JP6801587B2 - Voice dialogue device - Google Patents

Description

The present invention relates to a voice dialogue device.

Patent Document 1 describes that a prosodic feature is extracted from a voice waveform of a user's utterance, and whether or not the voice dialogue system has a utterance right is determined based on the prosodic feature.

Japanese Unexamined Patent Publication No. 2016-038501

The voice dialogue system described in Patent Document 1 outputs a filler (speaking) even when it is unclear whether or not the voice dialogue system has the right to speak, and claims that the voice dialogue system has the right to speak. In some cases. However, in reality, if the voice dialogue system does not have the right to speak, outputting the filler may result in interrupting the user's speech.

The present invention has been made to solve such a problem, and an object of the present invention is to provide a voice dialogue device capable of reducing the possibility of interrupting a user's utterance.

The voice dialogue device according to the present invention includes an input unit for inputting a user's voice and an output unit for outputting voice to the user. Further, the voice dialogue device includes a user voice analysis unit that extracts a feature amount from at least the voice waveform of the voice input by the input unit, and the output unit when the output unit outputs a filler to the user. A value indicating the reliability of outputting the filler is calculated based on the feature amount extracted by the user voice analysis unit, and the sound pressure level of the filler is determined based on the value indicating the reliability of outputting the filler. It is equipped with a sound pressure level determination unit.

According to the voice dialogue device according to the present invention, the sound pressure level determination unit calculates a value indicating the reliability of outputting the filler based on the feature amount extracted by the user voice analysis unit, and outputs the filler. The sound pressure level of the filler is determined based on the value indicating the reliability. Therefore, the output unit can output the filler at a sound pressure level corresponding to a value indicating the reliability of outputting the filler. The reliability of outputting the filler has a positive correlation with the possibility that the right to speak is in the voice dialogue device. That is, even when it is unclear whether or not the utterance right belongs to the voice dialogue device, the output unit can output the filler at a sound pressure level corresponding to the possibility that the utterance right belongs to the voice dialogue device. Therefore, when it is unlikely that the voice dialogue device has the right to speak, the output unit outputs the filler at a small sound pressure level, and it is possible to reduce the interruption of the user's utterance by the filler output by the output unit. This makes it possible to provide a voice dialogue device capable of reducing the possibility of interrupting the user's utterance.

It is a block diagram which shows the schematic structure of the robot which concerns on Embodiment 1 of this invention. It is a graph which shows the end part of the utterance section of the user which concerns on Embodiment 1 of this invention. It is a table which shows an example of each element of the feature quantity vector which concerns on Embodiment 1 of this invention. It is a figure explaining the creation of a subset in the creation of the determination model by offline learning which concerns on Embodiment 1 of this invention. It is a figure explaining the generation of a branch function candidate in the creation of the determination model by offline learning which concerns on Embodiment 1 of this invention. It is a figure explaining the determination of the branch function candidate in the creation of the determination model by offline learning which concerns on Embodiment 1 of this invention. It is a figure explaining the determination of the reliability which outputs a filler using the determination model which concerns on Embodiment 1 of this invention. This is an example of a graph showing the relationship between the user utterance length, which is a part of the feature vector according to the first embodiment of the present invention, and the frequency of silence or filler. This is an example of a table showing the relationship between the user utterance length, which is a part of the feature vector according to the first embodiment of the present invention, and the ratio of outputting the filler. It is a figure which shows an example of the conversation between a user and a robot which concerns on Example 1 of this invention. It is a table which shows an example of each element of the feature quantity vector which concerns on Example 1 of this invention. It is a graph which shows the voice waveform of the filler output by a robot in the example of conversation shown in FIG.

Embodiment 1
Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a schematic configuration of a robot 100 as a voice dialogue device according to the first embodiment of the present invention. As shown in FIG. 1, the robot 100 according to the first embodiment includes an input unit 110, an output unit 120, and a control unit 130. The control unit 130 includes a user voice analysis unit 131, a sound pressure level determination unit 132, and the like. The sound pressure level determination unit 132 includes a method selection unit 132A, a sound pressure level setting parameter calculation unit 132B, a learning database (learning DB) 132C, an offline learning unit 132D, a judgment model and a conditional branching database (judgment model and conditional branching). Formula DB) 132E and voice synthesis unit 132F are provided. The robot 100 outputs a voice as a response or outputs a filler according to the utterance of the user. Here, the utterance is a voice having a meaning as the content of the dialogue. Further, the filler is an aizuchi, and is a voice of a connection between one utterance and the next utterance of the user.

