 1 ≦ u <J when u (u: a positive integer) is set as the number of parameters (one mode of the first parameter) indicating non-use of the feature amount. Are generated to satisfy the condition. Thus, the information processing apparatus 140 does not need to use all of the high-dimensional (eg, 988) dimensions required in the related art for the necessary feature amounts of the learning sound data in the learning process for generating the learned model. Thus, the processing load at the time of executing the learning process can be reduced.

  In addition, the learning processing unit 152 of the information processing apparatus 140 generates a genetic code according to a probability distribution based on the feature amount reduction processing result for each of the N feature amount reduction blocks corresponding to the current generation of the genetic algorithm used for the feature amount reduction processing. N feature amount reduction blocks corresponding to the next generation of the genetic algorithm are generated. Thereby, the information processing device 140 generates a feature amount reduction block (that is, an individual) in which genes evolving for each generation (that is, one of the parameters “0” and “1”) are listed using evolutionary programming. Therefore, it is possible to select individuals having excellent genes (that is, individuals capable of further reducing the feature amount used in the learning process).

  Also, the information processing apparatus 140 performs a reduction process of J feature amounts based on the reduction information, and performs a learning process using the feature amount reduced by the reduction process from the data of the learning sound signal. To generate a trained model. Accordingly, the information processing device 140 can efficiently generate a highly accurate learned model suitable for the analysis processing of human emotions.

  Further, the information processing device 140 performs a predetermined data extension process on the data of the learning sound signal. Thereby, the information processing apparatus 140 can generate a highly accurate learned model that is suitable for the analysis processing of human emotions relatively and efficiently as compared with the case where the data expansion processing is not executed.

  In addition, the information processing device 140 further includes a detection unit that detects information related to human emotion from data of the sound signal collected by the microphone 110 using the generated learned model. Accordingly, the information processing apparatus 140 uses the learned model generated using the learning sound data to perform a high-level analysis process of the human emotion included in the detection sound data collected in real time by the microphone 110. Can be performed with precision.

  In addition, the information processing device 140 may further include a detection unit that detects information related to the type of the mechanical sound from the data of the sound signal collected by the microphone 110 using the generated learned model. Accordingly, the information processing device 140 uses the learned model generated using the learning sound data to generate a mechanical sound (for example, installed in a factory or the like) included in the detection sound data collected by the microphone 110 in real time. Of the type (sound generated from the machine to be performed) can be analyzed with high accuracy.

(Modification of First Embodiment)
FIG. 8 is a block diagram showing a configuration example of a sound data learning system 200A according to a modification of the first embodiment. In a modification of the first embodiment, learning is performed by a server device 340 communicably connected to the information processing device 140A via a network or a communication line instead of the information processing device 140 according to the first embodiment. An example in which learning processing and evaluation processing using sound data for use, generation processing of a learned model, and detection processing of detection sound data will be described.

  The sound data learning system 200A is configured to include the microphone 110, the audio interface 120, the information processing device 140A, and the server device 340. The information processing device 140A includes a communication unit 141, a processing unit 142A, a storage unit 143, an operation input unit 144, a display unit 145, and a communication unit 146, and the processing unit 142A has a function of the control unit 151. Have. The communication unit 146 has a wired or wireless communication interface and communicates with an external server device 340. The information processing device 140A is connected to the server device 340 via a communication path 300 such as a wired or wireless network or a communication line. In other respects, the configuration is the same as that of the sound data learning system 200 shown in FIG.

