JP6576968B2 - End-of-speech determination device, end-of-speech determination method, and program - Google Patents

Description

  The present invention relates to an end-of-speech determination device, an end-of-speech determination method, and a program for determining whether or not an utterance in a dialogue by a plurality of speakers is an end-of-speaker utterance.

  By detecting the end of the speaker's speech (utterance at the end of the speech) from the conversation between the customer and the person in charge at the call center or window, for example, after the customer has finished speaking, summarize the customer's speech Processing such as analysis by the system becomes possible.

  Information on whether the utterance in the dialogue is the end of the utterance as a method of determining whether the utterance in the dialogue by multiple speakers such as the dialogue between the customer and the agent in charge is the end of the utterance There is a method of using learning data (speech end learning data) to which is given (see Non-Patent Document 1). In this method, learning data to which information indicating whether or not the utterance at the end of the talk is added to the utterance in the dialogue is used. Then, by machine learning using the learning data, an end-of-speech determination model for determining whether or not the utterance in the dialogue is the end-of-speech utterance is generated.

R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin.LIBLINEAR: A library for large linear classification Journal of Machine Learning Research 9 (2008), 1871-1874.

  In general, there is a difference in the expression of each utterance at the end of the utterance. Therefore, in the method disclosed in Non-Patent Document 1, when the speech end determination model generated using the speech end learning data in a certain field is applied to another field, it is determined whether or not the speech is at the end of the speech. It may not be possible to determine with high accuracy. It is conceivable to prepare end-of-speech learning data for each field to be determined, but preparing end-of-speech learning data for each field causes an increase in cost.

  An object of the present invention, which has been made in view of the above problems, is a speech end determination device that can determine whether or not an utterance in a dialogue is an end of speech while suppressing an increase in cost. It is in providing the end determination method and program.

  In order to solve the above-described problem, the speech end determination device according to the present invention is a speech end determination device that determines whether or not an utterance in a dialogue by a plurality of speakers is an utterance at the end of a speaker's speech, A determination unit is provided for determining whether or not the utterance in the dialogue is an utterance at the end of the talk based on whether or not the speaker is changed in the dialogue.

  In addition, in order to solve the above-described problem, the speech end determination method according to the present invention is a speech end determination method for determining whether an utterance in a dialogue by a plurality of speakers is an utterance at the end of a speaker's speech. And determining whether or not the utterance in the dialogue is an end-of-speech utterance based on whether or not the speaker of the utterance in the dialogue is changed.

  Moreover, in order to solve the said subject, the program which concerns on this invention makes a computer function as the above-mentioned talking end determination apparatus.

  According to the talk end determination device, the talk end determination method, and the program according to the present invention, it is possible to determine whether or not the utterance in the dialogue is the utterance at the end of the talk while suppressing an increase in cost.

It is a block diagram which shows the structural example of the talking end determination apparatus which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the structural example of the talk end determination apparatus which concerns on the 2nd Embodiment of this invention. It is a figure for demonstrating the production | generation of a speech end determination model. It is a figure which shows an example of the dialogue between a customer and a reception person in charge. It is a figure which shows the structural example of speech end learning data. It is a figure which shows the other structural example of talking end learning data. It is a block diagram which shows the structural example of the talk end determination apparatus which concerns on the 3rd Embodiment of this invention. It is a figure which shows an example of the dialogue between a customer and a reception person in charge. It is a figure which shows an example of the determination result by the speech end determination model shown in FIG. It is a block diagram which shows the structural example of the speech end determination apparatus which concerns on the 4th Embodiment of this invention. It is a figure for demonstrating determination of the speech of the end of a talk by the determination part shown in FIG. It is a figure for demonstrating operation | movement of the learning data generation part shown in FIG. It is a figure which shows notionally about the machine learning which input the natural language. It is a figure which shows the specific example of bag-of-words. It is a figure for demonstrating the problem of the method of the machine learning which input the conventional natural language. It is a figure for demonstrating the method of the machine learning which input the natural language which concerns on this invention.

  Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings.

(First embodiment)
FIG. 1 is a block diagram illustrating a configuration example of the talking end determination device 10 according to the first embodiment of the present invention. Whether the utterance in the dialogue by a plurality of speakers such as the dialogue between the customer and the person in charge of the conversation is the utterance at the end of the story when the content that the speaker wants to convey is finished. It is to determine whether or not.

