AI voice resource intelligent calling method and platform

Abstract

The invention relates to the technical field of voice interaction, and particularly discloses an AI voice resource intelligent calling method and platform, and the method comprises the steps: obtaining the watching content and watching parameters of a user based on interaction authority; determining interestingness of the watching contents according to the watching parameters, and sorting the watching contents according to the interestingness; extracting a keyword in the watching content, and sending an audio acquisition request pointing to the keyword to the user based on the interaction permission to obtain an audio clip of the keyword; counting the audio clips of each keyword to obtain an audio clip library of each keyword, and sorting the audio clip libraries according to the TF-IDF values of the keywords in the watched contents and the sorting result of the watched contents; and when audio information sent by a user is received, responding to the audio information based on the sorted audio clip library. On the basis of an AI voice recognition scheme, a local recognition scheme is introduced, and the recognition response speed is greatly increased.

Description

Technical Field

[0001] The present invention relates to the field of voice interaction technology, and specifically to an AI voice resource intelligent calling method and platform. Background Art

[0002] AI voice resource intelligence refers to the analysis, processing and optimization of voice data through artificial intelligence technology, especially natural language processing (NLP) and speech recognition technology.

[0003] AI voice recognition solutions are used in many fields, and interactive reading scenarios are one of them. Reading service providers generally have their own library. Users may have some questions when reading. With the help of AI voice recognition solutions, users can ask questions verbally at any time and get answers. In this process, AI voice recognition solutions are indispensable.

[0004] The AI ​​speech recognition solution has extremely strong recognition capabilities, but the recognition response speed is very slow, which makes the user's willingness to interact very low. How to improve the recognition response speed is the technical problem that the technical solution of the present invention wants to solve. Summary of the Invention

[0005] The purpose of the present invention is to provide an AI voice resource intelligent calling method and platform to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: A method for intelligently calling AI voice resources, the method comprising: Based on the user terminal receiving the interaction authority granted by the user, obtaining the user's viewing content and viewing parameters based on the interaction authority; the viewing parameters include the number of displays and the duration of a single display; determining the interest level of the viewing content according to the viewing parameters, and sorting the viewing content according to the interest level; Extract keywords from the content being watched, send an audio request to the user based on the interaction permission, and obtain the audio clip of the keyword; Count the audio clips for each keyword to obtain an audio clip library for each keyword, and sort the audio clip library based on the TF-IDF value of the keyword in the viewing content and the ranking results of the viewing content itself; When receiving audio information sent by the user, respond to it based on the sorted audio clip library; Among them, the process of responding to it based on the sorted audio clip library includes an AI-based speech recognition solution for updating the audio clips; the sorting results of the audio clips are updated in real time.

[0007] As a further solution of the present invention, the step of receiving the interaction authority granted by the user at the user end and obtaining the user's viewing content and viewing parameters based on the interaction authority includes: When the user terminal is activated, it sends an interaction permission acquisition request to the user and receives the interaction permission granted by the user; After receiving the interaction permission granted by the user, the recording port built into the user side is activated; Based on the recording port, the viewing instruction is monitored in real time. When the viewing instruction is monitored, the viewing content pointed to by the viewing instruction is obtained, and the top time period of the viewing content is recorded in real time. All sticky time periods are counted to obtain viewing parameters of the viewing content; the number of sticky time periods corresponds to the number of displays, and the time span of the sticky time period is called a single display duration.

[0008] As a further solution of the present invention, the step of determining the interest level of the viewing content according to the viewing parameters and sorting the viewing content according to the interest level includes: For each viewing content, statistics are collected on the viewing parameters of the viewing content on all users within a preset period; Create an arithmetic progression of time points based on a preset tolerance, called a time series; For each time point in the time series, if the top period in the viewing parameters of a certain user terminal includes this time point, mark the user terminal; Determine the interest level of the content being viewed based on the total number of tags corresponding to each time point; sorting viewing content based on the interest level; The calculation process of the interest degree is as follows: Get the total number of markers at each time point and build an array; Calculate interestingness based on array; The relationship between interest and array is: ; Where, For interest, and is the preset weight; For the first numerical values, The total number of values ​​in the array.

