A multi-node collaborative hearing aid decision service method and system

Abstract

This invention discloses a multi-node collaborative hearing aid decision-making service method and system, belonging to the field of intelligent hearing aid service technology. The method comprises: receiving natural language text and hearing aid model parameters; performing scheme matching on the natural language text from the adjustment scheme library corresponding to the hearing aid model parameters to obtain several candidate matching schemes; selecting a target matching scheme from the candidate matching schemes based on a large language model, and generating a user intent status marker based on the target matching scheme; if the status marker is a first intent status marker, generating a structured instruction and sending it to the hearing aid to complete the hearing aid decision-making service; if the status marker is a second intent status marker, transferring the shared state in the shared state stream to an intelligent customer service node, causing the intelligent customer service node to generate a customer service response decision and send it to the hearing aid to complete the hearing aid decision-making service. This invention can improve the reliability of hearing aid decision-making services.

Description

Technical Field

[0001] This invention relates to the field of intelligent hearing aid services, and in particular to a multi-node collaborative hearing aid decision-making service method and system. Background Technology

[0002] Hearing aids are essential devices that help people with hearing loss compensate for their hearing loss and improve their speech communication. The actual listening experience of a hearing aid highly depends on the proper configuration of audio parameters such as noise reduction level, gain, and equalization. However, users frequently switch between home, outdoor, and noisy public environments, and individual hearing conditions and wearing preferences vary greatly, necessitating frequent and dynamic adjustments to audio parameters. Simultaneously, users have numerous inquiries regarding device pairing, function usage, and troubleshooting. Current hearing aid services involve multiple professional parameters and complex interfaces, making it difficult for ordinary users, especially the elderly, to understand and operate independently. Furthermore, adjustments and consultations often require switching between different channels, resulting in prominent problems such as high operational barriers, delayed response times, poor fit, and fragmented service. Therefore, ensuring timely and consistent responses to users' dynamic needs in different scenarios is a technical problem that requires further research.

[0003] Currently, existing hearing aid decision-making services primarily employ two types of technologies. The first is traditional manual fitting services, where professionals adjust parameters such as noise reduction level, gain, and equalization online or offline based on user descriptions or test data. The second is basic intelligent adjustment solutions, which provide users with parameter presets for limited, fixed scenarios. Some technologies also use model fine-tuning to build knowledge service modules to respond to inquiries. Existing solutions suffer from several drawbacks: the former heavily relies on human intervention, suffers from poor timeliness, and has high operating costs; the latter only provides limited configurations, lacks accurate natural language interaction and left / right ear differentiation capabilities, and model fine-tuning faces long knowledge update cycles and is prone to causing hallucinations. Furthermore, existing solutions fail to integrate high-frequency audio adjustment needs with numerous customer inquiries into a unified service framework, resulting in a disconnect between adjustment and consultation, and an inability to respond promptly and consistently to dynamically changing user needs. Therefore, existing technologies lack an integrated service that addresses both high-frequency audio adjustment needs and numerous customer inquiries, leading to poor reliability in hearing aid decision-making services when faced with dynamic user demands. Summary of the Invention

[0004] This invention proposes a multi-node collaborative hearing aid decision-making service method and system, which solves the problem of poor reliability in hearing aid decision-making services when faced with dynamic user needs due to the lack of integrated services that respond to the high-frequency audio adjustment needs and numerous customer service inquiries of existing technologies. This invention achieves seamless collaboration between audio adjustment services and intelligent customer service services, thereby improving the reliability of hearing aid decision-making services.

[0005] To achieve the above objectives, embodiments of the present invention provide a multi-node collaborative hearing aid decision service method, applied to a hearing aid, wherein the hearing aid includes an audio adjustment node and an intelligent customer service node; the audio adjustment node and the intelligent customer service node operate in the same shared state stream; the multi-node collaborative hearing aid decision service method is executed by the audio adjustment node; the multi-node collaborative hearing aid decision service method includes: receiving pre-acquired natural language text and hearing aid model parameters; performing scheme matching on the natural language text from an adjustment scheme library corresponding to the hearing aid model parameters to obtain several candidate matching schemes; and according to a preset large language module... The system selects a target matching scheme from the candidate matching schemes and generates a user intent status marker based on the target matching scheme. If the status marker is a first intent status marker, a structured instruction is generated based on the target matching scheme, and the structured instruction is sent to the hearing aid to complete the hearing aid decision service. If the status marker is a second intent status marker, the shared state in the shared state stream is transferred to the intelligent customer service node, which generates a customer service response decision through a preset retrieval enhancement generation technology, so that the intelligent customer service node sends the customer service response decision to the hearing aid to complete the hearing aid decision service.

[0006] This invention proposes a multi-node collaborative hearing aid decision-making service method. It sets up an audio adjustment node and an intelligent customer service node, both operating within the same shared state flow. The audio adjustment node performs solution matching, while a large language model filters target solutions and outputs intent state markers. Based on different state markers, hearing aid parameter adjustments are performed, or the complete shared state is transferred to the intelligent customer service node. Enhanced retrieval is used to generate and output customer service response decisions. Thus, the shared state flow between the two nodes enables real-time communication between the dialogue context and device parameters. The large language model differentiates between adjustment and non-adjustment requests for separate processing. Adjustment requests are closed-loop within the audio adjustment node, while non-adjustment requests are deterministically routed to the customer service node as a fallback. Users do not need to repeat descriptions or switch channels, and the state markers provide a clear exit point for each round of input, ensuring no request is missed. This improves the reliability of the hearing aid decision-making service under dynamically changing user needs.

[0007] Furthermore, the system receives pre-acquired natural language text and hearing aid model parameters, performs scheme matching on the natural language text from the adjustment scheme library corresponding to the hearing aid model parameters, and obtains several candidate matching schemes. This includes: encoding the pre-acquired natural language text into query vectors, and extracting trigger word vectors and applicable scenario vectors corresponding to each initial adjustment scheme from the adjustment scheme library according to the pre-acquired hearing aid model parameters; calculating the trigger similarity between each query vector and each trigger word vector, and the scenario similarity between each query vector and each applicable scenario vector; performing weighted fusion on the trigger similarity and the scenario similarity to obtain a comprehensive matching score for each initial adjustment scheme; sorting the comprehensive matching scores according to preset sorting requirements to obtain a comprehensive matching score sorting result, and selecting initial adjustment schemes whose comprehensive matching scores meet preset screening requirements as candidate matching schemes according to the comprehensive matching score sorting result, thereby obtaining several candidate matching schemes.

