JP2008181294A - Information processing apparatus, method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To more surely extract desired features from data. <P>SOLUTION: A next generation creation part 123 creates the several candidates of a filter for pre-processing to be used for pre-processing to be performed to data as the object of feature extraction in the case of extracting predetermined features from data. A gene evaluation part 122 evaluates the candidates of the filter for pre-processing, and a next generation creation part 123 creates new candidates by using the several candidates as the object of evaluation until predetermined conditions are satisfied by the evaluation value of the most highly evaluated candidate. When the predetermined conditions are satisfied by the most highly evaluated candidate, the gene evaluation part 122 outputs the candidate as a filter for pre-processing. This invention is applicable to an information processor. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an information processing device and method, and a program, and more particularly, to an information processing device and method, and a program that can extract a desired feature from data more reliably.

  Conventionally, a signal processing device that extracts a predetermined feature from input data is known (see, for example, Patent Document 1). An algorithm for such a signal processing device to extract features from data is constructed by a dedicated device or a human.

  In particular, when algorithms for extracting features are constructed by humans, humans can make use of foresight knowledge, that is, experience gained so far, so optimization can be performed using teacher data to improve extraction accuracy. It is possible to construct a simple algorithm, or to construct an algorithm that can extract features efficiently even if accuracy is low. Here, the teacher data refers to data necessary for constructing an algorithm, such as data to be extracted from features input to the signal processing apparatus and features extracted from the data.

  For example, as a method of improving accuracy using teacher data, a method of optimizing parameters used in an algorithm for extracting features by using GA (Genetic Algorithm) is known.

US Patent Application Publication No. 2004/0181401 A1

  However, in the above-described technique, it is not easy to reliably extract a desired feature from data. For example, when a feature extraction algorithm is constructed by a human, the accuracy of feature extraction by the algorithm depends greatly on the sense of the person who constructs the algorithm, and the work to improve the accuracy is a trial and error process. The construction of the algorithm required a lot of effort.

  Further, although the parameters used in the feature extraction algorithm can be optimized by GA, there is a limit to the parameter optimization, and thus the accuracy cannot be sufficiently improved.

  Furthermore, when an algorithm is constructed by a dedicated device, a huge amount of teachers are required to construct the algorithm when complicated processing is required for feature extraction because the device cannot use human foresight knowledge. In some cases, data is required, or even if a large amount of teacher data is used, the device cannot construct an algorithm.

  The present invention has been made in view of such a situation, and makes it possible to extract a desired feature from data more reliably.

  An information processing apparatus according to an aspect of the present invention is a candidate for a preprocessing filter used for preprocessing applied to the data, which is a target of feature extraction, when a predetermined feature is extracted from data. From the filter generation means for generating a candidate filter, the preprocessing means for performing the preprocessing on the teacher data used for the evaluation of the candidate filter using the candidate filter, and the teacher data subjected to the preprocessing An evaluation means for extracting the feature, and an evaluation value of the candidate filter indicating an evaluation of the extraction of the feature when the feature is extracted from the teacher data that has been preprocessed using the candidate filter. And an evaluation means for calculating on the basis of the feature extracted by the extraction means and a feature to be extracted from the teacher data obtained in advance.

  The filter generation means further generates a plurality of new candidate filters using some of the candidate filters based on each of the evaluation values of the plurality of candidate filters, and the evaluation means Each of the evaluation values of the new candidate filter can be calculated.

  In the evaluation means, when the highest evaluation value among the evaluation values calculated at the end satisfies a predetermined condition, the candidate filter having the highest evaluation value is used as a preprocessing filter for the preprocessing. Can be output as

  When the highest evaluation value among the evaluation values calculated last does not satisfy the condition, the filter generation means, based on the evaluation value calculated last, the last generated Several of the candidate filters can be used to further generate a plurality of new candidate filters.

  The preprocessing filter may be a filter that performs processing while maintaining the data format.

  The information processing apparatus uses the candidate filter based on the teacher data that has been subjected to the preprocessing using the candidate filter and the characteristics that should be extracted from the teacher data that has been obtained in advance. The discriminator creating means for creating a discriminator for extracting the feature from the preprocessed data by machine learning can be further provided.

  An information processing method or program according to one aspect of the present invention provides a candidate for a preprocessing filter that is used for preprocessing performed on the data that is a target of extraction of a feature when a predetermined feature is extracted from the data. The candidate filter is generated, the preprocessing is performed on the teacher data used for the evaluation of the candidate filter using the candidate filter, and the feature is extracted by the extraction unit from the teacher data subjected to the preprocessing. When the feature is extracted from the teacher data that has been preprocessed using the candidate filter, the extraction means extracts the evaluation value of the candidate filter that indicates the evaluation of the feature extraction And calculating based on the obtained feature and the feature to be extracted from the teacher data obtained in advance.

  In one aspect of the present invention, when a predetermined feature is extracted from data, a candidate filter that is a candidate for a preprocessing filter used for preprocessing applied to the data to be extracted of the feature is provided. The candidate filter is generated, the preprocessing is performed on the teacher data used for the evaluation of the candidate filter, and the feature is extracted from the teacher data subjected to the preprocessing by the extraction unit In the case where the feature is extracted from the teacher data that has been preprocessed using the candidate filter, the evaluation value of the candidate filter indicating the evaluation of the extraction of the feature is extracted by the extraction unit The calculation is performed based on the characteristics and the characteristics to be extracted from the teacher data obtained in advance.

  According to one aspect of the present invention, a preprocessing filter used for preprocessing for extracting features can be generated. In particular, according to one aspect of the present invention, a desired feature can be more reliably extracted from data.

  Embodiments of the present invention will be described below. Correspondences between the constituent elements of the present invention and the embodiments described in the specification or the drawings are exemplified as follows. This description is intended to confirm that the embodiments supporting the present invention are described in the specification or the drawings. Therefore, even if there is an embodiment which is described in the specification or the drawings but is not described here as an embodiment corresponding to the constituent elements of the present invention, that is not the case. It does not mean that the form does not correspond to the constituent requirements. Conversely, even if an embodiment is described here as corresponding to a configuration requirement, that means that the embodiment does not correspond to a configuration requirement other than the configuration requirement. It's not something to do.

  An information processing apparatus according to an aspect of the present invention is a candidate for a preprocessing filter used for preprocessing applied to the data, which is a target of feature extraction, when a predetermined feature is extracted from data. Using the filter generation means (for example, the initial generation generation unit 121 and the next generation generation unit 123 in FIG. 10) for generating the candidate filter and the candidate filter, the preprocessing is performed on the teacher data used for the evaluation of the candidate filter. Preprocessing means (for example, the preprocessing unit 143 in FIG. 10), extraction means (for example, the code determination unit 146 in FIG. 10) for extracting the features from the teacher data subjected to the preprocessing, When the feature is extracted from the teacher data that has been preprocessed using the candidate filter, an evaluation value of the candidate filter that indicates an evaluation of the feature extraction Comprising said features extracted by the extraction means, an evaluation means for calculating based on the features to be extracted from the teacher data obtained in advance (for example, the evaluation unit 147 of FIG. 10).

  The filter generation means further generates a plurality of new candidate filters using some of the candidate filters based on each of the evaluation values of the plurality of candidate filters (for example, FIG. 13 The process of step S44 to the process of step S47) can cause the evaluation means to calculate each evaluation value of the new candidate filter (for example, the process of step S97 in FIG. 14).

  In the evaluation means, when the highest evaluation value among the evaluation values calculated at the end satisfies a predetermined condition, the candidate filter having the highest evaluation value is used as a preprocessing filter for the preprocessing. (For example, the process in step S48 in FIG. 13).

  When the highest evaluation value among the evaluation values calculated last does not satisfy the condition, the filter generation means, based on the evaluation value calculated last, the last generated A plurality of new candidate filters can be further generated by using some of the candidate filters (for example, the processing from step S44 to step S47 in FIG. 13).

