JP7143591B2 - speaker estimation device - Google Patents

Description

The present invention relates to the technical field of a speaker estimation device for estimating a speaker.

As a device of this type, a device that estimates a speaker by voice recognition is known. For example, Patent Literature 1 discloses a technique of performing speech recognition and speech authentication in parallel, thereby performing speech recognition and speaker identification in parallel. Patent Literature 2 discloses a technique of confirming the identity of a speaker based on whether or not a specific keyword is included in the content of the speech.

As another related technique, Patent Literature 3 discloses a technique of determining whether or not there is an interest in speech content based on the response time to the speech content output from a speaker. Patent Document 4 discloses a technique of determining a user's interest based on a keyword extracted from a conversation and the mental state of the user when the content of the conversation was input. Patent Literature 5 discloses a technique of estimating the occupant composition based on the seating position in the vehicle, the speaker, and the content of the conversation.

JP 2016-071050 A JP 2010-109618 A JP 2017-111493 A JP 2009-294790 A JP 2012-133530 A

In the technique described in the above-mentioned Patent Document 1, the speaker is specified using voice authentication (for example, authentication using voiceprint data). However, it may be difficult to accurately identify the speaker only with voice authentication. In other words, the related art including the above-mentioned Patent Document 1 leaves ample room for improvement in accuracy in terms of accurately identifying the speaker.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a speaker estimation apparatus capable of accurately estimating a speaker.

In one aspect of the speaker estimating device according to the present invention, provision means for providing a topic to a user, acquisition means for acquiring at least one of the content of the user's utterance and voice features with respect to the topic, and the content of the utterance. and an analysis means for analyzing the feature amount of the user based on at least one of the features of the speech, and an estimation means for estimating the personal attribute of the user based on the feature amount.

1 is a block diagram showing the configuration of a speaker estimation device according to this embodiment; FIG. 4 is a flow chart showing the operation flow of the speaker estimation device according to the embodiment; FIG. 11 is a table showing an example of rules for analyzing patterns of endings of words of users; FIG. 4 is a table showing an example of rules for analyzing user conversation length patterns; FIG. 11 is a table showing an example of rules for analyzing user hesitation patterns; FIG. FIG. 4 is a table showing an example of rules for analyzing a user's word repetition pattern; FIG. FIG. 11 is a table showing an example of rules for analyzing characteristic hobbies and word classification patterns; FIG. FIG. 11 is a table showing an example of rules for analyzing POI and word classification patterns; FIG. FIG. 11 is a table showing an example of rules for analyzing classification patterns of restaurants and words; FIG. 4 is a table showing an example of the relationship between words representing user's emotional expressions and scores. It is a table|surface which shows an example of the relationship between the sentence which a user uttered, and a score. 4 is a table showing an example of a specific method of user collation processing;

An embodiment of a speaker estimation device will be described below with reference to the drawings.

<Device configuration>
First, the configuration of the speaker estimation device according to this embodiment will be described with reference to FIG. FIG. 1 is a block diagram showing the configuration of the speaker estimation device according to this embodiment.

As shown in FIG. 1 , the speaker estimation device according to the present embodiment includes an ECU (Electric Control Unit) 100 , a speaker 200 and a microphone 300 . The speaker estimation device is mounted on a vehicle (not shown) and executes processing for estimating (identifying) the driver of the vehicle.

The ECU 100 includes a topic provision unit 110, an utterance acquisition unit 120, a feature amount analysis unit 130, and a user verification unit 140 as processing blocks for realizing the functions.

The topic providing unit 110 is configured to be able to provide topics to the driver of the vehicle via the speaker 200 . That is, the topic providing unit 110 has a function of conversing with the driver of the vehicle. The topic provision part 110 should just provide a topic at random at the stage where the driver cannot be estimated. On the other hand, the topic providing unit 110 may provide a topic in which the user is interested after estimating the driver. The topic providing unit 110 is a specific example of the "providing means" in the appendix described later.

The speech acquisition unit 120 is configured to be able to acquire the speech of the driver of the vehicle via the microphone 300 . More specifically, the utterance acquisition unit 120 acquires the utterance content of the driver of the vehicle (i.e., what kind of content is being spoken) and voice features (for example, voice speed, voice pitch, intonation, etc.). etc.). The utterance contents and voice features acquired by the utterance acquisition unit 120 are configured to be output to the feature amount analysis unit. The utterance acquisition unit 120 is a specific example of the “acquisition unit” in the appendix described later.