The input unit 110 includes a microphone or the like, collects the user's voice, and inputs the user's voice to the user voice analysis unit 131.

The output unit 120 includes a speaker or the like, and outputs a voice uttered to the user from the robot 100 or outputs a filler. Specifically, the voice synthesized by the voice synthesis unit 132F described later is output. Further, in the present embodiment, the voice synthesis unit 132F synthesizes a sound pressure level filler based on the sound pressure level setting parameter Iv input from the sound pressure level determination unit 132, and outputs the filler to the output unit 120. The details of the voice synthesis unit 132F will be described later.

The control unit 130 includes a CPU (not shown), a storage unit (not shown), and the like. Then, when the CPU executes the program stored in the storage unit, all the processing in the control unit 130 is realized.
Further, the program stored in each storage unit of the control unit 130 includes a code for realizing the processing in each of the control units 130 by being executed by the CPU. The storage unit is configured to include, for example, this program and an arbitrary storage device capable of storing various information used for processing in the control unit 130. The storage device is, for example, a memory or the like.

Specifically, when the CPU executes the program stored in the storage unit, the control unit 130 functions as the user voice analysis unit 131 and the sound pressure level determination unit 132. Further, the storage unit stores a learning database 132C, a determination model, a conditional branching database 132E, and the like.

The user voice analysis unit 131 extracts the feature amount from the voice waveform of the voice input by the input unit 110. Further, the user voice analysis unit 131 extracts the feature amount from the history information of the response of the robot 100 to the user (past history of the device response). Then, the user voice analysis unit 131 generates a feature amount vector using the feature amount extracted from the user's voice waveform and the history information of the response to the user, and outputs the feature amount vector to the sound pressure level determination unit 132.

Specifically, the user voice analysis unit 131 divides the voice waveform of the voice input by the input unit 110 into one or more "utterance sections". Here, the "utterance section" means a section from the beginning to the end of the user's utterance, and the user voice analysis unit 131 determines from where to where the "utterance section" is based on the sound pressure of the user's utterance. decide. FIG. 2 is a graph showing the end portion of the user's utterance section, where the vertical axis represents sound pressure (dB) and the horizontal axis represents time. For example, as shown in FIG. 2, the user voice analysis unit 131 again exceeds the sound pressure threshold and becomes zero at a certain time T from the time when the sound pressure of the user's voice becomes smaller than the predetermined sound pressure threshold. When is N times or less, the relevant time point (the time point when the sound pressure of the user's voice becomes smaller than the predetermined sound pressure threshold) is detected as the end of the "speech section". In FIG. 2, T is 400 msec (milliseconds) and N is 0, but the values of T and N are values appropriately set depending on the experimental object and the environment. Similarly, in the user voice analysis unit 131, the number of times that the sound pressure of the user's voice becomes smaller than the sound pressure threshold and becomes zero again is N ₂ times in T ₂ for a certain period of time from the time when the sound pressure of the user's voice becomes larger than the predetermined sound pressure threshold. If it is the above, the time point (the time point when the sound pressure of the user's voice becomes larger than a predetermined sound pressure threshold) is detected as the start of the "speech section". Similarly, the values of T ₂ and N ₂ are values appropriately set depending on the experimental object and the environment.

Then, the user voice analysis unit 131 extracts the feature amount from the voice waveform of the i-th (i is an integer of 1 or more) “speech section”. Further, the user voice analysis unit 131 generates a feature amount vector vi by using the feature amount extracted from the voice waveform of the i-th "utterance section" and the history information of the response to the user, and the sound pressure level determination unit 132. Output to.