[0008] In the above scheme, natural language text is encoded into query vectors. The similarity between the query vector and the trigger word vector and the applicable scenario vector is calculated separately. The similarity is then weighted and fused to obtain a comprehensive matching score. Candidate matching schemes are selected based on the score ranking. Thus, by capturing the direct trigger semantics of user speech and the semantics of the environmental scenario through dual-channel vector similarity, the weighted fusion strategy prioritizes trigger words and provides a fallback of scenario semantics. This overcomes the shortcomings of traditional keyword matching in recognizing synonymous spoken expressions, and can adapt to user colloquial and abbreviated input. Furthermore, a highly relevant candidate set is converged from the adjustment scheme library corresponding to the hearing aid model, compressing the candidate processing range of the subsequent large language model, reducing the interference of irrelevant schemes, ensuring the recall quality of candidate matching schemes under varying user expressions, and providing a reliable input basis for the subsequent large model to select target schemes.

[0009] Furthermore, a target matching scheme is selected from the candidate matching schemes according to a preset large language model, and a state marker of the user intent is generated according to the target matching scheme. This includes: calculating the execution result and boundary conclusion based on the adjustment method in the candidate matching scheme, and writing the execution result and the boundary conclusion into the candidate matching scheme to obtain an updated candidate matching scheme; concatenating the updated candidate matching scheme, the natural language text, and the historical dialogue data in the shared state stream into a decision prompt and inputting it into the preset large language model; selecting the target matching scheme from the candidate matching schemes through the preset large language model, and outputting the intent classification result; wherein, there is one and only one target matching scheme; when the intent classification result is adjustment type... When the intent and the parameters of the target matching scheme are within the safety boundary corresponding to the hearing aid model parameters, a first intent state marker is generated; when the intent classification result is a non-adjustment type intent or the parameters of the target matching scheme exceed the safety boundary, a second intent state marker is generated; the multi-node collaborative hearing aid decision service method further includes: when the intent classification result is incomplete information or ambiguous request, a third intent state marker is generated; wherein, based on the third intent state marker, a follow-up question is generated without generating structured instructions; when the intent classification result is a consultation type intent, a fourth intent state marker is generated; wherein, based on the fourth intent state marker, a response is generated through candidate matching schemes without generating structured instructions.

[0010] In the above scheme, the adjustment execution results and boundary conclusions are pre-written into the candidate matching schemes. The candidate schemes, user text, and historical dialogue input are concatenated into a large model, and a unique target matching scheme is selected. Four different intention state labels are generated by combining intent classification and parameter safety boundaries. For different labels, instructions are issued, follow-up questions are asked, responses are given, and node handover processing is performed respectively. Thus, by pre-deducing the execution boundary, the large model can perceive the risk of parameter out-of-bounds in advance. Based on the historical dialogue of the shared state flow, multi-turn contextual understanding ability is ensured. Multiple state labels are generated to distinguish adjustment execution, missing information, simple consultation, and non-adjustment requests. This can not only intercept illegal adjustment requests with out-of-bounds parameters, but also proactively ask the user follow-up questions in scenarios with incomplete information, avoiding blindly issuing invalid adjustment instructions. In this way, it can cope with ambiguous, incomplete, and multi-intent user input, and improve the reliability of hearing aid decision-making services under dynamic user needs.

[0011] Furthermore, if the state marker is a first intent state marker, a structured instruction is generated according to the target matching scheme, and the structured instruction is sent to the hearing aid to complete the hearing aid decision service. This includes: mapping the adjustment parameters in the target matching scheme to the instruction format supported by the corresponding hearing aid model to generate the structured instruction; writing the structured instruction into the instruction queue of the shared state stream so that the hearing aid can read and execute the corresponding parameter adjustment; and writing the target matching scheme and the comparison information before and after adjustment into the shared state stream to update the historical dialogue data of the shared state stream to complete the hearing aid decision service.

[0012] In the above scheme, the target matching scheme parameters are mapped to structured instructions adapted to the hearing aid model and written into a shared state stream instruction queue for the hearing aid to read and execute. Simultaneously, adjustment scheme information, adjustment time, and status flags are written into the shared state stream, updating historical dialogue data. Thus, by taking into account the differences in instruction formats among different hearing aid models, parameter mapping operations complete the conversion from upper-level decision logic to device-recognizable control instructions. The shared state stream instruction queue further decouples the business nodes and the hearing aid terminal. Furthermore, complete metadata of the current adjustment is stored in the shared state stream. Whether the adjustment continues or the node is transferred to the intelligent customer service, the current adjustment record can be read, enabling cross-node perception of historical adjustment behavior, ensuring the continuity of multi-turn dialogue business logic, and avoiding logical contradictions in subsequent interactions.

[0013] This invention provides a multi-node collaborative hearing aid decision service method, which further includes: acquiring a target matching scheme and pre- and post-adjustment comparison information written into the shared state stream after each parameter adjustment; generating a visualization card based on the target matching scheme and the pre- and post-adjustment comparison information; receiving an interaction command triggered by a user on the visualization card; wherein the interaction command includes a cancel adjustment command or a readjustment command; in response to the cancel adjustment command, restoring the audio parameters of the hearing aid to the state before adjustment and updating the shared state stream; in response to the readjustment command, receiving natural language text and the hearing aid model parameters input by the user again, and re-executing the hearing aid decision service process.

[0014] In the above scheme, a visual card is generated based on the target matching scheme and before-and-after comparison information written into the shared state stream after each parameter adjustment. It receives user-triggered commands to cancel or readjust the adjustment. Upon cancellation, the hearing aid audio parameters are restored to their pre-adjustment state and the shared state stream is updated. Upon readjustment, the user-re-entered natural language text and model parameters are re-entered into the audio adjustment node to execute the decision-making process. Thus, the adjustment details are intuitively displayed through visual cards, lowering the understanding threshold for ordinary users. The cancellation function allows for quick reverting to unsatisfactory tuning effects, and the shared state stream is updated synchronously to ensure subsequent interactions read the correct device parameters. The readjustment command reuses the entire audio adjustment node business process without rebuilding the processing chain, incorporating the user feedback loop into the entire decision-making service system. This allows users to dynamically correct their adjustment requests, further improving the user experience in dynamic usage scenarios.

