Automatic voice response device, automatic voice response method, and automatic voice response program

Abstract

In automated voice response systems, the goal is to strike a good balance between accurately pinpointing the speaker's intent and quickly identifying it. [Solution] The automated voice response device performs the following steps: a first step of acquiring the content of the voice from the speaker; a second step of inputting the content into a first machine learning model that has been trained and requesting the first machine learning model to output which of the candidate intentions the content corresponds to; a third step of identifying the single intention as the speaker's intention if the first machine learning model outputs only one intention selected from the candidate intentions; and a fourth step of performing a question generation process and then returning to the first step if the first machine learning model outputs multiple intentions selected from the candidate intentions.

Description

Technical Field

[0001] The present invention relates to a voice automatic response device, a voice automatic response method, and a voice automatic response program.

Background Art

[0002] In a call center or the like, a voice automatic response system that automatically responds to a call from a customer using a voice guide prepared hierarchically in advance may be used. Such a system is set so that the customer can reach the target service by pressing the corresponding button on the telephone according to the voice guide.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in the above system, it is necessary to force the customer to perform repetitive operations such as listening to a voice guide, pressing a push button, listening to a voice guide again, and pressing a push button. It takes time until the customer's intention is specified.

[0005] On the other hand, if the voice guide is not performed hierarchically, the service may be provided without sufficiently narrowing down the customer's intention, so there is a possibility that the customer's expectation cannot be accurately met.

[0006] Therefore, an object of the present invention is to achieve a good balance between accurately narrowing down the intention of the speaker and quickly specifying the intention of the speaker in voice automatic response.

Means for Solving the Problems

[0007] An automated voice response device according to one aspect of the present invention is an automated voice response device that interacts with a speaker using voice, and includes a processing unit that can refer to a database in which a plurality of intent candidates are registered. The processing unit performs the following steps: a first step of acquiring the content of voice from the speaker; a second step of inputting the content into a first machine learning model that has been trained and requesting the first machine learning model to output which of the intent candidates the content corresponds to; a third step of identifying the single intent as the speaker's intent if the first machine learning model outputs only one intent selected from the intent candidates; and a fourth step of performing a question generation process and then returning to the first step if the first machine learning model outputs a plurality of intents selected from the intent candidates. The question generation process includes inputting the content into a second machine learning model that has been trained and requesting the second machine learning model to output a question to the speaker in order to narrow down the plurality of outputted intents to one intent; and outputting the question output by the second machine learning model to the speaker.

[0008] An automated voice response method according to one aspect of the present invention is an automated voice response method that engages in a voice dialogue with a speaker, and includes: a first step of acquiring the content of a voice from a speaker; a second step of inputting the content into a first machine learning model that has been trained and requesting the first machine learning model to output which of a plurality of intent candidates in a database the content corresponds to; a third step of identifying the single intent as the speaker's intent if the first machine learning model outputs only one intent selected from the intent candidates; and a fourth step of performing a question generation process and returning to the first step if the first machine learning model outputs a plurality of intents selected from the intent candidates. The question generation process includes inputting the content into a second machine learning model that has been trained and requesting the second machine learning model to output a question to the speaker in order to narrow down the plurality of intents output to one intent; and outputting the question output by the second machine learning model to the speaker.

[0009] An automated voice response program according to one aspect of the present invention causes a processor to execute the automated voice response method. The program can be stored in a computer-readable storage medium. The storage medium is a non-temporary, tangible medium. [Effects of the Invention]

[0010] According to the present invention, in automated voice response, it is possible to achieve a good balance between accurately narrowing down the service the speaker is looking for and quickly reaching the service the speaker is looking for. [Brief explanation of the drawing]

[0011] [Figure 1] Figure 1 is a schematic diagram of the automated voice response system according to the embodiment. [Figure 2] Figure 2 is a hardware block diagram of the automated voice response server shown in Figure 1. [Figure 3] Figure 3 is a block diagram showing the functional aspects of the automated voice response server shown in Figure 1. [Figure 4] Figure 4 is a diagram showing the database in Figure 2. [Figure 5] Figure 5 is a flowchart illustrating the processing of the automated voice response server shown in Figure 3. [Figure 6] Figure 6 is a flowchart illustrating the question generation process shown in Figure 3. [Figure 7] Figure 7 is a diagram showing an example of prompts to the generating AI in the question generation process shown in Figure 6. [Figure 8] Figure 8 is a diagram illustrating another example of the database shown in Figure 2. [Modes for carrying out the invention]

[0012] The embodiments will be described below with reference to the drawings.