FIG. 3 shows an example of the feature amount vector vi generated by the user voice analysis unit 131. Specifically, FIG. 3 shows each element of the feature vector vi and the value of the element. As shown in FIG. 3, each element of the feature amount vector vi is roughly classified into one belonging to the "i-th user utterance information" and one belonging to the "past history of device response". In the example shown in FIG. 3, as the elements belonging to the "i-th user utterance information", "phrase end ○ msec", "whole utterance section", "length of user utterance" and the like are listed. Further, as elements belonging to the "past history of device response", "duration of the same topic", "previous system utterance type", "previous system question type" and the like are listed. In FIG. 3, the value of each element is shown in the column surrounded by the thick frame. Further, in FIG. 3, the “system” refers to the robot 100.

In FIG. 3, the “phrase end ○ msec” means an utterance from the end of the user's utterance section to ○ msec (○ millisecond) before. In FIG. 3, the fundamental frequency f0 and the volume of the voice waveform of “end of phrase ○ msec” are listed as elements of the feature vector vi. Further, the fundamental frequency f0 of the voice waveform of "phrase end ○ msec", the average value of the volume, the variance value, the value of the slope of increase / decrease, the maximum value, etc. are set as the values of the elements of the feature vector vi. The volume is the volume (dB) of the voice of the user's utterance input from the input unit 110. The user voice analysis unit 131 normalizes the average value, the dispersion value, the increase / decrease slope value, and the maximum value of the fundamental frequency f0 and the volume for each user.

Further, in FIG. 3, the fundamental frequency f0 and the volume of the voice waveform of the “whole utterance section” are listed as the elements of the feature amount vector vi. Further, the fundamental frequency f0 of the voice waveform of the “whole utterance section”, the average value of the volume, the dispersion value, the value of the slope of increase / decrease, the maximum value, and the like are set as the values of the elements of the feature vector vi. The user voice analysis unit 131 normalizes the average value, the dispersion value, the increase / decrease slope value, and the maximum value of the fundamental frequency f0 and the volume for each user.

Further, in FIG. 3, "length of user utterance" is mentioned as an element of the feature amount vector vi. Further, the numerical value (sec) of the "user utterance length", that is, how many seconds the user utterance length was, is used as the value of the element of the feature amount vector vi. The "user utterance length" is the length (time (sec)) of the "utterance section" determined by the user voice analysis unit 131 by the above method.

Further, in FIG. 3, "duration of the same topic" is listed as an element of the feature vector vi. Further, the numerical value (sec) of the "duration of the same topic", that is, how many seconds the duration of the same topic was, is used as the value of the element of the feature vector vi. The "duration of the same topic" is, for example, the time from the time when the robot 100 outputs the voice of the "next topic guidance" last time to the time when the robot 100 outputs the voice of the "next topic guidance" this time. .. Further, the robot 100 outputs the voice of "next topic guidance" when, for example, the silence time of the user is longer than a predetermined time or when a predetermined time or more has passed since the last time when the voice of "next topic guidance" was output. To do.

Further, in FIG. 3, "the utterance type of the immediately preceding system" is listed as an element of the feature vector vi. Here, the "system" refers to the robot 100. Further, the "speech type of the immediately preceding system" is "aizuchi", "listening response", "question", etc., which are the values of the elements of the feature vector vi. It should be noted that "aizuchi", "listening response", "question" and the like are represented by discrete values such as "0", "1" and "2", respectively. That is, the user voice analysis unit 131 also digitizes the feature quantities that are not represented by numerical values, such as "aizuchi", "listening response", and "question", by expressing them as discrete values.

Further, in FIG. 3, the "question type of the immediately preceding system" is listed as an element of the feature vector vi. Further, the "question type of the immediately preceding system" is "Fukahori question", "next topic guidance", etc., which are the values of the elements of the feature vector vi. Similar to the above, the “Fukahori question”, the “next topic guidance”, etc. are expressed by discrete values such as “0” and “1”, respectively. That is, the user voice analysis unit 131 also digitizes the feature quantities that are not represented by numerical values, such as "Fukahori question" and "next topic guidance", by expressing them as discrete values.

Then, the user voice analysis unit 131 generates the feature amount vector vi from the elements shown in FIG. 2 and the values of the elements. The feature amount vector vi generated from the element and the value of the element shown in FIG. 2 is represented by, for example, vi = (..., 2.4, ..., 20, 1, 1, ...). ..