  The information processing apparatus uses the candidate filter based on the teacher data that has been subjected to the preprocessing using the candidate filter and the characteristics that should be extracted from the teacher data that has been obtained in advance. A discriminator creating means (for example, a code discriminator learning unit 145 in FIG. 10) for creating a discriminator for extracting the feature from the preprocessed data by machine learning can be further provided.

  An information processing method or program according to one aspect of the present invention provides a candidate for a preprocessing filter that is used for preprocessing performed on the data that is a target of extraction of a feature when a predetermined feature is extracted from the data. Is generated (for example, step S41 in FIG. 13), and the preprocessing is performed on the teacher data used for evaluation of the candidate filter by using the candidate filter (for example, step S82 in FIG. 14). The feature is extracted from the teacher data that has been subjected to the pre-processing by an extraction means (for example, step S90 and step S91 in FIG. 14), and the teacher data that has been pre-processed using the candidate filter In the case where the feature is extracted from, the evaluation value of the candidate filter indicating the evaluation of the feature extraction is extracted before the extraction unit extracts the evaluation value. Including features and is calculated based on the feature to be extracted from the teacher data obtained in advance (e.g., step S97 of FIG. 14) the step.

  Embodiments to which the present invention is applied will be described below with reference to the drawings.

  In the feature extraction system according to an embodiment to which the present invention is applied, as shown in FIG. 1, predetermined features are extracted from input data and output according to a constructed algorithm. This feature extraction system is configured, for example, as shown in FIG.

  The feature extraction system shown in FIG. 2 includes a filter generation device 11 and a signal processing device 12, and the signal processing device 12 includes a preprocessing unit 21 and a feature extraction processing unit 22.

  The filter generation device 11 generates a preprocessing filter composed of a combination of a filter that performs integration processing, a filter that performs predetermined processing such as a high-pass filter (hereinafter also referred to as a unit filter, as appropriate), and a signal processing device. 12 pre-processing units 21. Here, the preprocessing filter is a filter that performs processing while maintaining the format of the input data. Therefore, when preprocessing, which is filter processing using a preprocessing filter, is performed on predetermined data, data in the same format as before preprocessing is obtained.

  Further, the filter generation device 11 generates a discriminator that is data used for processing for extracting a desired feature from the data and supplies the discriminator to the feature extraction processing unit 22 of the signal processing device 12.

  The preprocessing unit 21 performs preprocessing on the input data using the preprocessing filter supplied from the filter generation device 11 and supplies the preprocessed data to the feature extraction processing unit 22. The feature extraction processing unit 22 extracts and outputs a predetermined feature from the data supplied from the preprocessing unit 21 using the discriminator supplied from the filter generation device 11.

  In this way, by applying pre-processing to the input data, components that are useful for feature extraction in the input data are further emphasized, or only unnecessary components, for example, components that adversely affect feature extraction are included. Therefore, it is possible to more reliably extract a desired feature from input data.

  More specifically, for example, in the feature extraction system, the data obtained by analyzing the waveform of music, etc., into two axes of energy for each time and pitch, that is, voice data for reproducing the music, Time-pitch data obtained by performing processing for analyzing the waveform into energy for each time and pitch can be input data, and extracted features can be music chords, that is, chords. Hereinafter, processing for analyzing a waveform into energy for each time and pitch will be referred to as twelve sound analysis processing.

  In the feature extraction system, when the input data is time-pitch data and the extracted feature is a music code, the feature extraction system is configured as shown in FIG.

  The feature extraction system shown in FIG. 3 includes a filter generation device 51 and a signal processing device 52, and the signal processing device 52 includes a preprocessing unit 61 and a code extraction unit 62. Here, the filter generation device 51 of FIG. 3 corresponds to the filter generation device 11 of FIG. 2, and the signal processing device 52 of FIG. 3 corresponds to the signal processing device 12 of FIG. Further, the preprocessing unit 61 and the code extraction unit 62 in FIG. 3 correspond to the preprocessing unit 21 and the feature extraction processing unit 22 in FIG.

  The filter processing device 51 generates a preprocessing filter by combining unit filters, and supplies the preprocessing filter to the preprocessing unit 61 of the signal processing device 52. Further, the filter generation device 51 generates a code discriminator used for extracting a chord from the pre-processed time-pitch data, and the generated code discriminator is used as the code extraction unit 62 of the signal processing device 52. To supply.

  The preprocessing unit 61 of the signal processing device 52 performs preprocessing on the input time-pitch data using the preprocessing filter supplied from the filter processing device 51, and supplies it to the chord extraction unit 62. The chord extraction unit 62 uses the chord discriminator supplied from the filter processing device 51 to extract and output the chord of the music from the time-pitch data supplied from the preprocessing unit 61.

  Thus, in the feature extraction system, the chord progression of the music can be obtained as the output feature in response to the time-pitch data input.

  FIG. 4 is a block diagram showing a more detailed configuration example of the code extraction unit 62 of FIG.

  The chord extraction unit 62 includes a beat detection unit 91, a beat-by-beat feature amount extraction unit 92, and a chord determination unit 93.

  The beat detection unit 91 detects a beat from the time-pitch data supplied from the preprocessing unit 61. Here, the beat is a hit point or a beat, and is a reference that can be heard as a basic unit in music. The beat is generally used in a plurality of meanings, but hereinafter, it is used in the meaning of the time that is the beginning of the basic unit of time in the music. The time that is the beginning of the basic time unit in the music is referred to as the beat position, and the basic time unit range in the music is referred to as the beat range. The beat length is a so-called tempo.

  That is, the beat detection unit 91 detects the position of the beat in the time-pitch data from the time-pitch data obtained from the sound data of the music. Then, the beat detection unit 91 supplies beat information indicating each position of the beat in the time-pitch data and the input time-pitch data to the beat feature quantity extraction unit 92.

  Since the range of beats from the position of the beat in the time-pitch data to the next beat position is the range of beats, the range of beats can be known if the position of the beat in the time-pitch data is known.

  The beat feature quantity extraction unit 92 uses a time-pitch data supplied from the beat detection unit 91 to determine a voice feature quantity within a predetermined range, for example, a voice feature quantity for each beat (hereinafter, a chord discrimination feature for each beat). Extract). In other words, the beat-by-beat feature quantity extraction unit 92 calculates the feature quantities indicating the characteristics of the sounds at the respective pitches of the 12 average rate pitches in the respective beat ranges of the time-pitch data based on the beat information. Extract. The beat feature quantity extraction unit 92 supplies the extracted chord discrimination feature quantity for each beat to the chord discrimination unit 93.

  The chord discriminating unit 93 discriminates the chord for each beat from the chord discriminating feature amount supplied from the beat feature amount extracting unit 92 using the chord discriminator supplied from the filter generating device 51. Output. That is, the chord discriminating unit 93 discriminates a chord in the beat range from the chord discrimination feature value for each beat. As will be described later, the code discriminator is created in advance by learning based on feature amounts, and a new code discriminator is supplied from the filter generation device 51 to the code discriminating unit 93 every time the preprocessing filter is changed. .

  Thus, the signal processing device 52 determines the chord for each beat of the music from the time-pitch data of the music. For example, as shown in FIG. 5, the signal processing device 52 determines a code that is C, a code that is E minor, a code that is A minor, a code that is C, and a code that is F from the time-pitch data of the music. , A chord that is E minor, a chord that is D minor, a chord that is G, and the like are determined for each beat. For example, the signal processing device 52 outputs the chord name of the chord for each beat.

  Next, with reference to the flowchart of FIG. 6, a chord determination process, which is a process in which the signal processing device 52 extracts and outputs chords for each beat from the input time-pitch data, will be described. This code discrimination process is started when time-pitch data is input to the signal processing device 52.

  In step S11, the pre-processing unit 61 performs pre-processing on the input time-pitch data using the pre-processing filter supplied from the filter generation device 51, and the pre-processed time-pitch data is processed. This is supplied to the beat detection unit 91.

  Here, the time-pitch data input to the preprocessing unit 61 is data indicating the energy of each sound at each time, for example, as shown in FIG. In FIG. 7, the vertical direction indicates the pitch, and the horizontal direction indicates time. Moreover, the density | concentration has shown the strength of energy.

