JP5749213B2 - Audio data analysis apparatus, audio data analysis method, and audio data analysis program - Google Patents

Description

  The present invention relates to a voice data analysis device, a voice data analysis method, and a voice data analysis program for detecting a conversation for the purpose of exchange from voice data including speech voice.

  Withdrawal due to the diminished daily communication with the surroundings has become a serious social problem. Being able to identify people with poor communication at an early stage leads to prevention of withdrawal problems.

  In order to grasp a person with poor communication, it is necessary to have a technique for managing an individual's daily face-to-face communication state. Daily communication includes, for example, a non-face-to-face type performed via a personal computer, a mobile phone, and the like, and a face-to-face type performed directly face to face. The former can be grasped by acquiring a log or the like of the communication device. However, since the latter is performed without the communication device, it is necessary to grasp by some means.

  In response to the above-described necessity, a technique for automatically detecting daily face-to-face communication using voice data recorded on a terminal owned by an individual has been proposed. Non-Patent Document 1 defines face-to-face communication as a state in which any two parties are in close proximity and speaking. A terminal equipped with a microphone is held by all the target users, and the proximity state is identified by the correlation of the respective audio data acquired from any two terminals. Further, the average power and pitch of the voice data acquired from the terminal held by the user who is recognized as the proximity state are calculated, and “the utterance likelihood of the terminal owner” is estimated from the former, and “the likelihood of voice” is estimated from the latter. And the conversation state is recognized by using both together.

Masayuki Okamoto, Naoki Ikeya, Junsuke Nishimura, Jun Kikuchi, Kenta Nagata, Masanori Hattori, Sogo Tsuboi, Taira Sasakawa, "Detection of ad hoc conversation based on cross-correlation of terminal speech", Transactions of the Database Society of Japan, Vol. 7, No .1, pp.163-168, 2008

  In the technique of Non-Patent Document 1, there is a problem that even one-way communication or monologue is confused with face-to-face communication. When face-to-face communication is to be detected from the perspective of withdrawal, the challenge is to distinguish one-way communication or monologue from face-to-face communication for exchange. As an example of one-way communication, there is a relationship with a store clerk in the case of accounting at a supermarket cash register. In this scene, the store clerk unilaterally tells the total amount and thank you. In other words, the purpose is to transmit information unilaterally, not to exchange. The cause of the withdrawal state is a decrease in the frequency of face-to-face communication, especially for the purpose of exchange. Therefore, in order to grasp the withdrawal state, it is necessary to be able to distinguish and detect those in the face-to-face communication for the purpose of exchange. In the following, face-to-face communication in a broad sense that does not distinguish between purposes is referred to as “utterance”, and face-to-face communication for the purpose of exchange is referred to as “conversation”.

  The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a voice data analysis apparatus and a voice data analysis method for detecting a conversation for the purpose of alternating current from data obtained by breaking voice data into morphemes. And providing an audio data analysis program.

  In order to achieve the above object, the present invention provides a speech data analysis apparatus, wherein the length of a non-vocal sound segment is determined for data in which speech data of a target user and others is classified into a morpheme segment and a non-vocal speech segment. The non-speech sound segment is determined to be a segment of the utterance interval, and the utterance segment extraction means for extracting the utterance segment, the nick detection unit for detecting the morpheme segment, and the utterance segment extraction For each utterance section cut out by the means, the Aichi utterance discriminating means for discriminating whether the speaker of the morphological section of Aizuchi is the target user or another person, When the speaker includes both the target user and another person, the speaker has a conversation identifying means for identifying the speech section as a conversation section, and the speech recognition means has a voice model of the target user , The speech data corresponding to any morpheme section is compared, and for the morpheme section of the speech data collated with the speech model, the target user is determined to be a speaker, and the speech data not collated with the speech model It is determined that the other person is the speaker for the morphological section of Aizuchi.