  The server device 340 is configured by an information processing device (computer) having a processor and a memory, and performs various processes related to learning processing and evaluation processing using learning sound data, generation processing of a learned model, and detection processing of detection sound data. Execute the process. The server device 340 includes a communication unit 341, a processing unit 342, and a storage unit 243. The communication unit 341 transmits and receives various data such as sound data and learning data to and from the information processing device 140A. The processing unit 342 has a processor such as a CPU (Central Processing Unit) and a DSP (Digital Signal Processor). The processing unit 342 executes processing according to a predetermined program, and realizes functions such as learning processing and evaluation processing using the above-described learning sound data, generation processing of a learned model, and detection processing of detection sound data. The processing unit 342 includes, as functional components, a control unit 351 for performing various controls, a learning processing unit 352 for performing learning processing, a detection processing unit 353 for performing detection processing, and a determination processing unit 354 for performing determination processing. . Here, the learning processing unit 352, the detection processing unit 353, and the determination processing unit 354 are the learning processing unit 152, the detection processing unit 153, and the determination processing unit 154 of the processing unit 142 of the information processing device 140 in the configuration example of FIG. The same processing is performed. Note that a part of the learning processing unit 352, the detection processing unit 353, and the determination processing unit 354 may be executed by the server device 340, and the rest may be executed by the processing unit 142A of the information processing device 140A.

  In the modified example of the first embodiment, the processing according to the first embodiment is configured to be executed in a distributed manner in a plurality of information processing apparatuses connected via a network or a communication line. In particular, the learning process and the evaluation process using the learning sound data, and the generation process of the learned model are performed using an information processing device such as the server device 340 having a high processing capability, so that a complicated algorithm operation or the like can be performed. It is easy to deal with high-speed processing. The processing by the learning processing unit, the detection processing unit, and the determination processing unit is appropriately assigned to each processing in a local information processing device connected to an audio interface or a remote information processing device connected through a communication path. May be performed. For example, each process according to the first embodiment can be executed by an appropriate information processing device according to various conditions such as a system configuration, a usage environment, a data processing algorithm, a data amount, a data characteristic, and an output mode. It is.

  Although various embodiments have been described with reference to the drawings, it is needless to say that the present disclosure is not limited to such examples. It is clear that those skilled in the art can conceive various changes, modifications, substitutions, additions, deletions, and equivalents within the scope of the claims. Naturally, it is understood that they belong to the technical scope of the present disclosure. Further, the components in the above-described various embodiments may be arbitrarily combined without departing from the spirit of the invention.

  In the above-described first embodiment or its modified example, a description will be given by exemplifying that the learned model is a learning model for analyzing the emotion of a person included in the learning sound data or the detection sound data. However, the present invention is not limited to analysis of human emotions. For example, the trained model indicates the type (type) of the machine sound generated from a machine installed in a factory or the like, or the type of the tapping sound when a person hits with a hammer or the like, included in the learning sound data or the detection sound data. A learning model used for analysis may be used.

  For example, a power spectrum is an example of a feature amount for analyzing a mechanical sound or a tapping sound. For example, when the learning sound data or the detection sound data input from the audio interface 120 is sampled at a frequency of 48000 Hz, the information processing apparatus 140 obtains a 512-dimensional feature amount per 0.021 seconds (1024/48000). be able to. In order to reduce the feature amount effective for analyzing the operation sound of the machine sound from 512 dimensions, an evolutionary algorithm (for example, a genetic algorithm) is used in the sound data learning systems 200 and 200A according to the first embodiment or its modification. ) Can be similarly reduced.

  The present disclosure relates to a sound data learning system, a sound data learning method, and a sound data learning that support both reduction of the processing load imposed at the time of performing a learning process and efficient and high-precision generation of a learned model according to the application. Useful as a device.

110 microphone 120 audio interface 121 input unit 122 AD converter 123 buffer 124, 141 communication unit 140 information processing unit 142 processing unit 143 storage unit 144 operation input unit 145 display unit 151 control unit 152 learning processing unit 153 detection processing unit 154 determination processing unit 200, 200A sound data learning system

Claims (10)