  In FIG. 7, twelve sound energies having a height of 12 average rates in each of a plurality of octave components, that is, octaves 1 to 7 are shown as time-pitch data.

  The preprocessing unit 61 performs preprocessing using, for example, a preprocessing filter illustrated in FIG. 8A or 8B on the time-pitch data illustrated in FIG.

  In the preprocessing filter shown in FIG. 8A, the portion of A11, that is, the character “Sqr” represents that the energy of each sound at each time of the time-pitch data is squared, and the portion of A12, that is, the character “ Time # LPF, 0.5 "has a coefficient of 0.5 in the time direction of the time-pitch data squared by the portion of A11, that is, for the energy at each time of a predetermined sound in a predetermined octave of the time-pitch data. It represents that the filter processing using the low-pass filter is. Further, the A13 portion, that is, the characters “Freq # HPF, 0.3” is in the frequency direction of the time-pitch data filtered by the A12 portion, that is, in the energy of each sound at a predetermined time of the time-pitch data. On the other hand, filter processing using a high-pass filter having a coefficient of 0.3 is performed.

  These A11 to A13 portions indicate the respective unit filters constituting the preprocessing filter. The time-pitch data is indicated by the filter processing by the unit filter indicated by the A11 portion to the A13 portion. Filter processing by the unit filter is performed in order.

  Therefore, as preprocessing using the preprocessing filter shown in FIG. 8A, arithmetic processing for squaring a value indicating each energy, filter processing using a low-pass filter, and high-pass filter are used for time-pitch data. Filter processing is performed in order.

  In the preprocessing filter shown in FIG. 8B, the portion A21, that is, the character “Time # Differential” is in the time direction of the time-pitch data, that is, each time of a predetermined sound in a predetermined octave of the time-pitch data. The A22 part, that is, the character “Abs” represents that the absolute value of the energy of each sound at each time of the time-pitch data is obtained. Further, the part A23, that is, the character “Freq # Order” is used to order the energy of each sound at a predetermined time in the time-pitch data, that is, at a predetermined time in the time-pitch data from the smallest value. A24, that is, the characters “Time # LPF, 0.4” are in the time direction of the time-pitch data, that is, with respect to the energy at each time of a predetermined sound in a predetermined octave of the time-pitch data. , The filter processing using a low-pass filter having a coefficient of 0.4 is performed.

  These A21 to A24 portions indicate the respective unit filters constituting the preprocessing filter. The time-pitch data is indicated by the filter processing by the unit filter indicated by the A21 portion to the A24 portion. Filter processing by the unit filter is performed in order.

  Therefore, as preprocessing using the preprocessing filter shown in FIG. 8B, time-pitch data differential processing in the time direction, calculation processing for obtaining absolute values, processing for ordering, and filter processing using a low-pass filter Are applied in order.

  In addition, as a unit filter constituting the preprocessing filter, a filter for normalizing with an absolute value, a filter for normalizing in the time direction or the frequency direction, with respect to the time direction or the frequency direction A filter for performing normalization so that the average value becomes a predetermined value, a filter for performing normalization so that the average value and variance become predetermined values in the time direction or the frequency direction, etc. Can be used.

  Furthermore, as a unit filter constituting the preprocessing filter, a filter that performs arithmetic processing using a logarithmic function, an exponential function, etc., a filter that performs arithmetic processing for calculating a power, addition, subtraction, multiplication, division, differentiation, integration, etc. A filter that performs arithmetic processing, a filter that performs arithmetic processing to obtain an absolute value, a square, a square root, or the like, a bandpass filter, a filter that removes a DC component, a filter that performs processing for ordering or peak value detection, or the like is used. be able to.

  In the preprocessing using the preprocessing filter, the format of the data to be processed, that is, the time-pitch data is maintained before and after the processing. That is, the pre-processed time-pitch data is also data indicating the energy of each sound at each time, similarly to the pre-process time-pitch data.

  Returning to the description of the flowchart of FIG. 6, when pre-processing is performed on the time-pitch data, the beat detection unit 91 detects a beat from the time-pitch data supplied from the pre-processing unit 61 in step S <b> 12. That is, the beat detection unit 91 detects the position of the beat in the time-pitch data from the time-pitch data, and obtains beat information indicating each position of the beat in the time-pitch data and the time-pitch data for each beat. This is supplied to the feature quantity extraction unit 92.

  For example, the beat detection unit 91 obtains a difference in energy in the time direction as a change in energy of 12 sounds at respective heights of 12 average rates in each octave indicated by the time-pitch data, and calculates the sound in each time. The energy change is integrated for each of the twelve sounds of the plurality of octaves, and the result is used as attack information. Here, the attack information is data obtained by converting the change in volume that makes a human to feel a beat over time.

  The beat detection unit 91 detects the most basic sound length in the music that is the target of chord detection. That is, for example, the beat detection unit 91 considers attack information, which is time-series information, as a normal waveform, performs a process such as a short time Fourier transform on the attack information, and extracts a basic pitch (pitch) from the attack information. Find the most basic sound length. For example, the most basic sound in music is a sound represented by a quarter note, an eighth note, or a sixteenth note. Hereinafter, the most basic sound length in the music is referred to as a basic beat cycle.

  Furthermore, the beat detection unit 91 applies predetermined signal processing to the data of each of the twelve sounds in each octave indicated by the time-pitch data, thereby distributing the energy in the pitch direction and the number of peaks per unit time. And the tempo is estimated from the extracted music feature amount based on data obtained in advance by learning with the music feature amount and the tempo. Then, the beat detection unit 91 determines the final tempo using the estimated tempo and the basic beat cycle, and uses the determined tempo, that is, beat information indicating the position of the beat and time-pitch data for each beat. This is supplied to the feature quantity extraction unit 92.

  In step S <b> 13, the beat feature quantity extraction unit 92 extracts the chord determination feature quantity from each beat range of the time-pitch data based on the beat information indicating the beat position supplied from the beat detection unit 91. And supplied to the code discriminating unit 93.

  For example, the feature extraction unit 92 for each beat cuts out only the data of the beat range from a predetermined beat position to the next beat position based on the beat position indicated by the beat information from the time-pitch data. Then, the energy indicated by the data of the extracted beat range is averaged over time. Thereby, the energy for every 12 sounds of each height of 12 average rates in each octave is calculated | required.

  Further, the beat-by-beat feature quantity extraction unit 92 weights, for example, the energy of every 12 notes at the height of 12 average rates in each octave of 7 octaves, and the same pitch name in each octave of 7 octaves. The energy of each of the twelve sounds identified by the pitch names is obtained. The beat-by-beat feature quantity extraction unit 92 arranges the energy of each of the twelve sounds in the order of the pitch of the pitch name, and generates a chord discrimination feature quantity indicating the energy of the sounds in the order of the scale. That is, for example, the beat feature quantity extraction unit 92 calculates the energy of the sound of the pitch name C by adding the energy of the sound of the pitch name C in each octave out of the weighted energy.

  The beat feature quantity extraction unit 92 uses a feature quantity used for discriminating a route (hereinafter referred to as a route discrimination feature quantity) as a chord discrimination feature quantity for each beat from the beat range of the time-pitch data. And a feature amount used to determine whether the chord is a major chord or a minor chord (hereinafter referred to as a major / minor distinguishing feature amount). In addition, a weight for weighting sound energy, which is used when generating a feature value for route discrimination, and a weight for weighting sound energy, which is used when a feature amount for major / minor discrimination, is used. Is different.

  When the chord discrimination feature value is extracted for each beat, the chord discrimination unit 93 uses the chord discriminator supplied from the filter generation device 51 to perform chord discrimination supplied from the beat feature amount extraction unit 92 in step S14. The chord for each beat is determined from the feature amount.

