AI live broadcast verbal skill dynamic generation and field control cooperative management system and method

Abstract

The invention discloses a live broadcast verbal skill generation and field control cooperative management system and method based on an AI large model, and belongs to the technical field of intelligent live broadcast. The system comprises a master control system which integrates a verbal skill generation engine and a field control cooperation engine. The verbal skill generation engine generates verbal skill through an AI large model based on commodity information and real-time live broadcast data, and performs dynamic optimization through verbal skill round number management and compliance filtering. And the field control cooperation engine captures multi-platform bullet screens and automatically executes window popup, interaction and process control according to a preset rule. The system is characterized in that the system is provided with a linkage arbitration module, unified scheduling and conflict judgment are carried out on output instructions of the two engines, and deep collaboration and closed-loop management of verbal skill content and operation actions are achieved. According to the invention, the problems of disjunction between content generation and field control execution, low automation degree and tedious operation in the prior art are solved, and high-quality and full-automatic live broadcast operation is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to an AI-powered live-streaming script dynamic generation and on-site control collaborative management system and method. Background Technology

[0002] With the rapid development of live-streaming e-commerce, improving the automation and intelligence of live-streaming operations has become a core demand of the industry. Existing technologies attempt to solve the problem of low efficiency in manual operations from different angles, mainly forming two technical routes: one focuses on automating the live-streaming format and interaction mode, and the other focuses on the intelligentization of pre-live-streaming preparation.

[0003] The first approach is represented by virtual digital human live streaming technology. As shown in Chinese patent document CN116248909A, it proposes a virtual digital human live streaming method based on takeover and hosting collaboration. The core of this method lies in the operator preparing scripts and materials in advance, creating video content, and configuring multiple response videos for preset questions. During the live stream, a hosting mode is primarily used. When encountering unpredictable questions, a real person can intervene to take over, or the system can attempt to search for answers online. This solution combines the advantages of stable output from virtual humans and flexible responses from real people through "hosting-takeover" collaboration. However, its scripts and responses are essentially pre-recorded static videos, lacking the ability to generate and adjust content in real time based on the dynamics of the live stream, such as user emotions and changes in the number of online viewers. There is also no intelligent linkage between scene control actions, such as pop-ups and interactions, and the degree of automation is limited.

[0004] The second approach focuses on using data intelligence to simplify pre-broadcast preparations. For example, Chinese patent document CN120223914A discloses a live-streaming processing method. This method intelligently recommends products to be explained based on the live-streaming account's historical data and product attributes, and automatically generates corresponding live-streaming scripts based on product descriptions, forming a live-streaming plan. This solution significantly improves the efficiency of product selection and script preparation. However, its technical focus lies in generating the plan before the live stream begins. Once the live stream starts, the generated scripts are relatively fixed text and cannot be dynamically optimized and adjusted based on real-time interactive data, such as bullet comments and user behavior, during the live stream. Furthermore, this solution does not address live-streaming management, such as automatic pop-ups, intelligent replies, and process control; script generation and live-streaming execution remain separate processes.

[0005] This reveals a common limitation in existing technologies: the generation of live-stream content (i.e., the script) is separate from the execution of live-stream operations (i.e., the control of the live stream), failing to form a closed-loop intelligent system. This leads to: 1. Rigid content: The scripts cannot be dynamically generated and optimized based on real-time interactive context, resulting in a monotonous user experience; 2. Execution disconnect: Marketing actions, such as pop-ups and promotions, cannot be accurately synchronized with the real-time generated script content, resulting in a break in the conversion process; 3. Cumbersome operation: Although the efficiency of individual points has improved, the anchor or operation staff still need to perform high-intensity manual coordination and operation in multiple links such as script guidance, interactive response, and process control, resulting in low automation of the entire process.

[0006] Therefore, there is an urgent need in this field for a technical solution that can deeply couple and integrate real-time intelligent script generation with precise automated scene control execution throughout the entire live streaming process, so as to truly achieve end-to-end automation from content creation to operation execution. Summary of the Invention

[0007] Therefore, it is necessary to provide an integrated system and method that can deeply couple intelligent content generation with precise operation execution, in order to overcome the technical problems of the disconnect between live broadcast script generation and scene control management and the low degree of automation in existing technologies.

[0008] To address the aforementioned technical problems, this invention proposes a collaborative management system for live-streaming script generation and event control based on a large AI model. The core of this AI-powered live-streaming script generation and event control collaborative management system lies in a master control system. This master control system is not simply an integration of two independent modules, but rather a unified decision-making and execution center. Specifically, the system includes: A script generation engine is used to dynamically generate and optimize live streaming scripts based on multi-dimensional information. The live streaming control and collaboration engine is used to automatically execute live streaming operation actions according to preset rules; The linkage arbitration module is used to coordinate the command output of the two engines mentioned above.