[0015] Furthermore, if the state marker is a second intent state marker, the shared state in the shared state stream will be transferred to the intelligent customer service node. The intelligent customer service node will then generate a customer service response decision using a preset retrieval enhancement generation technology. This decision will be sent to the hearing aid to complete the hearing aid decision service, including: transferring historical dialogue data from the shared state stream to the intelligent customer service node; and instructing the intelligent customer service node to perform the hearing aid decision service action: obtaining the corresponding natural language text from the historical dialogue data and modifying the natural language text. Write and generate a retrieval query; based on the current query and the retrieval query, perform multi-path parallel retrieval from a pre-built knowledge base to obtain a candidate knowledge set; reorder the candidate knowledge set through a cross-encoder to obtain a reordering result, and select knowledge items that meet preset screening requirements from the reordering result as reference knowledge; submit the reference knowledge and the natural language text to the preset large language model to generate a customer service response decision constrained by the reference knowledge; write the customer service response decision into the shared state stream and send it to the hearing aid for output, ending the hearing aid decision service action.

[0016] In the above scheme, after the shared state is transferred to the intelligent customer service node, the intelligent customer service node rewrites the natural language text based on historical dialogue data to obtain the retrieval query. A multi-path parallel recall of candidate knowledge sets is performed, and a cross-encoder re-ranks and filters reference knowledge. The reference knowledge constrains the large model to generate customer service response decisions and writes them into the shared state stream for output. Therefore, considering that user questions in multi-turn dialogues often contain ellipsis, query rewriting completes the contextual semantics, multi-path parallel recall broadens the knowledge retrieval coverage, and a cross-encoder addresses the shortcomings of coarse matching in vector recall, filtering out semantically mismatched knowledge items. This allows the pre-set large language model to generate answers constrained by reference knowledge, no longer relying on internal model parameters to memorize knowledge, effectively suppressing the large language model illusion problem. All processing results are written into the shared state stream, maintaining a consistent data view with the audio adjustment node, ensuring state consistency throughout the entire dialogue interaction process.

[0017] Furthermore, in the step of performing multi-path parallel retrieval from a pre-built knowledge base based on the current query and the retrieval query to obtain a candidate knowledge set, the pre-construction process of the knowledge base includes: acquiring pre-written question-and-answer pairs and product documents, as well as historical dialogue data; de-identifying the historical dialogue data and extracting structured question-and-answer pairs through the preset large language model; performing similarity deduplication and repetition classification on the structured question-and-answer pairs, and vectorizing them together with the pre-written question-and-answer pairs and product documents and storing them in the knowledge base.

[0018] In the above solution, pre-written question-and-answer pairs, product documents, and historical dialogue data are obtained. The historical dialogue data is first anonymized, and then a large language model extracts structured question-and-answer pairs. After similarity deduplication and repetition classification, these pairs are vectorized and stored in a knowledge base along with the pre-written question-and-answer pairs and product documents. Thus, the anonymization process filters user privacy information at the source, avoiding the risk of personal data leakage. The pre-defined large language model then transforms fragmented, multi-turn customer service dialogues into standardized question-and-answer pairs, enabling the knowledge base to self-learn and expand. Similarity deduplication and classification eliminate duplicate and similar question-and-answer entries, reducing interference from repetitive knowledge in subsequent retrieval stages. Finally, the pre-written documents and dialogue-extracted questions and answers are combined to form the knowledge base, balancing the authority of official knowledge with coverage of real user scenarios, while also supporting the stable completion of the retrieval enhancement generation process by intelligent customer service nodes.

[0019] This invention also provides a multi-node collaborative hearing aid decision service system, including: a candidate matching scheme generation module, a status marker generation module, a first hearing aid decision service module, and a second hearing aid decision service module; wherein: the candidate matching scheme generation module is used to receive pre-acquired natural language text and hearing aid model parameters, and perform scheme matching on the natural language text from the adjustment scheme library corresponding to the hearing aid model parameters to obtain several candidate matching schemes; the status marker generation module is used to select a target matching scheme from the candidate matching schemes according to a preset large language model, and generate a user based on the target matching scheme. The first hearing aid decision service module is configured to generate a structured instruction based on the target matching scheme and send the structured instruction to the hearing aid if the state marker is a first intent state marker, thereby completing the hearing aid decision service; the second hearing aid decision service module is configured to transfer the shared state in the shared state stream to the intelligent customer service node if the state marker is a second intent state marker, so that the intelligent customer service node generates a customer service response decision through a preset retrieval enhancement generation technology, and sends the customer service response decision to the hearing aid, thereby completing the hearing aid decision service.

[0020] This invention proposes a multi-node collaborative hearing aid decision service system. It sets up an audio adjustment node and an intelligent customer service node, both operating within the same shared state flow. The audio adjustment node performs solution matching, while a large language model filters target solutions and outputs intent state markers. Based on different state markers, hearing aid parameter adjustments are performed, or the complete shared state is transferred to the intelligent customer service node. Enhanced retrieval is used to generate and output customer service response decisions. Thus, the shared state flow between the two nodes enables real-time communication between the dialogue context and device parameters. The large language model differentiates between adjustment and non-adjustment requests, handling them accordingly. Adjustment requests are closed-loop within the audio adjustment node, while non-adjustment requests are deterministically routed to the customer service node as a fallback. Users do not need to repeat descriptions or switch channels, and the state markers provide a clear exit point for each round of input, ensuring no request is missed. This improves the reliability of the hearing aid decision service under dynamically changing user needs.

[0021] Furthermore, the candidate matching scheme generation module includes a vector extraction unit, a similarity calculation unit, a comprehensive matching score calculation unit, and a matching scheme filtering unit; wherein: the vector extraction unit is used to encode the pre-acquired natural language text into query vectors, and extract the trigger word vectors and applicable scenario vectors corresponding to each initial adjustment scheme from the adjustment scheme library according to the pre-acquired hearing aid model parameters; the similarity calculation unit is used to calculate the trigger similarity between each query vector and each trigger word vector, and the scenario similarity between each query vector and each applicable scenario vector; the comprehensive matching score calculation unit is used to perform weighted fusion of the trigger similarity and the scenario similarity to obtain the comprehensive matching score of each initial adjustment scheme; the matching scheme filtering unit is used to sort the comprehensive matching scores according to preset sorting requirements to obtain a comprehensive matching score sorting result, and select the initial adjustment schemes whose comprehensive matching scores meet the preset filtering requirements as candidate matching schemes according to the comprehensive matching score sorting result, thereby obtaining several candidate matching schemes. Attached Figure Description