  In addition, human beings are not always able to talk about what they want to convey, and they can think and talk in the middle of the story. For this reason, the voice may be interrupted (a silent section occurs) while the speaker is speaking. An utterance is a speaker's story separated by such breaks in speech.

  The talking end determination device 10 illustrated in FIG. 1 includes a determination unit 11.

  The determination unit 11 determines whether or not the utterance in the dialog is an end-of-speaking utterance based on the dialog structure of a dialog by a plurality of speakers (a dialog between the customer and the person in charge). Specifically, the determination unit 11 detects the presence or absence of a speaker change in which a speaker changes in a conversation, and determines that the utterance immediately before the speaker change is the utterance at the end of the talk.

  In general, in a dialogue between a customer and a person in charge, for example, after the customer has finished talking about the content that the customer wants to inquire, the person in charge answers the inquiry, and after the person in charge has finished talking about the answer, the customer There are many dialog structures that make further inquiries. That is, when a speaker change occurs, there is a tendency that the utterance just before that is often the utterance at the end of the talk before the speaker change occurs. Based on this tendency, the determination unit 11 determines whether or not the utterance in the dialogue is the utterance at the end of the conversation.

  Note that the determination unit 11 determines whether or not the speaker change has occurred after removing the utterances of only fillers that are not related to the content of the dialogue, such as “Ah”, “Ehto”, “Yes”, among the utterances in the dialogue. Determine. The filler-only utterance is likely to be an interaction between the customer and the customer who is speaking, and if such an utterance is included in the determination of whether or not a speaker change has occurred, Nevertheless, it may be determined that a speaker change has occurred. Therefore, in this embodiment, it is determined whether or not a speaker change has occurred after removing the filler-only utterance.

  As described above, in the present embodiment, the speech end determination device 10 includes the determination unit 11 that determines whether or not the utterance in the dialogue is the utterance at the end of the speaker based on the presence or absence of the change of the speaker in the dialogue. Prepare.

  The dialogue structure in which the utterance just before the speaker change is the utterance at the end of the talk is often seen regardless of the field in which the dialogue is being conducted. By using this dialog structure to determine whether or not the utterance in the dialog is an end-of-speech utterance, the dialog end without increasing the cost of preparing the end-of-speech learning data for each field to be determined It can be determined whether or not the utterance at is the utterance at the end of the talk.

(Second Embodiment)
FIG. 2 is a diagram illustrating a configuration example of a speech end determination device 10A according to the second embodiment of the present invention. The talk end determination device 10A according to the present embodiment generates a talk end determination model for determining whether or not an utterance in a dialogue by a plurality of speakers, such as a dialogue between a customer and an agent in charge, is an utterance at the end of the talk. It generates end-of-speech learning data. In FIG. 2, the same components as those in FIG.

  The talk end determination device 10A shown in FIG. 2 is different from the talk end determination device 10 shown in FIG. 1 in that a learning data generation unit 12 is added.

  The learning data generation unit 12 uses the learning end learning data 13 for generating a speech end determination model by machine learning to determine whether or not the utterance in the dialogue is an end of speech based on the determination result of the determination unit 11. Generate. For example, as illustrated in FIG. 3, the generated end-of-speech learning data 13 is input to the determination model generation unit 14, and whether or not the utterance in the dialogue is an end-of-speech utterance by machine learning by the determination model generation unit 14. An end-of-speech determination model 15 is generated to determine whether or not. The end-of-speech determination model 15 determines, for example, whether or not the utterance in the dialogue between the customer and the reception staff at the call center is the end-of-speech utterance. Note that the end-of-speech determination model 15 and the determination model generation unit 14 may be included in the end-of-speech determination device 10A, or may be included in an external device different from the end-of-speech determination device 10A.

  FIG. 4 is a diagram illustrating an example of a dialogue between a customer and a person in charge.

  In the example shown in FIG. 4, the customer performs an utterance # 2 “I want to ask you for a moment”, followed by an utterance # 2 “I bought it on the Internet”. In response to customer utterances # 1 and # 2, the person in charge responds to utterance # 3 of “Yes” as a response to the customer utterance.