[0009] The script generation engine further includes a multi-source information fusion unit, an AI large-scale model service unit, a script turn management unit, and a compliance filtering unit. Its technical approach lies in not only utilizing the AI ​​large-scale model to ensure the creativity and contextual relevance of the script, but also proactively preventing content duplication through algorithmic script turn management, and controlling risks in real time through compliance filtering.

[0010] The field control collaboration engine further includes a rule-driven executor. Its technical approach lies in abstracting operational actions into structured, configurable rules, such as executing action A when condition C is met. This allows complex field control logic to be flexibly defined through configuration files, achieving low-code implementation of operational strategies.

[0011] Specifically, the core technical means of this invention lies in the linkage arbitration module. This module, acting as the central nervous system of the system, receives content instructions from the script generation engine in real time, such as pushing a script, and response events from the scene control collaboration engine, such as user inquiries and rule events, such as timed pop-ups. It can arbitrate and sort these parallel or conflicting instructions according to a preset global priority strategy, such as prioritizing real-time user interaction over preset tasks, outputting a unified and orderly execution sequence. This ensures that in complex and ever-changing live streaming scenarios, actions such as script pushing, bullet screen replies, and product pop-ups can occur in a coordinated and orderly manner, thus truly achieving a seamless closed loop from content generation to action execution.

[0012] Accordingly, the present invention also provides a method and a computer-readable storage medium based on the above system.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Achieve full-process closed-loop automation: By linking the arbitration module, real-time script generation is deeply integrated with automated field control, realizing a complete closed loop of perception, decision-making, generation and execution, and greatly reducing manual intervention.

[0014] 2. Dynamic, intelligent, and compliant content: Based on AI big data models and real-time data, the scripts are dynamically generated, and the deduplication algorithm ensures the freshness of the content, while real-time filtering ensures the safety of the content.

[0015] 3. Precise and coordinated operational actions: The script content can precisely trigger corresponding scene control actions, such as pop-up windows when the product is mentioned. The marketing actions and the explanation content are highly synchronized, improving conversion efficiency.

[0016] 4. High flexibility and configurability: The field control rules are defined through configuration files, and even non-technical personnel can easily adjust the operation strategy, making the system highly adaptable. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall system architecture and data flow of the AI ​​live streaming script dynamic generation and scene control collaborative management system of the present invention; Figure 2 This is a schematic diagram of the speech round management and compliance filtering process of the AI ​​live broadcast speech dynamic generation and scene control collaborative management system of the present invention; Figure 3 This is a schematic diagram of the scene control rule parsing and execution logic of the AI ​​live broadcast script dynamic generation and scene control collaborative management system of the present invention; Figure 4 This is a flowchart of the decision-making process of the linkage arbitration module in the AI ​​live broadcast script dynamic generation and on-site control collaborative management system of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0019] Example 1: Taking an intelligent e-commerce live stream for clothing as an example, the implementation method of the present invention is described in detail.

[0020] 1. System Architecture and Initialization refer to Figure 1 The main control system of this system is deployed on a server and connects to live streaming clients such as OBSStudio and live streaming platforms such as Douyin Open Platform via API.

[0021] During system initialization, operations personnel perform the following configurations through the web management backend: (1) Import the product list for this live stream, products.yaml. Example content is as follows: #products.yaml example

[0022] (2) Write the field control rule configuration file rules.yaml, refer to... Figure 3 Example content is as follows: #rules.yaml example

[0023] (3) Load the general compliant thesaurus and the peer brand words that are prohibited in this session.

[0024] 2. Script Generation and Optimization Process The live stream has started; the product being discussed is SKU123. For the workflow of the sales script generation engine, please refer to [link / reference]. Figure 2 : (1) Multi-source information fusion: The engine obtains information about product SKU123, such as keywords: drape, retro, etc., and real-time data, such as online users: 5000, and popular bullet comments: "Is it suitable for petite people?"

[0025] (2) AI Large Model Generation: AI Large Model Service Unit, which can call the domestically optimized open source large model API, and receive the prompt: "Generate a 30-second recommendation script for 'Retro Flowy Dress' that highlights the 'flowy and slimming' feature and responds to the concerns of 'petite'." Generate the initial script text A.

[0026] (3) Repetition of dialogue rounds: a. Dialogue A is preprocessed to remove stop words, and its text fingerprint FP_A is generated using the MinHash algorithm.