[0009] As a further solution of the present invention, the step of counting the audio clips for each keyword to obtain an audio clip library for each keyword, and sorting the audio clip library according to the TF-IDF value of the keyword in the viewing content and the ranking result of the viewing content itself includes: Count the audio clips with the keyword as the index to obtain the audio clip library for each keyword; Query the TF-IDF value of the keyword in the viewing content, query the interest of the viewing content, and determine the feature value of the audio clip library based on the TF-IDF value and interest; The audio segment library is sorted according to the feature values.

[0010] As a further solution of the present invention, upon receiving the audio information sent by the user, the step of responding to the audio information based on the sorted audio clip library includes: When receiving audio information sent by the user, read the audio clip library in sequence; Identify audio information based on the audio clip library and generate response instructions; When the audio information recognition based on the audio clip library fails, the audio information is recognized based on AI and the recognition result is recorded; The audio information and recognition results recognized by AI are used as samples to train a lightweight model. When the error rate of the lightweight model is less than the preset threshold, the lightweight model is used as the AI's pre-application model.

[0011] As a further solution of the present invention, upon receiving the audio information sent by the user, the step of responding to the audio information based on the sorted audio clip library further includes: Reading the error rate of the lightweight model, and determining the AI ​​application interval based on the error rate; After each application interval, an audio clip is generated based on AI and input into the lightweight model to verify the recognition accuracy of the lightweight model; Update the error rate of the lightweight model according to the recognition accuracy.

[0012] The technical solution of the present invention also provides an AI voice resource intelligent calling platform, which includes: A viewing information acquisition module is configured to receive the interaction authority granted by the user at the user terminal and acquire the user's viewing content and viewing parameters based on the interaction authority; the viewing parameters include the number of displays and the duration of a single display; an interest level determination module, configured to determine the interest level of viewing content based on the viewing parameters, and sort the viewing content based on the interest level; The audio clip acquisition module is used to extract keywords from the viewing content, send an audio acquisition request pointing to the keyword to the user based on the interaction permission, and obtain the audio clip of the keyword; A library construction module is used to count the audio clips of each keyword, obtain the audio clip library of each keyword, and sort the audio clip library according to the TF-IDF value of the keyword in the viewing content and the ranking result of the viewing content itself; An information response module, configured to respond to an audio message sent by a user based on the sorted audio clip library upon receiving the message; Among them, the process of responding to it based on the sorted audio clip library includes an AI-based speech recognition solution for updating the audio clips; the sorting results of the audio clips are updated in real time.

[0013] As a further solution of the present invention: the viewing information acquisition module includes: The permission acquisition unit is used to send an interaction permission acquisition request to the user when the user terminal is activated, and receive the interaction permission granted by the user; A port activation unit is used to activate the recording port built into the user terminal after receiving the interaction permission granted by the user; A time period recording unit is used to monitor the viewing instruction in real time based on the recording port. When the viewing instruction is monitored, the viewing content pointed to by the viewing instruction is obtained, and the top viewing time period of the viewing content is synchronously recorded in real time. The time period statistics unit is used to count all top time periods to obtain viewing parameters of the viewing content; the number of top time periods corresponds to the number of displays, and the time span of the top time period is called the single display duration.

[0014] As a further solution of the present invention: the interest determination module includes: A viewing parameter statistics unit, configured to collect statistics on viewing parameters of each viewing content on all user terminals within a preset period; The time series creation unit is used to create an arithmetic progression of time points based on a preset tolerance, called a time series; A user terminal marking unit is configured to mark a user terminal for each time point in the time series if the top time period in the viewing parameters of the user terminal includes the time point; an interest calculation unit, configured to determine the interest of the viewing content based on the total number of marks corresponding to each time point; a first sorting unit, configured to sort viewing contents based on the interest level; The calculation process of the interest degree is as follows: Get the total number of markers at each time point and build an array; Calculate interestingness based on array; The relationship between interest and array is: ; Where, For interest, and is the preset weight; For the first numerical values, The total number of values ​​in the array.

[0015] As a further solution of the present invention: the library construction module includes: A segment library acquisition unit is used to count audio segments using keywords as indexes to obtain an audio segment library for each keyword; A feature value calculation unit is used to query the TF-IDF value of the keyword in the viewing content, query the interest level of the viewing content, and determine the feature value of the audio clip library based on the TF-IDF value and the interest level; The second sorting unit is configured to sort the audio segment library according to the feature value.