The sound pressure level determination unit 132 calculates a value indicating the reliability of outputting the filler based on the feature amount extracted by the user voice analysis unit 131 when the output unit 120 outputs the filler to the user. The sound pressure level of the filler is determined based on the value indicating the reliability of outputting the filler.
Specifically, the sound pressure level determination unit 132 calculates a value Iv indicating the reliability of outputting the filler based on the feature amount vector vi input from the user voice analysis unit 131, and sets the value Iv as “sound”. Pressure level setting parameter ". In the present embodiment, the sound pressure level setting parameter Iv determined by the sound pressure level determination unit 132 satisfies 0.5 ≦ Iv ≦ 1. That is, the sound pressure level setting parameter calculation unit 132B, which will be described later, determines that the value of Iv is 0.5 when the value of Iv obtained by the calculation is less than 0.5.
Then, the sound pressure level determination unit 132 outputs the determined sound pressure level setting parameter Iv to the voice synthesis unit 132F.

More specifically, the sound pressure level determination unit 132 includes a method selection unit 132A, a sound pressure level setting parameter calculation unit 132B, a learning database (learning DB) 132C, an offline learning unit 132D, a judgment model, and a conditional branching database. (Determining model and conditional branching DB) 132E is provided.

The method selection unit 132A selects a method in which the sound pressure level setting parameter calculation unit 132B calculates the sound pressure level setting parameter Iv. In the present embodiment, the sound pressure level setting parameter calculation unit 132B uses the determination model to calculate the sound pressure level setting parameter Iv (hereinafter referred to as “first method”) and the feature amount vector vi. The sound pressure level setting parameter Iv is calculated using the conditional branching equation created by using some of the features of the above (hereinafter referred to as the "second method"), or the sound is used. Calculate the pressure level setting parameter Iv. Therefore, the method selection unit 132A selects whether the sound pressure level setting parameter calculation unit 132B uses the first method or the second method to calculate the sound pressure level setting parameter Iv. Specifically, the method selection unit 132A uses either the first method or the second method based on the amount of data used to calculate the sound pressure level setting parameter Iv, the specifications of the robot 100, and the like. Select whether. Then, the method selection unit 132A outputs an instruction as to which of the first method and the second method is used to calculate the sound pressure level setting parameter Iv to the sound pressure level setting parameter calculation unit 132B.

The sound pressure level setting parameter calculation unit 132B reads either the judgment model or the conditional branching formula from the judgment model and the conditional branching formula database 132E according to the instruction input from the method selection unit 132A. Then, the sound pressure level setting parameter calculation unit 132B calculates the sound pressure level setting parameter Iv based on the feature amount vector vi input from the user voice analysis unit 131 by using the determination model or the conditional branching expression. The details of the determination model and the conditional branching formula will be described later.

The learning database 132C stores data necessary for creating a determination model and a conditional branching expression. Specifically, the learning database 132C stores the voice data of the simulated dialogue collected in advance. In addition, a feature vector and a teacher label are attached to each utterance included in the voice data. More specifically, the voice waveform of the voice of the simulated dialogue is divided into one or more utterance sections by the above-mentioned method or the like, and the feature amount vector and the teacher label are attached to the voice waveform of each utterance section. Here, the feature vector attached to the i-th (i is an integer of 1 or more) utterance section is vi, and the teacher label is ci. That is, the learning database 132C associates the voice waveform of the i-th utterance section of the simulated dialogue collected in advance with the feature amount vector vi of the utterance section and the teacher label ci of the utterance section. It is stored.

The offline learning unit 132D reads the data necessary for creating the judgment model from the learning database 132C, creates the judgment model, and outputs the created judgment model to the judgment model and the conditional branching database 132E. Further, the offline learning unit 132D reads the data necessary for creating the conditional branching expression from the learning database 132C, creates the conditional branching expression, and uses the created conditional branching expression in the determination model and the conditional branching expression database 132E. Output. Before the robot 100 actually starts the dialogue with the user, the offline learning unit 132D creates a determination model and a conditional branching expression, and the determination model and the conditional branching expression database 132E creates the determination model and the conditional branching expression. It is stored.