[0013] FIG. 1 is a schematic diagram of the voice automatic response system 1 according to the embodiment. As shown in FIG. 1, the voice automatic response system 1 is a system for automatically responding to a call from an interlocutor C in voice. That is, in this embodiment, a call on the phone is exemplified as a dialogue, and the interlocutor C is an unspecified large number of customers.

[0014] The voice automatic response system 1 includes a voice automatic response server 2, a voice recognition model 3, and a generative AI 4. The voice automatic response server 2 is an example of a voice automatic response device and is connected to a communication network N. The communication network N is, for example, the Internet, but may also be an intranet or the like.

[0015] The voice recognition model 3 converts voice into text using known voice recognition technology. The voice recognition model 3 accumulates, for example, the features of voice obtained from a large amount of recorded utterances in learning data, compares the features extracted from the input voice to be recognized with the accumulated features, and outputs the language text having the features closest to the features of the input voice as the recognition result.

[0016] The voice recognition model 3 is connected to the communication network N so as to be accessible from the voice automatic response server 2. Note that the voice recognition model 3 may be provided inside the voice automatic response server 2 or may be connected to the voice automatic response server 2 via a communication path different from the communication network N.

[0017] The generative AI 4 is a machine learning model that generates new data not included in the learning data by learning the regularity and structure of a large amount of learning data. The generative AI 4 can be constructed using a large language model (LLM). The generative AI 4 can function as a first machine learning model and a second machine learning model described later.

[0018] The generation AI 4 is connected to the communication network N so as to be accessible from the voice automatic response server 2. Note that the generation AI 4 may be provided inside the voice automatic response server 2, or may be connected to the voice automatic response server 2 through a communication path different from the communication network N. The generation AI 4 may also兼 have the voice recognition model 3.

[0019] The voice automatic response server 2 is connected to the information processing terminal 10 operated by the user U via the communication network N. That is, the information processing terminal 10 can access the voice automatic response server 2 via the communication network N. The information processing terminal 10 can be, for example, a notebook computer, a desktop computer, a smartphone, a tablet terminal, etc.

[0020] The voice automatic response server 2 is set to automatically respond to calls coming in to the user organization to which the user U belongs. The user U can operate the information processing terminal 10 to edit the database DB of the voice automatic response server 2 described later.

[0021] The caller C makes a call to the user organization to which the user U belongs using the telephone 20. The telephone 20 may be connected to the communication network N via the telephone line M, or may be connected to the communication network N using an Internet telephone function such as VoIP. The telephone 20 can be a mobile phone or a landline phone.

[0022] Figure 2 is a block diagram of the voice automatic response server 2 in terms of hardware in Figure 1. As shown in Figure 2, the voice automatic response server 2 includes a processor 131, a system memory 132, a storage memory 133, and a communication interface 134. The processor 131 includes a CPU. The system memory 132 includes a volatile memory such as RAM. The storage memory 133 includes a non-volatile memory such as a hard disk and a flash memory.

[0023] The storage memory 133 stores the voice automatic response program P. The storage memory 133 also has a database DB. The communication interface 134 is an interface for connecting to the communication network N, and includes, for example, an Ethernet interface. The processor 131, system memory 132, storage memory 133, and communication interface 134 are connected to each other via a communication bus.

[0024] The functions disclosed in this embodiment can be performed using processing circuits, including general-purpose processors, dedicated processors, integrated circuits, ASICs, conventional circuits, and / or combinations thereof, configured to perform the functions.

[0025] Figure 3 is a functional block diagram of the voice automatic response server 2 shown in Figure 1. As shown in Figure 3, the voice automatic response server 2 comprises a processing unit 31, a storage unit 32, and a communication unit 33. The processing unit 31 is implemented by a processor 131 that executes the voice automatic response program P read from the storage memory 133 shown in Figure 2 into the system memory 132. The storage unit 32 is implemented by the storage memory 133. The storage unit 32 has the aforementioned database DB. The communication unit 33 is implemented by a communication interface 134.