The feature amount analysis unit 130 is configured to be able to analyze feature amounts for estimating the user based on the utterance contents and voice features acquired by the utterance acquisition unit 120 . Note that the feature amount analyzed by the feature amount analysis unit 130 and a specific analysis method will be described in detail later. The feature amount analyzed by the feature amount analysis unit 130 is configured to be output to the user verification unit 140 . The feature quantity analysis unit 130 is a specific example of the “analyzing means” in the appendix described later.

The user verification unit 140 is configured to be able to execute user verification processing based on the feature quantity analyzed by the feature quantity analysis unit 130 . That is, the user verification unit 140 has a function of estimating (identifying) who is the current driver of the vehicle. The result of matching by the user matching unit 140 is used to determine the subsequent provision of in-vehicle services to the driver (for example, the content of topics to be provided). The user collation unit 140 is a specific example of the “estimating means” in the appendix described later.

<Flow of operation>
Next, the operation flow of the speaker estimation device according to this embodiment will be described with reference to FIG. FIG. 2 is a flow chart showing the operation flow of the speaker estimation device according to this embodiment.

As shown in FIG. 2, when the speaker estimation device according to the present embodiment operates, it is first determined whether or not the driver has gotten into the vehicle (step S11). When it is determined that the driver is not in the vehicle (step S11: NO), the process of step S11 is executed again after a predetermined period of time.

If it is determined that the driver is in the vehicle (step S11: YES), the topic providing unit 110 starts talking with the driver (step S12). When the conversation starts, topic providing unit 110 randomly provides topics to the driver (step S13).

Subsequently, the speech acquisition unit 120 acquires the content of the driver's speech on the provided topic (step S14). After that, the speech acquisition unit 120 determines whether or not a sufficient amount of speech content for analysis has been acquired (step S15). If it is determined that a sufficient amount of utterance content for analysis has not been acquired (step S15: NO), the utterance acquisition unit 120 continues the process of acquiring the utterance content of the driver. Alternatively, the topic providing unit 110 may provide another topic to the driver in order to acquire new utterance content.

If it is determined that a sufficient amount of speech content for analysis has been acquired (step S15: YES), the feature quantity analysis unit 130 analyzes the feature quantity for estimating the driver based on the acquired speech content. do. Specifically, the feature amount analysis unit 130 analyzes phrases used in the driver's response (step S16), analyzes words used in the driver's response (step S17), and analyzes the driver's emotion with respect to the topic (step S18). Run. A specific method of analyzing the feature amount will be described in detail later.

While performing the above-described processing based on the content of the driver's utterance (that is, steps S14 to S18), the utterance acquisition unit 120 also acquires the features of the driver's speech for the provided topic (step S19). After that, the speech acquisition unit 120 determines whether or not a sufficient amount of speech features for analysis has been acquired (step S20). If it is determined that a sufficient amount of voice features for analysis has not been acquired (step S20: NO), the speech acquisition unit 120 continues the process of acquiring the driver's voice features. Alternatively, the topic providing unit 110 may provide another topic to the driver in order to acquire new utterance content.

If it is determined that a sufficient amount of speech features for analysis has been acquired (step S20: YES), the feature quantity analysis unit 130 acquires feature quantities for estimating the driver based on the acquired speech features. to parse Specifically, the feature amount analysis unit 130 analyzes the tone of the user's voice on the provided topic (step S21). A specific method of analyzing the feature amount will be described in detail later.

After the feature amount analysis unit 130 finishes analyzing the feature amount, the user verification unit 140 executes user verification processing based on the feature amount obtained as the analysis result (step S22). That is, a process for estimating who the current driver of the vehicle is is executed. Specific contents of the user collation processing will be described in detail later.

<Analysis of phrases used in responses>
Next, the analysis of phrases frequently used in responses by users (that is, the process of step S16 in FIG. 2) will be specifically described with reference to FIGS. 3 to 6. FIG. FIG. 3 is a table showing an example of rules for analyzing user's ending patterns. FIG. 4 is a table showing an example of rules for analyzing patterns of conversation lengths of users. FIG. 5 is a table showing an example of rules for analyzing user hesitation patterns. FIG. 6 is a table showing an example of rules for analyzing a user's word repetition pattern.

Note that rules for analyzing phrases used in responses are created in advance for each pattern indicating a predetermined characteristic. A model may also be created by performing machine learning on the created rules. For example, a support vector machine may be used to create a classification model, pre-classify predetermined patterns as correct data, and automatically determine feature quantities from the determined classification. Alternatively, the learning data may be input to a DNN (Deep Neural Network) to automatically output the feature amount.