First, the creation of the determination model by the offline learning unit 132D will be described. Here, a method in which the offline learning unit 132D creates a random forest as a determination model will be described. As shown in FIG. 4, the offline learning unit 132D reads data necessary for creating a determination model from the learning database 132C, and prepares a sample set S having a plurality of samples. Specifically, the i-th sample is data including the voice waveform of the i-th utterance section, the feature amount vector vi attached to the utterance section, and the teacher label ci attached to the utterance section. .. In FIG. 4, the hatched sample contains the teacher label ci (filler) and the non-hatched sample contains the teacher label ci (silence). Further, the teacher label ci (filler) is a label that outputs the filler, and the teacher label ci (silence) is a label that silences. Then, as shown in FIG. 4, the offline learning unit 132D is a subset S _j (j is an integer of 1 or more and T or less) of T sample sets S (T is an integer of 1 or more). ), Randomly divide. Here, T corresponds to the number of decision trees included in the random forest. Incidentally, the offline learning unit 132D is a subset of the sample set S S _1, S _2, when divided into · · · S _T, there is a sample containing one sample different subsets or are distributed respectively, not be distributed to any of the subset You may do it.

Next, the offline learning unit 132D generates a branch function for each node of the decision tree of the random forest. For example, as shown in FIG. 5, the offline learning unit 132D uses the feature amount vector vi of the sample included in the subset _Sj before branching to determine the type of feature amount and the feature amount that are elements of the branching function _fk . K combinations with the threshold value (k is an integer of 1 or more) are randomly selected to generate the branch function f _k . As for the number of candidates k, assuming that the number of features included in the feature vector vi is m (m is an integer of 1 or more), it is desirable that the following equation (1) is satisfied.
In the example shown in FIG. 5, since the feature amount vector vi includes 17 types of feature amounts (m = 17), k is about 4. Therefore, in FIG. 5, the offline learning unit 132D is, for example, the type and threshold value of the first feature amount (0.4), the type and threshold value of the third feature amount (0.3), and the fifth feature amount. Four combinations of the type and threshold value (0.6) and the 17th feature type and threshold value (0.4) are selected to generate the branch function _fk . Further, in the example shown in FIG. 5, the generated branch function is represented by the following equation (2). In the equation (2), x _m means the value of the m-th feature quantity.

Next, the offline learning unit 132D calculates the entropy of each node in the T decision tree of random forest, calculates the information gain I _j. For example, in the example shown in FIG. 6, the entropy H (S _j ) of the parent node having the subset S _j is expressed by the following equation (3).
Entropy H of the branch to the left of the candidate node 1 ^(S _{L 1)} and of the branch to the right H ^(S _{R 1)} is represented by the following formulas (4) and (5),
Entropy H of the branch to the left of the candidate node 2 ^(S _{L 2)} and of a branch to the right H ^(S _{R 2)} is expressed by the following equation (6) and (7).
The information gain I _j of each terminal node is expressed by the following equation (8). In equation (8), H (S _j ) means the entropy before branching (that is, the entropy of the parent node), H (S _l ) means the entropy of branching to the left, and H (S _r). ) Means the entropy of the branch to the right.
Therefore, in the example shown in FIG. 6, the information gain I ₁ of the node of candidate ₁ is represented by the following equation (9).
For information gain I ₂ nodes of the candidate 2, represented by the following formula (10),
The information gain I ₂ of the candidate 2 is larger than the information gain I ₁ of the candidate 1. Then, the offline learning unit 132D determines the shape of the decision tree so that the information gain of each decision tree is maximized. That is, in the example shown in FIG. 6, the branch from the parent node to the candidate 2 is selected. In other words, the offline learning unit 132D classifies the subset Sj (subset before classification) of the parent node so that the information gain is maximized. Then, the offline learning unit 132D outputs the T number of decision trees of the random forest created in this way to the determination model and the conditional branching database 132E as the determination model.