[0022] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating the steps of a multi-node collaborative hearing aid decision service method according to a certain embodiment of the present invention; Figure 2 This is a schematic diagram of the module structure of a multi-node collaborative hearing aid decision service system provided in one embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0026] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0029] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0030] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0031] To address the issue of unreliable hearing aid decision-making services in response to dynamic user needs, which stems from the lack of integrated services that address the high-frequency audio adjustment requirements of hearing aids and the large volume of customer inquiries in existing technologies, see [reference needed].Figure 1 , Figure 1 This is a flowchart illustrating the steps of a multi-node collaborative hearing aid decision service method according to a certain embodiment of the present invention. Figure 1 As shown in the figure, this embodiment of the invention proposes a multi-node collaborative hearing aid decision service method, applied to a hearing aid, wherein the hearing aid is equipped with an audio adjustment node and an intelligent customer service node; the audio adjustment node and the intelligent customer service node operate in the same shared state flow; the multi-node collaborative hearing aid decision service method is executed by the audio adjustment node; the multi-node collaborative hearing aid decision service method includes steps 101 to 104, each step being as follows: Step 101: Receive the pre-acquired natural language text and hearing aid model parameters, and perform scheme matching on the natural language text from the adjustment scheme library corresponding to the hearing aid model parameters to obtain several candidate matching schemes; Step 102: Select a target matching scheme from the candidate matching schemes according to the preset large language model, and generate a status marker of the user intent according to the target matching scheme; Step 103: If the status marker is the first intention status marker, then generate a structured instruction according to the target matching scheme, and send the structured instruction to the hearing aid to complete the hearing aid decision service; Step 104: If the state marker is the second intent state marker, the shared state in the shared state stream will be transferred to the intelligent customer service node, and the intelligent customer service node will generate a customer service response decision through a preset retrieval enhancement generation technology, so that the intelligent customer service node will send the customer service response decision to the hearing aid to complete the hearing aid decision service.

[0032] One possible implementation involves the hearing aid reporting its Media Access Control address via Bluetooth Low Energy connection, which is then automatically identified by a preset address range and model mapping table to obtain the hearing aid model parameters. User input, either voice or text, is used as natural language text, which can carry left / right ear orientation information and increment amplitude information. The audio adjustment node then receives the user's input natural language text and hearing aid model parameters, and retrieves candidate schemes from an adjustment scheme library corresponding to the hearing aid model. Each model in the adjustment scheme library corresponds to a standardized adjustment scheme. Each scheme is characterized by a scheme identifier, scheme name, applicable scenario, trigger word, supported ear side, and parameter range description. It covers various adjustable parameters, including noise reduction level and on / off, equalization of three frequencies, volume, input gain, overall output gain, tone, call enhancement, directivity, mode switching, and factory reset. These adjustments are organized according to relative increment adjustment, fixed levels, on / off, mode switching, and factory reset methods, and support independent configuration for the left and right ears. The audio adjustment node then encodes the natural language text into a query vector, calculates the cosine similarity with the pre-stored trigger word vectors and applicable scenario vectors in the scheme library, and then merges them into a comprehensive matching score according to the hierarchical weighting rule. Based on the preset screening threshold, the adjustment schemes that meet the requirements are selected as candidate matching schemes, resulting in several candidate matching schemes.

[0033] Then, the audio adjustment node concatenates candidate matching schemes, natural language text, and historical dialogue data from the shared state stream into a decision prompt, which is then input into a preset large language model. This preset large language model can be interpreted using a fusion large language model (LLM). By combining the LLM with contextual information, four tasks can be completed: scheme selection, intent determination, numerical calculation, and instruction output. During the candidate matching scheme generation stage, the results after execution are calculated according to each scheme type. Relative adjustments are calculated based on the current value plus a step size, fixed gears are calculated based on the preset target value, and on / off and mode adjustments are calculated based on the target state. Furthermore, scenarios such as exceeding limits, reaching upper or lower limits, illegal gears, and already being in the target state are predicted and written into the candidate matching scheme. The updated candidate matching schemes are obtained. After deduction and boundary calculation, the audio adjustment node selects a unique target matching scheme from several updated candidate matching schemes. Based on the target matching scheme, it distinguishes between adjustment, consultation, unrelated to adjustment, incomplete information, or multiple intentions, and assigns a status label to each type of intention. For example, when the intention classification result is an adjustment intention and the parameters of the target matching scheme are within the safety boundary corresponding to the hearing aid model parameters, a first intention status label is generated; when the intention classification result is a non-adjustment intention or the parameters of the target matching scheme exceed the safety boundary, a second intention status label is generated. It is worth mentioning that when the intention classification result is incomplete information or ambiguous request, a third intention status label is generated, and a follow-up question is generated based on the third intention status label without generating structured instructions; when the intention classification result is a consultation intention, a fourth intention status label is generated, and a response is generated based on the fourth intention status label using the candidate matching scheme without generating structured instructions. Furthermore, the first intent status marker guides the issuance of instructions, the second intent status marker guides the handover to the intelligent customer service node, and the third and fourth intent status markers guide follow-up questions and answers and consultations respectively, without issuing instructions, so that each round of user input has a clear exit.

[0034] If the status flag is the first intent status flag, the audio adjustment node maps the parameters of the target matching scheme to the instruction format supported by the hearing aid, generates structured instructions, and writes them into the instruction queue of the shared state stream. The hearing aid reads these instructions and performs the parameter adjustment. Simultaneously, a visual card can be generated to show the comparison before and after adjustment, and the user can undo or redo the adjustment. After adjustment, the target matching scheme information, adjustment time, and first intent status flag are written to the shared state stream, historical dialogue data is updated, and the service loop is completed.

[0035] If the status marker is the second intent status marker, the audio adjustment node transfers the shared state stream to the intelligent customer service node. This shared state stream includes dialogue history, device parameters, and intent markers. The intelligent customer service node then uses a preset retrieval enhancement generation technology to rewrite the user's current text into a self-contained query based on historical dialogue data. It then retrieves candidate knowledge from a pre-built knowledge base in parallel, and uses a cross-encoder to reorder and filter the top N reference knowledge points. N is a preset quantity requirement that can be adjusted according to actual application needs. Finally, the reference knowledge and the user's question are input into a large language model, enabling the intelligent customer service node to generate a knowledge-based constrained response decision and send it to the hearing aid output, completing the service loop. Notably, if the problem cannot be resolved, a human customer service transfer entry is provided, synchronizing the dialogue history to ensure service continuity.