[0026] Figure 4 is a diagram showing the database DB in Figure 2. As shown in Figure 4, the database DB has multiple candidate query content and multiple candidate answers corresponding to each of those candidate query content entries pre-registered. The candidate query content is an example of the speaker C's intention. The candidate query content is a list of possible inquiries made by speaker C over the phone. The candidate answers are the answer sentences that will be automatically responded to by voice guidance in response to each of those candidate query content entries.

[0027] The database DB is, for example, a non-hierarchical database. User U can register new information in the database DB and edit information already registered in the database DB by accessing the voice automated response server 2 via the communication network N from the information processing terminal 10.

[0028] If the user organization to which user U belongs is a bank, the information shown in Figure 4 will be registered in the database DB. For example, if the inquiry is "Transfer fee, to another bank, less than 30,000 yen", the answer will be "For transfers to other banks of less than 30,000 yen, the fee is 250 yen." If the inquiry is "Transfer fee, to another bank, 30,000 yen or more", the answer will be "For transfers to other banks of 30,000 yen or more, the fee is 330 yen." Other cases are as shown in Figure 4.

[0029] Figure 5 is a flowchart illustrating the processing of the automated voice response server 2 shown in Figure 3. The processing of the automated voice response server 2 will be explained below following the flow shown in Figure 5, with reference to Figures 1-4 as appropriate. Note that the processing of the automated voice response server 2 is executed by the processing unit 31.

[0030] When the automated voice response server 2 receives a call from the telephone 20, it sends a pre-prepared initial guidance voice message to the telephone 20 and outputs it to the speaker C (step S1). The automated voice response server 2 determines whether or not it has received voice from the speaker C via the communication network N from the telephone 20 (step S2).

[0031] When the automated voice response server 2 receives speech from speaker C, it converts the content of that speech into text using speech recognition technology (step S3). Specifically, the speech recognition server 2 inputs the acquired speech into the speech recognition model 3 and obtains text from the speech recognition model 3 that represents the content of that speech. This corresponds to the first step in which the automated voice response server 2 receives the content of speech from speaker C.

[0032] The automated voice response server 2 inputs the transcribed voice content and multiple query content candidates registered in the database DB to the generating AI 4 and requests the generating AI 4 to output which of the multiple query content candidates the voice content corresponds to. This corresponds to the second step of inputting the voice content from speaker C into a trained first machine learning model and requesting the first machine learning model to output which of the multiple query content candidates the voice content corresponds to. In other words, the generating AI 4 acts as a first machine learning model that takes the voice content from speaker C as input and outputs which of the multiple query content candidates the content corresponds to.

[0033] The automated voice response server 2 receives a response from the generating AI 4 (step S5). Specifically, the generating AI 4 may select only one query that matches the input voice content from multiple query content candidates and output only that selected query content as the response. Alternatively, the generating AI may select multiple query content that may match the input voice content from multiple query content candidates and output all of those selected query content as the response. In addition, the generating AI 4 may output a message indicating that it is unclear which of the multiple query content candidates the input voice content matches.

[0034] In step S6, if the number of inquiry content selected by the generating AI4 from among multiple inquiry content candidates is zero, the automated voice response server 2 outputs an error (step S7) and returns to step S2. Specifically, the automated voice response server 2 sends a voice message to the telephone 20 prompting speaker C to repeat the inquiry content, and then returns to step S2.

[0035] In step S6, if the number of query content selected by the generating AI4 from among multiple query content candidates is one, the automated voice response server 2 identifies the query content output by the generating AI4 as the query content intended by the speaker C (step S8).

[0036] The automated voice response server 2 refers to the database DB and outputs a response corresponding to the identified inquiry to speaker C (step S9). This corresponds to the third step in which, when the first machine learning model outputs only one inquiry selected from multiple candidate inquiries, the server identifies that one inquiry as the inquiry intended by speaker C. Specifically, the automated voice response server 2 extracts a response corresponding to the identified inquiry from multiple candidate responses registered in the database DB, converts the extracted response into speech, and sends the speech-converted response to telephone 20.