  The present invention relates to a speech data analysis method performed by a speech data analysis apparatus, wherein the length of a non-vocal sound segment is the length of data in which speech data of a target user and others is classified into a morpheme segment and a non-vocal speech segment. When the predetermined time is exceeded, the non-speech sound segment is determined as a segment of the utterance interval, an utterance segment extraction step for extracting the utterance segment, an nick detection step for detecting the morpheme segment, and the utterance segment extraction step For each utterance section cut out in step 1, the audible morpheme section in the utterance section and the utterance of the morpheme section in one utterance section are determined. When the person includes both the target user and the other person, a conversation identifying step for identifying the speech section as a conversation section is performed. The user's speech model is compared with speech data corresponding to the speech morpheme section, and for the speech morpheme segment of speech data matched with the speech model, the target user is determined to be a speaker, and the speech It is determined that the other person is the speaker for the morpheme section of speech data that is not collated with the model.

  The present invention is an audio data analysis program for causing a computer to function as the audio data analysis apparatus.

  ADVANTAGE OF THE INVENTION According to this invention, the audio | voice data analysis apparatus, the audio | voice data analysis method, and the audio | voice data analysis program which detect the conversation aiming at alternating current from the data by which audio | voice data were decomposed | disassembled into the morpheme can be provided.

It is a block diagram of the audio | voice data analyzer which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the procedure of the process in 1st Embodiment. It is an example of data decomposed | disassembled into the morpheme in 1st Embodiment. It is an example of data in which the delimiter tag in the first embodiment is set. It is an example of data in which the identification tag in the first embodiment is set. It is an example of data in which the role tag in 1st Embodiment was set. It is an example of data identified as a “conversation” section in the first embodiment. It is explanatory drawing for demonstrating data in 1st Embodiment. It is a block diagram of the audio | voice data analyzer which concerns on the 2nd Embodiment of this invention. It is a flowchart which shows the procedure of the process in 2nd Embodiment. It is an example of the data decomposed | disassembled into the morpheme in 2nd Embodiment. It is an example of data in which the delimiter tag in the second embodiment is set. It is an example of data in which the identification tag in 2nd Embodiment was set. It is an example of data in which the speaker tag in 2nd Embodiment was set. It is an example of data identified as a “conversation” section in the second embodiment.

  Embodiments of the present invention will be described below with reference to the drawings.

  In this embodiment, in order to detect a “conversation” section, which is a face-to-face communication intended for exchange, from speech data including speech speech, the following two points are focused on as features unique to “conversation” .

  One is that “conversation” has two roles, the listener and the speaker. The other is that when you play the role of the listener, you will hit it, but you won't hit it when you are the speaker.

  Based on these, in the present embodiment, the features of “speech” and “conversation” are defined as follows, and by determining the presence or absence of these features, the “conversation” section that is face-to-face communication for the purpose of exchange Is detected.

  “Speech” is a group of utterances continuous in time series. The “conversation” is a “speech” in which both a section that utters Aichi and a section that utters other than Aichi are mixed.

<First Embodiment>
FIG. 1 is a configuration diagram illustrating the configuration of the audio data analysis device 1 according to the first embodiment. The voice data analyzing apparatus 1 shown in the figure includes a voice data receiving unit 1a, a morpheme analyzing unit 1b, a speech segment extracting unit 1c, an Aichi detector unit 1d, an Aichi dictionary unit 1e, a role classification unit 1f, a conversation An identification unit 1g, an output unit 1h, and a storage unit 1i are provided.

  The audio data receiving unit 1a inputs audio data. The audio data input in the first embodiment is assumed to be data in which noise such as environmental sounds is small and only the voice of the target user can be clearly obtained. That is, it is voice data from which the voice of one target user can be acquired. This voice data may include a voice that is so small that it can be ignored when a voice other than the target user (person) is processed, and may also include an environmental sound other than a human voice.