[0036] This invention proposes a multi-node collaborative hearing aid decision-making service method. It sets up an audio adjustment node and an intelligent customer service node, both operating within the same shared state flow. The audio adjustment node performs solution matching, while a large language model filters target solutions and outputs intent state markers. Based on different state markers, hearing aid parameter adjustments are performed, or the complete shared state is transferred to the intelligent customer service node. Enhanced retrieval is used to generate and output customer service response decisions. Thus, the shared state flow between the two nodes enables real-time communication between the dialogue context and device parameters. The large language model differentiates between adjustment and non-adjustment requests for separate processing. Adjustment requests are closed-loop within the audio adjustment node, while non-adjustment requests are deterministically routed to the customer service node as a fallback. Users do not need to repeat descriptions or switch channels, and the state markers provide a clear exit point for each round of input, ensuring no request is missed. This improves the reliability of the hearing aid decision-making service under dynamically changing user needs.

[0037] One possible implementation involves receiving pre-acquired natural language text and hearing aid model parameters, performing scheme matching on the natural language text from a library of adjustment schemes corresponding to the hearing aid model parameters to obtain several candidate matching schemes, including: encoding the pre-acquired natural language text into query vectors, and extracting trigger word vectors and applicable scenario vectors corresponding to each initial adjustment scheme from the adjustment scheme library according to the pre-acquired hearing aid model parameters; calculating the trigger similarity between each query vector and each trigger word vector, and the scenario similarity between each query vector and each applicable scenario vector; performing weighted fusion on the trigger similarity and the scenario similarity to obtain a comprehensive matching score for each initial adjustment scheme; sorting the comprehensive matching scores according to preset sorting requirements to obtain a comprehensive matching score sorting result, and selecting initial adjustment schemes whose comprehensive matching scores meet preset screening requirements as candidate matching schemes according to the comprehensive matching score sorting result to obtain several candidate matching schemes.

[0038] For example, the audio adjustment node receives pre-acquired natural language text and hearing aid model parameters. The natural language text can be obtained through user speech transcription, such as "I can't hear clearly with my left ear in the market" in a certain scenario. Then, based on the hearing aid model parameters, the corresponding adjustment scheme library is loaded. The adjustment scheme library pre-stores multiple initial adjustment schemes for that model of hearing aid, and each scheme contains two types of semantic materials: trigger words and applicable scenarios.

[0039] Then, the audio adjustment node uses a text embedding model such as OpenAI's text-embedding-3-large to encode the user's natural language text into a high-dimensional query vector. Simultaneously, it extracts trigger word vectors and applicable scenario vectors corresponding to all initial adjustment schemes from the adjustment scheme library; these vectors have been pre-calculated and loaded into memory. Then, it calculates the cosine similarity between the query vector and the trigger word vector and applicable scenario vector of each initial adjustment scheme in parallel, obtaining the trigger similarity and scenario similarity respectively. The trigger similarity is taken as the highest similarity value between all trigger word vectors and the query vector for that scheme. In this embodiment, the calculation process of cosine similarity can be implemented using existing technology and will not be elaborated here.

[0040] The audio adjustment node then calculates the comprehensive matching score for each initial adjustment scheme based on the trigger similarity level using a hierarchical weighted fusion strategy. In this embodiment, the specific rules are as follows: if the trigger similarity is not lower than 0.85, it is considered a strong hit, and the comprehensive matching score is directly taken from the trigger similarity; if the trigger similarity is greater than or equal to 0.70 and less than 0.85, it is considered a weak hit, and the comprehensive matching score is calculated using a weighted average of trigger similarity (0.75) and scene similarity (0.25); if the trigger similarity is lower than 0.70, it is considered a miss, and the score is calculated using an equal weighted average of trigger similarity and scene similarity (0.50 each).

[0041] Finally, the audio adjustment node filters out initial adjustment schemes with a comprehensive matching score below 0.5, and sorts the remaining schemes according to their comprehensive matching scores from highest to lowest. This ranking from highest to lowest represents the preset ranking requirements. To take into account the context of multi-turn dialogues, a three-way parallel recall is performed, combining the user's original input, the self-contained query rewritten by the large language model, and the user's original speech from the previous round. A minimum-seat strategy is used to merge the results, and finally, the top N initial adjustment schemes are selected as candidate matching schemes. Each candidate matching scheme also includes a pre-calculated post-execution rehearsal result. The selection of the top N initial adjustment schemes represents the initial adjustment schemes that meet the preset screening requirements.

[0042] In the above scheme, natural language text is encoded into query vectors. The similarity between the query vector and the trigger word vector and the applicable scenario vector is calculated separately. The similarity is then weighted and fused to obtain a comprehensive matching score. Candidate matching schemes are selected based on the score ranking. Thus, by capturing the direct trigger semantics of user speech and the semantics of the environmental scenario through dual-channel vector similarity, the weighted fusion strategy prioritizes trigger words and provides a fallback of scenario semantics. This overcomes the shortcomings of traditional keyword matching in recognizing synonymous spoken expressions, and can adapt to user colloquial and abbreviated input. Furthermore, a highly relevant candidate set is converged from the adjustment scheme library corresponding to the hearing aid model, compressing the candidate processing range of the subsequent large language model, reducing the interference of irrelevant schemes, ensuring the recall quality of candidate matching schemes under varying user expressions, and providing a reliable input basis for the subsequent large model to select target schemes.

[0043] One possible implementation involves selecting a target matching scheme from candidate matching schemes based on a preset large language model, and generating a state marker for the user intent based on the target matching scheme. This includes: calculating the execution result and boundary conclusion based on the adjustment method in the candidate matching schemes, and writing the execution result and boundary conclusion into the candidate matching schemes to obtain an updated candidate matching scheme; concatenating the updated candidate matching scheme, the natural language text, and historical dialogue data in the shared state stream into a decision prompt and inputting it into the preset large language model; selecting a target matching scheme from the candidate matching schemes using the preset large language model, and outputting an intent classification result; wherein there is one and only one target matching scheme; when the intent classification result is adjustment... When the intent classification result is a non-adjustment type intent and the parameters of the target matching scheme are within the safety boundary corresponding to the hearing aid model parameters, a first intent state marker is generated; when the intent classification result is a non-adjustment type intent or the parameters of the target matching scheme exceed the safety boundary, a second intent state marker is generated; the multi-node collaborative hearing aid decision service method further includes: when the intent classification result is incomplete information or ambiguous request, a third intent state marker is generated; wherein, based on the third intent state marker, a follow-up question is generated without generating a structured instruction; when the intent classification result is a consultation type intent, a fourth intent state marker is generated; wherein, based on the fourth intent state marker, a response is generated through a candidate matching scheme without generating a structured instruction.