  For example, as shown in FIG. 9, the code discriminating unit 93 uses a code discriminator to calculate a C major code, a C minor chord from a route discriminating feature amount and a major / minor discriminating feature amount as a chord discriminating feature amount. Minor code, C # major code, C # minor code, D major code, D minor code, D # major code, D # minor code, E major code, E minor code, F major Code, minor code of F, major code of F #, minor code of F #, major code of G, minor code of G, major code of G #, minor code of G #, major code of A, minor code of A , A # major code, A # minor code, B major code, and B minor code, A code indicating a correct chord range, to determine for each beat (range beats).

  In step S15, the chord discrimination unit 93 outputs the chord obtained by the discrimination as a chord for each beat, and the chord discrimination process ends. In more detail, the chord discriminating unit 93 outputs the chord name of the chord as a chord for each beat.

  In this way, the signal processing device 52 performs preprocessing on the input time-pitch data, and extracts and outputs a chord for each beat from the preprocessed time-pitch data.

  In this way, by applying pre-processing to the input time-pitch data, components useful for feature extraction in the time-pitch data are more emphasized, and components that hinder feature extraction are removed. Can do. This makes it possible to extract a desired feature more reliably from the pre-processed time-pitch data.

  Next, creation of a preprocessing filter and a code discriminator will be described. In the feature extraction system shown in FIG. 3, an algorithm for extracting a song code as a feature from time-pitch data is constructed by a human, but a part of the algorithm, for example, pre-processing is performed by a filter generation device 51. And GP (Genetic Programming).

  For example, a preprocessing filter used for preprocessing is generated by genetically evolving a combination of unit filters by GP, and a code discriminator corresponding to the preprocessing filter is generated by machine learning. .

  FIG. 10 is a block diagram illustrating a configuration example of a filter generation device 51 that creates a preprocessing filter and a code discriminator.

  The filter generation device 51 includes an initial generation generation unit 121, a gene evaluation unit 122, and a next generation generation unit 123.

  The initial generation generation unit 121 generates several preprocessing filters that are candidates for the preprocessing filter supplied to the preprocessing unit 61 of the signal processing device 52. In the following description, one preprocessing filter that is a candidate for the preprocessing filter supplied to the preprocessing unit 61 is referred to as a gene, and in particular, the gene generated by the initial generation generation unit 121 is referred to as the initial generation. The gene is called. The initial generation unit 121 generates several initial generation genes by combining unit filters, and supplies the genes to the gene evaluation unit 122.

  The gene evaluation unit 122 evaluates the gene supplied from the initial generation generation unit 121 and supplies the evaluation result to the next generation generation unit 123. In addition, the gene evaluation unit 122 evaluates the next generation gene supplied from the next generation generation unit 123, that is, the gene of the next generation of the gene of the generation that was last evaluated, and the result of the evaluation is used as the next generation gene. It supplies to the production | generation part 123. FIG. Then, the gene evaluation unit 122 repeatedly evaluates several generations of genes to select a final preprocessing filter, a selected preprocessing filter, and a code discriminator corresponding to the preprocessing filter Is supplied to the signal processing device 52.

  Here, the next-generation gene refers to a gene that is generated using the genes that are the targets of evaluation in the gene evaluation unit 122 as a reference. Therefore, for example, the next generation gene for the early generation, that is, the gene of the next generation of the early generation, is the gene to be evaluated next to the gene of the early generation generated using the gene of the early generation. The

  The gene evaluation unit 122 includes a teacher data holding unit 141, an analysis unit 142, a preprocessing unit 143, a teacher data division unit 144, a code discriminator learning unit 145, a code discrimination unit 146, and an evaluation unit 147.

  The teacher data holding unit 141 holds, as teacher data, voice data of music used to create a code discriminator and evaluate genes. The teacher data holding unit 141 supplies the held teacher data to the analyzing unit 142 and the teacher data dividing unit 144.

  The analysis unit 142 performs twelve-tone analysis processing on the teacher data supplied from the teacher data holding unit 141, and supplies the time-pitch data obtained as a result to the preprocessing unit 143. For each gene supplied from the initial generation generation unit 121 or the next generation generation unit 123, the preprocessing unit 143 applies each of the time-pitch data supplied from the analysis unit 142 using a preprocessing filter that is a gene. Pre-processing is performed, and the result is supplied to the code discriminator learning unit 145 and the code discriminating unit 146. Therefore, for example, when the gene G11 and the gene G12 are supplied to the preprocessing unit 143, the preprocessing unit 143 performs preprocessing using the gene G11 for each of all the time-pitch data supplied from the analysis unit 142. To the chord discriminator learning unit 145 and the chord discriminating unit 146, and the pre-processing using the gene G12 is performed on all of the time-pitch data supplied from the analyzing unit 142 to perform the chord discrimination. To the machine learning unit 145 and the code determination unit 146.

  The teacher data dividing unit 144 uses the teacher data supplied from the teacher data holding unit 141 for learning data that is teacher data used for learning by the code discriminator and evaluation data that is teacher data used for gene evaluation. Divide it randomly. Then, the teacher data dividing unit 144 supplies the teacher data set as learning data to the code discriminator learning unit 145 and supplies the teacher data set as evaluation data to the evaluation unit 147.

  The code discriminator learning unit 145 uses a machine learning based on the time-pitch data supplied from the preprocessing unit 143 and the learning data supplied from the teacher data dividing unit 144 for each gene, and the code corresponding to the gene. A discriminator is created and supplied to the code discriminating unit 146 and the evaluation unit 147.

  The chord discriminating unit 146 discriminates the chord for each beat from the time-pitch data supplied from the pre-processing unit 143 using the chord discriminator supplied from the chord discriminator learning unit 145, and the discrimination result is evaluated. 147. That is, the chord discriminating unit 146 performs pre-processing on the time-pitch data that has been pre-processed by using a predetermined gene among the time-pitch data supplied from the pre-processing unit 143. The chord is discriminated using a chord discriminator created by using the time-pitch data.

  The evaluation unit 147 uses the evaluation data supplied from the teacher data division unit 144 and the code discrimination result supplied from the code discrimination unit 146 to evaluate each gene for accuracy of code discrimination, The evaluation result is supplied to the next generation generation unit 123. In other words, since the correct code name that is the code name of the correct code for each beat of the music is known in advance for the audio data of the music that is the teacher data, the evaluation unit 147 is determined by the code discriminator. The gene is evaluated by comparing the code of the voice data corresponding to the evaluation data in the code with the correct code name known in advance.

  The evaluation unit 147 repeatedly evaluates several generations of genes, and the highest evaluation value among the evaluation values indicating the evaluation result of the gene of the last evaluated generation, that is, the evaluation value of the highest evaluation gene. When a predetermined condition such as being equal to or greater than a predetermined threshold value is satisfied, the gene having the highest evaluation value and the code discriminator corresponding to the gene are designated as the preprocessing unit 61 and the code of the signal processing device 52. It supplies to the discrimination | determination part 93.

  The next generation generation unit 123 supplies from the evaluation unit 147 using the gene to be evaluated, that is, the gene of the initial generation supplied from the initial generation generation unit 121 or the gene generated by the next generation generation unit 123. Based on the result of the evaluation of those genes, a next generation gene is generated and supplied to the preprocessing unit 143 and the evaluation unit 147.

  The next generation generation unit 123 includes a selection unit 151, a mutation processing unit 152, an intersection processing unit 153, and a random generation unit 154.

  The selection unit 151 of the next generation generation unit 123 uses the gene of the last generation evaluated by the evaluation unit 147 to select some genes with the highest evaluation among the genes of the generation evaluated last. And select the selected gene as the next generation gene.

  The mutation processing unit 152 generates a next-generation gene by performing mutation processing on some of the highly evaluated genes based on the evaluation result of the gene evaluated last by the evaluation unit 147. That is, for each of several highly evaluated genes, the mutation processing unit 152 converts some of the genes, that is, some of the unit filters constituting the genes, to other randomly selected unit filters. A new gene is generated by changing, and the generated gene is set as the next generation gene.

  Based on the result of the evaluation of the gene of the last generation evaluated by the evaluation unit 147, the cross processing unit 153 performs cross processing on some of the highly evaluated genes to generate the next generation gene. That is, for each of several highly evaluated genes, the intersection processing unit 153 replaces part of the genes, that is, some of the unit filters constituting the genes with unit filters constituting the other genes of the same generation. A new gene is generated by changing to, and the generated gene is used as the next generation gene.