Next, a method of determining the sound pressure level setting parameter Iv using the determination model created as described above will be described. First, the sound pressure level setting parameter calculation unit 132B reads out the determination model and the random forest (determination model) stored in the conditional branching database 132E. Next, as shown in FIG. 7, the sound pressure level setting parameter calculation unit 132B has T number of decision trees (treet) based on the feature amount vector vi of the i-th utterance section input from the user voice analysis unit 131. _1. Traversal ..., treat _T ). Then, the sound pressure level setting parameter calculation unit 132B gives a value indicating the ratio of the subset of the parent nodes to the terminal nodes of each decision tree reached during learning by the offline learning unit 132D as a conditional probability. Obtained as _PT (c | v). In addition, c in the _PT (c | v) is a "label for outputting a filler". Then, as shown in FIG. 7, the sound pressure level setting parameter calculation unit 132B calculates the average value P (c | v) of the conditional probabilities _PT (c | v) obtained in each of the T decision trees. , The sound pressure level setting parameter Iv. The average value P (c | v) of the conditional probability P _T (c | v) is represented by the following equation (11). The average value P (c | v) is a value obtained by integrating the results of identifying the feature vector vi by each of the T decision trees in the random forest.
Then, the sound pressure level setting parameter calculation unit 132B outputs the determined sound pressure level setting parameter Iv to the voice synthesis unit 132F.

Next, the creation of the conditional branching expression by the offline learning unit 132D will be described. The offline learning unit 132D reads the data necessary for creating the conditional branching expression from the learning database 132C, and creates the conditional branching expression. Specifically, the offline learning unit 132D creates a decision tree of a random forest in the same manner as described above.

Next, the offline learning unit 132D reads out the feature quantity vector vi of the i-th speech section of the simulated dialogue collected in advance from the learning database 132C, and reads the feature quantity vector vi of the feature quantity vector vi element and the feature quantity of the element. Traverse the random forest decision tree based on the value. For example, the offline learning unit 132D traverses the decision tree of the random forest based on the element "length of user utterance" of the feature vector vi and the value of the feature of the element, thereby showing the element ". Create a graph showing the relationship between "user utterance length" and the frequency of silence or filler. In FIG. 8, the vertical axis indicates the frequency of outputting silence or filler, and the horizontal axis indicates the user utterance length (sec). The unit "sec" means "second".

Further, the offline learning unit 132D creates a table showing the relationship between the element "user utterance length" shown in FIG. 9 and the ratio of outputting the filler from the graph shown in FIG. For example, when the length of the user's utterance is 0 seconds or more and less than 1 second, the ratio of outputting the filler is 2 ÷ (20 + 2) = 0.11 from FIG. However, since 0.5 ≦ Iv ≦ 1, the ratio of outputting the filler when the length of the user utterance is 0 seconds or more and less than 1 second is set to 0.5. Further, when the length of the user's utterance is 3 seconds or more and less than 4 seconds, the ratio of outputting the filler is 12 ÷ (12 + 4) = 0.75 from FIG. Then, the offline learning unit 132D sets the ratio of outputting the filler to the sound pressure level setting parameter Iv, and creates a conditional branching equation from the graph shown in FIG. 8 and the table shown in FIG. In the case of the examples shown in FIGS. 8 and 9, the conditional branching equation is represented by the following equation (12). In the following equation (12), t is the length (sec) of the user's utterance.
Then, the offline learning unit 132D outputs the conditional branching expression created in this way to the determination model and the conditional branching expression database 132E.

Next, a method of determining the sound pressure level setting parameter Iv using the conditional branching equation created as described above will be described. First, the sound pressure level setting parameter calculation unit 132B reads the conditional branch expression from the discrimination model and the conditional branch expression database 132E, and the feature quantity vector vi of the i-th utterance section input from the user voice analysis unit 131 and the condition. The sound pressure level setting parameter Iv is calculated based on the branching equation. For example, when the feature amount of the element "user utterance length" included in the feature amount vector vi of the i-th utterance section input from the user voice analysis unit 131 is 2.4 (sec), the sound pressure level setting parameter. The calculation unit 132B determines the value of the sound pressure level setting parameter Iv to be 0.69 based on the conditional branching equation represented by the equation (12).
Then, the sound pressure level setting parameter calculation unit 132B outputs the determined sound pressure level setting parameter Iv to the voice synthesis unit 132F.