As shown in FIG. 3, in the analysis of phrases used in responses, patterns of word endings may be analyzed as feature amounts. Specifically, whether or not to use polite words such as "~desu" and "~masu" three times or more within one minute, and dialects such as "~jan" and "~dabe" three times within one minute. It is sufficient to determine whether or not they use more than that, and whether or not they use simple expressions such as "--sho" and "--daro" three or more times within one minute.

As shown in FIG. 4, in the analysis of phrases used in responses, the pattern of the length of conversation may be analyzed as a feature amount. Specifically, whether or not one person talks intently on the same topic (whether or not one person talks for five minutes twice or more within an hour), whether or not one person talks about the same topic (one hour It is sufficient to determine whether or not the speaker speaks twice or more for five minutes at a time within the same period or not, and whether or not the speaker speaks briefly (whether one talk lasts about 20 seconds). In addition, in cases other than the above conditions, it may be judged to be a general average.

As shown in FIG. 5, in the analysis of phrases used in responses, hesitation patterns may be analyzed as feature quantities. Specifically, whether or not you use "eh~" three or more times within one minute, whether or not you use "ah~" three or more times within one minute, and whether or not you stutter three or more times within one minute It should be determined whether or not there is

As shown in FIG. 6, in the analysis of phrases used in responses, patterns of repeated words may be analyzed as feature quantities. Specifically, whether or not the demonstrative words such as "that" and "that" are used three or more times within one minute, and whether or not proper nouns (for example, the name of a baseball player, the name of a place, etc.) are used three times or more within one minute It is sufficient to determine whether or not the pronouns such as "that guy" and "that guy" are used three or more times within one minute. In addition, in cases other than the above conditions, it may be judged to be a general average.

<Analysis of words used in responses>
Next, the analysis of words frequently used in responses by the user (that is, the process of step S17 in FIG. 2) will be specifically described with reference to FIGS. 7 to 9. FIG. FIG. 7 is a table showing an example of rules for analyzing characteristic hobbies and word classification patterns. FIG. 8 is a table showing an example of rules for analyzing POI and word classification patterns. FIG. 9 is a table showing an example of rules for analyzing restaurant and word classification patterns.

Note that rules for analyzing words used in responses are created in advance for each word of a predetermined genre. A model may also be created by performing machine learning on the created rules. For example, a support vector machine may be used to create a classification model, pre-classify predetermined patterns as correct data, and automatically determine feature quantities from the determined classification.

As shown in FIG. 7, in the analysis of the words used in the response, the hobby, which is a feature amount, may be determined from the number of times the word related to the hobby is used. Specifically, if the name of a baseball player or the name of a team is used five times or more within one hour, it may be determined that baseball is a hobby. If the soccer player's name or team's name is used five times or more within one hour, it can be determined that soccer is a hobby. If the name of a place frequently used for hiking is used five times or more within one hour, it may be determined that hiking is a hobby. If the work name of the book is used five times or more within one hour, it may be determined that reading is a hobby. If the name of a place or a tourist attraction that is frequently used for travel is used five times or more within one hour, it may be determined that travel is a hobby.

As shown in FIG. 8, in the analysis of words used in responses, POIs, which are feature amounts, may be determined from the number of times words are used regarding POIs (Point Of Interest). Specifically, if the name of the movie theater is used five times or more within one hour, it may be determined that the movie theater is the POI. If the name of a work exhibited in the museum or the name of the genre of the exhibition is used five times or more within one hour, the museum may be determined to be the POI.

As shown in FIG. 9, in analyzing the words used in the response, the restaurant, which is a feature amount, may be determined from the number of times the words related to the restaurant are used. Specifically, when the name of an Italian restaurant or the name of a dish is used five times or more within one hour, the Italian restaurant may be determined as the feature quantity. If the name of a Japanese restaurant or the name of a dish is used five times or more within one hour, the Japanese restaurant may be determined as the feature quantity.

<Analysis of emotions toward topics>
Next, the analysis of the user's emotion with respect to the topic (that is, the process of step S18 in FIG. 2) will be specifically described with reference to FIGS. 10 and 11. FIG. FIG. 10 is a table showing an example of the relationship between words representing user's emotional expressions and scores. FIG. 11 is a table showing an example of the relationship between sentences uttered by users and scores.

As shown in FIG. 10, when judging an emotion about a topic, a score may be given to words expressing the emotion, and the total value or average value of the scores may be used for determination. Specifically, when using the word "wonderful", "+2", when using the word "worst", "-2", when using the word "so-so" should be given a score of "+1". "+" is the score for positive emotions, and "-" is the score for negative emotions.