  The customer makes an utterance # 4 "What will the delivery fee be?" Following the utterance # 3 of the person in charge. The customer's utterances # 1, # 2, and # 4 are contents for inquiring about a delivery fee at the time of purchase on the Internet, and the inquiry is finished with the utterance # 4. Therefore, the customer's utterance # 4 corresponds to the utterance at the end of the conversation.

  In response to the customer's utterance # 4, the person in charge performs the utterance # 5 that answers the customer's inquiry that “the delivery fee on the Internet is currently free”. The response to the customer's inquiry is completed by the utterance # 5 of the person in charge. Therefore, the utterance # 5 of the person in charge corresponds to the utterance at the end of the conversation.

  The conversation end learning data 13 for generating the conversation end determination model 15 will be described with reference to FIG. 5 by taking the dialogue between the customer and the person in charge shown in FIG. 4 as an example.

  As described above, when the dialogue includes only the filler utterance, it may be determined that the speaker change has occurred even though the speaker has not actually finished speaking. Therefore, in the end-of-speech learning data 13, the utterance of only the filler is removed. Therefore, as the end-of-speech learning data 13, as shown in FIG. 5, utterances # 1, # 2, # 4, and # 5 excluding the utterance # 3 of only the filler (“Yes”) by the person in charge are extracted. The Then, information (speaking end flag) indicating whether or not each utterance is an ending speech is given. In the example shown in FIG. 5, when the talk end flag is “0”, it is not an end-of-speech utterance, and when the talk end flag is “1”, it is an end-of-speech utterance. Therefore, “1” is set to the talk end flags of the utterances # 4 and # 5, which are utterances of the talk end, and “0” is set to the talk end flags of the other utterances # 1 and # 2. Thus, the end-of-speech learning data 13 is data in which the utterance of the customer or the person in charge of the customer is associated with the information indicating whether or not the utterance is the end-of-speech utterance.

  In addition, although the example which removes the speech only of a filler was demonstrated in FIG. 5, it is not restricted to this. For example, when a filler is included in an utterance other than the utterance of only the filler, the filler may be removed or the filler may be left as it is.

  In addition, in FIG. 5, the description has been given using the example of setting the end-of-speech flag for each utterance. However, the present invention is not limited to this, and the end-of-speech flag is set for utterances in which the utterances until the end of the speech are sequentially connected. Also good.

  For example, as shown in FIG. 6, since the utterance # 1 is not the utterance at the end of the talk, “0” is set to the talk end flag of the utterance # 1. Next, an utterance connecting customer utterance # 1 and customer utterance # 2 following utterance # 1 is added to the end-of-speech learning data 13. Since the utterance # 2 is not the utterance at the end of the talk, the utterance connecting the utterance # 1 and the utterance # 2 is not the utterance at the end of the talk. Therefore, “0” is set to the speech end flag of the speech connecting speech # 1 and speech # 2.

  Next, the utterance connecting the customer utterance # 1, the utterance # 2 following the utterance # 1, and the customer utterance # 4 following the utterance # 2 (excluding the filler only utterance # 3) is the end of the learning data. 13 is added. Since the utterance # 4 is the utterance at the end of the talk, the utterance connecting the utterance # 1, the utterance # 2, and the utterance # 4 is the utterance at the end of the talk. Therefore, “1” is set to the speech end flag of the speech connecting speech # 1, speech # 2, and speech # 4. In this way, an utterance in which utterances up to the end of the talk are sequentially connected and a talk end flag of the utterance may be added to the talk end learning data 13.

  The end-of-speech learning data 13 as shown in FIGS. 5 and 6 can be manually generated from the dialogue between the customer and the person in charge. However, it will be costly to generate such end-of-speech learning data 13 for each field for which end-of-speech determination is desired.

  Therefore, in the present embodiment, it is determined from the dialog structure whether or not the utterance in the dialog is an utterance at the end of the talk, and the determination result is used as the talk end learning data 13. In this way, the speech end learning data 13 can be automatically generated from the dialogue in the field where the end of speech determination is desired. Then, by generating the end-of-speech determination model 15 using the generated end-of-speech learning data 13, it is possible to determine whether or not the utterance in the conversation is the end-of-speech utterance while suppressing an increase in cost. .

