JP2015172736A - Voice analysis device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To estimate the property of an object person for analysis easily with the use of a singing voice.SOLUTION: The voice analysis device includes: an indicator calculation unit 32 which, for each of mutually different K properties, uses a recognition model M[k] (k=1-k) indicating a trend of a singing voice corresponding to the property to calculate an estimation indicator E[k] which is an indicator of likelihood of coincidence of a singing voice V of an object person for analysis with the singing voice of the property; and an estimation processing unit 34 for estimating a property of the object person for analysis according to the estimation indicator E[k] calculated by the indicator calculation unit 32 for each property.

Description

  The present invention relates to a technique for analyzing speech.

  Patent Document 1 discloses a technique for analyzing consumer trends (consumption trends) regarding products. Sing history information of a plurality of consumers is used for analysis of consumption trends. According to the technique of Patent Document 1, there is an advantage that a consumption trend can be analyzed with high accuracy even when there is not enough purchase history information.

JP 2011-221595 A

  With the technology of Patent Document 1, it is possible to analyze the overall trend of consumption trends by a large number of consumers, but it is not possible to estimate the nature of each individual's activity trend. In view of the above circumstances, an object of the present invention is to simply estimate the characteristics of a person to be analyzed using a singing voice.

  In order to solve the above problems, the speech analysis apparatus of the present invention uses a recognition model that represents the tendency of the singing speech corresponding to the attribute for each of a plurality of different attributes to analyze the singing speech of the person to be analyzed. An index calculation unit that calculates an evaluation index that is an accuracy index corresponding to the voice of the attribute, and an estimation processing unit that estimates the characteristics of the analysis target person according to the evaluation index calculated for each attribute by the index calculation unit It has. In the above configuration, since the evaluation index, which is an index of accuracy corresponding to each voice of multiple attributes, is calculated by the recognition model for each attribute, the property of the analysis target can be easily estimated according to each evaluation index. Is possible.

  In a preferred aspect of the present invention, the estimation processing unit estimates the property of the analysis target person according to a combination of two or more attributes selected according to the evaluation index calculated by the index calculation unit among a plurality of attributes. For example, a configuration for estimating the characteristics of the person to be analyzed according to a combination of two or more attributes selected in numerical order of evaluation indices from among a plurality of attributes, or two of the plurality of attributes whose evaluation indices exceed a predetermined value A configuration that estimates the characteristics of the person to be analyzed according to the combination of the above attributes is preferable. In the above aspect, since the property of the person to be analyzed is estimated according to the combination of two or more attributes selected according to each evaluation index, for example, only one attribute is selected according to the evaluation index; In comparison, there is an advantage that the estimation results of the properties of the analysis subject are diversified.

  In a preferred aspect of the present invention, the estimation processing unit calculates, for each of a plurality of weight value series corresponding to different estimation results, a weighted sum of each evaluation index to which the plurality of weight values included in the weight value series are applied. An estimation result corresponding to a weight value series selected according to the weighted sum among a plurality of weight value series is determined. In the above aspect, since the estimation result is specified according to the weight value of each evaluation index to which the weight value of each weight value series is applied, the weight of each evaluation index with respect to the estimation result can be adjusted according to each weight value. There are advantages.

  In a preferred aspect of the present invention, each recognition model of a plurality of attributes is generated by machine learning using each reference voice of a set corresponding to the attribute among a plurality of sets obtained by classifying a plurality of reference voices for each attribute. Is done. In the above aspect, the recognition model is generated by machine learning using each reference speech classified for each attribute. Therefore, there is an advantage that high-accuracy estimation that reflects a realistic tendency of the relationship between the characteristics of the speech and the nature of the speaker is possible. Specifically, the plurality of reference sounds are classified into a plurality of sets according to at least one of a time zone in which the reference sound is generated, a place where the reference sound is generated, and a character of the speaker of the reference sound. . It is also possible to classify each reference sound into a plurality of sets according to the purchase history of the product by the speaker of the reference sound.

  The speech analysis apparatus according to a preferred aspect of the present invention includes a singing evaluation unit that evaluates the skill of the person to be analyzed by an evaluation method according to the estimation result by the estimation processing unit. In the above aspect, since the skill of singing of an analysis object person is evaluated with the evaluation method according to the estimation result by an estimation process part, appropriate singing evaluation according to the property of an analysis object person is implement | achieved. The “evaluation method according to the estimation result” is the one related to the singing evaluation including the contents of the evaluation process, the points (weights) contributing to the evaluation result, the evaluation criteria and the evaluation items, and the variables applied to the evaluation process. It means that more than one kind of matter changes according to the estimation result.

1 is a configuration diagram of a speech analysis apparatus according to a first embodiment of the present invention. It is explanatory drawing of learning data. It is explanatory drawing of a learning process. It is a flowchart of a learning process. It is a block diagram of an audio | voice analysis part. It is a flowchart of an estimation process. It is explanatory drawing of reference data. It is a flowchart of the estimation process in 2nd Embodiment. It is a block diagram of the audio | voice analysis system which concerns on 4th Embodiment. It is a block diagram of the audio | voice analysis apparatus which concerns on 5th Embodiment. It is an example of a display of the song evaluation result in 5th Embodiment. It is an example of a display of the song evaluation result in 5th Embodiment.

<First Embodiment>
FIG. 1 is a configuration diagram of a speech analysis apparatus 100 according to the first embodiment of the present invention. The speech analysis apparatus 100 is an information processing apparatus that estimates the characteristics (personality, propensity, behavior pattern, etc.) of an analysis target person by analyzing a singing voice that is pronounced by an arbitrary user (hereinafter referred to as “analysis target person”). And a computer system that includes an arithmetic processing unit 12, a storage device 14, a sound collection device 16, and a display device 18.