[0037] Thus, when the content of the inquiry from speaker C accurately expresses the speaker's intent, the automated voice response server 2 can quickly identify the content of the inquiry intended by speaker C and quickly provide speaker C with the answer they expect.

[0038] In step S6, if the voice automated response server 2 has selected multiple query options from the multiple query options generated by the generation AI 4, it performs a question generation process described later to narrow down the query options to one (step S10). After performing the question generation process, the voice automated response server 2 returns to step 1. This corresponds to step 4, when the first machine learning model outputs multiple query options selected from multiple query options, and the system returns to step 1 after performing the question generation process.

[0039] Figure 6 is a flowchart illustrating the question generation process in Figure 3. As shown in Figure 6, the automated voice response server 2 executes steps S21 to S23 as the question generation process in step S10. First, the automated voice response server 2 inputs the content of the voice from speaker C and multiple candidate inquiry contents into the generation AI 4, and in step S5, requests the generation AI 4 to output a question to narrow down the multiple inquiry contents obtained from the generation AI 4 into a single inquiry contents (step S21).

[0040] Figure 7 is a diagram showing an example of a prompt to the generating AI4 in the question generation process shown in Figure 6. As shown in Figure 7, the voice automated response server 2 causes the generating AI4 to refer to a query content candidate file showing query content candidates registered in the database DB, and an audio content file showing the content of the audio of speaker C transcribed into text. It also generates a prompt request such as, "Based on the following query content candidates and the speaker's audio content, please generate an effective question to identify the query content intended by the speaker," and inputs the generated request to the generating AI4.

[0041] This Generative AI4 also functions as a second machine learning model, taking the content of the audio from speaker C and multiple candidate inquiry questions as input, and outputting a question to narrow down the multiple intentions output by Generative AI4 into a single intention. In other words, Generative AI4 functions as a machine learning model that combines the roles of both the first and second machine learning models mentioned above.

[0042] Note that the second machine learning model used in step S21 may be a different machine learning model from the first machine learning model used in step S4. In that case, the first machine learning model may be a model trained using training data in which the input is speech content and multiple candidate inquiry content, and the output is the inquiry content to which the speech content corresponds. The second machine learning model may be a model trained using training data in which the input is speech content and multiple candidate inquiry content, and the output is a question to narrow down the multiple intentions output by the first machine learning model to a single intention.

[0043] Next, the automated voice response server 2 obtains the question output by the generating AI 4 (step S22). The automated voice response server 2 converts the question output by the generating AI 4 into speech and sends the converted question to the telephone 20 to be output to speaker C (step S23). For example, if the voice content of speaker C obtained in step S2 above was "Please tell me about the transfer fee," the question output to speaker C in step S23 could be a question such as "Is the transfer amount less than 30,000 yen?" or "Is the recipient bank another bank?"

[0044] The question output to speaker C in step S23 may include multiple questions. For example, the question output to speaker C in step S23 may be "Is the recipient bank another bank? Also, is the transfer amount less than 30,000 yen?". In this way, the number of inquiries obtained from generating AI4 in step S5 can be narrowed down to one early on.

[0045] To that end, in the prompt shown in Figure 6, the automated voice response server 2 may make a request such as, "Generate an effective question to quickly identify the inquiry intended by speaker C," or "Generate a question that includes multiple questions simultaneously as needed to quickly identify the inquiry intended by speaker C." In other words, the request from the automated voice response server 2 to the generating AI 4 may be a request to generate a question that simultaneously addresses multiple decision-making factors necessary to identify the intent of speaker C.

[0046] Returning to Figure 5, once the question generation process (step S10) outputs a question to speaker C, speaker C speaks in response to the question, and the automated voice response server 2 obtains the second speech from speaker C (step S2). That is, the automated voice response server 2 returns to step S2 and repeats steps S2 to S6 described above.

[0047] Thus, if the content of the audio from speaker C does not adequately express the content of the inquiry, it is possible to narrow down the content of the inquiry intended by speaker C by asking questions to elicit information from speaker C.