  For example, a mobile terminal having a recording function (voice acquisition function) including a microphone (for example, a directional microphone) is always provided from the neck to the target user. The mobile terminal stores voice data uttered by the target user acquired by the microphone in a storage unit such as a memory of the mobile terminal together with time data (time stamp or the like). Then, the mobile terminal transmits audio data for a predetermined period (for example, one day, several hours, etc.) stored in the storage unit at a predetermined timing to the audio data analysis device 1. The voice data receiver 1a of the voice data analyzer 1 receives voice data transmitted from the mobile terminal, and sends the received voice data to the morpheme analyzer 1b. The voice data analysis apparatus 1 collectively processes voice data for a period that includes a plurality of “utterances” and “conversations”. For example, a mode in which the mobile terminal transfers data for one day to the voice data analysis apparatus 1 while the user is sleeping, and the voice data analysis apparatus 1 processes the transferred voice data may be considered.

  Further, the voice data analyzing apparatus 1 may be a mobile terminal having a recording function including a microphone (for example, a directional microphone). The target user always has the voice data analysis device 1 as a mobile terminal from the head. The voice data receiving unit 1a of the voice data analyzing apparatus 1 may input the target user's voice data acquired by a microphone and stored in a storage unit (not shown) together with time data for a predetermined period.

  The morpheme analysis unit 1b analyzes the speech data into morphemes, and classifies the speech data into morpheme sections including morphemes and other non-voicing sections. When the length of the non-speech sound segment exceeds a predetermined time for the data classified into the morpheme segment and the non-speech sound segment, the utterance segment cutout unit 1c determines that the non-speech sound segment is a delimiter of the utterance segment. Cut out the utterance section.

  The Aichichi detection unit 1d detects the Amorphic morpheme section. Specifically, when the morpheme of each morpheme section matches the any one of the dictionary data in the dictionary unit 1e, it is determined that the morpheme section is the selected morpheme section. In the Aizuchi dictionary unit 1e (Aizuchi storage means), a plurality of Aizuchi data is stored. The nickname dictionary unit 1e is a database that is referred to when the nickname detection unit 1d detects nickname. The exclamations that are frequently used are registered as text data in the dictionary section 1e in advance.

  For each utterance section cut out by the utterance section cutout section 1c, the role classification unit 1f determines that the target user is the section of the listener for the morpheme section, and the target user for the other morpheme sections. Is determined to be a speaker role section. The conversation identifying unit 1g identifies the utterance section as the conversation section when both the section of the listener role and the section of the speaker role are included in one utterance section. The output unit 1h (calculation means) calculates the degree of communication of the target user based on the total time of conversation sections or the number of conversation sections identified by the conversation identification section 1g. In the storage unit 1i, the morpheme section and the non-voicing section included in the voice data are stored in time series together with the start time and the end time.

  The voice data analysis apparatus 1 described above can use, for example, a general-purpose computer system including a CPU, a memory, an external storage device such as an HDD, an input device, and an output device. In this computer system, the CPU executes the program for the voice data analysis apparatus 1 loaded on the memory, whereby each function of the voice data analysis apparatus 1 is realized. The program for the audio data analysis apparatus 1 can be stored in a computer-readable recording medium such as a hard disk, a flexible disk, a CD-ROM, an MO, or a DVD-ROM, or can be distributed via a network.

  Next, the processing of this embodiment will be described.

  FIG. 2 is a flowchart showing a processing flow of the voice data analysis apparatus 1 of the present embodiment.

  In S100, the voice data receiving unit 1a acquires the voice data of the target user and sends it to the morphological analysis unit 1b. As described above, time data is added to the audio data.

  In S101, the morphological analysis unit 1b first converts speech data into text data by performing speech recognition (see Document 1) for the received series of speech data for a predetermined period. Next, the morpheme analysis unit 1b divides the text data into morpheme columns by performing morpheme analysis (see Document 2).

Reference 1: “MeCab: Yet Another Part-of-Speech and Morphological Analyzer”, http://mecab.sourceforge.net/#format
Reference 2: "SpeechRec Solution Package", http://www.ntt-it.co.jp/goods/vcj/v-series/speechrec/solution_package.html
Here, the morpheme analysis unit 1b refers to the start time of each morpheme and sets one morpheme section from the start time of one morpheme to the start time of the next morpheme. From the start time to the end time of a non-speech sound sandwiched between two morpheme sections, where the part not included in the morpheme section is a non-speech sound, such as a scene that is not uttered or a “between” part seen during utterance Is one non-voiced sound interval. Then, the morpheme analysis unit 1b generates a table including the start time and the end time of each of the morpheme section and the non-voiced sound section, and stores the table in the storage unit 1i.