As shown in FIG. 11, when judging an emotion about a topic, it may be determined by learning features of sentences by machine learning. For example, it is possible to assign a correct emotion value to a sentence, perform machine learning based on the result, and create a model. For the learning logic at this time, for example, a support vector machine or DNN may be used. According to such a model, a score of "+2" is assigned to a sentence such as "Yesterday's XX was amazing." A score of "-2" is given when using the sentence "The pasta I ate yesterday was the worst." A score of “+1” is given when the sentence “Today's □□□ is quite good” is used.

<Analysis of tone for topic>
Next, the analysis of the tone of the user's voice on the topic (that is, the process of step S21 in FIG. 2) will be specifically described.

Note that the rules for analyzing the tone of voice are created in advance based on the features of the voice. A model may also be created by performing machine learning on the created rules. For example, a support vector machine may be used to create a classification model, pre-classify predetermined patterns as correct data, and automatically determine feature quantities from the determined classification.

More specifically, in tone analysis for a topic, it is sufficient to determine whether the tone of voice is positive or negative based on the speed of speech, pitch (frequency) of speech, intonation, etc. . For example, if the voice is fast and high, it may be determined that the tone is positive. Also, if the voice is slow and low, it may be determined that the tone is negative.

<User verification process>
Next, the user verification process (that is, the process of step S22 in FIG. 2) will be specifically described with reference to FIG. FIG. 12 is a table showing an example of a specific method of user verification processing.

The user collation processing may be performed by creating a determiner using a combination of feature amounts obtained as a result of the analysis processing described above. Note that machine learning may be performed when creating the determiner.

When using the determiner shown in FIG. 14, the frequently used phrase pattern is "use polite language often", the frequently spoken topic is "baseball", the frequently spoken word is "baseball player's name", The user pattern is determined to be "A" when the sentiment of the topic of frequent discussion is "positive" and the tone of the topic is "positive." If the frequently talked topic is "soccer", the frequently talked word is "soccer player name", the sentiment of the frequently talked topic is "negative", and the topic tone is "positive", then the user pattern is: It is determined to be "B". The frequently used phrase pattern is "many simple expressions", the frequently talked about topic is "reading", the frequently talked about word is "title of work", and the frequently talked about topic sentiment is "positive" If the tone is "dark (negative)", the user pattern is determined to be "C".

As described above, according to the speaker estimating device according to the present embodiment, a user pattern (that is, who the driver is is ) is estimated. Therefore, it is possible to estimate who the current vehicle driver (that is, the speaker) is with extremely high accuracy.

<Appendix>
Various aspects of the invention derived from the embodiments described above will be described below.

(Appendix 1)
The speaker estimating device described in appendix 1 includes providing means for providing a topic to a user, acquisition means for acquiring at least one of the content of the user's utterance and voice features for the topic, and the content and voice of the utterance. and an estimating means for estimating personal attributes of the user based on the feature amount.

According to the speaker estimation device described in appendix 1, the feature amount of the user is analyzed based on at least one of the content of the user's utterance and the feature of the voice with respect to the provided topic. Then, the user's personal attribute is estimated based on the analyzed feature amount. Note that the "feature amount" is a parameter for estimating the user's personal attributes, and includes, for example, information on phrases and words that the user frequently uses, the user's emotion and tone of voice regarding the provided topic, and the like. . “Individual attribute” is attribute information for specifying an individual user, and is, for example, information indicating the user's identity (identity).

According to the above-described configuration, the personal attribute is estimated based on the feature amount analyzed from the utterance of the user (that is, the speaker). It is possible to estimate (in other words, identify) the user with higher accuracy.

The present invention is not limited to the above-described embodiments, and can be modified as appropriate within the scope not contrary to the gist or idea of the invention that can be read from the scope of claims and the entire specification. is also included in the technical scope of the present invention.

100 ECUs
110 Topic Providing Unit 120 Speech Acquisition Unit 130 Feature Amount Analysis Unit 140 User Verification Unit 200 Speaker 200 Microphone

Claims (1)

a providing means for providing a topic to a user;
Acquisition means for acquiring the contents of the user's utterances and voice features for the topic;
analysis means for analyzing the emotion on the topic based on the content of the utterance and analyzing the tone of voice on the topic based on the characteristics of the voice ;
a speaker estimation device, comprising: estimation means for estimating personal attributes of the user based on the topic, feelings about the topic, and tone of voice about the topic .

Images