[0016] Compared with the existing technology, the beneficial effects of the present invention are: based on the AI ​​speech recognition solution, the present invention introduces a local recognition solution, by identifying existing content, extracting keywords, and then obtaining the audio clip corresponding to each keyword to obtain a local database, and using the recognition process based on the local database as a prerequisite for the AI ​​speech recognition solution, greatly improving the recognition response speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention.

[0018] Figure 1 This is a flowchart of the AI ​​voice resource intelligent calling method.

[0019] Figure 2 This is a structural diagram of the AI ​​voice resource intelligent call platform. DETAILED DESCRIPTION

[0020] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0021] Figure 1 This is a flowchart of a method for intelligently calling AI voice resources. In an embodiment of the present invention, a method for intelligently calling AI voice resources includes: Step S100: receiving the interaction authority granted by the user at the user terminal, and obtaining the user's viewing content and viewing parameters based on the interaction authority; the viewing parameters include the number of displays and the duration of a single display; The application scenario of the technical solution of the present invention is an information interaction platform, such as a reading platform. The platform provides some books for user terminals to read. The user terminals are smart devices that can access the database of this platform, such as mobile phones and computers. Users can interact with the platform while reading, ask some questions through voice, and the platform will quickly answer them to deepen the user's understanding of the book.

[0022] In the above content, the user interacts with this platform with the help of the user terminal. During the interaction process, the user's information needs to be obtained. Therefore, the user needs to grant interaction permissions. A pop-up window will appear during use, and the user's authorization instruction will be received. The interaction permissions in the technical solution of the present invention also include reading process recording permissions, that is, this platform will record the user's viewing process. Of course, the user can also choose not to grant it, then the voice service provided by this application cannot be carried out, but the reading service and conventional voice service can still be provided, which is specifically selected by the user; for the sake of convenience, this application assumes that the user has granted interaction permissions.

[0023] Step S200: determining the interest level of the viewing content according to the viewing parameters, and sorting the viewing content according to the interest level; The viewing parameters of a viewing content refer to how many times and how long a user has pinned a certain content when watching it. Pinning refers to pinning at the display level, which is a display method. By analyzing the viewing parameters, the user's interest in the viewing content can be determined. The higher the interest, the more the user likes the viewing content. The above analysis process is for all user terminals. The viewing parameters of the same viewing content in all user terminals are obtained, and then the interest of the viewing content is calculated. The viewing content is arranged in descending order according to the interest. The higher the interest, the higher the viewing content is.

[0024] Step S300: extracting keywords from the viewing content, sending an audio acquisition request directed to the keyword to the user based on the interaction authority, and obtaining an audio clip of the keyword; In the technical solution of the present invention, the viewing content is generally a book, which is a text file. Keywords are extracted from the text file, and then an audio acquisition request pointing to the keyword is sent to the user. If the user feeds back an audio information, an audio clip of the keyword can be obtained.

[0025] Step S400: Count the audio clips for each keyword to obtain an audio clip library for each keyword, and sort the audio clip library according to the TF-IDF value of the keyword in the viewing content and the ranking result of the viewing content itself; The audio clips obtained for each keyword are not unique. For each keyword, all its audio clips are counted to obtain the audio clip library of each keyword. Since keywords themselves can be divided into regular and unconventional, they are distinguished by TF-IDF values. Combining the TF-IDF values ​​of keywords and the order of corresponding viewing content, keywords can be sorted to distinguish the importance of keywords. Generally speaking, the more common and important keywords are ranked higher. Accordingly, after the keyword sorting is completed, the audio clip library is sorted based on the same order.

[0026] Step S500: upon receiving the audio information sent by the user, responding to it based on the sorted audio clip library; Step S500 is a specific application process. This application constructs a local library of commonly used audio clips and their usage order. When receiving audio information sent by the user, the audio clip library is read in sequence, and then the audio information and the audio clips in the audio clip library are compared and matched. If the match is successful, the keywords of the audio information sent by the user are identified, and subsequent responses can be made based on the keywords. For the audio itself, the recognition has been completed.

[0027] It should be noted that the present application provides a comparison and recognition process based on the local audio library, which only needs to be traversed once to respond, and the speed is extremely fast, but it is easy to have mismatches. At this time, the recognition fails. Therefore, an external AI-based recognition process is required for intelligent recognition of audio. However, the model used by AI is extremely large and the response speed is very slow. The specific limitations are: the process of responding to it based on the sorted audio clip library contains an AI-based speech recognition solution for updating the audio clips; the sorting results of the audio clips are updated in real time.