(Third embodiment)
FIG. 7 is a diagram illustrating a configuration example of the talking end determination device 10B according to the third embodiment of the present invention. In FIG. 7, the same components as those in FIGS.

  The talk end determination device 10B shown in FIG. 7 differs from the talk end determination device 10A shown in FIG. 2 in that a determination model generation unit 14 and a talk end determination model 15 are added. That is, in the present embodiment, the talk end determination device 10B generates the talk end learning data 13 from the determination result of whether or not the talk is ended using the dialog structure, and uses the generated talk end learning data 13 to end the talk. A determination model 15 is generated. Then, a determination result of whether or not the utterance in the dialogue is the utterance at the end of the talk by the talk end determination model 15 is output.

  Next, operation | movement of the speech end determination apparatus 10B which concerns on this embodiment is demonstrated in detail.

  As pre-processing, the end-of-speech determination model 15 for determining whether or not the utterance in the dialog is the end-of-speak utterance is generated by machine learning using the end-of-speech learning data 13 in which the end-of-speech flag is added to the utterance in the dialog. Is done. Note that as the end-of-speech learning data 13 used in the pre-processing, for example, data to which the end-of-speech flag is manually added, data to which the end-of-speech flag is given by the determination using the above-described dialog structure, and the like can be used.

  The machine learning method is not particularly limited as long as an appropriate model can be generated based on the learning data, and various methods such as deep learning and support vector machine can be used. There is no particular limitation on the information (feature) used for determining whether or not the speech is at the end of the speech, and various information can be used so that correct determination is possible.

  Next, as online processing (real-time processing according to speech), voice data (dialog data) of dialogue between the customer and the person in charge is input to the speech end determination device 10A, and the speech end determination model 15 is used. Thus, it is determined whether or not the utterance in the dialogue indicated by the dialogue data is the utterance at the end of the talk.

  In the following, it is assumed that the dialogue data between the customer and the person in charge as shown in FIG. 8 is input as the dialogue data. Note that the utterance of the customer and the utterance of the person in charge of the customer are input to the talk end determination device 10A through different channels (two channels).

  In the example shown in FIG. 8, the customer performs the utterance # 12 “I don't know which is better” after the utterance # 11 “I don't know much about it”.

  After the customer's utterance # 12, the person in charge performs the utterance # 13 of only the filler of “Yes” as a consideration for the customer's utterance. The customer utters “which product is recommended” after the utterance # 13 of the person in charge. Since the utterance of the customer who inquires about the recommended product is over, the person in charge of the response has made utterance # 15 to list the recommended product.

  As shown in FIG. 9, the conversation end determination model 15 determines that the customer utterance # 11 is not the end of the utterance and the customer utterances # 12 and # 14 are the end of the utterance. Suppose that it is determined. The talk end determination device 10B outputs the determination result of the talk end determination model 15.

  In the present embodiment, the description has been given using the example in which the talking end determination device 10B includes the learning data generation unit 12 and the determination model generation unit 14. However, the present invention is not limited to this, and the talking end determination device 10B. Another external device may include the learning data generation unit 12 and the determination model generation unit 14. In this case, the talk end determination device 10B acquires the talk end determination model 15 generated by the external device, and outputs the determination result of the acquired talk end determination model 15.

(Fourth embodiment)
FIG. 10 is a diagram illustrating a configuration example of a talking end determination device 10C according to the fourth embodiment of the present invention. In FIG. 10, the same components as those in FIG.

  The talking end determination device 10C shown in FIG. 10 differs from the talking end determination device 10B shown in FIG. 7 in that the learning data generation unit 12 is changed to a learning data generation unit 12C.

  The learning data generation unit 12C receives the determination result of the determination unit 11 and the determination result of the speech end determination model 15, and generates the end of learning data 13 based on these determination results.

  Next, the operation of the talking end determination device 10C according to the present embodiment will be described. In the talk end determination device 10C according to the present embodiment, pre-processing and online processing are performed in the same manner as the talk end determination device 10B according to the third embodiment.

  Next, as post-processing, the determination unit 11 determines whether or not the utterance in the dialog indicated by the input dialog data is an end-of-speak utterance based on the dialog structure.