[0044] For example, the audio adjustment node, based on the adjustment method of each candidate scheme (such as relative incremental adjustment and fixed-level adjustment), and combined with the current real-time parameters of the hearing aid, pre-calculates the execution result if the scheme is implemented. For example, the noise reduction level changes from level 3 to level 4, as well as boundary conclusions, such as whether the parameters have reached their upper limit or whether the target state has been reached. Subsequently, the calculated execution result and boundary conclusions are written into the corresponding candidate matching scheme as supplementary information, forming a more complete and updated candidate matching scheme.

[0045] The audio adjustment node combines the updated candidate matching scheme, the natural language text input by the user in this round, and the historical dialogue data stored in the shared state stream to construct a structured decision prompt, and inputs it into the preset large language model. The preset large language model can be interpreted using a fusion large language model (LLM), and the historical dialogue data includes previous dialogue records and device status, etc.

[0046] After receiving decision prompts, the pre-defined large language model performs a series of decision-making tasks, including accurately selecting a unique target matching scheme from candidate matching schemes and identifying specific parameters specified by the user, such as not adjusting "overall gain" if the user explicitly says "volume"; then, it performs intent determination, distinguishing whether the user's intent is to perform adjustment, seek information, access incomplete information, or is unrelated to adjustment; for schemes that require the user to provide a numerical value, the pre-defined large language model calculates the absolute target value based on the context; finally, the pre-defined large language model generates natural language scripts for the user and outputs adjustment instructions in the form of structured instructions.

[0047] The audio adjustment node generates corresponding user intent status markers based on the target matching scheme output by the preset large language model and the user intent, specifically in the following cases: When the intent classification result is a clear adjustment intent, and the selected target matching scheme has no constraints, the required values ​​are complete, and its parameters are within the safety boundary corresponding to the hearing aid model, a first intent state label is generated.

[0048] When the intent classification result is a non-regulatory intent such as general consultation, or the parameters of the target matching scheme exceed the safety boundary, or it is unrelated to regulation and has no suitable scheme, a second intent status label is generated.

[0049] When the intent classification result indicates that the user has not provided the necessary numerical information, or the request is vague and cannot be matched with a unique solution, a third intent state marker is generated. At this time, follow-up questions will be asked to the user based on the pre-defined large language model, but no structured instructions will be issued.

[0050] When the intent classification result is a consultation intent such as inquiring about the differences between solutions and the range of values, a fourth intent status marker is generated. At this time, a response script will be directly generated based on the information in the candidate solutions, and no structured instructions will be issued.

[0051] In the above scheme, the adjustment execution results and boundary conclusions are pre-written into the candidate matching schemes. The candidate schemes, user text, and historical dialogue input are concatenated into a large model, and a unique target matching scheme is selected. Four different intention state labels are generated by combining intent classification and parameter safety boundaries. For different labels, instructions are issued, follow-up questions are asked, responses are given, and node handover processing is performed respectively. Thus, by pre-deducing the execution boundary, the large model can perceive the risk of parameter out-of-bounds in advance. Based on the historical dialogue of the shared state flow, multi-turn contextual understanding ability is ensured. Multiple state labels are generated to distinguish adjustment execution, missing information, simple consultation, and non-adjustment requests. This can not only intercept illegal adjustment requests with out-of-bounds parameters, but also proactively ask the user follow-up questions in scenarios with incomplete information, avoiding blindly issuing invalid adjustment instructions. In this way, it can cope with ambiguous, incomplete, and multi-intent user input, and improve the reliability of hearing aid decision-making services under dynamic user needs.

[0052] One possible implementation involves, if the state marker is a first intent state marker, generating a structured instruction based on the target matching scheme and sending the structured instruction to the hearing aid to complete the hearing aid decision service. This includes: mapping the adjustment parameters in the target matching scheme to an instruction format supported by the corresponding hearing aid model to generate the structured instruction; writing the structured instruction into the instruction queue of the shared state stream so that the hearing aid can read and execute the corresponding parameter adjustment; and writing the target matching scheme and before-and-after adjustment comparison information into the shared state stream to update the historical dialogue data of the shared state stream, thus completing the hearing aid decision service.

[0053] For example, when the determination state is marked as the first intent state, it indicates that the user's intent is clear and the adjustment scheme is safe and feasible. The audio adjustment node will execute the corresponding hearing aid decision service. Specifically, it will map the adjustment parameters in the target matching scheme selected by the large language model to the low-level instruction format supported by the specific hearing aid model. The adjustment parameters include noise reduction level and volume value, etc. This mapping process transforms the high-level adjustment intent into structured instructions that the hearing aid hardware can recognize and execute, conforming to its private communication protocol.

[0054] The audio adjustment node then writes the generated structured instructions into the instruction queue in the shared state stream. The hearing aid device continuously listens to and reads this instruction queue through the established Bluetooth connection. Once a new instruction is read, the hearing aid immediately executes the corresponding parameter adjustment operation, thereby realizing the personalized settings required by the user.

[0055] Finally, to update the context and support subsequent interactions, the target matching scheme corresponding to each parameter adjustment and the comparison information before and after the adjustment are written into the shared state stream to update the historical dialogue data. This completes a closed loop of hearing aid decision-making services, from intent recognition to command execution.

[0056] In the above scheme, the target matching scheme parameters are mapped to structured instructions adapted to the hearing aid model and written into a shared state stream instruction queue for the hearing aid to read and execute. Simultaneously, adjustment scheme information, adjustment time, and status flags are written into the shared state stream to update historical dialogue data. Thus, by taking into account the differences in instruction formats among different hearing aid models, parameter mapping operations complete the conversion from upper-level decision logic to device-recognizable control instructions. The shared state stream instruction queue further decouples the business nodes and the hearing aid terminal. Furthermore, complete metadata of the current adjustment is stored in the shared state stream. Whether the adjustment continues or the node is transferred to the intelligent customer service, the current adjustment record can be read, enabling cross-node perception of historical adjustment behavior, ensuring the continuity of multi-turn dialogue business logic, and avoiding logical contradictions in subsequent interactions.