[0048] In step S4, the automated voice response server 2 may input a series of voice messages from speaker C in one dialogue session into the generating AI 4. Here, one dialogue session means the period from when speaker C starts a call using the telephone 20 until when the call is ended. That is, in step S4, the automated voice response server 2 may input both the content of the first voice message from speaker C and the content of the second voice message from speaker C into the generating A14.

[0049] In this way, the generating AI 4 can accurately narrow down the inquiry content intended by speaker C. If the generating AI 4 also retains audio content entered before the currently entered audio content in a single dialogue session, then in step 4, the voice automatic response server 2 may input only the current audio content to the generating AI 4.

[0050] Furthermore, in step S22, the automated voice response server 2 may input a series of voice messages from speaker C in a single dialogue session to the generating AI 4. In this way, the generating AI 4 can accurately generate questions that narrow down the inquiry intended by speaker C. Note that if the generating AI 4 also retains voice messages input before the currently inputted voice message in a single dialogue session, then, similar to step S4, the automated voice response server 2 may input only the current voice message to the generating AI 4 in step S22 as well.

[0051] Furthermore, when the automated voice response server 2 inputs the candidate inquiry content into the generating AI 4 in step S4 after the question generation process in step S10, it may narrow down the candidate inquiry content to the multiple inquiry content output by the generating AI 4 in the previous step S4, and input the narrowed-down candidate inquiry content into the generating AI 4. In this way, the inquiry content intended by the speaker C can be narrowed down step by step.

[0052] If, after the question generation process in step S10, the generating AI4 outputs multiple inquiry contents in the current step S4, and the multiple inquiry contents output by the generating AI4 in the current step S4 are not included in the multiple candidate inquiry contents narrowed down by the output of the generating AI4 in the previous step S4, the automated voice response server 2 may undo the narrowing down of the candidate inquiry contents. That is, the automated voice response server 2 may reset the narrowed candidate inquiry contents to the multiple candidate inquiry contents registered in the database DB. By doing so, a smooth response can be achieved even if the speaker C changes their intention.

[0053] Furthermore, when the automated voice response server 2 inputs the voice content to the generating AI 4 in step S4 after the question generation process in step S10, the candidate query content that the generating AI 4 refers to may remain as one of the multiple candidate query content registered in the database DB. In this way, the AI ​​can ask questions to the speaker C and, based on the information extracted from the speaker C, identify the query content intended by the speaker from a wide range of candidate query content.

[0054] According to the configuration described above, the automated voice response system 1 can achieve a good balance between accurately narrowing down the inquiry content intended by speaker C and quickly reaching the answer intended by speaker C. Furthermore, since the database DB registers multiple answer candidates corresponding to multiple inquiry content candidates, it can easily and accurately identify the answer corresponding to speaker C's inquiry content.

[0055] Furthermore, since the generating AI4 outputs questions rather than answers, it prevents hallucination from occurring when presenting answers. Also, the generating AI4 is used to narrow down the content of the inquiry intended by speaker C, and since there is no need to make the database DB hierarchical, the database DB can be easily constructed.

[0056] Figure 8 is a diagram showing another example of the database DB in Figure 2. While Figure 4 explained an example where speaker C's intention is a query, Figure 8 explains an example where speaker C's intention is a request. As shown in Figure 8, the database DB is pre-registered with multiple request candidates and multiple message candidates corresponding to each of those request candidates.

[0057] For example, speaker C's request may include making a reservation, ordering, or canceling a product or service. The options may include menus, courses, or plans. The options may also include the person providing the service. The options may also include the name of the business or store providing the service or product. In the example in Figure 8, speaker C's request may be a reservation for one of several courses offered by the user organization's restaurant. Message candidates are candidate response sentences that are automatically responded to each of these request candidates via voice guidance.

[0058] In this case, the automated voice response server 2 replaces "inquiry content" with "request" and "answer" with "message" in the flowcharts shown in Figures 5 and 6, and performs the same processing.

[0059] [Note] The embodiments described above are specific examples of the following disclosures.