  First, the determination unit 11 removes the utterance of only the filler (utterance # 13) from the utterances in the dialogue. Then, the determination unit 11 determines whether or not a speaker change has occurred between each utterance and the utterance following the utterance. As described above, the customer's utterance and the response person's utterance are input through different channels. The determination unit 11 can determine whether or not a speaker change has occurred by monitoring the input of each channel. When the determination unit 11 determines that the speaker change has occurred, the determination unit 11 determines that the utterance immediately before the speaker change is the utterance at the end of the talk, and determines that the speaker change has not occurred, Determine that the utterance is not the end of the talk.

  As shown in FIG. 11, between the utterance # 11 and the utterance # 12 following the utterance # 11, the utterance remains the customer, and no speaker change has occurred. In addition, between the utterance # 12 and the utterance # 14 following the utterance # 12 (excluding the filler-only utterance # 3), the utterance remains the customer, and no speaker change occurs. Also, between the utterance # 14 and the utterance # 15 following the utterance # 14, the utterance is changed from the customer to the person in charge, and a speaker change occurs. Therefore, as shown in FIG. 8, the determination unit 11 determines that the customer's utterances # 11 and # 12 are not utterances at the end of the conversation, and determines that the customer's utterance # 14 is an utterance at the end of the conversations.

  The learning data generation unit 12C compares the determination result by the speech end determination model 15 with the determination result by the determination unit 11. Then, as shown in FIG. 12, the learning data generation unit 12 </ b> C adds, to the speaking end learning data 13, an utterance in which the determination result by the speaking end determination model 15 matches the determination result by the determination unit 11.

  In the example shown in FIG. 12, the determination result of utterance # 11 (not the utterance at the end of the talk) and the determination result of utterance # 14 (the utterance at the end of the talk) match between the talk end determination model 15 and the determination unit 11. ing. The learning data generation unit 12C adds the utterances # 11 and # 14 and the determination result as the talking end learning data 13. When the determination result is the same between the speech end determination model 15 and the determination unit 11, the reliability of the determination result is considered high. For this reason, by determining the coincidence result between the speech end determination model 15 and the determination unit 11 as the speech end learning data 13, the reliability of the speech end determination model 15 is improved by machine learning using the speech end learning data 13. Improvements can be made.

  Note that the learning data generation unit 12C may add the determination result by the determination unit 11 to the speech end learning data 13 without comparing the determination result by the speech end determination model 15 and the determination result by the determination unit 11. .

  As described above, in the present embodiment, the speech end determination device 10 </ b> C gives a determination result indicating whether or not the speech in the dialogue is the speech end speech, in the determination unit 11 and the speech end determination model 15. A learning data generation unit 12C to be added to the learning data 13 is provided.

  By making the determination result coincided with the end-of-speech determination model 15 and the determination unit 11 into the end-of-speech learning data 13, the reliability of the end-of-speech determination model 15 can be improved by machine learning using the end-of-speech learning data 13. You can plan.

  As described above, the end-of-speech learning data 13 may be generated not in units of utterances but in units in which a plurality of continuous utterances are collected (accumulated). For example, if one utterance is not an end-of-speech utterance, an end-of-speech flag is assigned to that utterance, and an end-of-speech flag is also attached to an utterance that connects that utterance with the next utterance, thereby learning the end of the talk. Data 13 may be generated. In this case, the utterances are sequentially connected until it is determined that the utterance is the end of the talk. When it is determined that the utterance is the end utterance, the accumulation of the utterance is reset, and it is determined whether the utterance next to the utterance determined to be the end utterance is the end utterance. The

  In addition, the conversation between the customer and the person in charge is converted into text by voice recognition processing and then processed by the talking end determination model 15. Here, an error may occur in the speech recognition process. Therefore, a process using the N-best method of preparing a plurality of candidates as a recognition result by a speech recognition process of a word included in an utterance may be performed.

  FIG. 13 is a diagram conceptually showing machine learning using a natural language as an input.

  In machine learning using natural language as input, feature calculation is performed on input text in order to perform numerical vectorization in accordance with the input of machine learning (support vector machine (SVM)). That is, at the time of learning or determination, it is necessary to perform numerical vectorization corresponding to the input of machine learning for the input text. As such feature calculation, for example, a model (bag-of-words) that considers only whether words are included in a sentence and does not consider how words are arranged is generally used.