  The random generation unit 154 generates a new gene (preprocessing filter) by combining randomly selected unit filters, and uses the generated gene as the next generation gene. The next generation generation unit 123 supplies the next generation gene generated by the selection unit 151 to the random generation unit 154 in this way to the evaluation unit 147 and the preprocessing unit 143.

  FIG. 11 is a block diagram illustrating a more detailed configuration example of the code discriminator learning unit 145 of FIG. The chord discriminator learning unit 145 includes a beat-by-beat feature quantity extracting unit 181, an adding unit 182, a shift processing unit 183, and a chord discriminator generating unit 184.

  The beat feature quantity extraction unit 181 detects a beat from the time-pitch data supplied from the preprocessing unit 143, extracts a chord discrimination feature quantity for each beat, and supplies it to the addition unit 182. The adding unit 182 adds the previously known correct code name to the teacher data, which is the learning data supplied from the teacher data dividing unit 144, and the chord determination feature amount supplied from the beat-by-beat feature amount extraction unit 181. Is added to the shift processing unit 183.

  The shift processing unit 183 shifts the chord discrimination feature amount added to (added to) the teacher data supplied from the adding unit 182 and the correct code name by one sound at a time, and adds them to the teacher data. It supplies to the production | generation part 184. The code discriminator generating unit 184 creates (generates) a code discriminator corresponding to each gene by machine learning from the teacher data as learning data supplied from the shift processing unit 183, and the code discriminating unit 146 and the evaluation unit 147.

  FIG. 12 is a block diagram illustrating a more detailed configuration example of the code determination unit 146 of FIG. The chord discriminating unit 146 includes a beat-by-beat feature amount extracting unit 211 and a discriminating unit 212.

  The beat feature quantity extraction unit 211 detects a beat from the time-pitch data supplied from the preprocessing unit 143, extracts a chord discrimination feature quantity for each beat, and supplies it to the discrimination unit 212. The discriminator 212 uses the chord discriminator supplied from the chord discriminator generation unit 184 of the chord discriminator learning unit 145, and uses the chord discriminating feature amount supplied from the beat feature amount extracting unit 211 to calculate the voice data. The chord for each beat is discriminated and supplied to the evaluation unit 147. Here, the discriminating unit 212 includes the chord discriminating feature amount extracted from the time-pitch data that has been preprocessed by using a predetermined gene out of the supplied chord discriminating feature amount, and the gene The code is discriminated from the corresponding code discriminator.

  Next, a preprocessing filter generation process, which is a process in which the filter generation device 51 generates a preprocessing filter, will be described with reference to the flowchart of FIG.

  In step S <b> 41, the initial generation generation unit 121 generates an initial generation gene and supplies the gene to the preprocessing unit 143 and the next generation generation unit 123. That is, the initial generation generation unit 121 generates several candidates for preprocessing filters as genes by combining one or a plurality of unit filters, and sets each of the generated genes as an initial generation gene.

  In step S42, the filter generation device 51 performs an evaluation process on each of the genes to be evaluated. The details of the evaluation process will be described later. In this evaluation process, for each generation gene to be evaluated, the code discrimination accuracy when the gene is used as a preprocessing filter is shown. An evaluation value is obtained.

  When the evaluation process is performed and the evaluation value of each gene is obtained, in step S43, the evaluation unit 147 determines whether or not the evaluation value has changed. For example, the evaluation unit 147 obtains a difference between the highest evaluation value of the gene evaluation values and the evaluation value obtained last time, that is, the highest evaluation value of the evaluation values of the genes of the previous generation, and obtains this time. Evaluation when the difference between the evaluation values obtained, the difference between the evaluation values obtained last time, and the difference between the evaluation values obtained last time is less than or equal to a predetermined threshold value, that is, the evaluation values hardly change. It is determined that the value no longer changes.

  In addition, for example, when the highest evaluation value among gene evaluation values is equal to or higher than a predetermined threshold, the evaluation value does not change as a gene for extracting a code with sufficient accuracy is obtained. It may be determined that

  If it is determined in step S43 that the evaluation value has changed, that is, if it has been determined that a gene for extracting a code with sufficient accuracy has not yet been obtained, a next-generation gene is generated. The unit 147 supplies an evaluation value, which is an evaluation result of each gene, to the next generation generation unit 123, and the process proceeds to step S44.

  In step S44, the selection unit 151 of the next generation generation unit 123 selects and selects some genes having a high evaluation value among the evaluated genes based on the evaluation value of the gene supplied from the evaluation unit 147. This gene will be the next generation gene.

  For example, when the last evaluated gene is an initial generation gene, the selection unit 151 selects some genes having a high evaluation value from the initial generation genes supplied from the initial generation generation unit 121. And the next generation gene. Further, for example, when the last evaluated gene is not the gene of the initial generation, that is, when it is a gene generated by the next generation generation unit 123, the selection unit 151 performs the previous selection by the selection unit 151 to the random generation unit 154. Among the generated genes, some genes with high evaluation values are selected as the next generation genes.

  In step S45, based on the gene evaluation values supplied from the evaluation unit 147, the mutation processing unit 152 performs mutation processing on some genes having high evaluation values to generate next-generation genes. That is, for each of several genes having a high evaluation value, the mutation processing unit 152 newly replaces some of the unit filters constituting the gene with another unit filter selected at random. To generate next-generation genes.

  In step S46, based on the gene evaluation values supplied from the evaluation unit 147, the cross processing unit 153 performs cross processing on some genes having high evaluation values to generate next-generation genes. That is, the intersection processing unit 153 changes, for each of several genes having a high evaluation value, some of the unit filters constituting the genes to unit filters constituting other genes of the same generation. Generate a new gene and make it the next generation gene.

  In step S47, the random generation unit 154 generates a new gene (pre-processing filter candidate) by combining randomly selected unit filters, and sets it as the next generation gene.

  In the process of step S45 and the process of step S46, as in the case of the process of step S44, when the last evaluated gene is an early generation gene, the initial stage supplied from the initial generation generation unit 121 is used. When the next generation gene is generated using the generation gene and the last evaluated gene is the gene generated by the next generation generation unit 123, that is, the selection unit 151 to the random generation unit 154, the next The gene generated by the generation generation unit 123 is used to generate the next generation gene.

  When the next generation gene is generated by the selection unit 151 to the random generation unit 154 in this way, the next generation generation unit 123 supplies the generated next generation gene to the evaluation unit 147 and the preprocessing unit 143. Then, when the next generation gene is supplied to the evaluation unit 147 and the preprocessing unit 143, the process returns to step S42, and each of the newly generated next generation gene is evaluated.

  If it is determined in step S43 that the evaluation value of the gene has ceased to change, that is, if a gene for extracting a code with sufficient accuracy is obtained, the evaluation unit 147 finally evaluates in step S48. The gene with the highest evaluation value is selected from the selected genes, and the selected gene, that is, the preprocessing filter and the code discriminator corresponding to the preprocessing filter are output.

  That is, the evaluation unit 147 preprocesses the selected gene (pre-processing filter) having the highest evaluation value among the genes of the generation to be evaluated last supplied from the next generation generation unit 123. And the pre-processing is performed using the selected gene of the code discriminator corresponding to the gene of the generation to be evaluated last supplied from the code discriminator learning unit 145. A chord discriminator for discriminating a chord from the received time-pitch data, that is, a chord discriminator created from the time-pitch data that has been preprocessed using the selected gene Supply. When the selected preprocessing filter and code discriminator are output, the preprocessing filter generation process ends.

  In this way, the filter generation device 51 creates a preprocessing filter and a code discriminator. In this way, an algorithm for extracting the chord of music from time-pitch data is constructed by humans, and the pre-processing filter used in the algorithm and the chord discriminator are generated by the filter generator 51, so that the music can be accurately reproduced. It is possible to generate an algorithm that can extract the codes.