[0028] Regarding step S100, the step of receiving the interaction authority granted by the user at the user terminal and obtaining the user's viewing content and viewing parameters based on the interaction authority includes: When the user terminal is activated, it sends an interaction permission acquisition request to the user and receives the interaction permission granted by the user; After receiving the interaction permission granted by the user, the recording port built into the user side is activated; Based on the recording port, the viewing instruction is monitored in real time. When the viewing instruction is monitored, the viewing content pointed to by the viewing instruction is obtained, and the top time period of the viewing content is recorded in real time. All sticky time periods are counted to obtain viewing parameters of the viewing content; the number of sticky time periods corresponds to the number of displays, and the time span of the sticky time period is called a single display duration.

[0029] The process of obtaining and applying permissions is not complicated. The above content provides a specific explanation of it. When the user terminal is used (the user enters the account and activates it), a request for interactive permission acquisition is sent to the user, and the interactive permission granted by the user is received. After receiving the interactive permission granted by the user, the recording port built into the user terminal is activated. The recording port obtains what content the user has viewed, that is, the viewing content pointed to by the viewing instruction entered by the user. At the same time, the recording port obtains the top time period of the viewing content, that is, at which time the viewing content is displayed on the top.

[0030] Finally, all sticky time periods are counted, the number of sticky time periods is used as the number of displays, and the time span of each sticky time period is used as the single display duration.

[0031] Regarding step S200, the step of determining the interest level of the viewing content according to the viewing parameters and sorting the viewing content according to the interest level includes: For each viewing content, statistics are collected on the viewing parameters of the viewing content on all users within a preset period; Create an arithmetic progression of time points based on a preset tolerance, called a time series; For each time point in the time series, if the top period in the viewing parameters of a certain user terminal includes this time point, mark the user terminal; Determine the interest level of the content being viewed based on the total number of tags corresponding to each time point; The viewing contents are ranked based on the interestingness.

[0032] The purpose of the above content is to sort the viewing content. For each viewing content, it may be viewed by users on different user terminals. Therefore, it is necessary to consider its viewing parameters on each user terminal. The specific consideration method is as follows: The platform determines a period, such as a month, and divides the month into multiple time points based on a preset tolerance, such as one minute, which are called time series. For a piece of viewing content, it may be displayed by any user terminal within this month. The time points are read in chronological order to obtain the top time period including the time point. If the user terminal displays the viewing content at that time point, the top time period is the one including the time point. Therefore, the number of top time periods indicates how many user terminals displayed the viewing content at that time point.

[0033] After the above processing, a number can be obtained for each viewed content at each time point, thereby constructing an array. By analyzing the array, it can be determined on which user terminals the viewed content was displayed within this month, how often it was displayed, and how long it was displayed. Based on the array, a parameter called interest level can be calculated, and the viewed content can be sorted based on the interest level.

[0034] Specifically, the relationship between the interest level and the array is: ; Where, For interest, and is the preset weight; For the first numerical values, The total number of values ​​in the array.

[0035] In actual applications, the more times R&D personnel browse and the longer they browse, will be bigger, and in addition, Indicates the change of browsing behavior over time. The more times you browse, the larger this value will be, and the higher the final interest level will be.

[0036] It should be noted that In the calculation function of When this term is used as the independent variable, the entire calculation function is an increasing function in the range of zero to one, that is, the interest level is limited to the range of zero to one. On this basis, by introducing a constant in the numerator, such as 100, the value range of the interest level can be adjusted at will (no constant is introduced in the formula).

[0037] Regarding step S400, the step of counting the audio clips for each keyword to obtain an audio clip library for each keyword, and sorting the audio clip library according to the TF-IDF value of the keyword in the viewing content and the ranking result of the viewing content itself includes: Count the audio clips with the keyword as the index to obtain the audio clip library for each keyword; Query the TF-IDF value of the keyword in the viewing content, query the interest of the viewing content, and determine the feature value of the audio clip library based on the TF-IDF value and interest; The audio segment library is sorted according to the feature values.

[0038] The keyword-based audio clip statistics process is not complicated, but the sorting process needs to be explained. The keywords in multiple viewing contents may be the same. Therefore, for all viewing contents, the TF-IDF value of each keyword can be calculated. Furthermore, in the technical solution of the present invention, the viewing content itself is sorted based on interest. This means that the keyword has both TF-IDF value and interest. In turn, a parameter reflecting whether the keyword is important can be determined, which is called a feature value. The audio clip library can be sorted with the help of the feature value.