  The sound collection device 16 is a device (microphone) that collects ambient sounds. The sound collection device 16 of the first embodiment collects the singing voice V in which the analysis target person sang the music. The display device 18 (for example, a liquid crystal display panel) displays an image instructed from the arithmetic processing device 12. For example, the analysis result of the property of the person to be analyzed is displayed on the display device 18. The analysis result can also be output as sound from a sound emitting device (speaker or earphone).

  The arithmetic processing unit 12 performs overall control of each element of the speech analysis device 100 by executing a program stored in the storage device 14. Specifically, as illustrated in FIG. 1, the arithmetic processing device 12 includes a learning processing unit 22 that generates a recognition model used for analyzing the characteristics of the analysis target person, and a recognition model generated by the learning processing unit 22. And the voice analysis unit 24 that estimates the characteristics of the person to be analyzed by analyzing the singing voice V using the voice. A configuration in which each function of the arithmetic processing device 12 is distributed to a plurality of devices, or a configuration in which a dedicated electronic circuit realizes a part of the function of the arithmetic processing device 12 may be employed.

  The storage device 14 stores a program executed by the arithmetic processing device 12 and various data used by the arithmetic processing device 12. A known recording medium such as a semiconductor recording medium and a magnetic recording medium or a combination of a plurality of types of recording media is arbitrarily employed as the storage device 14. The storage device 14 of the first embodiment stores a learning data group GL. The learning data group GL is a set (big data) of a large number of learning data DL collected in advance. The learning processing unit 22 generates a recognition model by machine learning using the learning data group GL stored in the storage device 14.

  As illustrated in FIG. 2, any one learning data DL in the learning data group GL is sample data relating to a singing voice (hereinafter referred to as “reference voice”) used for machine learning of a recognition model, and voice data. It includes DA, feature data DB, and related data DC. The audio data DA is an audio file representing the time waveform of the reference audio. On the other hand, the feature data DB and the related data DC are used for classification of the reference voice represented by the voice data DA.

  The feature data DB is data representing the feature amount of the reference voice (particularly a feature amount specific to singing). For example, in addition to basic features such as the pitch and volume of the reference speech, multiple types such as features that characterize the voice quality of the reference speech (for example, MFCC, singing formants) and reference speech evaluation results (scores) The feature amount is included in the feature data DB.

  The related data DC is information related to the singer or singing of the reference voice. As illustrated in FIG. 2, the related data DC of the first embodiment includes personal information DC1, song information DC2, and trend information DC3. The personal information DC1 is information (for example, age, sex, address, occupation, etc.) related to the individual singer of the reference voice. The singing information DC2 is information (singing environment, history, tendency) related to the singing of the reference voice. For example, singing time, type of singing place (for example, distinction between high-class stores / karaoke boxes, etc.), singing frequency, number of singing (number of singing in a specific group), number of singing (number of customers in store), score (for example, specific group) The score information DC2 can be included in the singing information DC2.

  The trend information DC3 is information relating to the personality and preferences of the reference voice singer. For example, information such as the personality of the reference voice singer and emotions at the time of singing is extracted from the result of the questionnaire for the singer and included in the trend information DC3. Further, for example, information related to the singer's preference and behavior pattern estimated from the use status (for example, the contents of a post or profile) of a blog or SNS (Social Networking Service) may be included in the trend information DC3.

  Specific examples of each learning data DL are as described above. A learning data group GL is configured by collecting learning data DL for a large number of singers. Each learning data DL is collected from a singer, for example, when using a karaoke shop. For example, the voice data DA and the characteristic data DB are generated from the singing voice of the singer, and the personal information DC1 and the singing information DC2 of the related data DC are extracted from the membership information of the service card issued at the time of membership registration of the karaoke shop. obtain. Note that the method for acquiring the attribute of the reference voice (for example, related data DC) is not limited to the above example. For example, the personal information of a singer's karaoke service card and the purchase history of each user registered at various stores are collated to extract behavior patterns such as the purchase history of the singer as trend information DC3 Alternatively, a configuration in which personal information of a karaoke shop service card and personal information registered in a profile such as a blog or SNS are collated to extract the singer's preference and behavior pattern as the trend information DC3 may be employed.

<Learning processing unit 22>
As illustrated in FIG. 3, the learning processing unit 22 of the first embodiment uses a plurality (K) of recognition models M [1] to M [M] by machine learning using the learning data group GL stored in the storage device 14. Generate [K]. One arbitrary recognition model M [k] (k = 1 to K) classifies a plurality of reference sounds (speech data DA) for each attribute by using the feature data DB and the related data DC of each learning data DL. Generated by machine learning using a plurality of reference speech belonging to the k-th set C [k] among the K sets C [1] to C [K]. Therefore, the recognition model M [k] of the first embodiment is a statistical model that represents the tendency of feature quantities of a plurality of reference speech classified into one set C [k]. For example, a statistical model using a mixture distribution such as GMM (Gaussian Mixture Model) or a probability model such as HMM (Hidden Markov Model) is preferably used as the recognition model M [k].

  FIG. 4 is a flowchart of the learning process in which the learning processing unit 22 generates K recognition models M [1] to M [K] from the learning data group GL. When the learning process of FIG. 4 is started, the learning processing unit 22 converts a plurality of speech data DA (reference speech) of each learning data DL into K sets C [1] to C [K] as illustrated in FIG. ] (SA1). That is, a plurality of reference voices are classified into K types of attributes. Note that one reference voice may belong to a plurality of sets C [k]. For the classification of each reference voice, as illustrated below, feature data DB and related data DC of each learning data DL are used.