  FIG. 14 is a diagram illustrating a specific example of bag-of-words.

  In bag-of-words, if a sentence contains a word, the numerical value corresponding to the word is set to 1, and an input vector expressing whether or not the sentence contains a word is calculated. If the input text is “Can I cancel my time deposit on the Internet?”, Morphological analysis is performed on the input text.

  Specifically, as shown in FIG. 14, a morphological analysis of a large amount of text generates a list in advance so that words appearing in the text are covered and a word number is assigned to each word. The Then, among the words included in the word list, the value of the input vector corresponding to the word number of the word appearing in the input text is “1”. The value of the input vector corresponding to the word number of a word that does not appear in the input text is “0”.

  Instead of morphological analysis, feature calculation methods such as bag-of-pos using parts of speech, or feature calculation combining bag-of-words and bag-of-pos may be used.

  For the processing by the speech end determination model 15, the speech recognition result text obtained by converting the voice of the dialogue into the text by the speech recognition processing is subjected to morphological analysis, and the feature calculation is performed by bag-of-words or the like from the result of the morphological analysis. Conceivable.

  Here, it is assumed that an error has occurred in the speech recognition process. For example, in response to a voice saying “Can I cancel my time deposit on the Internet?”, As shown in FIG. . When morphological analysis is performed on such an erroneous speech recognition result text and an input vector is calculated, errors are accumulated, and the feature calculation reflecting the correct input speech cannot be performed.

  As described above, the conventional method using the speech recognition result as an input for machine learning has a problem that it is likely to be an erroneous morpheme unit. The cause of such a problem is that the speech recognition dictionary used for speech recognition processing is often different from the morpheme analysis dictionary used for morpheme analysis, and inconsistencies are caused by differences in words registered in these dictionaries. May occur. Another reason is that the morphological analysis is performed on a normal sentence that can be read by humans, and therefore an erroneous morphological analysis may be performed due to an error in the speech recognition result text.

  Further, in the conventional method in which the speech recognition result is input to machine learning, if an error occurs in the speech recognition processing, the feature calculation is performed while the error is included, and thus there is a problem that the feature does not reflect the correct input speech. .

  Therefore, in the present invention, as shown in FIG. 16, a word sequence (N-best result) of the N-th candidate obtained as a result of the speech recognition processing is used to calculate a feature (eg, bag-of-words) in machine learning. Used for.

  In the speech recognition process, a process of searching for a word string closest to the input speech among combinations of registrations (including part-of-speech information and the like) registered in the speech recognition dictionary is performed. Therefore, as a result of the speech recognition process, a sequence of words (including part-of-speech information) can be obtained. In addition, it is possible to obtain up to N-th candidates other than the first candidate in order of proximity to the input voice. Therefore, even if the first candidate is incorrect, there is a high possibility that a correct word is included in the N candidate.

  As described above, in the present invention, the word sequence of the N-th candidate obtained as a result of the speech recognition process is used without performing the morphological analysis process. For this reason, an error due to morphological analysis does not occur, and the result of the speech recognition process including the error is directly reflected in the feature. In addition, since morphological analysis is not performed, the amount of processing can be reduced. Further, since morphological analysis is not performed, it is not necessary to prepare a morphological analysis dictionary. In addition, since the speech recognition results up to the Nth candidate are reflected in the feature, even if a speech recognition error occurs in the first candidate, there is a high possibility that the Nth candidate contains a correct word, and these are calculated. Can be reflected.

  Although not particularly mentioned in the embodiment, a program for executing each process performed by a computer functioning as the speech end determination device 10 may be provided. The program may be recorded on a computer readable medium. If a computer-readable medium is used, it can be installed on a computer. Here, the computer-readable medium on which the program is recorded may be a non-transitory recording medium. The non-transitory recording medium is not particularly limited, but may be a recording medium such as a CD-ROM or a DVD-ROM.

  Although the above embodiments have been described as representative examples, it will be apparent to those skilled in the art that many changes and substitutions can be made within the spirit and scope of the invention. Therefore, the present invention should not be construed as being limited by the above-described embodiments, and various modifications and changes can be made without departing from the scope of the claims. For example, it is possible to combine a plurality of constituent blocks described in the configuration diagram of the embodiment into one, or to divide one constituent block.