A microphone that picks up the learning sound signal,
An extracting unit for extracting J (J: a positive integer) feature amounts of the learning sound signal;
A generation unit configured to generate N (N: an integer equal to or greater than 2) feature amount reduction blocks each including J parameters that define whether or not each of the J feature amounts is used;
A preparation process of preparing data sets of k (k: 2 or more specified values) different patterns using the data of the learning sound signal;
For each of the N data sets of the different patterns for each of the N feature amount reduction blocks, learning using the corresponding feature amount reduction block and a partial data set of the data of the learning sound signal. Processing, using the learning result of the learning processing, evaluation processing on the use of each of the J feature amounts for the remaining data set of the data of the learning sound signal,
A feature amount reduction unit that performs a feature amount reduction process that repeats
One of the feature amount reduction blocks is selected as J feature amount reduction information based on the feature amount reduction processing results for each of the N feature amount reduction blocks, each of which has k types of evaluation results. A learning control unit;
Sound data learning system. The learning control unit stores the selected feature amount reduction block in a memory as the reduction information.
The sound data learning system according to claim 1. When the number of first parameters indicating non-use of the feature amount is u (u: a positive integer), the generation unit generates N feature amount reduction blocks so as to satisfy 1 ≦ u <J. Generate,
The sound data learning system according to claim 1. The generation unit generates a next generation of the genetic algorithm according to a probability distribution based on a feature amount reduction processing result for each of the N feature amount reduction blocks corresponding to a current generation of the genetic algorithm used for the feature amount reduction processing. Generating N feature amount reduction blocks corresponding to
The sound data learning system according to claim 1. Performing a reduction process of J feature amounts based on the reduction information, and performing the learning process using the feature amount reduced by the reduction process from the data of the learning sound signal; A learning unit that generates a learned model by using
The sound data learning system according to claim 1. The learning unit performs a predetermined data extension process on the data of the learning sound signal,
The sound data learning system according to claim 5. Using the generated learned model, further comprising a detection unit that detects information related to human emotions from data of a sound signal collected by the microphone,
The sound data learning system according to claim 5. Using the generated learned model, further comprising a detection unit that detects information about the type of mechanical sound from the data of the sound signal collected by the microphone,
The sound data learning system according to claim 5. A sound data learning method in a sound data learning system having a microphone,
Collecting the learning sound signal by the microphone;
Extracting J (J: positive integer) feature amounts of the learning sound signal;
Generating N (N: an integer equal to or greater than 2) feature amount reduction blocks, each of which includes J parameters defining whether or not each of the J feature amounts is used;
A preparation process of preparing data sets of k (k: 2 or more specified values) different patterns using the data of the learning sound signal;
For each of the N data sets of the different patterns for each of the N feature amount reduction blocks, learning using the corresponding feature amount reduction block and a partial data set of the data of the learning sound signal. Processing, using the learning result of the learning processing, evaluation processing on the use of each of the J feature amounts for the remaining data set of the data of the learning sound signal,
Performing a feature amount reduction process of repeating
One of the feature amount reduction blocks is selected as J feature amount reduction information based on the feature amount reduction processing results for each of the N feature amount reduction blocks, each of which has k types of evaluation results. Having a step and
Sound data learning method. Connected to a microphone that picks up the learning sound signal,
An extracting unit for extracting J (J: a positive integer) feature amounts of the learning sound signal;
A generation unit configured to generate N (N: an integer equal to or greater than 2) feature amount reduction blocks each including J parameters that define whether or not each of the J feature amounts is used;
A preparation process of preparing data sets of k (k: 2 or more specified values) different patterns using the data of the learning sound signal;
For each of the N data sets of the different patterns for each of the N feature amount reduction blocks, learning using the corresponding feature amount reduction block and a partial data set of the data of the learning sound signal. Processing, using the learning result of the learning processing, evaluation processing on the use of each of the J feature amounts for the remaining data set of the data of the learning sound signal,
A feature amount reduction unit that performs a feature amount reduction process that repeats
One of the feature amount reduction blocks is selected as J feature amount reduction information based on the feature amount reduction processing results for each of the N feature amount reduction blocks, each of which has k types of evaluation results. A learning control unit;
Sound data learning device.

Images