  The feature data DB of each learning data DL is generally used for classification focusing on the acoustic features of the reference speech. For example, a reference voice in which the pitch expressed by the feature data DB accurately follows the expected note string of the music, a monotonous reference voice with little fluctuation in the volume or voice quality within the note, a low-volume voice quality A reference voice with a small number of voices and a reference voice with a small number of singing techniques such as shaku and vibrato are classified into a set C [k] of “serious” attributes.

  The related data DC is generally used for classification focusing on the preference and propensity (behavior pattern) of the singer of the reference voice. For example, reference sounds belonging to a time zone in which the singing time specified by the related data DC belongs to midnight are classified into a set C [k] of “nocturnal” attributes, and the type of the singing store specified by the related data DC is high-class. The reference speech that is a store is classified into a set C [k] of attributes of “rich people”. In addition, a reference voice in which the character of the singer designated by the related data DC is “adhesive” or a reference voice having a high number of singing times specified by the related data DC is a set C [k ]are categorized. In addition, a well-known data analysis technique is arbitrarily employ | adopted for the classification | category (clustering) of each reference audio | voice. For example, clustering by the k-means method disclosed in Japanese Patent Laid-Open No. 2005-222138 can be used for classification of each reference speech.

  When the plurality of reference voices are classified, the learning processing unit 22 selects one set C [k] from the K sets C [1] to C [K] (SA2). Then, as illustrated in FIG. 3, the learning processing unit 22 generates a recognition model M [k] by machine learning using a plurality of audio data DA belonging to the set C [k] (SA3). Specifically, the learning processing unit 22 generates a recognition model M [k] so that a statistical tendency of a plurality of reference sounds belonging to one set C [k] is expressed. For machine learning of the recognition model M [k], for example, a known machine learning technique such as decision tree learning is arbitrarily employed. For example, decision tree learning using C4.5 (J. Rose Quinlan, “C4.5 Programs for machihne learning,” Morgan Kaufmann Publishers, 1993) is suitable.

  The recognition model M [k] of the first embodiment is a statistical model for calculating an evaluation index (likelihood) E [k] for the singing voice V to be recognized. The evaluation index E [k] calculated by the recognition model M [k] has an accuracy that the singing voice V corresponds to the statistical tendency of each reference voice belonging to the set C [k] (the singing voice V is set to the set C [k]. Is an index of likelihood). That is, the evaluation index E [k] is set to a larger numerical value as the degree to which the feature amount of the singing voice V matches the tendency of the feature amount of the reference speech expressed by the recognition model M [k]. The recognition model M [k] generated by the learning processing unit 22 is stored in the storage device 14. Specifically, a variable that defines the recognition model M [k] is stored in the storage device 14.

  As illustrated in FIG. 3, the learning processing unit 22 adds the attribute information A [k] to the recognition model M [k] (SA4). The attribute information A [k] is information (label) expressing the attribute corresponding to the set C [k]. For example, the attribute information A [k] for designating the attribute of “night play” is added to the recognition model M [k] of the set C [k] of the reference voice belonging to the midnight time zone, and the type of the singing store Attribute information A [k] designating the attribute of the “rich people” is added to the recognition model M [k] of the reference speech set C [k] where is a high-end store.

  The learning processing unit 22 determines whether or not the generation of the recognition model M [k] has been completed for the K sets C [1] to C [K] (SA5). When the determination result is negative (SA5: NO), the learning processing unit 22 newly selects an ungenerated set C [k] of the recognition model M [k] (SA2), and then recognizes the recognition model M [k. ] Generation (SA3) and addition of attribute information A [k] (SA4). When the generation of the K recognition models M [1] to M [K] is completed (SA5: YES), the learning processing unit 22 ends the learning process of FIG.

<Speech analysis unit 24>
The voice analysis unit 24 in FIG. 1 estimates the characteristics of the person to be analyzed by analyzing the singing voice V collected by the sound collection device 16. FIG. 5 is a specific configuration diagram of the voice analysis unit 24. As illustrated in FIG. 5, the speech analysis unit 24 according to the first embodiment includes an index calculation unit 32 and an estimation processing unit 34.

  The index calculation unit 32 applies the singing voice V of the person to be analyzed to each of the K recognition models M [1] to M [K] generated by the learning processing unit 22, thereby different attributes (recognition model M K evaluation indexes E [1] to E [K] corresponding to [k]) are calculated. The estimation processing unit 34 estimates the property of the analysis target person according to the evaluation index E [k] calculated for each attribute by the index calculation unit 32. The estimation processing unit 34 of the first embodiment generates analysis information Q indicating the estimation result of the property of the analysis target person.

  FIG. 6 is a flowchart of an estimation process in which the voice analysis unit 24 generates the analysis information Q from the singing voice V of the person to be analyzed. For example, the estimation process of FIG. 6 is started with an instruction given from the analysis subject before the start of singing.

  When the estimation process is started, the index calculation unit 32 selects one recognition model M [k] from the K recognition models M [1] to M [K] stored in the storage device 14 (SB1). Then, the index calculation unit 32 calculates the evaluation index E [k] by applying the singing voice V supplied from the sound collection device 16 to the recognition model M [k] (SB2). Specifically, the index calculation unit 32 extracts feature quantities such as pitch, volume, and voice quality from the singing voice V, and the feature quantities are represented in the set C [k] represented by the recognition model M [k]. The observed likelihood (accuracy) is calculated as an evaluation index E [k]. For example, the index calculation unit 32 generates a probability density distribution (mixed distribution model such as GMM) of the feature amount extracted from the singing voice V, and the likelihood between the probability density distribution and the recognition model M [k] ( Accuracy) is calculated as an evaluation index E [k]. The probability density distribution of each feature amount is generated from, for example, 13-dimensional MFCC for voice quality, and is generated from an approximate curve of a trajectory within a unit time (for example, 1 second) for pitch and volume.