  FIG. 3 is a diagram illustrating an example of the table generated in S101. In the illustrated table, each morpheme and non-voiced sound section (indicated by “◯” in FIG. 3), and their start time and end time are associated with each other and stored in time series. The start time and end time shown in FIG. 3 are expressed in hours: minutes: seconds: milliseconds (h: m: s: ms). The millisecond (ms) is usually three digits, but is represented by two digits here for the sake of simplicity. 4 to 7 and FIGS. 11 to 15 also indicate the start time and the end time as in FIG.

  In S102, the utterance section cutout unit 1c first determines the time during which the utterance is occurring in the voice data input in S100 in order to determine whether or not the conversation is for each scene where the utterance is being performed. Divide by scene. Here, the period from the start time to the end time of one scene where an utterance is occurring is referred to as an “utterance” section. For example, between 10:33 and 10:35, a person asks for directions and guides them, and then meets and talks with a friend from 13:10 to 13:30. In this example, the time from 10:33 to 13:30 is divided into two scenes, the scene where the route was guided and the scene where the conversation was held with a friend. The “speech” section, from 13:10 to 13:30, is the second “speech” section.

  The speech segment cutout unit 1c is generated by the morpheme analysis unit 1b and refers to a non-speech sound segment in a table (see FIG. 3) stored in the storage unit 1i, and is longer than a predetermined time (for example, 5 minutes). The utterance sound section is defined as a “speech” section break, and the morpheme section between the breaks is regarded as one “utterance” section. For example, the utterance section cutout unit 1c updates the table by newly providing a delimiter tag column in the table of the storage unit 1i. Then, for each non-voiced sound section, it is determined whether or not a predetermined time is exceeded. For a non-voiced sound section that exceeds the predetermined time, a delimiter tag “1” is set, and the non-voiced sound does not exceed the predetermined time. Separation tag “0” is set in the section. The delimiter tag “0” is set for all morpheme sections other than the non-voiced sound section.

  FIG. 4 is an example of a table in which a delimiter tag column is provided in the table of the storage unit 1i by the utterance section cutout unit 1c. The “speech” section delimited by the delimiter tag “1” starts from the end time of the non-speech sound segment determined as the previous (previous) delimiter, and the non-speech sound segment determined as the subsequent delimiter Let the start time be the end time. In the example shown in FIG. 4, for the “speech” section delimited by the non-speech sound sections 41 and 42 of the delimiter tag “1”, the start time is the end time of the previous non-speech sound section 41 “13:10”. 03:00 ”and the end time is“ 13: 30: 00: 00 ”which is the start time of the non-voiced sound section 42 later.

  In S103, the nickname detection unit 1d refers to the table stored in the storage unit 1i for each “speech” segment cut out in S102, detects a morpheme corresponding to the morpheme segment, and detects the nickname. Add a tag. As for detection of AIZUCHI, the AIZIC dictionary unit 1e is used in the present embodiment. In other words, the excitement verbs that are frequently used such as “Yes”, “Fun”, “Hey”, “Yes”, “Yes”, “I see”, and so on are registered in the dictionary part 1e as text data in advance. . The Aichichi detection unit 1d uses text matching technology to compare each morpheme with the emotional verb registered in the Aichichi dictionary unit 1e, and matches one of the emotional verbs registered in the Aichichi dictionary unit 1e. A morpheme is detected as an identification, and an identification tag is added.

  FIG. 5 is an example of a table in which a column of the identification tag is provided in the table of the storage unit 1i by the identification detector 1d. The nickname detection unit 1d sets the nickname tag “1” in the morpheme section corresponding to the morpheme section from among the morpheme sections, sets the nickname tag “0” in the other sections, and stores in the storage unit 1i. Update the table. In the example shown in the figure, “Yes” and “I see” morphemes are detected as nicks.