[0039] Specifically, the calculation process of one of the eigenvalues ​​is: Receive the weights set by the platform, multiply them by the TF-IDF value and the interest level, and then add them up to get the final feature value. In addition to introducing coefficient weights, function weights can also be introduced. For example, a logarithmic function can be applied to the TF-IDF value. In this case, its weight will be lower (compared with the linear function, the logarithmic function changes more smoothly within the range greater than one).

[0040] Regarding step S500, upon receiving the audio information sent by the user, the step of responding to it based on the sorted audio clip library includes: When receiving audio information sent by the user, read the audio clip library in sequence; Identify audio information based on the audio clip library and generate response instructions; When the audio information recognition based on the audio clip library fails, the audio information is recognized based on AI and the recognition result is recorded; The audio information and recognition results recognized by AI are used as samples to train a lightweight model. When the error rate of the lightweight model is less than the preset threshold, the lightweight model is used as the AI's pre-application model.

[0041] In one example of the technical solution of the present invention, when audio information sent by a user is received, local recognition is performed first, and then AI recognition is performed; the process of local recognition is to read the audio clip library in sequence, recognize the audio information based on the audio clip library, and generate a response instruction; however, there is not much content in the audio clip library. Although in the application scenario of the present invention (interactive reading scenario), the keywords are limited, there will still be some other uncounted audios. At this time, it is necessary to use AI to recognize the audio information; on this basis, the present application provides another solution, which is based on AI training of a local lightweight recognition model. Compared with AI, it has a fast response speed, but a slightly reduced capability. However, since the application scenario of the present invention is actually a limited private domain scenario, when the lightweight model is good enough, it can be used as a front model of AI. At this time, the recognition process becomes: first local audio library recognition, then lightweight model recognition, and finally AI recognition.

[0042] Furthermore, when receiving the audio information sent by the user, the step of responding to it based on the sorted audio clip library also includes: Reading the error rate of the lightweight model, and determining the AI ​​application interval based on the error rate; After each application interval, an audio clip is generated based on AI and input into the lightweight model to verify the recognition accuracy of the lightweight model; Update the error rate of the lightweight model according to the recognition accuracy.

[0043] The recognition ability of the lightweight model cannot reach the level of AI. Therefore, an adversarial architecture is needed to verify the lightweight model. At preset time intervals, an audio clip is randomly generated based on AI and input into the lightweight model to verify the recognition accuracy of the lightweight model. When the recognition accuracy is low, the error rate is lowered. At this time, the lightweight model is no longer used and degraded to the training stage. If the recognition accuracy is high enough, the error rate is kept unchanged and continues to be used.

[0044] Figure 2 This is a structural block diagram of an AI voice resource intelligent call platform. In an embodiment of the present invention, an AI voice resource intelligent call platform is provided. The platform 10 includes: The viewing information acquisition module 11 is configured to receive the interaction authority granted by the user at the user terminal and acquire the user's viewing content and viewing parameters based on the interaction authority; the viewing parameters include the number of displays and the duration of a single display; an interest level determination module 12, configured to determine the interest level of the viewing content according to the viewing parameters, and sort the viewing content according to the interest level; The audio segment acquisition module 13 is used to extract keywords from the viewing content, send an audio acquisition request pointing to the keyword to the user based on the interaction authority, and obtain the audio segment of the keyword; A library construction module 14 is used to count the audio clips of each keyword to obtain an audio clip library for each keyword, and sort the audio clip library according to the TF-IDF value of the keyword in the viewing content and the ranking result of the viewing content itself; An information response module 15 is configured to respond to an audio message sent by a user based on the sorted audio segment library upon receipt of the audio message; Among them, the process of responding to it based on the sorted audio clip library includes an AI-based speech recognition solution for updating the audio clips; the sorting results of the audio clips are updated in real time.