  The index calculation unit 32 determines whether or not the calculation of the evaluation index E [k] (E [1] to E [K]) is completed for each of the K recognition models M [1] to M [K]. (SB3). When the determination result is negative (SB3: NO), the index calculation unit 32 newly selects an uncalculated recognition model M [k] for the evaluation index E [k] (SB1), and the recognition model M [k ] Is used to calculate the evaluation index E [k] (SB2). On the other hand, when calculation of K evaluation indexes E [1] to E [K] corresponding to different recognition models M [k] is completed (SB3: YES), generation of analysis information Q by the estimation processing unit 34 is performed. It is executed (SB4, SB5).

  Specifically, the estimation processing unit 34 includes two or more attributes (attribute information A) corresponding to the evaluation indexes E [1] to E [K] among the K attributes corresponding to each recognition model M [k]. [k]) is selected (SB4). For example, the estimation processing unit 34 selects a predetermined number of recognition models M [k] positioned higher in the numerical order (descending order) of the evaluation index E [k] among the K recognition models M [1] to M [K]. Each attribute information A [k] added to each is selected. As understood from the above description, an attribute having a high likelihood of conforming to the characteristics of the analysis target person is selected. For example, a singing voice V in which the pitch accurately follows the intended note sequence of the music, a monotonous singing voice V with less fluctuation in volume or voice quality within the note, and a singing voice V with a small volume and less voice quality. For the singing voice V with few singing techniques such as shakukuri and vibrato, attribute information A [k] designating “serious” is selected. In addition, the method of selecting the attribute (attribute information A [k]) of the singing voice V according to the evaluation index E [k] is not limited to the above examples. For example, a configuration in which the evaluation index E [k] exceeds the predetermined threshold (fixed value or variable value) among the K attributes, or one attribute that maximizes the evaluation index E [k] is selected. The structure to do can also be employ | adopted.

  The estimation processing unit 34 generates analysis information Q according to a combination of two or more attributes (attribute information A [k]) selected with reference to the evaluation index E [k] (SB5). Specifically, a sentence including the character string of the attribute specified by the attribute information A [k] is generated as the analysis information Q. For example, if attribute information A [k1] that specifies “night play” and attribute information A [k2] that specifies “sticky” are selected, “Your personality is sticky that you like to play at night.” Is generated, and attribute information A [k3] specifying “serious” and attribute information A [k4] specifying “rich” are selected, “you are serious The analysis information Q of the sentence “It is a wealthy class” is generated. The estimation processing unit 34 causes the display device 18 to display the analysis information Q generated by the above procedure (SB6).

  As described above, in the first embodiment, the evaluation index E [k] (E [1] to E [K]), which is an index of accuracy corresponding to the speech of each of the K attributes, is used to recognize each attribute. Since it is calculated by the model M [k], it is possible to easily estimate the properties (personality, propensity, behavior pattern) of the person to be analyzed according to each evaluation index E [k]. In the first embodiment, in particular, since the property of the person to be analyzed is estimated according to a combination of two or more attributes selected according to each evaluation index E [k] among the K attributes, for example, K attributes Compared with a configuration in which only one attribute is selected according to the evaluation index E [k], there is an advantage that various properties of the analysis target person can be estimated. In the first embodiment, a plurality of reference voices are classified into K sets C [1] to C [K] according to the singing time of the reference voice, the singing place (type of singing store), and the character of the singer. Therefore, there is an advantage that the property of the analysis target person can be estimated from various viewpoints.

  In the first embodiment, a recognition model M [k] is generated by machine learning using each reference voice (voice data DA) classified into the set C [k] among a plurality of reference voices recorded in advance. Therefore, for example, when comparing each recognition model M [k] with an artificially selected configuration without using actual reference speech, it reflects a realistic trend of the relationship between the characteristics of speech and the nature of the singer There is also an advantage that high-precision estimation is possible.

Second Embodiment
A second embodiment of the present invention will be described below. In addition, about the element which an effect | action and function are the same as that of 1st Embodiment in each form illustrated below, the reference | standard referred by description of 1st Embodiment is diverted, and each detailed description is abbreviate | omitted suitably.

  The speech analysis unit 24 according to the second embodiment uses the K recognition models M [generated by the learning process for each of the K sets C [1] to C [K] corresponding to different singing time zones. 1] to M [K] are used to estimate the properties of the analysis target person. That is, the recognition model M [k] expresses a statistical tendency of the reference speech sounded in the k-th time zone out of K pieces.

  Specifically, the index calculation unit 32 of the speech analysis unit 24 applies K singing voices V of the analysis target person to each of the K recognition models M [k] as in the first embodiment. Evaluation indices E [1] to E [K] are calculated. The evaluation index E [k] corresponds to the degree to which the singing voice V matches the statistical tendency of the reference voice sounded in the k-th time zone. The estimation processing unit 34 generates analysis information Q according to the K evaluation indexes E [1] to E [K] calculated by the index calculation unit 32. In the second embodiment, the reference data DR shown in FIG. 7 is used to generate the analysis information Q (estimation of the character of the person to be analyzed) by the estimation processing unit 34.