The value of Iv determined by the sound pressure level determination unit 132 is prepared by the robot 100 so that the user and the robot 100 can interact with each other even if the same feature amount vector vi is input from the user voice analysis unit 131. The value will differ depending on the content of the question being asked, the language used by the user and the robot 100, and the like. In other words, the offline learning unit 132D has a determination model and a conditional branching formula in advance for each content such as dialogue between the user and the robot 100, question content prepared by the robot 100, language used by the user and the robot 100, and the like. To create.

Next, a method in which the voice synthesis unit 132F determines the sound pressure level according to the sound pressure level setting parameter Iv input from the sound pressure level determination unit 132 will be described.
For example, when the sound pressure level setting parameter Iv = 0.5, the voice synthesis unit 132F is output from the robot 100 in the user's ear in an environment where the distance between the user and the robot 100 is 50 cm and the room is a private room in the hospital. When the volume of the voice is 50 dB and the sound pressure level setting parameter Iv = 1, the sound pressure level capable of synthesizing the voice such that the volume of the voice output from the robot 100 in the user's ear is 60 dB under the environment. Stores the calculation formula. For example, the sound pressure level calculation formula is expressed by the following formula (13). In the following equation (13), P is an adjustment variable in the speech synthesis unit 132F.
That is, when Iv is a value close to 0.5, the reliability of outputting the filler is low, so that the voice synthesis unit 132F sets the sound pressure level of the filler to a relatively small sound pressure level (for example, 50 dB). On the other hand, when Iv is a value close to 1, the reliability of outputting the filler is high, so that the voice synthesis unit 132F sets the sound pressure level of the filler to a relatively large sound pressure level (for example, 60 dB).

Alternatively, the sound synthesis unit 132F stores in advance a table in which the sound pressure level setting parameter Iv and the sound pressure level (dB) are associated with each other, and the sound pressure level setting parameter Iv input from the sound pressure level determination unit 132. And the table, the sound pressure level at which the filler is output may be determined.

Example 1
Next, an example of dialogue between the robot 100 and the user according to the first embodiment will be described as the first embodiment. FIG. 10 shows an example of a conversation between the user and the robot 100.
For example, when the user's i-th utterance content is the question "Have you been to Robo-kun?" In FIG. 10 (when the user is asking the robot 100), the feature amount of the i-th utterance section. The values of each element and the feature amount of the vector vi are as shown in FIG. 3, and the feature amount vector vi of the i-th utterance section is vi = (..., 2.4, ..., 20, 1, 1, ...). Then, the value of Iv calculated by the sound pressure level setting parameter calculation unit 132B using the feature amount vector vi and the above-mentioned determination model (random forest) is 0.7 (see FIG. 7). Further, the value of Iv calculated by the sound pressure level setting parameter calculation unit 132B using the feature amount vector vi and the above-mentioned conditional branching equation is 0.69 (see FIG. 9).
Then, the sound synthesis unit 132F is 60 dB based on the sound pressure level setting parameter Iv and the above-mentioned sound pressure level calculation formula or a table in which the sound pressure level setting parameter Iv and the sound pressure level (dB) are associated with each other. The filler "um" is synthesized at a sound pressure level close to, and output to the output unit 120.

On the other hand, when the i-th utterance content of the user is the utterance "I went to the park yesterday" in FIG. 10 (when the user does not ask the robot 100), the feature amount vector of the i-th utterance section. The values of each element and the feature amount of vi are as shown in FIG. 11, and the feature amount vector vi of the i-th utterance section is vi = (..., 1.7, ..., 20, 1, 1). , ...). Then, the value of Iv calculated by the sound pressure level setting parameter calculation unit 132B using the feature amount vector vi and the above-mentioned determination model (random forest) is 0.55. Further, the value of Iv calculated by the sound pressure level setting parameter calculation unit 132B using the feature amount vector vi and the above-mentioned conditional branching equation is 0.5 (see FIG. 9).
Then, the sound synthesis unit 132F makes a sound based on the sound pressure level setting parameter Iv and the above-mentioned sound pressure level calculation formula or a table in which the sound pressure level setting parameter Iv and the sound pressure level (dB) are associated with each other. The filler "um" is synthesized at a pressure level of 50 dB and output to the output unit 120.