  In S104, the role classification unit 1f determines whether the target user is a listener or a speaker at that time by determining whether the content of the utterance is likely to be obvious. In the present embodiment, the role classification unit 1f refers to the table stored in the storage unit 1i, and within each “utterance” section cut out by the utterance section cutout unit 1c, is a section serving as a listener and a speaker role. The table is updated by adding a role tag.

  Specifically, the role classification unit 1f refers to the delimiter tag of the “utterance” section and the nickname tag set in the table, and the nick tag “1” from the start time to the end time of the “spoken” section. "Are searched in order. The time when the tag “1” is detected immediately after the search is started and the time when the tag “1” is detected for the first time in the state of the speaker role are set as the start of the listener role. Further, the time when the tag “0” is detected immediately after the search is started and the time when the tag “0” is detected for the first time in the state of the listener are set as the start of the speaker. In the “speech” section, it is assumed that the role is always one, and the current role is continued until one of the roles starts. It should be noted that only the identification tag of the morpheme section is referred to, and the identification tag “0” of the non-vocal sound section is ignored. In this way, the role classification unit 1f determines that the target user is the section of the listener role for the Aizuchi morpheme section, and determines that the target user is the section of the speaker role for the other morpheme sections. The determination result is set in the table of the storage unit 1i and updated.

  FIG. 6 is an example of a table in which a role tag column is provided in the table of the storage unit 1i by the role classification unit 1f. In the example shown in the figure, since the tag “1” 61 is detected immediately after the search of the “speech” section is started (13: 10: 00: 00), the role classification unit 1f reads the corresponding morpheme section as a listener. The listener is set in the role tag until the start of the speaker is detected. Then, the role classification unit 1f detects the time point 62 when the tag “0” is first detected in the state of the listener role as the start of the speaker role, and the role tag is detected until the start of the listener is detected. Set up a speaker.

  In S105, the conversation identifying unit 1g sets the role of the speaker to be a speaker or a listener in the “speech” section and the role changes with time, and the condition is satisfied. An utterance interval is identified as a “conversation” interval. The conversation identifying unit 1g refers to the delimiter tag and the role tag of the “utterance” section of the table stored in the storage unit 1i, and both the roles of the listener and the speaker role are mixed in one “utterance” section. If it is, the “speech” section is determined as the “conversation” section.

  FIG. 7 is a table showing an example of the “conversation” section, and the roles of the listener and the speaker are included in one “utterance” section.

  In S106, the output unit 1h refers to the table stored in the storage unit 1i, calculates the degree of communication of the target user, and outputs the calculation result to an output device such as a display. The degree of communication is a numerical value (index) indicating the degree of communication of the target user. Regarding the calculation of the degree of communication, for example, the following various calculation methods are conceivable. The higher the degree of communication, the more active the communication with the surroundings, and conversely, the lower the degree of communication, the less communication. It can be estimated that there is.

For example, the output unit 1h calculates, as the communication degree, the proportion of time during which conversation is performed in the time (for example, one day) of the input voice data. That is, the output unit 1h calculates by dividing the total time of the section identified as the “conversation” section by the conversation identifying unit 1g by the time of the entire audio data.
Further, the output unit 1h calculates the ratio of the number of “conversation” sections to the number of “speech” sections as the degree of communication. That is, the output unit 1h calculates the degree of communication by dividing the number of “conversation” sections identified by the conversation identification unit 1g by the number of “utterance” sections cut out by the utterance section cutout unit 1c.
FIG. 8 is an explanatory diagram for explaining data in the processing of the present embodiment.

  The audio data receiving unit 1a receives (inputs) audio data. The morpheme analysis unit 1b converts speech data into text data by speech recognition, and decomposes it into morpheme sections and non-vocal sections by morpheme analysis. The utterance section cutout unit 1c separates a non-speech sound section exceeding a predetermined length, and cuts out an “utterance” section using a delimiter tag. The nick detection unit 1d detects nick from the morpheme section. The role classification unit 1f classifies the “utterance” section into a listener role and a speaker role. When the “utterance” section includes both the listener role and the speaker role, the conversation identifying unit 1g determines the “utterance” zone as the “conversation” zone.