  In other words, the pre-processing filter candidate is used as a gene, the next generation gene is repeatedly generated and evaluated, and a gene with a high evaluation value is used as the pre-processing filter, so that the music code can be extracted more accurately. Preprocessing filters and code discriminators that can be generated can be generated. And, by using the generated preprocessing filter and code discriminator in an algorithm constructed from a small amount of teacher data by utilizing human foresight knowledge, the accuracy of code extraction by the algorithm can be improved. As a result, an algorithm that can extract features more reliably can be easily obtained.

  In the feature extraction system, a pre-processing filter and a code discriminator used in an algorithm constructed by a human are generated by the filter generation device 51. Therefore, a highly accurate algorithm is semi-automatically constructed by the filter generation device 51. It can be said that.

  Next, an evaluation process corresponding to the process of step S42 of FIG. 13 will be described with reference to the flowchart of FIG. This evaluation process is performed for each gene to be evaluated. That is, an evaluation process is performed for each gene.

  In step S <b> 81, the analysis unit 142 performs 12-tone analysis processing on the voice data that is the teacher data supplied from the teacher data holding unit 141, and supplies the time-pitch data obtained as a result to the preprocessing unit 143. That is, the analysis unit 142 divides the sound of the sound data into a plurality of octave components, and further obtains the energy of twelve sounds at each height of the twelve average rate in each octave, so that twelve of each octave is obtained. Time-pitch data indicating the energy of each sound is obtained.

  In step S <b> 82, the preprocessing unit 143 performs preprocessing on the time-pitch data supplied from the analysis unit 142 using the gene supplied from the initial generation generation unit 121 or the next generation generation unit 123. Then, the preprocessing unit 143 supplies the pre-processed time-pitch data to the beat feature quantity extraction unit 181 of the chord discriminator learning unit 145 and the beat feature quantity extraction unit 211 of the chord discrimination unit 146. To do.

  For example, when an initial generation gene is evaluated, since the initial generation gene is supplied from the initial generation generation unit 121 to the preprocessing unit 143, the preprocessing unit 143 is supplied from the initial generation generation unit 121. Pretreatment is performed using early generation genes. In addition, when the gene of the generation after the initial generation, that is, the gene generated by the next generation generation unit 123 is evaluated, the next generation generation unit 123 sends a new generation to be evaluated to the preprocessing unit 143. Therefore, the preprocessing unit 143 performs preprocessing using the gene supplied from the next generation generation unit 123.

  More specifically, the preprocessing unit 143 performs preprocessing on all the time-pitch data supplied from the analysis unit 142 by using each of the genes in the time-pitch data. Therefore, for example, when the gene G41 and the gene G42 are supplied to the preprocessing unit 143, the preprocessing unit 143 performs preprocessing using the gene G41 on all time-pitch data, and uses the gene G42. Pre-processing is performed. As a result, for one time-pitch data, time-pitch data pre-processed using the gene G41 and time-pitch data pre-processed using the gene G42 are obtained. It is done.

  In step S83, the teacher data dividing unit 144 divides the teacher data supplied from the teacher data holding unit 141 into learning data and evaluation data. In other words, the teacher data dividing unit 144 randomly selects whether the teacher data supplied from the teacher data holding unit 141 is to be learning data or evaluation data. The data is divided into learning data and evaluation data.

  Then, the teacher data dividing unit 144 supplies the teacher data determined as the learning data as a result of the division to the adding unit 182 of the code discriminator learning unit 145, and the teacher data set as the evaluation data is evaluated by the evaluation unit 147. To supply.

  In step S <b> 84, the beat-by-beat feature quantity extraction unit 181 of the chord discriminator learning unit 145 extracts the chord discrimination feature quantity from the time-pitch data supplied from the preprocessing unit 143, and supplies it to the addition unit 182.

  For example, the beat-by-beat feature quantity extraction unit 181 performs the same processing as the beat detection unit 91 to obtain the basic beat cycle, thereby detecting the beat position from the time-pitch data, and further, the beat-by-beat feature quantity extraction unit 92. And for each range of beats in the time-pitch data, weight the energy of every 12 notes at each height of the 12 average rate of each octave, and the same note name in each octave By adding the energy of the sound and obtaining the energy of each of the 12 sounds specified by the pitch name, a feature quantity for route discrimination and a feature quantity for major / minor discrimination are generated, and these feature quantities are code-determined. It is used as a feature amount.

  In step S85, the adding unit 182 uses the chord determination feature amount supplied from the beat feature amount extraction unit 181 and the correct code name for each beat known in advance for the learning data as the learning data. Add to teacher data.

  For example, as shown in FIG. 15, the adding unit 182 is informed in advance of the chord progression of the music for the music reproduced by the audio data as the teacher data. FIG. 15 shows music 1 to music 3 as music reproduced by audio data as teacher data, and the chord progression of those music. In the figure, the chord name located on the left side is the correct chord name in the beat range closer to the start position of the music.

  The chord progression of the song 1 is that the chord name of the first beat range of the song 1 is C, then the chord that is E minor, the chord that is A minor, the chord that is C, the chord that is F, and the E minor , A code that is a D minor, and a code that is a G.

  The chord progression of the music piece 2 is that the chord name in the first beat range of the music piece 2 is D minor, then the chord is B ♭ (flat), the chord is C, the chord is A minor, B The code is ♭, the code is F, the code is G, and the code is C. Further, the chord progression of the music 3 is that the chord name of the first beat range of the music 3 is G, and then the chord D, the chord F, the chord C, the chord D, G A code is a code that is D, a code that is F, a code that is C #, and a code that is E ♭.

  Furthermore, the adding unit 182 also knows in advance the start time and end time of the chord in the music, that is, the start time and end time of each beat range for each music. Therefore, for example, in the adding unit 182, the start time of the first chord of the music piece 1 (code whose code name is C) is the start time of the music piece 1, and the end time is the time 13 seconds after the start time. This is known beforehand.

  For the teacher data as the supplied learning data, the adding unit 182 adds the chord discriminating feature amount of the teacher data supplied from the beat feature amount extracting unit 181 and the correct answer of the teacher data known in advance. Add a code name. Here, the adding unit 182 performs a process of adding the code discrimination feature quantity and the correct code name to the teacher data for each code discrimination feature quantity extracted from the teacher data as the learning data. In addition, since the teacher data set as the evaluation data is not supplied to the adding unit 182, the code discrimination feature extracted from the teacher data as the evaluation data among the code discrimination feature values supplied to the adding unit 182. As for the feature quantity, the code discrimination feature quantity is not added to the teacher data.

  For example, the teacher data D1 and the teacher data D2 as learning data are supplied to the adding unit 182 and the teacher data D1 preprocessed using the gene G41 (more specifically, obtained from the teacher data D1) Code discrimination feature F11 extracted from time-pitch data) and preprocessing using chord discrimination feature F12 and gene G42 extracted from teacher data D2 preprocessed using gene G41 Assume that the code discrimination feature quantity F21 extracted from the teacher data D1 and the code discrimination feature quantity F22 extracted from the teacher data D2 preprocessed using the gene G42 are supplied.

  In this case, the adding unit 182 has the teacher data D1 to which the code identification feature quantity F11 and the correct code name are added, the teacher data D2 to which the code discrimination feature quantity F12 and the correct code name are added, and the code discrimination feature. The teacher data D1 to which the amount F21 and the correct code name are added and the teacher data D2 to which the code identification feature amount F22 and the correct code name are added are supplied to the shift processing unit 183.

  Returning to the description of the flowchart of FIG. 14, when the code determination feature amount and the correct code name are added to the learning data, the shift processing unit 183 determines each teacher data supplied from the adding unit 182 in step S86. Then, the chord discrimination feature and the correct chord name for each beat of the teacher data are shifted by one sound.

  In step S87, the shift processing unit 183 adds the chord determination feature amount shifted by one sound and the correct code name to the teacher data.

  In step S88, the shift processing unit 183 performs the process performed in steps S86 and S87, that is, the process of shifting the chord discrimination feature value and the correct code name for each beat by one sound and adding it to the teacher data 11 times. It is determined whether or not the process has been repeated.