[0027] b. Retrieve the Historic_FP_Set, a set of historical verbal fingerprints with the topic “SKU123”, from the cache, containing 3 fingerprints from the past 20 minutes.

[0028] c. Calculate the Jaccard similarity between FP_A and each fingerprint in Historic_FP_Set, assuming the highest value is 0.8.

[0029] d. The system's preset threshold T = 0.7. Since 0.8 > 0.7, the repetition is judged to be too high, triggering a reconstruction.

[0030] e. The system sends a new prompt to the AI ​​model: "Please keep the core selling points, but avoid using overly similar expressions as before. You can try to elaborate from the perspective of 'fabric composition' or 'styling suggestions'." f. The AI ​​large model generates new text B. After calculating FP_B, its maximum similarity with the historical set is verified to be 0.65, which is required to be <0.7, thus passing the deduplication check.

[0031] (4) Compliance filtering: The compliance filtering unit scans statement B and finds no prohibited words. Statement B is marked as "available".

[0032] (5) Output: The script B and its fingerprint FP_B are stored in the history cache of “SKU123”. At the same time, the script B, as a “medium priority” “push script” event P, is sent to the pending queue of the linkage arbitration module.

[0033] 3. Site control coordination and event monitoring process The field control collaboration engine continues to run: (1) When the rule-driven executor monitors the system time and reaches 300 seconds after the broadcast starts, it triggers the rule rule_send_welcome, generates a "low priority" "send welcome barrage" event E1, and submits it to the linkage arbitration module.

[0034] (2) Almost simultaneously, the bullet screen analysis submodule captures the user's bullet screen: "What size should I wear if I am 165cm tall and weigh 100kg?" The intent recognition model determines that it is an intent to "ask for size", triggers the rule rule_auto_reply_size, generates a "high priority" "reply to size inquiry" event E2, and submits it to the linkage arbitration module.

[0035] 4. Linked Arbitration and Enforcement Process refer to Figure 4The linkage arbitration module receives three events at this time: event P, dialogue B, medium priority; event E1, low priority; and event E2, high priority.

[0036] (1) Arbitration decision: According to the preset high > medium > low priority strategy, the arbitration module first processes the high priority event E2.

[0037] (2) Execute high-priority events: Immediately call the live streaming platform API through the external execution interface to send the reply template content configured in the rules to the public screen of the live streaming room to reply to the user.

[0038] (3) Execution of a medium-priority event: Next, the arbitration module processes the medium-priority event P. The content of the script B is pushed to the live streaming client, i.e., the OBS teleprompter, for the broadcaster to read.

[0039] (4) Triggering a new event: The anchor begins to read script B. When saying "Its drape fabric is very smooth...", the speech-to-text module recognizes the keyword "drape". The rule-driven executor of the scene control collaboration engine listens to this event, finds that "drape" matches the keyword list of product SKU123, thereby triggering the rule rule_popup_on_keyword, generating a new medium-priority event E3, displaying the SKU123 pop-up window, and submitting it to the arbitration module.

[0040] (5) Execute collaborative actions: After receiving E3, the arbitration module immediately executes the command and control via API to pop up a product card pop-up window for SKU123 at a specified location on the screen. At this time, the sales pitch and the product pop-up window are precisely synchronized.

[0041] (6) Execute low priority events: Finally, low priority events E1, such as sending a welcome message, are executed in a short idle window after its trigger time, such as 2 seconds after the previous action is completed.

[0042] 5. Process complete When the live stream reached 22:00, the rule rule_end_stream was triggered, and the system automatically executed the end-of-stream process.

[0043] Throughout the live stream, the aforementioned cycle of real-time perception, intelligent generation, arbitration decision-making, and collaborative execution continues to operate. The script generation engine continuously generates personalized and unique content based on the latest interactions, the scene control and collaboration engine precisely executes rule-based operational actions, and the linkage arbitration module ensures that all actions occur at the right time and in the right order, achieving highly automated operation of the entire live stream process without human intervention.

[0044] To further optimize the decision-making intelligence and execution reliability of the linkage arbitration module, this invention continues to provide Embodiment 2, which improves the core arbitration strategy of the arbitration module as follows: 1. Arbitration model based on dynamic priority weight calculation The core decision-making mechanism of the joint arbitration module is based on a dynamic priority weight calculation model. This model believes that the urgency of an event should not be determined solely by a preset static label, but rather by a dynamic value jointly determined by basic priority, timeliness decay, and real-time business value.

[0045] For each event to be arbitrated, the arbitration module calculates a dynamic execution weight, W, for it when it enters the queue. The formula for calculating the W value is defined as follows:

[0046] in: P_base, or basic static priority, corresponds to the labels "high", "medium", "low" defined in the rule configuration and is mapped to fixed values ​​in the system, such as 90, 60, and 30; it reflects the inherent importance of the event.