<Second Embodiment>
FIG. 9 is a configuration diagram showing the configuration of the audio data analysis device 2 in the second embodiment. The voice data analyzing apparatus 2 shown in the figure includes a voice data receiving unit 2a, a morpheme analyzing unit 2b, a speech segment cutout unit 2c, an Aichi detector unit 2d, an Aichi dictionary unit 2e (Aiichi storage means), A speech recognition unit 2f, a conversation identification unit 2g, an output unit 2h (calculation means), a storage unit 2i, and a personal voice model recording unit 2j are provided.

  The voice data analysis apparatus 2 according to the second embodiment is provided with an audible speaker discriminating unit 2f and a person voice model recording unit 2j. The voice data analysis apparatus 1 according to the first embodiment (see FIG. 1). ) Is different. Other voice data receiving unit 2a, morphological analysis unit 2b, utterance section cutout unit 2c, nicking detection unit 2d, nicking dictionary unit 2e, conversation identification unit 2g, output unit 2h and storage unit 2i are the first implementation. Voice data receiving unit 1a, morphological analysis unit 1b, speech segment cutout unit 1c, voice detection unit 1d, voice search dictionary unit 1e, conversation identification unit 1g, output unit 1h and storage unit 1i Since it is the same, description is abbreviate | omitted here.

  In the second embodiment, the audio data acquired by the audio data receiving unit 2a is assumed to be audio data including not only the voice of the target user but also the voice of the other party. For example, a mobile terminal having a recording function including a sound collecting microphone or the like (or a voice data analyzing apparatus 2 that is a mobile terminal having a recording function including a sound collecting microphone) is always carried by the target user. It can be realized with this. This sound data may also include environmental sounds other than the sound of the speaker (target user, conversation partner).

  The Aichichi speaker discriminating unit 2f discriminates whether the speaker in the Amorphous morpheme segment is the target user or another person for each utterance segment extracted by the utterance segment extracting unit 2c. Specifically, the Aichi speaker determination unit 2f compares the target user's voice model recorded in the person's voice model recording unit 2j with the voice data corresponding to the morphological section of Aichi and matches the voice model. It is determined that the target user is the speaker for the speech morpheme section of the voice data, and the other person is the speaker for the speech morpheme section of the speech data that is not matched with the speech model.

  The person voice model recording unit 2j is a database that is referred to when the speaker identification unit 2f performs speaker verification. The speech model of the target user to be referred to is created in advance by learning and registered in the speech model recording unit 2j.

  Note that the conversation identifying unit 2g of the present embodiment, when the speaker of the morpheme section of the previous one includes both the target user and another person in one utterance section, the utterance section is the conversation section. Identify.

  FIG. 10 is a flowchart showing the processing of this embodiment.

  Processing from S200 to S203 is performed in the same manner as S100 to S100 in the first embodiment. That is, in S200, the audio data receiving unit 2a acquires audio data. In S201, the morphological analysis unit 2b converts the speech data into text data, decomposes the converted text data into morphemes, and stores the processing result in the storage unit 2i. FIG. 11 is an example of a table generated by the morphological analysis unit 2b and stored in the storage unit 2i. In the illustrated table, morpheme sections and non-vocal sections are stored in association with their start times and end times.

  In S202, the utterance section cutout unit 2c refers to the table in the storage unit 2i, sets a delimiter tag based on the length of the non-vocal sound section, and cuts out the “speech” section. FIG. 12 shows an example of a table updated by the speech segment cutout unit 2c, to which a delimiter tag is added.

  In S203, the Aichi detector 2d detects the Aichi from the morpheme section of the table of the storage unit 2i using the Aichi dictionary unit 2e, and sets the Aichi tag. FIG. 13 is an example of a table updated by the identification detector 2d, and an identification tag is added.