  For example, as shown in FIG. 16, when the correct code name that is the name of the correct code indicated by the code for each beat corresponding to the chord discrimination feature value for each beat is D, C, C #, D, D #, E, F, F #, G, G #, A, A #, and B data for indicating the energy of each of the pitch names are arranged in order. The feature amount is added to the teacher data by the adding unit 182 along with the correct code name D.

  The shift processing unit 183 receives data indicating the energy of the sounds of the pitch names of C #, D, D #, E, F, F #, G, G #, A, A #, B, and C. In order to arrange them sequentially, the arrangement of data indicating the energy of the sound in the feature quantity for route discrimination and the feature quantity for major / minor discrimination is shifted, and the correct code name is shifted to C #. The shift processing unit 183 sequentially arranges data indicating the sound energy of the pitch names of C #, D, D #, E, F, F #, G, G #, A, A #, B, and C. The route discrimination feature quantity and major / minor discrimination feature quantity are added to the teacher data together with the correct code name C #.

  Further, the shift processing unit 183 receives data indicating the energy of the sounds of the respective pitch names D, D #, E, F, F #, G, G #, A, A #, B, C, and C #. In order to arrange them sequentially, the arrangement of data indicating the energy of the sound in the feature quantity for route discrimination and the feature quantity for major / minor discrimination is further shifted, and the correct code name is shifted to D. The shift processing unit 183 sequentially arranges data indicating the sound energy of the pitch names of D, D #, E, F, F #, G, G #, A, A #, B, C, and C #. The route discrimination feature quantity and major / minor discrimination feature quantity are added to the teacher data together with the correct code name D.

  As described above, the shift of the arrangement of the data indicating the sound energy in the route distinguishing feature amount and the major / minor distinguishing feature amount is repeated 11 times, so that 12 pieces of data from one route distinguishing feature amount are teacher data. 12 data are added to the teacher data from one major / minor discrimination feature quantity.

  The adding unit 182 and the shift processing unit 183 perform processing for adding the chord determination feature quantity and the correct code name to the teacher data as learning data for all beat ranges of all supplied learning data.

  Returning to the description of the flowchart of FIG. 14, if it is determined in step S88 that the process has not been repeated 11 times, the process returns to step S86, and the above-described process is repeated until it is determined that the process has been repeated 11 times.

  On the other hand, if it is determined in step S88 that the repetition has been performed 11 times, the shift processing unit 183 supplies the teacher data to which the code discrimination feature amount and the correct code name are added to the code discriminator generation unit 184. The process proceeds to step S89.

  In step S89, the code discriminator generation unit 184 generates (creates) a code discriminator by machine learning using the teacher data supplied from the shift processing unit 183 for each gene, and codes the generated code discriminator as a code. The data is supplied to the determination unit 212 of the determination unit 146 and the evaluation unit 147.

  That is, the chord discriminator generating unit 184 uses the teacher data to which the chord discriminating feature amount extracted from the time-pitch data pre-processed by using the same gene among the supplied teacher data is used. Then, a code discriminator corresponding to the gene, that is, a code discriminator for extracting a correct code from the time-pitch data using the gene and pre-processed is generated.

  For example, the code discriminator generator 184 sets kNN (k-Nearest Neighbor), SVM (Support Vector Machine), Naive Bayes, Mahalanobis distance with the closest code as the correct answer, or the code with the highest probability as the correct answer. A code discriminator is generated (created) by machine learning using GMM (Gaussian Mixture Model) or the like.

  In step S 90, the beat feature quantity extraction unit 211 of the chord discrimination unit 146 extracts the chord discrimination feature quantity from the time-pitch data supplied from the preprocessing unit 143, and supplies it to the discrimination unit 212.

  For example, the beat-by-beat feature quantity extraction unit 211 performs the same processing as the beat-by-beat feature quantity extraction unit 181, detects the beat position from the time-pitch data by obtaining the basic beat period, and further, the time-pitch data. For each range of beats, weight the energy of each twelve note at the height of each twelve average rate in each octave, add the energy of the same note name in each octave, and By obtaining the energy of each of the 12 specified sounds, a route discrimination feature quantity and a major / minor discrimination feature quantity are generated, and these feature quantities are used as the chord discrimination feature quantity.

  In step S91, the determination unit 212 uses the chord discriminator supplied from the chord discriminator generation unit 184 to discriminate the chord for each beat from the chord discrimination feature quantity supplied from the beat feature quantity extraction unit 211. .

  For example, the discriminating unit 212 uses a chord discriminator corresponding to the predetermined gene for the chord discriminating feature amount extracted from the time-pitch data that has been preprocessed using the predetermined gene. To determine the chord for each beat.

  In step S92, the beat-by-beat feature quantity extraction unit 211 determines whether or not chord discrimination has been performed for all time-pitch data supplied from the preprocessing unit 143.

  If it is determined in step S92 that chord determination has not been performed for all time-pitch data, the process returns to step S90 and the above-described process is repeated.

  On the other hand, if it is determined in step S92 that chord discrimination has been performed for all time-pitch data, the discrimination unit 212 supplies the chord name for each beat obtained by the discrimination to the evaluation unit 147. The process proceeds to step S93.

  In step S93, the evaluation unit 147 selects one code name supplied from the determination unit 212, and determines whether or not the code name is correctly determined.

  That is, the code name extracted from the teacher data set as learning data and the code name extracted from the teacher data set as evaluation data are supplied from the determination unit 212 to the evaluation unit 147. The evaluation unit 147 evaluates each gene using the code name extracted from the teacher data that is the evaluation data among the code names supplied from the determination unit 212.

  For example, the evaluation unit 147 uses the chord name extracted from the time-pitch data that has been pre-processed by using a predetermined gene to be evaluated among the chord names that have been supplied. A predetermined gene to be evaluated is evaluated. Since the evaluation unit 147 knows in advance the correct code of the song data of the teacher data that is the evaluation data, the evaluation unit 147 evaluates a predetermined gene with the code name supplied from the determination unit 212. One code name is selected from the code names used for the purpose, and the selected code name is compared with the corresponding correct code name to determine whether or not the code is correctly determined.

  If it is determined in step S93 that the determination is correct, in step S94, the evaluation unit 147 increments the number of correct answers for determination of the held code by one.

  For example, for each gene, the evaluation unit 147 is the code name supplied from the determination unit 212, and among the code names used for evaluating the gene, the number of correct answers indicating the number of correctly identified code names And the number of incorrect answers indicating the number of code names that were not correctly determined. If it is determined in step S93 that the selected one code name has been correctly determined, the evaluation unit 147 increments the number of correct answers held.

  On the other hand, if it is determined in step S93 that it has not been correctly determined, in step S95, the evaluation unit 147 increments the number of incorrect answers held by one.

  When the number of correct answers is incremented in step S94 or the number of incorrect answers is incremented in step S95, in step S96, the evaluation unit 147 evaluates all the genes that are currently used for evaluation. It is determined whether or not the code name has been evaluated, that is, whether or not the code name has been correctly identified for all the evaluation data used for gene evaluation.

  If it is determined in step S96 that all code names have not been evaluated, the process returns to step S93, and the next one code name supplied from the determination unit 212 is selected for evaluation. Note that the processing from step S93 to step S95 is performed for all genes to be evaluated.

  On the other hand, if it is determined in step S96 that all the code names have been evaluated, in step S97, the evaluation unit 147 determines that each gene is based on the number of correct answers and the number of incorrect answers for each gene. The evaluation value is calculated, and the process proceeds to step S43 in FIG. For example, the evaluation unit 147 calculates the correct answer rate of code discrimination as an evaluation value of the gene based on the number of correct answers and the number of incorrect answers for one gene.

  In this way, the filter generation device 51 evaluates each of the generated genes. In this way, by evaluating the generated gene, it is possible to generate the next generation gene using the highly evaluated gene, so as the generation of the gene progresses, the code of the music can be extracted more reliably It is possible to generate a preprocessing filter and a code discriminator.