[0047] F_time, the timeliness decay factor, measures the risk of an event becoming invalid due to waiting. Its value decays over time; for example, for a timed "welcome message" event, its F_time can be designed to decay exponentially with the delay time (t).

[0048] Where λ is the decay coefficient. This ensures that periodic tasks that are excessively delayed will automatically have their weight reduced, avoiding accumulation.

[0049] F_business, or Real-Time Business Value Factor: This is key to model intelligence. This factor is dynamically evaluated and assigned a value by the system based on the real-time context when an event is generated. For example: A user inquiry event identified as having "strong purchase intent" can have its F_business value set to 100.

[0050] A push event for a "promotional message" generated based on currently best-selling products, whose value can be set to 80.

[0051] A typical "product feature explanation" presentation might have a value of 50.

[0052] The evaluation of this factor can be combined with real-time data from multiple dimensions, such as user history, current product conversion rate, and conversation sentiment analysis.

[0053] α, β, and γ are configurable weighting coefficients used to adjust the relative importance of the three dimensions in the final decision. System administrators can flexibly configure these coefficients according to different live streaming scenarios, such as new product launches, clearance sales, and routine presentations. The default settings are α=0.5, β=0.2, and γ=0.3, emphasizing basic priority and business value.

[0054] In each scheduling cycle, for example, every 100 milliseconds, the arbitration module recalculates the W value of all events in the queue and generates a sequence of execution instructions strictly according to the W value in descending order. This model enables the system to transcend fixed priority rules and achieve adaptive scheduling based on maximizing real-time utility.

[0055] 2. Conflict resolution and guarantee rules based on finite state machines To ensure the determinism and robustness of system behavior under the aforementioned dynamic scheduling, the linkage arbitration module also incorporates a conflict resolution rule base based on a finite-state machine (FSM). These rules define how event states transition and the system's response strategies in specific conflict scenarios. Examples of key rules are as follows: Rule R1, High-Priority Interruption and Compensation: When a high-priority real-time interactive event (e.g., Event_User Emergency Inquiry) conflicts with a medium-priority planned event (e.g., Event_Product Pop-up A) that is about to be executed, the state machine-driven arbitration module immediately executes the interactive event; simultaneously, the planned event's state is changed from ready to suspended. After the high-priority event is completed, the module does not simply discard the suspended event, but checks whether its triggering conditions are still valid. For example, it checks whether the current broadcaster's speech-to-text still contains related product keywords. If the conditions are valid, the event's state is restored to ready, and a temporary weight boost is given so that it can be executed as compensation in the next scheduling cycle; if the conditions have expired, its state is set to discarded.

[0056] Rule R2, Intelligent Sorting of Priority Queues: For multiple events with similar W values ​​and the same priority, such as multiple explanatory scripts to be pushed, the state machine adopts a sub-strategy that prioritizes the shortest waiting time, with an additional topic coherence check. The algorithm will prioritize events with longer waiting times, but will also evaluate their semantic relevance to already executed events. If switching topics would cause too much contextual jump, such as suddenly switching from "clothing fabric" to "electronic product parameters," execution will be suspended, and a transitional script event will be inserted to ensure the smoothness and naturalness of the live stream content.

[0057] Rule R3, Strong Control of the Process Terminal: When a "process control event," such as Event_Stop Live Streaming," is triggered, the state machine immediately switches the system to the "closing state." In this state, the arbitration module sends a "termination signal" to the script generation engine and the scene control collaboration engine, and clears all unnecessary marketing and entertainment events from the queue. Only predefined key event sequences related to "Live Stream End" are retained and prioritized for execution, such as "acknowledgment messages," "follow-up prompts," and "end-of-stream animation," ensuring the integrity and professionalism of the process.

[0058] Through the collaborative operation of the aforementioned dynamic weight calculation model and finite state machine rule base, the linkage arbitration module provided by this invention not only achieves intelligent event sorting but also possesses advanced decision-making capabilities for handling complex conflicts, ensuring critical business processes, and maintaining user experience consistency. This constitutes a technological leap from "simple instruction forwarding" to "intelligent scheduling hub," and is the core guarantee for this system to achieve high-quality automated live streaming throughout the entire process.