  As illustrated in FIG. 7, the reference data DR is a data table in which weight value series W [1] to W [R] are registered for a plurality of (R) pieces of analysis information Q [1] to Q [R]. It is. Any one weight value series W [r] (r = 1 to R) is composed of K weight values w [1] to w [1] to w corresponding to different evaluation indices E [k] (time zone of singing time). [K] (K = 10 in the illustration of FIG. 7). In addition, although the case where each weight value w [k] was set to binary (0/1) was illustrated for convenience in FIG. 7, the specific numerical value of each weight value w [k] is arbitrary. For example, the weight value w [k] is not limited to an integer, and the difference between the weight values w [k] is not questioned.

  Each analysis information Q [r] expresses a sentence (comment) of the estimation result of the personality of the person to be analyzed, as in the first embodiment. In the weight value series W [r] of the analysis information Q [r] expressing the character of a singer who tends to sing in a specific time zone, one or more weight values w [k] corresponding to the time zone are obtained. It is set to a large value compared to other time zones. For example, as illustrated in FIG. 7, in the weight value series W [1] of the analysis information Q [1] “nocturnal”, the weight value w [7] to the night time zone (after 20:00) w [10] is set to 1, and the remaining weight w [k] corresponding to the daytime time zone is set to 0. The reference data DR exemplified above is prepared in advance and stored in the storage device 14.

  FIG. 8 is a flowchart of the estimation process in the second embodiment. A part of the operation (SB4-SB6) of the estimation processing unit 34 of the first embodiment (FIG. 6) selects and displays the analysis information Q [r] using the reference data DR in the second embodiment. Replaced by (SC4-SC6).

  When the index calculation unit 32 calculates K evaluation indexes E [1] to E [K] by applying the singing voice V of the analysis target person to each recognition model M [k] (SB3: YES), the estimation process The unit 34 calculates an evaluation index X [r] for each of the R weight value series W [1] to W [R] (each of the R pieces of analysis information Q [1] to Q [R]) ( SC4). Each evaluation index X [r] is a weighted sum (X [r] = w [1] E [1] of each evaluation index E [k] to which each weight value w [k] of the weight value series W [r] is applied. ] + W [2] E [2] + ...... + w [K] E [K]). Therefore, the evaluation index E [k] becomes a larger numerical value as the analysis target person's singing voice V is approximated to the reference voice sounded in the time zone where the weighted value w [k] is large. As understood from the above description, the evaluation index E [k] corresponds to the probability that the singing voice V corresponds to the statistical tendency of the reference voice belonging to the single set C [k], and the evaluation index X [r ] Corresponds to the probability that the singing voice V corresponds to the overall tendency of the reference voices of a plurality of different sets C [k].

  The estimation processing unit 34 selects one piece of analysis information Q [r] corresponding to the evaluation index X [r] from among the R pieces of analysis information Q [1] to Q [R] of the reference data DR (SC5). . Specifically, one piece of analysis information Q [r] corresponding to the weight sequence W [r] in which the evaluation index X [r] has the maximum value among the R pieces of analysis information Q [1] to Q [R]. Is selected. The estimation processing unit 34 acquires the analysis information Q [r] selected from the reference data DR from the storage device 14 and displays it on the display device 18 (SC6).

  In the second embodiment, the same effect as in the first embodiment is realized. In the second embodiment, the evaluation index X [r] is calculated as a weighted sum of the evaluation indices E [k] using K weighted values w [1] to w [K] corresponding to different time zones. Therefore, it is possible to estimate the property (particularly the behavior pattern) of the person to be analyzed from the viewpoint of the statistical tendency of the voice for each time zone.

<Third Embodiment>
The related data DC of each learning data DL in the third embodiment includes a history (hereinafter referred to as “purchase history”) in which a singer of the reference voice has purchased a specific product. The purchase history is extracted from, for example, a service card (for example, a point card) presented to the store by the purchaser of the product. In the classification (SA1) of each reference voice by the learning processing unit 22, the purchase history specified by the related data DC is referred to. Therefore, a recognition model M [k] (a statistical model expressing a statistical tendency of the reference voice of the purchaser of the product) is generated, which is an attribute of the reference speech pronounced by the purchaser who tends to purchase the specific product. .

  The operations of the voice analysis unit 24 (the index calculation unit 32 and the estimation processing unit 34) are the same as those in the first embodiment. For an analysis target person who is estimated to purchase a specific product, the evaluation index E [k] of the recognition model M [k] corresponding to the trend set C [k] is likely to be a large numerical value. Therefore, according to the third embodiment, by referring to the evaluation index E [k], there is an advantage that it is possible to estimate the characteristics of the analysis target person including a tendency to purchase a specific product (consumption trend). In addition, the analysis information Q generated in the third embodiment can be used for effective proposal of specific products for each analysis target person or for market research (marketing) by a store or company. Specifically, the estimation processing unit 34 uses the analysis information Q (estimated result of the person to be analyzed) generated according to the K evaluation indexes E [1] to E [K] for marketing (for advertising purposes). ) In the database. With the above configuration, it is possible to realize efficient advertisement and advertisement using a marketing database. For example, for an analysis subject whose tendency to like a specific item (for example, a hamburger, a drink, an alcoholic beverage, or a food item such as a bowl) is estimated from the analysis information Q, an image that advertises the item is displayed in the time zone between songs. It is displayed on the display device 18. The database for marketing is constructed in a server device that can communicate with the voice analysis device 100, for example. However, it is also possible to construct a database in the storage device 14 of the speech analysis apparatus 100.

<Fourth embodiment>
FIG. 9 is a configuration diagram of the speech analysis system 200 in the fourth embodiment. As illustrated in FIG. 9, the speech analysis system 200 according to the fourth embodiment includes a management device 52 and a plurality of speech analysis devices 54. Each of the plurality of voice analysis devices 54 is realized by a communication terminal such as a communication karaoke device, and communicates with the management device 52 via a communication network 58 such as the Internet.