[0060] (Note 1) An automated voice response device that interacts with the speaker using voice, It includes a processing unit that can refer to a database containing multiple candidate intentions, The aforementioned processing unit, The first step is to obtain the content of the voice from the aforementioned speaker, A second step involves inputting the aforementioned content into a first machine learning model that has already been trained, and requesting the first machine learning model to output which of the candidate intentions the aforementioned content corresponds to. If the first machine learning model outputs only one intention selected from the aforementioned intention candidates, a third step is taken to identify the aforementioned one intention as the intention of the speaker, If the first machine learning model outputs multiple intentions selected from the aforementioned intention candidates, the fourth step is executed, which involves performing a question generation process and then returning to the first step. The aforementioned question generation process is: The above content is input into a second machine learning model that has already been trained, and the second machine learning model is instructed to output questions to the speaker in order to narrow down the multiple intentions output into a single intention. An automated voice response device, which includes outputting a question output by the second machine learning model to the speaker.

[0061] (Note 2) The first machine learning model is a generative AI that also serves as the second machine learning model, as described in Appendix 1 of the automated voice response device.

[0062] (Note 3) The automatic voice response device according to Appendix 1, wherein when the processing unit inputs the content to the first machine learning model, it inputs the content of a series of voices of the speaker in one dialogue session to the first machine learning model.

[0063] (Note 4) The automatic voice response device according to Appendix 1, wherein when the processing unit inputs the content to the second machine learning model, it inputs the content of a series of voices of the speaker in one dialogue session to the second machine learning model.

[0064] (Note 5) The automatic voice response device described in Appendix 1, wherein when the processing unit inputs the content to the first machine learning model after the question generation process, it narrows down the candidate intentions to the selected intentions.

[0065] (Note 6) The automated voice response device as described in Appendix 5, wherein the processing unit resets the narrowed-down candidate intentions to the multiple candidate intentions in the database if the intention output by the first machine learning model this time, after the question generation process, is not included in the multiple intentions output by the first machine learning model last time.

[0066] (Note 7) The automated voice response device as described in Appendix 1, wherein when the processing unit inputs the content to the first machine learning model after the question generation process, the intent candidates remain as the multiple intent candidates in the database.

[0067] (Note 8) The aforementioned database is a non-hierarchical database, as described in Appendix 1 of the automated voice response system.

[0068] (Note 9) The aforementioned intent is the content of the inquiry, The processing unit outputs a response to the identified inquiry to the speaker, as described in Appendix 1.

[0069] (Note 10) The aforementioned candidate intent is a candidate inquiry content, The automated voice response system described in Appendix 9 has a database which registers multiple answer candidates corresponding to each of the multiple inquiry candidate items.

[0070] (Note 11) The aforementioned intention is a request, The processing unit outputs a message corresponding to the identified request to the speaker, as described in Appendix 1.

[0071] (Note 12) The aforementioned requests include reservations, orders, or cancellations of product or service options, as described in Appendix 11 of the automated voice response system.

[0072] (Note 13) The aforementioned options include menus, courses, or plans, as described in Appendix 12, for the automated voice response system.

[0073] (Note 14) The aforementioned options include the name of the person in charge, the business office, or the store name, as described in Appendix 12 of the automated voice response system.

[0074] (Note 14) A voice-activated automated response method that engages in voice-based interaction with a speaker, The first step is to obtain the content of the audio from the speaker, A second step involves inputting the aforementioned content into a first machine learning model that has already been trained, and requesting the first machine learning model to output which of the multiple intent candidates in the database the aforementioned content corresponds to. If the first machine learning model outputs only one intention selected from the aforementioned intention candidates, a third step is taken to identify the aforementioned one intention as the intention of the speaker, If the first machine learning model outputs multiple intentions selected from the aforementioned intention candidates, the fourth step is executed, which involves performing a question generation process and then returning to the first step. The aforementioned question generation process is: The above content is input into a second machine learning model that has already been trained, and the second machine learning model is instructed to output questions to the speaker in order to narrow down the multiple intentions that are output into a single intention. A speech automated response method comprising: outputting a question output by the second machine learning model to the speaker.