[0059] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0060] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An AI-powered live-streaming script dynamic generation and on-site control collaborative management system, characterized in that, include: A main control system, the main control system integrating: The script generation engine is used to generate live streaming scripts based on the input structured product information and real-time live streaming data through a large AI model, and to perform round-by-round deduplication and compliance filtering on the generated scripts. The live streaming control collaboration engine is used to capture live stream comments in real time and perform intent recognition through a multi-platform interface adaptation layer. At the same time, it loads and executes preset live streaming control rules to automatically control product pop-ups, interactive comment sending, and the live streaming process. The linkage arbitration module is connected to the script generation engine and the scene control collaboration engine respectively. It is used to receive and arbitrate instruction events from the two engines, generate a unified execution sequence according to the preset priority, and drive the live broadcast client to perform corresponding operations.

2. The AI ​​live streaming script dynamic generation and scene control collaborative management system according to claim 1, characterized in that: The script generation engine includes: A multi-source information fusion unit is used to access data streams from product databases and live streaming platforms; The AI ​​large model service unit calls the generative pre-trained large model to generate the initial speech text based on the fused information; The script turn management unit uses text fingerprinting technology based on minimum hashing to calculate the Jaccard similarity between the new script and the set of historical scripts on the same topic. If the similarity exceeds a set threshold, script reconstruction is triggered. The compliance filtering unit has a built-in dynamically updated prohibited word library for scanning and replacing dialogue text.

3. The AI ​​live streaming script dynamic generation and scene control collaborative management system according to claim 2, characterized in that: The real-time live stream data includes the ID of the product being explained, the number of people online at the moment, the word cloud of trending words in the comments, and the frequency of user interaction.

4. The AI ​​live streaming script dynamic generation and scene control collaborative management system according to claim 1, characterized in that: The field control collaboration engine includes: A rule-driven executor is used to parse field control rules defined in a structured configuration file, wherein the rules include at least triggering conditions, execution actions, and action parameters; The triggering condition is a logical judgment expression of the live room status variable, and the execution action includes at least one of displaying a product pop-up window, sending system bullet comments, and performing live broadcast start / stop operations.

5. The AI ​​live streaming script dynamic generation and scene control collaborative management system according to claim 4, characterized in that: The structured configuration file is in YAML or JSON format, and the live stream status variables include system time, current product ID, specific keyword hit status, and online user range.

6. The AI ​​live streaming script dynamic generation and scene control collaborative management system according to claim 1, characterized in that, The preset priority strategy of the linkage arbitration module is as follows: real-time user consultation response instructions have higher priority than preset timed task instructions, and preset timed task instructions have higher priority than regular script push instructions.

7. The AI ​​live streaming script dynamic generation and scene control collaborative management system according to claim 1, characterized in that, The linked arbitration module includes: The dynamic weight calculation unit is used to calculate the dynamic execution weight value W for each instruction event to be processed. The calculation formula is as follows: Where P_base is the preset basic priority score based on the event type; F_time is the timeliness factor that decays with the event waiting time; F_business is the business value factor that is dynamically evaluated based on the real-time live broadcast context data; α, β, γ are configurable weight coefficients. The linkage arbitration module is configured to periodically sort events from high to low according to their dynamic execution weight values ​​W, and generate the unified execution sequence.

8. A method for dynamic generation of AI live streaming scripts and collaborative management of live streaming control, wherein the method applies the AI ​​live streaming script dynamic generation and collaborative management system as described in any one of claims 1-7, characterized in that, Includes the following steps: S1: System initialization, loading product information, preset site control rules and compliance terminology; S2: The script generation engine continuously acquires real-time data, calls the AI ​​big model to generate scripts, performs deduplication and filtering, and submits the usable scripts to the linkage arbitration module. S3: The live stream control collaboration engine captures and analyzes bullet screen events in real time, while monitoring the status of the live stream to determine the triggering conditions of the live stream control rules, and submits the events to be responded to to the linkage arbitration module. S4: The linkage arbitration module arbitrates and sorts the various instruction events received according to the preset strategy, and generates a unified execution instruction sequence for the current moment; S5: The system drives the external interface to perform corresponding message push, bullet screen reply, pop-up control or process operation according to the execution instruction sequence.

9. The AI-powered live-streaming script dynamic generation and collaborative management method for live-streaming control as described in claim 8, characterized in that, In step S2, the deduplication specifically involves: calculating the text fingerprint for the generated speech, querying the historical fingerprint set corresponding to the speech topic, and if the maximum similarity exceeds the threshold T, then modifying the prompt word parameters to require the AI ​​large model to reconstruct the speech until the similarity between the new speech fingerprint and all fingerprints in the historical set is lower than the threshold T.

10. The AI ​​live streaming script dynamic generation and scene control collaborative management method according to claim 8, characterized in that, In step S4, when a high-priority event is triggered, the linkage arbitration module can delay or suspend low-priority periodic task instructions.

Images