  The management device 52 is a distribution server device (typically a web server) that includes the learning processing unit 22 similar to the above-described embodiments. The learning processing unit 22 of the management device 52 generates K recognition models M [1] to M [K] by learning processing (FIG. 4) similar to the above-described embodiments, and then distributes the recognition models to each speech analysis device 54. To do. The recognition model M [k] distributed from the management device 52 is transmitted and held to each voice analysis device 54 via the communication network 58.

  Each speech analysis device 54 includes the speech analysis unit 24 similar to the above-described embodiments. The voice analysis unit 24 of each voice analysis device 54 uses the K recognition models M [1] to M [K] distributed from the management device 52 to sing the analysis target person in the same manner as in the above-described embodiments. By analyzing the voice V, the property of the person to be analyzed is estimated and displayed.

  In the fourth embodiment, the same effects as those of the above-described embodiments are realized. In the fourth embodiment, since each recognition model M [k] generated by the management device 52 separate from each speech analysis device 54 is distributed to each of the plurality of speech analysis devices 54, each speech It is not necessary for the analysis device 54 to include the learning processing unit 22. Therefore, there is an advantage that the configuration and processing of the voice analysis device 54 are simplified.

<Fifth Embodiment>
FIG. 10 is a configuration diagram of the speech analysis apparatus 100 according to the fifth embodiment. As illustrated in FIG. 10, the arithmetic processing device 12 of the speech analysis device 100 according to the fifth embodiment executes the program stored in the storage device 14, thereby providing the same elements (learning processing unit) as those in the first embodiment. 22, in addition to the voice analysis unit 24), it functions as a singing evaluation unit 26.

  The singing evaluation unit 26 evaluates the skill of singing by the person to be analyzed by analyzing the singing voice V collected by the sound collecting device 16. Specifically, the singing evaluation unit 26 of the fifth embodiment analyzes the singing voice V by an evaluation method according to the result of the voice analysis unit 24 estimating the property of the person to be analyzed (that is, the analysis information Q). The skill of singing is evaluated, and an index value (hereinafter referred to as “singing evaluation value”) S indicating the evaluation result is calculated. A known technique can be arbitrarily employed for the singing evaluation process itself by the singing evaluation unit 26. In the following description, it is assumed that the singing evaluation value S is set to a larger value as the analysis target person sings better. An image corresponding to the song evaluation value S generated by the song evaluation unit 26 is displayed on the display device 18. For example, the singing evaluation value S and an evaluation comment corresponding to the numerical value are displayed on the display device 18. Specific examples of the evaluation method of the singing voice V according to the analysis information Q are listed below.

[Evaluation Example 1]
The K recognition models M [1] to M [K] include a recognition model M having a “male” attribute and a recognition model M having a “female” attribute. The singing evaluation unit 26 evaluates the singing voice V in the singing evaluation process for men when the analysis information Q indicates the attribute of “male”, while the woman evaluates when the analysis information Q indicates the attribute of “female”. The singing voice V is evaluated by the singing evaluation process.

  For example, in the singing evaluation process, the person to be analyzed accurately matches (or approximates) the pitch of each musical note of a musical piece in a predetermined range related to the pitch (hereinafter referred to as “reference vocal range”). The singing voice V is analyzed so that the singing evaluation value S becomes larger as the singing is performed. The reference vocal range applied to the singing evaluation process for men is located on the lower side compared to the reference vocal range applied to the singing evaluation process for women. Therefore, when the pitch of the singing voice V is high when the analysis information Q indicates the attribute “male”, the numerical value of the singing evaluation value S is larger than when the analysis information Q indicates the attribute “female”. It becomes.

[Evaluation Example 2]
Similarly to the first evaluation example, the K recognition models M [1] to M [K] include a recognition model M having a “male” attribute and a recognition model M having a “female” attribute. The singing evaluation unit 26 makes the evaluation items of the voice quality different when the analysis information Q indicates the attribute of “male” and when the analysis information Q indicates the attribute of “female”. For example, when the analysis information Q indicates an attribute of “male”, the singing evaluation unit 26 determines the degree of masculine voice quality in the singing voice V (for example, a husky voice quality in which the non-harmonic component is dominant over the harmonic component). The singing voice V is evaluated so that the singing evaluation value S increases as the degree) increases. On the other hand, when the analysis result indicates the attribute of “female”, the singing evaluation unit 26 determines the degree of feminine voice quality in the singing voice V (for example, the degree of clear voice quality in which the harmonic component is dominant over the non-harmonic component). ), The singing voice V is evaluated so that the singing evaluation value S becomes a larger numerical value.

  A configuration in which the singing evaluation unit 26 analyzes the voice quality (male voice / female voice) of the singing voice V is also assumed. When the analysis information Q indicates the attribute of “male”, when the singing voice V is analyzed as a male voice (that is, the analysis target estimated using the gender of the analysis target person estimated from the singing voice V and the recognition model M) Singing evaluation value S is large in the case where the person's gender matches), and when the analysis information Q indicates the attribute of “female”, the singing evaluation value S is large when the singing voice V is analyzed as a female voice. The singing evaluation unit 26 evaluates the singing voice V so that On the other hand, when the analysis subject's gender estimated from the singing speech V and the analysis subject's gender estimated using the recognition model M are different, the singing speech is such that the singing evaluation value S is a small value. V is evaluated.