[0075] (Note 15) An automated voice response program that causes a processor to execute the automated voice response method described in Appendix 14. [Explanation of Symbols]

[0076] 1. Automated Voice Response System 2. Automated voice response server (automated voice response device) 4. Generative AI (First Machine Learning Model, Second Machine Learning Model) 31 Processing Unit 131 processors C Speaker DB Database P Voice Automated Response Program

Claims

1. An automated voice response device that interacts with the speaker using voice, It includes a processing unit that can refer to a database containing multiple candidate intentions, The aforementioned processing unit, The first step is to obtain the content of the voice from the aforementioned speaker, A second step involves inputting the aforementioned content into a first machine learning model that has already been trained, and requesting the first machine learning model to output which of the candidate intentions the aforementioned content corresponds to. If the first machine learning model outputs only one intention selected from the aforementioned intention candidates, the third step is to identify the aforementioned one intention as the intention of the speaker, If the first machine learning model outputs multiple intentions selected from the aforementioned intention candidates, the fourth step is executed, which involves performing a question generation process and then returning to the first step. The aforementioned question generation process is: The above content is input into a trained second machine learning model, which is a generation AI, and the second machine learning model is instructed to output questions to the speaker in order to narrow down the multiple intentions output into a single intention. An automated voice response device, which includes outputting a question output by the second machine learning model to the speaker.

2. The voice automatic response device according to claim 1, wherein the first machine learning model is a generative AI that also serves as the second machine learning model.

3. The automatic voice response device according to claim 1, wherein when the processing unit inputs the contents to the first machine learning model, it inputs the contents of a series of voices of the speaker in one dialogue session to the first machine learning model.

4. The automatic voice response device according to claim 1, wherein when the processing unit inputs the contents to the second machine learning model, it inputs the contents of a series of speeches of the speaker in one dialogue session to the second machine learning model.

5. The automatic voice response device according to claim 1, wherein when the processing unit inputs the content to the first machine learning model after the question generation process, it narrows down the candidate intentions to the selected plurality of intentions.

6. The automated voice response device according to claim 5, wherein if the processing unit determines that the intention output by the first machine learning model this time, after the question generation process, is not included in the plurality of intentions output by the first machine learning model last time, it resets the narrowed-down intention candidates to the plurality of intention candidates in the database.

7. The automated voice response device according to claim 1, wherein when the processing unit inputs the content to the first machine learning model after the question generation process, the intent candidates remain as the plurality of intent candidates in the database.

8. The voice automatic response device according to claim 1, wherein the database is a non-hierarchical database.

9. The intent identified above is the content of the inquiry. The automatic voice response device according to claim 1, wherein the processing unit outputs a response to the inquiry to the speaker.

10. The aforementioned multiple candidate intentions are multiple candidate query content, The automated voice response device according to claim 9, wherein the database registers a plurality of answer candidates corresponding to each of the plurality of inquiry content candidates.

11. The aforementioned identified intention is a request. The automatic voice response device according to claim 1, wherein the processing unit outputs a message corresponding to the request to the speaker.

12. The voice automated response device according to claim 11, wherein the aforementioned requests include making reservations, orders, or cancellations for options relating to goods or services.

13. The voice automated response device according to claim 12, wherein the aforementioned options include a menu, a course, or a plan.

14. The aforementioned options include the name of the person in charge, the business office, or the store name, as described in claim 12.

15. A voice-activated automated response method that engages in voice-based interaction with a speaker, The first step is to obtain the content of the speech from the speaker, A second step involves inputting the aforementioned content into a first machine learning model that has already been trained, and requesting the first machine learning model to output which of the multiple intent candidates in the database the aforementioned content corresponds to. If the first machine learning model outputs only one intention selected from the aforementioned intention candidates, the third step is to identify the aforementioned one intention as the intention of the speaker, If the first machine learning model outputs multiple intentions selected from the aforementioned intention candidates, the fourth step is executed, which involves performing a question generation process and then returning to the first step. The aforementioned question generation process is: The above content is input into a trained second machine learning model, which is a generation AI, and the second machine learning model is instructed to output questions to the speaker in order to narrow down the multiple intentions output into a single intention. A voice automated response method comprising: outputting a question output by the second machine learning model to the speaker.

16. A voice automatic response program that causes a processor to execute the voice automatic response method described in claim 15.

Images