10, 10A, 10B, 10C Talk End Determination Device 11 Judgment Unit 12, 12C Learning Data Generation Unit 13 Talk End Learning Data 14 Judgment Model Generation Unit 15 Talk End Determination Model

Claims (11)

An end-of-speech determination device that determines whether or not an utterance in a dialogue by a plurality of speakers is an end-of-speaker utterance,
A determination unit that determines whether or not the utterance in the dialogue is an utterance at the end of the speech based on the presence or absence of a change of speakers in the dialogue ;
A learning data generation unit that generates learning data used for machine learning of a speech end determination model for determining whether or not an utterance in a dialogue is a speech at the end of a speech, based on a determination result of the determination unit Talking end judgment device. In the talk end judging device according to claim 1 ,
The learning data generation unit is configured to add, to the learning data, a determination result in which whether the utterance in the dialogue is an utterance at the end of a conversation matches between the determination unit and the end-of-speech determination model. End of talk determination device. In the talk end judging device according to claim 1 or 2 ,
An end-of-speech determination device that determines whether or not an utterance in the dialogue is an end-of-speaker utterance using the end-of-speech determination model and outputs a determination result.   An end-of-speech determination device that determines whether or not an utterance in a dialogue by a plurality of speakers is an end-of-speaker utterance,
  A determination unit that determines whether or not the utterance in the dialog is an end-of-speech utterance based on the presence or absence of a speaker change in the dialog;
  The determination unit is a speech end determination device characterized in that an utterance obtained by removing only the filler utterance from the utterance in the dialogue is the target of the determination.   An end-of-speech determination device that determines whether or not an utterance in a dialogue by a plurality of speakers is an end-of-speaker utterance,
  A determination unit that determines whether or not the utterance in the dialog is an end-of-speech utterance based on the presence or absence of a speaker change in the dialog;
  An utterance in the dialogue is generated by using a speech end determination model for determining whether the utterance in the dialogue is an utterance at the end of the conversation, which is generated by machine learning of the learning data generated based on the determination result of the determination unit. A speech end determination device characterized by determining whether or not the speech is at the end of a speaker's speech and outputting a determination result. An end-of-speech determination method in an end-of-speech determination device that determines whether an utterance in a dialogue by a plurality of speakers is an end-of-speaker utterance,
Determining whether the utterance in the dialogue is an end-of-speech utterance based on the presence or absence of the change of the speaker in the dialogue; and
Generating end-of-speech data used for machine learning of an end-of-speech determination model for determining whether or not an utterance in a dialogue is an end-of-speech utterance. Method. The method for determining the end of a conversation according to claim 6.
Whether or not the utterance in the dialogue is an utterance at the end of the talk is added to the learning data, a judgment result that matches the judgment based on the presence or absence of the change of the speaker and the talk end judgment model. End of talk determination method.   The method for determining an end of speech according to claim 7,
  A speech end determination method, further comprising: determining whether or not an utterance in the dialogue is an utterance at the end of a speaker's speech using the speech end determination model and outputting a determination result.   An end-of-speech determination method in an end-of-speech determination device that determines whether an utterance in a dialogue by a plurality of speakers is an end-of-speaker utterance,
  Determining whether the utterance in the dialogue is an end-of-speech utterance based on the presence or absence of a speaker change in the dialogue; and
  A speech end determination method characterized in that an utterance obtained by removing an utterance of only a filler from an utterance in the dialog is the target of the determination.   An end-of-speech determination method in an end-of-speech determination device that determines whether an utterance in a dialogue by a plurality of speakers is an end-of-speaker utterance,
  Determining whether the utterance in the dialogue is an end-of-speech utterance based on the presence or absence of a change of speakers in the dialogue;
  The utterance in the dialog is a speaker by using a speech end determination model for determining whether or not the utterance in the dialog is an utterance at the end of the conversation, which is generated by machine learning of the learning data generated based on the determination result. Determining whether or not the speech is at the end of the speech, and outputting a determination result. The program for functioning a computer as a speech end determination apparatus as described in any one of Claims 1-5 .

Images