  In S204, the Aichi speaker determination unit 2f determines whether the speaker of the Aichi (Aizuchi tag “1”) included in the “utterance” section is the target user himself / herself or the other person. At that time, it is determined which of the target user and the other person is the listener. Learn the voice data of the target user in advance to create a voice model, and register it in the person's voice model recording unit 2j. The speech recognition unit 2f performs speaker verification between the speech data corresponding to the morpheme section detected by the identity detection unit 2d and the speech model of the target user registered in the subject speech model recording unit 2j. (Refer to Document 3). The voice data is input in S200 and is stored in a storage unit such as a memory (not shown).

Reference 3: Matsui, Tanabe, “Speaker Identification Using dPLRM”, Statistical Mathematics Research Institute, Special Issue “Computational Reasoning: Modeling, Mathematical, Algorithm”, Vol. 53, No. 2, pp.201-210, 2005.
When the voice data of Aichi is collated with the voice model of the person's voice model recording unit 2j (when collation is obtained), the Aichi speaker determining unit 2f determines that the speaker of the Aichi is the target user. If it is discriminated and the collation is not performed (if collation is not obtained), the speaker at the moment is discriminated from the other person. Then, the identification result “person” or “other” is tagged, and the table stored in the storage unit 2i is updated.

  FIG. 14 shows an example of a table in which a speaker tag column is provided in the table of the storage unit 2i by the speaker identification unit 2f. “Person” or “Other” is set for the speaker tag of the data for which the tag “1” is set.

  In S205, the conversation identifying unit 2g changes the time-speaking speaker with the time transition, such as the person who is hitting a symmetric user or another person in the “speech” section. This is the “conversation” condition, and the section in which this condition is satisfied is identified as the “conversation” section.

  When the roles of the listener and the speaker are switched between the target user and others in one “utterance” section, both play the role of a listener, that is, when both of them utter a whisper, the “speech” section is changed to the “conversation” section. Is determined. The conversation identifying unit 2g refers to the delimiter tag added in S202 and the speaker tag added in S204, and gives both the target user himself / herself and others to the utterer in one “utterance” section. Is mixed, it is determined as a “conversation” section.

  FIG. 15 is a table showing an example of a “conversation” section, in which one and another speaker tags are included in one “utterance” section.

  In S206, as in S106 of the first embodiment, the output unit 2h refers to the table stored in the storage unit 2i, calculates the communication level of the target user using the identified “conversation” section, The calculation result is output to an output device such as a display.

  In the first and second embodiments described above, it is possible to automatically extract the “conversation” section by distinguishing between “utterance” and “conversation”. That is, in the first embodiment, the “speech” section in which the target user plays both roles of “speaker” and “listener” is identified as the “conversation” section, and in the second embodiment, The “speech” section in which the user includes both the target user and the other person is identified as the “conversation” section.

  Thereby, in this embodiment, it is possible to extract only a communication section for the purpose of exchange from various communication (one-way communication, monologue, communication for the purpose of exchange, etc.) by voice. In other words, it is possible to grasp the state of communication with higher accuracy by eliminating communication that is not intended for one-way communication or communication such as monologue. For example, when the local government or the like manages the “conversation” situation of individual users on a daily basis, it is possible to grasp a person with whom daily communication with the surroundings is rare, leading to prevention of withdrawal.

  In the present embodiment, “speech” and “conversation” can also be discriminated for face-to-face communication performed by three or more persons including the target user. For example, in a scene in which “conversation” is being performed by three people, the target user himself / herself cannot enter the talk, and “conversation” may be configured by members other than the target user himself / herself. In the prior art, one-to-one processing is performed and all the users are recognized to be in a conversation state. However, according to the present embodiment, the target user himself / herself only speaks, so this is not regarded as “conversation”. .

  In this embodiment, since processing is performed using voice data that can be acquired from a terminal having one voice acquisition function as an input, only the target user himself / herself can hold the terminal. Therefore, it is easier to introduce than the conventional technique in which not only the target user but also everyone needs to hold the terminal.