FIG. 12 shows the voice waveform of the filler output by the robot 100 in the first embodiment. On the upper side of FIG. 12, the voice waveform of the filler "um" output by the robot 100 when the i-th utterance content of the user is the question "Have you been to Robo?" In FIG. 10 is shown. Further, on the lower side of FIG. 12, when the user's i-th utterance content is the utterance "I went to the park yesterday" in FIG. 10, the voice waveform of the filler "um" output by the robot 100 is displayed. Shown. Further, in FIG. 12, the vertical axis represents the amplitude of the voice waveform, and the horizontal axis represents the time (sec).

As shown in FIG. 12, when the user's i-th utterance content is the question "Have you been to Robo?", The robot 100 needs to answer to the user, and the reliability of outputting the filler. Can be said to be expensive. In the first embodiment, in such a case, the sound pressure level of the filler can be increased.
On the other hand, if the user's i-th utterance is "I went to the park yesterday", it can be said that the reliability of the robot 100 outputting the filler is low. In the first embodiment, in such a case, the sound pressure level of the filler can be reduced.

According to the robot 100 according to the first embodiment described above, the sound pressure level determining unit 132 has a value Iv indicating the reliability of outputting the filler based on the feature amount extracted by the user voice analysis unit 131. The sound pressure level of the filler is determined based on the calculated value Iv indicating the reliability of outputting the filler. Therefore, the output unit 120 can output the filler at a sound pressure level corresponding to the value Iv indicating the reliability of outputting the filler. The reliability of outputting the filler has a positive correlation with the possibility that the robot 100 has the right to speak. That is, even when it is unclear whether or not the utterance right belongs to the robot 100, the output unit 120 can output the filler at a sound pressure level corresponding to the possibility that the utterance right belongs to the robot 100. Therefore, when it is unlikely that the robot 100 has the right to speak, the output unit outputs the filler at a small sound pressure level, and it is possible to reduce the interruption of the user's utterance by the filler output by the output unit 120. As a result, it is possible to provide the robot 100 that can reduce the possibility of interrupting the user's utterance.

In the robot 100 according to the present embodiment, the sound pressure level of the filler is determined by using information such as the basic frequency f0 of the voice waveform of the user's voice and the past history of the utterance of the robot 100, so that the language of the user's voice is determined. The sound pressure level of the filler can be determined by a calculation that is relatively lighter than the processing using information.

The present invention is not limited to the above embodiment, and can be appropriately modified without departing from the spirit. For example, as the determination model, not only the above-mentioned random forest but also other machine learning methods such as a support vector machine (SVM) may be used. Further, in the present embodiment, the user voice analysis unit 131 extracts the feature amount of the feature amount vector vi from the information of the past history of the utterance of the robot 100, but the feature amount is only from the voice waveform of the user's voice. The feature quantity of the quantity vector vi may be extracted. Further, it goes without saying that the robot 100 may not emit the filler between the utterance sections of the user. The robot 100 may not only output the filler between the utterance sections of the user and the utterance section, but may also be silent or output the utterance (voice having meaning as the dialogue content). Good. Further, in the present embodiment, the robot 100 has been described as the voice dialogue device, but the voice dialogue device according to the present invention may be any device capable of interacting with the user, for example, an application for interacting with the user. It may be a smartphone or the like in which is incorporated.

100 robot (voice dialogue device)
110 Input unit 120 Output unit 130 Control unit 131 User voice analysis unit 132 Sound pressure level determination unit 132A Method selection unit 132B Sound pressure level setting parameter calculation unit 132C Learning database 132D Offline learning unit 132E Judgment model and conditional branching database 132F Voice Synthesis part

Claims (1)

A voice dialogue device including an input unit for inputting a user's voice and an output unit for outputting voice to the user.
A user voice analysis unit that extracts a feature amount including at least the fundamental frequency of the voice waveform as an element from at least the voice waveform of the voice input by the input unit.
When the output unit outputs the filler to the user, a value indicating the reliability of outputting the filler is calculated based on the feature amount extracted by the user voice analysis unit, and the filler is output. A sound pressure level determining unit that determines the sound pressure level of the filler based on a value indicating reliability,
A voice dialogue device.