[0057] This invention provides a multi-node collaborative hearing aid decision service method, which further includes: acquiring a target matching scheme and pre- and post-adjustment comparison information written into the shared state stream after each parameter adjustment; generating a visualization card based on the target matching scheme and the pre- and post-adjustment comparison information; receiving an interaction command triggered by a user on the visualization card; wherein the interaction command includes a cancel adjustment command or a readjustment command; in response to the cancel adjustment command, restoring the audio parameters of the hearing aid to the state before adjustment and updating the shared state stream; in response to the readjustment command, receiving natural language text and the hearing aid model parameters input by the user again, and re-executing the hearing aid decision service process.

[0058] For example, after each adjustment, a visualization card is automatically generated based on the specific parameters in the target matching scheme and the comparison information before and after the adjustment. This visualization card presents the adjustment content in an intuitive graphical way, such as displaying the changes in each parameter through bar charts or numerical comparisons, and includes the name and brief description of the scheme corresponding to this adjustment, such as "Market Noise Reduction Enhancement Mode." This visualization card is displayed to the user through the mobile application or device interface accompanying the hearing aid, allowing the user to clearly understand the specific content and effect of this adjustment.

[0059] Then, interactive controls are provided on the visual card to receive user-triggered interactive commands. These commands mainly include two types: undo adjustment commands and readjust adjustment commands. The undo adjustment command allows users to restore the previous state with one click if they are dissatisfied with the adjustment result output by the hearing aid decision service. The readjust adjustment command allows users to make new adjustment requests based on the current adjustment without restarting the entire dialogue process.

[0060] When a user triggers a cancellation adjustment command, the audio adjustment node responds by reading a snapshot of the original parameters saved before the adjustment from the shared state stream. This snapshot was automatically written to the shared state stream when the first intent state flag was generated. The node then restores the hearing aid's audio parameters to their pre-adjustment state. After restoration, the shared state stream is updated synchronously, writing the timestamp of the cancellation operation, the cancellation reason flag, and the restored parameter state into the historical dialogue data. This ensures that subsequent multi-turn dialogues can correctly perceive changes in user preferences and avoid repeatedly recommending canceled solutions.

[0061] When the user triggers a readjustment command, the audio adjustment node responds to the readjustment command by re-entering the natural language text input by the user, such as "lower the volume a little more" or "the right ear also needs adjustment," along with the hearing aid model parameters, into the audio adjustment node, and re-executes the complete hearing aid decision service process from scheme matching and intent recognition to command issuance.

[0062] In the above scheme, a visual card is generated based on the target matching scheme and before-and-after comparison information written into the shared state stream after each parameter adjustment. It receives user-triggered commands to cancel or readjust the adjustment. Upon cancellation, the hearing aid audio parameters are restored to their pre-adjustment state and the shared state stream is updated. Upon readjustment, the user-re-entered natural language text and model parameters are input into the audio adjustment node to execute the decision-making process. Thus, the adjustment details are intuitively displayed through visual cards, lowering the understanding threshold for ordinary users. The cancellation function allows for quick reverting to unsatisfactory tuning effects, and the shared state stream is updated synchronously to ensure subsequent interactions read the correct device parameters. The readjustment command reuses the entire audio adjustment node business process without rebuilding the processing chain, incorporating the user feedback loop into the entire decision-making service system. This allows users to dynamically correct their adjustment requests, further improving the user experience in dynamic usage scenarios.

[0063] One possible implementation involves, if the state marker is a second intent state marker, transferring the shared state in the shared state stream to the intelligent customer service node. The intelligent customer service node then generates a customer service response decision using a preset retrieval enhancement generation technology. This decision is then sent to the hearing aid to complete the hearing aid decision service. The implementation includes: transferring historical dialogue data from the shared state stream to the intelligent customer service node; and having the intelligent customer service node perform the hearing aid decision service action: obtaining the corresponding natural language text from the historical dialogue data and processing the natural language text. The query is rewritten to generate a search query; based on the current query and the search query, a multi-path parallel retrieval is performed from a pre-built knowledge base to obtain a candidate knowledge set; the candidate knowledge set is reordered using a cross-encoder to obtain a reordering result, and knowledge items that meet preset screening requirements are selected from the reordering result as reference knowledge; the reference knowledge and the natural language text are submitted to the preset large language model to generate a customer service response decision constrained by the reference knowledge; the customer service response decision is written into the shared state stream and sent to the hearing aid for output, ending the hearing aid decision service action.

[0064] For example, when the determination state flag is the second intent state flag, it indicates that the user intent is non-adjustment type or the adjustment request exceeds the security boundary. At this time, the audio adjustment node will seamlessly transfer all context information in the shared state stream to the intelligent customer service node.

[0065] In the intelligent customer service node, through the Retrieval Enhanced Generation (RAG) technology, combined with historical dialogue data in the shared state stream, the user's current natural language text, which may contain references or omissions, such as "Why can't it connect?", is rewritten into a complete and independently understandable retrieval query, such as "What should I do if I can't connect to ELEHEAR Beyond Pro pairing?" Then, based on the rewritten search query, a multi-path parallel retrieval is performed from a pre-built knowledge base to obtain an initial candidate knowledge set. The pre-built knowledge base includes manually pre-written question-and-answer pairs, product documents, and self-learning question-and-answer pairs. Subsequently, a cross-encoder is used to refine and re-rank the candidate knowledge set, resulting in a re-ranked list. The cross-encoder enables deep interaction between the query and the document, thereby more accurately evaluating relevance. Finally, knowledge items that meet preset screening requirements are selected from the re-ranked list as the final reference knowledge. In this embodiment, the top N knowledge items are selected as those that meet the preset screening requirements.

[0066] The selected reference knowledge, along with the user's original natural language text, is then submitted to a pre-defined large language model. This model is strictly constrained, capable only of generating customer service response decisions based on the provided reference knowledge. For example, questions concerning the company's products, functions, services, or user manuals must be answered strictly according to the information in the external knowledge base; external knowledge or fabricated content is prohibited. If no corresponding basis exists in the knowledge base, the response should be truthfully stated rather than speculated. This ensures the accuracy and traceability of the answers.

[0067] Finally, the generated customer service response decision is written to the shared state stream and sent to the hearing aid for voice or text output, completing the decision-making service. If the problem cannot be resolved, a human customer service transfer point can be provided, synchronizing all conversation history to ensure a closed-loop service.