[0045] Furthermore, the viewing information acquisition module 11 includes: The permission acquisition unit is used to send an interaction permission acquisition request to the user when the user terminal is activated, and receive the interaction permission granted by the user; A port activation unit is used to activate the recording port built into the user terminal after receiving the interaction permission granted by the user; A time period recording unit is used to monitor the viewing instruction in real time based on the recording port. When the viewing instruction is monitored, the viewing content pointed to by the viewing instruction is obtained, and the top viewing time period of the viewing content is synchronously recorded in real time. The time period statistics unit is used to count all top time periods to obtain viewing parameters of the viewing content; the number of top time periods corresponds to the number of displays, and the time span of the top time period is called the single display duration.

[0046] Specifically, the interest determination module 12 includes: A viewing parameter statistics unit, configured to collect statistics on viewing parameters of each viewing content on all user terminals within a preset period; The time series creation unit is used to create an arithmetic progression of time points based on a preset tolerance, called a time series; A user terminal marking unit is configured to mark a user terminal for each time point in the time series if the top time period in the viewing parameters of the user terminal includes the time point; an interest calculation unit, configured to determine the interest of the viewing content based on the total number of marks corresponding to each time point; a first sorting unit, configured to sort viewing contents based on the interest level; The calculation process of the interest degree is as follows: Get the total number of markers at each time point and build an array; Calculate interestingness based on array; The relationship between interest and array is: ; Where, For interest, and is the preset weight; For the first numerical values, The total number of values ​​in the array.

[0047] In addition, the library construction module 14 includes: A segment library acquisition unit is used to count audio segments using keywords as indexes to obtain an audio segment library for each keyword; A feature value calculation unit is used to query the TF-IDF value of the keyword in the viewing content, query the interest level of the viewing content, and determine the feature value of the audio clip library based on the TF-IDF value and the interest level; The second sorting unit is configured to sort the audio segment library according to the feature value.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for intelligently calling AI voice resources, characterized in that: The method comprises: Receiving the interaction authority granted by the user at the user end, and obtaining the viewing content and viewing parameters of the user based on the interaction authority; the viewing parameters include the number of displays and the duration of a single display; Determining the interest level of the viewing content according to the viewing parameter, and sorting the viewing content according to the interest level; Extract keywords from the viewing content, send an audio acquisition request pointing to the keywords to the user based on the interaction authority, and obtain the audio clip of the keyword; Count the audio clips of each keyword to obtain an audio clip library of each keyword, and sort the audio clip library according to the TF-IDF value of the keyword in the viewing content and the sorting result of the viewing content itself; When receiving an audio message sent by a user, respond to it based on the sorted audio clip library; Among them, the process of responding to it based on the sorted audio clip library includes an AI-based speech recognition solution for updating the audio clips; the sorting results of the audio clips are updated in real time.

2. The AI ​​voice resource intelligent calling method according to claim 1 is characterized in that: The step of receiving the interaction authority granted by the user at the user end and obtaining the viewing content and viewing parameters of the user based on the interaction authority comprises: When the user terminal is activated, a request for obtaining interaction permissions is sent to the user, and the interaction permissions granted by the user are received; After receiving the interaction permission granted by the user, the recording port built into the user side is activated; Based on the recording port, the reading instruction is monitored in real time. When the reading instruction is monitored, the viewing content pointed to by the reading instruction is obtained, and the top time period of the viewing content is synchronously recorded in real time; All sticky time periods are counted to obtain viewing parameters of the viewing content; the number of sticky time periods corresponds to the number of displays, and the time span of the sticky time period is called a single display duration.

3. The AI ​​voice resource intelligent calling method according to claim 1 is characterized in that: The step of determining the interest level of the viewing content according to the viewing parameters and sorting the viewing content according to the interest level comprises: For each viewing content, count the viewing parameters of the viewing content in all users within a preset period; Create an arithmetic progression of time points based on a preset tolerance, called a time series; For each time point in the time series, if the top time period in the viewing parameters of a certain user terminal includes the time point, the user terminal is marked; Determine the interest of viewing content based on the total number of marks corresponding to each time point; sorting the viewing content based on the interest level; The calculation process of the interest degree is as follows: Get the total number of markers at each time point and construct an array; Calculate interestingness based on array; The relationship between interest and array is: ； In the formula, For interest, and is the preset weight; For the first numerical values, The total number of values ​​in the array.

4. The AI ​​voice resource intelligent calling method according to claim 1, characterized in that: The step of counting the audio segments of each keyword to obtain the audio segment library of each keyword, and sorting the audio segment library according to the TF-IDF value of the keyword in the viewing content and the sorting result of the viewing content itself comprises: Count the audio clips with the keyword as the index to obtain the audio clip library for each keyword; Query the TF-IDF value of the keyword in the viewing content, query the interest of the viewing content, and determine the feature value of the audio clip library according to the TF-IDF value and the interest; The audio clip library is sorted according to the feature value.