[Evaluation Example 3]
The K recognition models M [1] to M [K] include an “adult” attribute recognition model M and a “children” attribute recognition model M. When the analysis information Q indicates the attribute “child”, the singing evaluation unit 26 sings the singing voice so that the singing evaluation value S is likely to be a larger value than when the analysis information Q indicates the attribute “adult”. Evaluate V. Specifically, when the analysis information Q indicates the attribute “child”, the evaluation criterion is lowered as compared with the case where the analysis information Q indicates the attribute “adult”.

[Evaluation Example 4]
The K recognition models M [1] to M [K] include a plurality of recognition models M having attributes related to age (such as “20s” and “50s”). The singing evaluation unit 26 analyzes the voice quality (husky degree) of the singing voice V, and when the analysis information Q indicates the attribute of “20s”, the singing voice is compared with the case of indicating the attribute of “50s”. The singing voice V is evaluated so that the singing evaluation value S becomes larger as the voice quality of V is husky (specifically, the non-harmonic component is dominant over the harmonic component).

[Evaluation Example 5]
The K recognition models M [1] to M [K] include a recognition model M having a “pops” attribute and a recognition model M having an “enka” attribute. The “pops” attribute recognition model M is generated from speech audio data DA suitable for “pops” music, and the “enka” attribute recognition model M is speech sound suitable for “enka” music. Generated from data DA. The singing evaluation unit 26 evaluates the frequency of singing techniques such as bars, shackles, and stuff in the singing voice V. Compared to “pops”, “enka” tends to use singing techniques such as bars and shackles. In consideration of the above tendency, the singing evaluation unit 26, when the analysis information Q indicates the “enka” attribute, compared to the case where the analysis information Q indicates the “pop” attribute, Set the specific gravity of the evaluation technique high. Therefore, when the analysis information Q indicates the “enka” attribute, the singing evaluation value S is set to a larger value as the singing technique is used more frequently than when the “pops” attribute is indicated. For detection of various singing techniques, a known technique can be arbitrarily adopted. No. 2007 (long tone) and Japanese Patent Application Laid-Open No. 2007-334364 (vibrato, inflection, properties, timing, and shackle) can be suitably employed. The significance of each singing technique is also disclosed in the above publications.

[Evaluation Example 6]
The K recognition models M [1] to M [K] include a recognition model M having a “passionate” attribute and a recognition model M having a “cool” attribute. When the analysis information Q indicates an attribute of “passionate”, the singing evaluation unit 26 determines the resonance (for example, the amplitude of the volume synchronized with the beat point of the music) and the clarity (for example, the harmonic component relative to the inharmonic component). The higher the dominance level, the higher the singing evaluation value S, and when the analysis information Q indicates the attribute of “cool”, the frequency of the singing technique is high and the voice quality is dark (in the high range compared to the low range) The singing voice V is evaluated so that the singing evaluation value S becomes a larger value as the harmonic component becomes superior to the harmonic component).

  The fifth embodiment exemplified above also achieves the same effect as that of the first embodiment. Further, in the fifth embodiment, the skill of singing by the analysis target person is evaluated by an evaluation method according to the analysis result by the voice analysis unit 24 (estimation processing unit 34). Singing evaluation can be realized. The configurations of the second to fourth embodiments can be applied to the fifth embodiment. For example, as illustrated in the fourth embodiment, the learning processing unit 22 may be omitted from the speech analysis device 100 of the fifth embodiment.

  In the above example, the skill of singing of the analysis target person is evaluated by the evaluation method according to the estimation result by the voice analysis unit 24, but the singing evaluation part 26 presents to the analysis target person according to the singing evaluation value S. A configuration in which the evaluation comment is changed according to the analysis information Q can also be adopted. For example, a case where K recognition models M [1] to M [K] include a recognition model M having a “male” attribute and a recognition model M having a “female” attribute, as in Evaluation Example 1 described above. Suppose. When the analysis information Q indicates the attribute of “male”, as illustrated in FIG. 11, an evaluation comment having contents suitable for the evaluation of the male song is displayed on the display device 18 together with the song evaluation value S, and the analysis information When Q indicates the attribute of “female”, as illustrated in FIG. 12, an evaluation comment having a content suitable for evaluating a female song is displayed on the display device 18 together with the song evaluation value S. That is, even if the singing evaluation value S is the same, if the attributes of the analysis information Q are different, the contents of the evaluation comments are different.

  It is also possible to apply information other than the analysis information Q to singing evaluation. For example, the structure which reflects attributes, such as an analysis object person's past song history (the tendency of the music sung in the past) and an analysis object person's age, in song evaluation may be employ | adopted. It is also possible for the song evaluation unit 26 to evaluate the song of the person to be analyzed for each of a plurality of sections obtained by dividing the music on the time axis. The section which is a unit for singing evaluation is a section corresponding to a predetermined number of bars or musical notes, for example. That is, the evaluation result by the singing evaluation unit 26 is updated at any time (for example, every predetermined number of bars or every predetermined number of notes) in parallel with the progress of the singing of the music.

<Modification>
Each of the above forms can be variously modified. Specific modifications are exemplified below. Two or more aspects arbitrarily selected from the following examples may be appropriately combined.

(1) The content of the related data DC of each learning data DL is not limited to the above example. For example, if the singing information DC2 includes whether or not the singer of the reference voice has a tendency to sing first in a group of a plurality of people, for example, the reference voice in which the tendency is designated is “active” (or It is possible to classify it into a set C [k] of attributes called “conspicuous”.

(2) In each of the above-described forms, the case where each recognition model M [k] generated in advance by the learning process is used for the estimation process is exemplified. However, from the user (analysis target person) who confirmed the analysis information Q It is also possible to correct each recognition model M [k] afterwards according to the instruction.