  In the present embodiment, the degree of communication of the target user is calculated using the extracted “conversation” section. Thereby, the degree of communication of the target user can be estimated.

  In addition, this invention is not limited to the said embodiment, Many deformation | transformation are possible within the range of the summary.

1: Voice data analyzer
1a: Audio data receiver
1b: Morphological analyzer
1c: Speech segment cutout part
1d: Aizuchi detector
1e: Aizuchi Dictionary
1f: Role classification part
1g: Conversation identifier
1h: Output section
1i: Storage unit 2: Audio data analyzer
2a: Audio data receiver
2b: Morphological analyzer
2c: Speech segment cutout part
2d: Aizuchi detector
2e: Aizuchi Dictionary
2f: Aizuchi speaker discriminator
2g: Conversation identifier
2h: Output section
2i: Memory unit
2j: Personal voice model storage

Claims (8)

For the data in which the voice data of the target user and others is classified into a morpheme section and a non-vocal sound section, if the length of the non-voice sound section exceeds a predetermined time, the non-voice sound section An utterance section cutout means for determining and cutting out the utterance section;
Aizuchi detection means for detecting Aizuchi morpheme sections;
For each utterance section cut out by the utterance section cutout means, an Aichi speaker determination means for determining whether the speaker of the morpheme section is a target user or another person,
A conversation identifying means for identifying a speech section as a conversation section when the speaker of the morpheme section of Aizuchi includes both the target user and another person in one speech section;
The nickname speaker discriminating means compares the target user's voice model with the voice data corresponding to the morpheme section of the nickname, and the target user determines the morpheme section of the voice data matched with the voice model. A speech data analysis apparatus, characterized in that it is determined that the speaker is a speaker and that another person is a speaker for a morpheme section of speech data that is not collated with the speech model. The voice data analysis apparatus according to claim 1,
An audio data analysis apparatus, further comprising a calculation unit that calculates a degree of communication of a target user based on a total time or the number of conversation segments identified by the conversation identification unit. The voice data analysis device according to claim 1 or 2,
The speech data analysis apparatus further comprising: a morpheme analysis unit that analyzes the speech data into morphemes and classifies the speech data into the morpheme section including the morpheme and the other non-vocal sound section. The voice data analysis device according to any one of claims 1 to 3,
A further feature is provided for storing a plurality of identification data.
The speech detection unit determines that the morpheme section is an morpheme section when the morpheme of each morpheme section matches any of the correlation data in the storage unit. Data analysis device. A voice data analysis method performed by a voice data analysis device,
For the data in which the voice data of the target user and others is classified into a morpheme section and a non-vocal sound section, if the length of the non-voice sound section exceeds a predetermined time, the non-voice sound section An utterance section cut-out step for discriminating and cutting out the utterance section;
An Aichi detection step for detecting a morphological section of the Aizichi;
For each utterance section cut out in the utterance section cutout step, an Aichi speaker determination step for determining whether the speaker of the morpheme section is a target user or another person,
In a single utterance section, when the speaker of the morpheme section of Aizuchi includes both the target user and another person, a conversation identifying step for identifying the utterance section as a conversation section is performed.
In the speech recognition determination step, the target user's speech model is compared with speech data corresponding to the speech morpheme section. A speech data analysis method characterized by discriminating that a speaker is a speaker and determining that another person is a speaker for a morpheme section of speech data that is not collated with the speech model. The voice data analysis method according to claim 5,
An audio data analysis method, further comprising a calculation step of calculating a degree of communication of the target user based on the total time of conversation sections or the number of conversation sections identified in the conversation identification step. The voice data analysis method according to claim 5 or 6,
The voice data analysis device further includes an identification memory unit that stores a plurality of identification data,
In the speech detection step, if the morpheme of each morpheme section coincides with any of the correlation data in the storage unit, the morpheme section is determined to be the morpheme section of the speech. Data analysis method.
  An audio data analysis program for causing a computer to function as the audio data analysis apparatus according to any one of claims 1 to 4.

Images