5. The AI ​​voice resource intelligent calling method according to claim 1 is characterized in that: When receiving the audio information sent by the user, the step of responding to it based on the sorted audio clip library includes: When receiving the audio information sent by the user, read the audio clip library in sequence; Recognize audio information based on the audio clip library and generate response instructions; When the audio information recognition based on the audio clip library fails, the audio information is recognized based on AI and the recognition result is recorded; The audio information and recognition results recognized by AI are used as samples to train the lightweight model. When the error rate of the lightweight model is less than the preset threshold, the lightweight model is used as the front-end application model of AI.

6. The AI ​​voice resource intelligent calling method according to claim 1, characterized in that: When receiving the audio information sent by the user, the step of responding to it based on the sorted audio clip library also includes: Reading the error rate of the lightweight model, and determining the application interval of the AI ​​according to the error rate; At each application interval, an audio clip is generated based on AI and input into the lightweight model to verify the recognition accuracy of the lightweight model; Update the error rate of the lightweight model according to the recognition accuracy.

7. An AI voice resource intelligent calling platform, characterized in that: The platform includes: A viewing information acquisition module, used to receive the interaction authority granted by the user based on the user terminal, and acquire the user's viewing content and viewing parameters based on the interaction authority; the viewing parameters include the number of displays and the single display duration; An interest level determination module, configured to determine the interest level of the viewing content according to the viewing parameters, and sort the viewing content according to the interest level; An audio clip acquisition module is used to extract keywords from the viewing content, send an audio acquisition request pointing to the keyword to the user based on the interaction authority, and obtain the audio clip of the keyword; A library construction module is used to count the audio clips of each keyword, obtain the audio clip library of each keyword, and sort the audio clip library according to the TF-IDF value of the keyword in the viewing content and the sorting result of the viewing content itself; An information response module, for responding to an audio message sent by a user based on the sorted audio clip library when receiving the audio message; Among them, the process of responding to it based on the sorted audio clip library includes an AI-based speech recognition solution for updating the audio clips; the sorting results of the audio clips are updated in real time.

8. The AI ​​voice resource intelligent calling platform according to claim 7, characterized in that: The viewing information acquisition module includes: The permission acquisition unit is used to send an interaction permission acquisition request to the user when the user terminal is activated, and receive the interaction permission granted by the user; A port activation unit, used to activate a recording port built into the user terminal after receiving the interaction permission granted by the user; A time period recording unit is used to monitor the viewing instruction in real time based on the recording port, and when the viewing instruction is monitored, obtain the viewing content pointed to by the viewing instruction, and synchronously record the top time period of the viewing content in real time; The time period statistics unit is used to count all top time periods to obtain viewing parameters of the viewing content; the number of top time periods corresponds to the number of displays, and the time span of the top time period is called a single display duration.

9. The AI ​​voice resource intelligent calling platform according to claim 7, characterized in that: The interest level determination module comprises: A viewing parameter statistics unit, used for counting the viewing parameters of each viewing content in all user terminals within a preset period; The time series creation unit is used to create an arithmetic progression of time points based on a preset tolerance, called a time series; A user terminal marking unit, for marking a user terminal for each time point in the time series if the top time period in the viewing parameters of a user terminal includes the time point; An interest calculation unit, used to determine the interest of viewing content according to the total number of marks corresponding to each time point; A first sorting unit, configured to sort the viewing content based on the interest level; The calculation process of the interest degree is as follows: Get the total number of markers at each time point and construct an array; Calculate interestingness based on array; The relationship between interest and array is: ； In the formula, For interest, and is the preset weight; For the first numerical values, The total number of values ​​in the array.

10. The AI ​​voice resource intelligent calling platform according to claim 7, characterized in that: The library construction module includes: A segment library acquisition unit is used to count audio segments with keywords as indexes to obtain an audio segment library for each keyword; A feature value calculation unit, used to query the TF-IDF value of the keyword in the viewing content, query the interest of the viewing content, and determine the feature value of the audio clip library according to the TF-IDF value and the interest; The second sorting unit is used to sort the audio segment library according to the feature value.

Images