  In other words, as the gene generation progresses, it is possible to obtain a preprocessing filter that can emphasize components useful for feature extraction in audio data and remove unnecessary components.

  In the above, the preprocessing using the preprocessing filter is processing in which the format of the data to be processed, that is, the time-pitch data is maintained before and after the processing. As described above, the process in which the data format changes before and after the process may be set as the preprocess. In such a case, the filter generator 51 generates a chord discriminator for extracting the chord of music from pre-processed time-pitch data whose data format has changed.

  In the above description, the example of extracting the chord from the time-pitch data of the music has been described. In addition, still images, moving images, texts (character strings), and the like are input to the signal processing device 12 as input data. In the processing device 12, a desired feature can be extracted from the input data. In such a case, the filter generation device 11 generates a preprocessing filter for extracting features from input data and a discriminator.

  The series of processes described above can be executed by hardware or can be executed by software. When a series of processing is executed by software, a program constituting the software executes various functions by installing a computer incorporated in dedicated hardware or various programs. For example, it is installed from a program recording medium in a general-purpose personal computer or the like.

  FIG. 17 is a block diagram illustrating an example of the configuration of a personal computer that executes the above-described series of processing using a program. A CPU (Central Processing Unit) 301 executes various processes according to a program recorded in a ROM (Read Only Memory) 302 or a recording unit 308. A RAM (Random Access Memory) 303 appropriately stores programs executed by the CPU 301 and data. The CPU 301, ROM 302, and RAM 303 are connected to each other by a bus 304.

  An input / output interface 305 is also connected to the CPU 301 via the bus 304. The input / output interface 305 is connected to an input unit 306 including a keyboard, a mouse, and a microphone, and an output unit 307 including a display and a speaker. The CPU 301 executes various processes in response to commands input from the input unit 306. Then, the CPU 301 outputs the processing result to the output unit 307.

  The recording unit 308 connected to the input / output interface 305 includes, for example, a hard disk, and records programs executed by the CPU 301 and various data. The communication unit 309 communicates with an external device via a network such as the Internet or a local area network.

  Alternatively, the program may be acquired via the communication unit 309 and recorded in the recording unit 308.

  The drive 310 connected to the input / output interface 305 drives a removable medium 331 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, and drives the program or data recorded therein. Get etc. The acquired program and data are transferred to the recording unit 308 and recorded as necessary.

  As shown in FIG. 17, a program recording medium that stores a program that is installed in a computer and is ready to be executed by the computer includes a magnetic disk (including a flexible disk), an optical disk (CD-ROM (Compact Disc-Read Only). Memory, DVD (Digital Versatile Disc), a magneto-optical disk, a removable medium 331 which is a package medium made of a semiconductor memory, or the like, a ROM 302 in which a program is temporarily or permanently stored, or a recording unit 308 It is comprised by the hard disk etc. which comprise. The program is stored in the program recording medium using a wired or wireless communication medium such as a local area network, the Internet, or digital satellite broadcasting via a communication unit 309 that is an interface such as a router or a modem as necessary. Done.

  In the present specification, the step of describing the program stored in the program recording medium is not limited to the processing performed in time series in the described order, but is not necessarily performed in time series. Or the process performed separately is also included.

  The embodiment of the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the gist of the present invention.

It is a figure explaining the feature extraction system of one embodiment of this invention. It is a block diagram which shows the structure of the feature extraction system of one embodiment of this invention. It is a block diagram which shows the structure of the feature extraction system of one embodiment of this invention. It is a block diagram which shows the structural example of a code extraction part. It is a figure explaining the chord name for every beat extracted from time-pitch data. It is a flowchart explaining a code | cord | chord discrimination | determination process. It is a figure explaining time-pitch data. It is a figure explaining the filter for pre-processing. It is a figure explaining discrimination | determination of the code | cord | chord using a code discriminator. It is a block diagram which shows the structure of a filter production | generation apparatus. It is a block diagram which shows the structure of a code discriminator learning part. It is a block diagram which shows the structure of a code discrimination | determination part. It is a flowchart explaining the filter production | generation process for pre-processing. It is a flowchart explaining an evaluation process. It is a figure explaining the chord progression of a music. It is a figure explaining the feature-value for code discrimination | determination and the shift of a correct code. It is a block diagram which shows the structure of a personal computer.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 11 Filter production | generation apparatus, 12 Signal processing apparatus, 21 Pre-processing part, 22 Feature extraction processing part, 51 Filter production | generation apparatus, 52 Signal processing apparatus, 61 Pre-processing part, 62 Code extraction part, 92 Feature-value extraction part for every beat, 93 Code discriminating unit, 121 Initial generation unit, 122 Gene evaluation unit, 123 Next generation generation unit, 141 Teacher data holding unit, 143 Preprocessing unit, 145 Code discriminator learning unit, 146 Code discriminating unit, 147 Evaluation unit, 151 selection Part, 152 mutation processing part, 153 crossing processing part, 154 random generation part, 181 beat feature quantity extraction part, 182 addition part, 183 shift processing part, 184 code discriminator generation part, 211 beat feature quantity extraction part, 212 Discriminator

Claims (8)

A filter generating means for generating a candidate filter that is a candidate for a preprocessing filter used for preprocessing applied to the data to be extracted from the data when a predetermined feature is extracted from the data;
Using the candidate filter, pre-processing means for performing the pre-processing on teacher data used for evaluation of the candidate filter;
Extraction means for extracting the features from the pre-processed teacher data;
When the feature is extracted from the teacher data that has been pre-processed using the candidate filter, the feature extracted by the extraction unit is an evaluation value of the candidate filter that indicates an evaluation of the feature extraction And an evaluation means for calculating based on the characteristics to be extracted from the teacher data obtained in advance. The filter generation means further generates a plurality of new candidate filters using some of the candidate filters based on each of the evaluation values of the plurality of candidate filters,
The information processing apparatus according to claim 1, wherein the evaluation unit calculates each evaluation value of the new candidate filter. When the highest evaluation value among the evaluation values calculated at the end satisfies a predetermined condition, the evaluation unit uses the candidate filter having the highest evaluation value as a preprocessing filter used for the preprocessing. The information processing apparatus according to claim 2. If the highest evaluation value among the evaluation values calculated last does not satisfy the condition, the filter generation means may generate the candidate generated last based on the evaluation value calculated last. The information processing apparatus according to claim 3, further generating a plurality of new candidate filters by using some of the filters. The information processing apparatus according to claim 1, wherein the preprocessing filter is a filter that performs processing while maintaining a format of the data. Based on the teacher data that has been subjected to the preprocessing using the candidate filter and the characteristics to be extracted from the teacher data that have been obtained in advance, the candidate filter is used to perform the preprocessing. The information processing apparatus according to claim 1, further comprising: a discriminator creating unit that creates a discriminator for extracting the feature from the data by machine learning. When a predetermined feature is extracted from data, a candidate filter that is a candidate for a preprocessing filter used for preprocessing applied to the data to be extracted is generated.
Using the candidate filter, the preprocessing is performed on the teacher data used for the evaluation of the candidate filter,
The feature is extracted from the teacher data subjected to the preprocessing by an extraction unit,
When the feature is extracted from the teacher data that has been pre-processed using the candidate filter, the feature extracted by the extraction unit is an evaluation value of the candidate filter that indicates an evaluation of the feature extraction And an information processing method including a step of calculating based on the characteristics to be extracted from the teacher data obtained in advance. When a predetermined feature is extracted from data, a candidate filter that is a candidate for a preprocessing filter used for preprocessing applied to the data to be extracted is generated.
Using the candidate filter, the preprocessing is performed on the teacher data used for the evaluation of the candidate filter,
The feature is extracted from the teacher data subjected to the preprocessing by an extraction unit,
When the feature is extracted from the teacher data that has been pre-processed using the candidate filter, the feature extracted by the extraction unit is an evaluation value of the candidate filter that indicates an evaluation of the feature extraction And a program for causing a computer to execute a process including a step of calculating based on the characteristics to be extracted from the teacher data obtained in advance.

Images