(3) In the second embodiment, the weights of the evaluation indexes E [1] to E [K] to which a plurality of weight values w [1] to w [K] (that is, time series) corresponding to different time zones are applied. Although the sum is calculated as the evaluation index X [r], the attribute corresponding to each weight value w [k] is not limited to the time zone. For example, evaluation is performed using a weighted sum of evaluation indexes E [1] to E [K] to which a plurality (K) of weight values w [1] to w [K] corresponding to different attribute sets C [k] are applied. It is also possible to calculate the index X [r].

  In the second embodiment, the analysis information Q [r] corresponding to the weight sequence W [r] having the maximum evaluation index X [r] is selected, but the K pieces of information calculated by the index calculation unit 32 are selected. It is also possible to specify the analysis information Q from the distribution characteristics of the evaluation indexes E [1] to E [K]. For example, a typical distribution of K evaluation indexes E [1] to E [K] (for example, a statistic such as an average value or a variance value) is represented by each of R pieces of analysis information Q [1] to Q [R]. Of the R pieces of analysis information Q [1] to Q [R], the distribution closest to the distribution of the K evaluation indices E [1] to E [K] (for example, the peak numerical value approximates). It is also possible for the estimation processing unit 34 to select analysis information Q [r] corresponding to a distribution or a distribution whose distribution value approximates.

(4) The voice analysis device 100 can be realized by a server device that communicates with a communication terminal such as a communication karaoke device. For example, the speech analysis apparatus 100 applies the singing speech V received from the communication terminal via the communication network to each of the K recognition models M [1] to M [K] (estimation process), and analyzes the analysis target person. Analysis information Q indicating the analysis result of the property is transmitted to the communication terminal.

(5) The synthesized voice synthesized by the voice synthesis technique can be analyzed as the singing voice V. That is, the analysis target person includes a virtual singer assumed in speech synthesis in addition to a real singer.

(6) The speech analysis apparatus according to each of the above-described embodiments is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to estimation of the characteristics of a person to be analyzed, and a CPU (Central Processing Unit). ) Etc., and can also be realized by the cooperation of a program. The program which concerns on the suitable aspect of this invention uses the recognition model showing the tendency of the singing voice corresponding to the said attribute about each of a plurality of different attributes, and the analysis subject's singing voice becomes the voice of the said attribute. The computer functions as an index calculation unit that calculates an evaluation index that is an index of the corresponding accuracy, and an estimation processing unit that estimates the characteristics of the person being analyzed according to the evaluation index calculated for each of multiple attributes by the index calculation unit Let The program of the present invention can be provided in a form stored in a computer-readable recording medium and installed in the computer. The recording medium is, for example, a non-transitory recording medium, and an optical recording medium (optical disk) such as a CD-ROM is a good example, but a known arbitrary one such as a semiconductor recording medium or a magnetic recording medium This type of recording medium can be included. For example, the program of the present invention can be provided in the form of distribution via a communication network and installed in a computer.

  The present invention is also specified as an operation method (voice analysis method) of the voice analysis device according to each aspect described above. The speech analysis method according to a preferred aspect of the present invention uses, for each of a plurality of different attributes, a recognition model representing a tendency of singing speech corresponding to the attribute, and the singing speech of the analysis target person is that of the attribute. Includes an index calculation process that calculates an evaluation index, which is an index of accuracy corresponding to speech, and an estimation process that estimates the characteristics of the person being analyzed according to the evaluation index calculated for each of multiple attributes in the index calculation process To do.

100, 54... Voice analysis device, 12... Arithmetic processing device, 14... Storage device, 16... Sound collection device, 18. …… Singing evaluation unit, 32 …… Index calculation unit, 34 …… Estimation processing unit, 52 …… Management device.

Claims (8)

For each of a plurality of different attributes, a recognition model representing the tendency of the singing voice corresponding to the attribute is used to calculate an evaluation index that is an index of the probability that the singing voice of the analysis target person corresponds to the voice of the attribute. An index calculation unit to
A speech analysis apparatus comprising: an estimation processing unit that estimates a property of the person to be analyzed according to the evaluation index calculated by the index calculation unit for each of the plurality of attributes. The estimation processing unit estimates the property of the analysis target person according to a combination of two or more attributes selected according to the evaluation index calculated by the index calculation unit among the plurality of attributes. Voice analysis device. The estimation processing unit calculates, for each of a plurality of weight value series corresponding to different estimation results, calculates a weighted sum of the evaluation indexes to which a plurality of weight values included in the weight value series are applied, and The speech analysis apparatus according to claim 1, wherein an estimation result corresponding to a weight value series selected according to the weighted sum among weight value series is specified. The recognition model of each of the plurality of attributes is generated by machine learning using each reference voice of a set corresponding to the attribute among a plurality of sets obtained by classifying a plurality of reference voices for each attribute. Item 4. The voice analysis device according to any one of items 3 to 4. The plurality of reference voices are classified into the plurality of sets according to at least one of a time zone in which the reference voice is pronounced, a place where the reference voice is pronounced, and a character of a speaker of the reference voice. Voice analysis device. The speech analysis apparatus according to any one of claims 1 to 5, wherein the estimation processing unit registers a result of estimating the property of the analysis target person in a marketing database. The speech analysis apparatus according to claim 1, further comprising: a song evaluation unit that evaluates the skill of the person to be analyzed by an evaluation method according to an estimation result by the estimation processing unit. The voice according to any one of claims 1 to 5, further comprising: a singing evaluation unit that presents an evaluation comment according to a result of evaluating the skill of singing by the person to be analyzed and an estimation result by the estimation processing unit. Analysis device.

Images