[0068] In the above scheme, after the shared state is transferred to the intelligent customer service node, the intelligent customer service node rewrites the natural language text based on historical dialogue data to obtain the retrieval query. A multi-path parallel recall of candidate knowledge sets is performed, and a cross-encoder re-ranks and filters reference knowledge. The reference knowledge constrains the large model to generate customer service response decisions and writes them into the shared state stream for output. Therefore, considering that user questions in multi-turn dialogues often contain ellipsis, query rewriting completes the contextual semantics, multi-path parallel recall broadens the knowledge retrieval coverage, and a cross-encoder addresses the shortcomings of coarse matching in vector recall, filtering out semantically mismatched knowledge items. This allows the pre-set large language model to generate answers constrained by reference knowledge, no longer relying on internal model parameters to memorize knowledge, effectively suppressing the large language model illusion problem. All processing results are written into the shared state stream, maintaining a consistent data view with the audio adjustment node, ensuring state consistency throughout the entire dialogue interaction process.

[0069] One possible implementation involves a step of performing multi-path parallel retrieval from a pre-built knowledge base based on the current query and the retrieval query to obtain a candidate knowledge set. The pre-building process of the knowledge base includes: acquiring pre-written question-and-answer pairs and product documents, as well as historical dialogue data; de-identifying the historical dialogue data and extracting structured question-and-answer pairs using the preset large language model; performing similarity deduplication and repetition classification on the structured question-and-answer pairs, and then vectorizing them together with the pre-written question-and-answer pairs and product documents before storing them in the knowledge base.

[0070] For example, two types of raw data are acquired as the basis for building the knowledge base. The raw data includes high-confidence question-and-answer pairs (FAQs) and product documents pre-written by professionals, as well as historical dialogue data accumulated from real service scenarios, such as the history of customer service ticket platforms and the text transcribed from call recordings.

[0071] Then, the historical dialogue data undergoes text normalization and rule-based anonymization. Deterministic rules such as regular expression matching are used to replace personally identifiable information (PII) such as phone numbers, email addresses, and order numbers in the text with uniform placeholders, ensuring data security from the source. Subsequently, the anonymized dialogue text is input into a pre-defined large language model. Through structured output constraints such as JSON Schema, the pre-defined large language model automatically understands the multi-turn dialogue content and extracts structured question-answer pairs. The structure of a structured question-answer pair typically includes a general question, a fact-based objective answer, and keywords for retrieval. Specifically, the structured output constraints ensure that the pre-defined large language model can only produce a valid structure of {question, answer, keyword} triples. The question is a general interrogative sentence with third-person paraphrasing and personal information removed; the answer is an objective statement based on the facts of the dialogue; and the keywords are 3-5 words used for retrieval.

[0072] Then, a semantic vector-based similarity deduplication algorithm is used to calculate the similarity between the newly extracted structured question-answer pairs and the question parts of existing questions and answers in the library. Based on the preset repetition threshold, the pairs are graded and processed accordingly. Specifically, pairs with extremely high repetition are automatically removed, pairs that are suspected to be similar are marked for manual review and confirmation, and pairs that are completely new are retained.

[0073] Finally, the structured question-and-answer pairs that have passed quality verification, the pre-written question-and-answer pairs, and the parsed product document content are processed into high-dimensional vectors using a unified text embedding model to generate semantic vectors. These semantic vectors are then stored in a vector database, and an Approximate Nearest Neighbor (ANN) index is built, thus completing a pre-built knowledge base that integrates structured and unstructured data and supports efficient semantic retrieval.

[0074] In the above solution, pre-written question-and-answer pairs, product documents, and historical dialogue data are obtained. The historical dialogue data is first anonymized, and then a large language model extracts structured question-and-answer pairs. After similarity deduplication and repetition classification, these pairs are vectorized and stored in a knowledge base along with the pre-written question-and-answer pairs and product documents. Thus, the anonymization process filters user privacy information at the source, avoiding the risk of personal data leakage. The pre-defined large language model then transforms fragmented, multi-turn customer service dialogues into standardized question-and-answer pairs, enabling the knowledge base to self-learn and expand. Similarity deduplication and classification eliminate duplicate and similar question-and-answer entries, reducing interference from repetitive knowledge in subsequent retrieval stages. Finally, the pre-written documents and dialogue-extracted questions and answers are combined to form the knowledge base, balancing the authority of official knowledge with coverage of real user scenarios, while also supporting the stable completion of the retrieval enhancement generation process by intelligent customer service nodes.

[0075] Based on the above method embodiments, corresponding apparatus embodiments are provided; see [link to apparatus embodiments]. Figure 2 , Figure 2This is a schematic diagram of the module structure of a multi-node collaborative hearing aid decision service system provided in one embodiment of the present invention. Figure 2 As shown, this embodiment of the invention also provides a multi-node collaborative hearing aid decision service system, including: a candidate matching scheme generation module 201, a state marker generation module 202, a first hearing aid decision service module 203, and a second hearing aid decision service module 204; wherein: the candidate matching scheme generation module 201 is used to receive pre-acquired natural language text and hearing aid model parameters, and perform scheme matching on the natural language text from the adjustment scheme library corresponding to the hearing aid model parameters to obtain several candidate matching schemes; the state marker generation module 202 is used to select a target matching scheme from the candidate matching schemes according to a preset large language model, and according to the target... The matching scheme generates a status marker for the user's intent; the first hearing aid decision service module 203 is used to generate a structured instruction according to the target matching scheme if the status marker is a first intent status marker, and send the structured instruction to the hearing aid to complete the hearing aid decision service; the second hearing aid decision service module 204 is used to transfer the shared state in the shared state stream to the intelligent customer service node if the status marker is a second intent status marker, so that the intelligent customer service node generates a customer service response decision through a preset retrieval enhancement generation technology, so that the intelligent customer service node sends the customer service response decision to the hearing aid to complete the hearing aid decision service.

[0076] This invention proposes a multi-node collaborative hearing aid decision service system. It sets up an audio adjustment node and an intelligent customer service node, both operating within the same shared state flow. The audio adjustment node performs solution matching, while a large language model filters target solutions and outputs intent state markers. Based on different state markers, hearing aid parameter adjustments are performed, or the complete shared state is transferred to the intelligent customer service node. Enhanced retrieval is used to generate and output customer service response decisions. Thus, the shared state flow between the two nodes enables real-time communication between the dialogue context and device parameters. The large language model differentiates between adjustment and non-adjustment requests, handling them accordingly. Adjustment requests are closed-loop within the audio adjustment node, while non-adjustment requests are deterministically routed to the customer service node as a fallback. Users do not need to repeat descriptions or switch channels, and the state markers provide a clear exit point for each round of input, ensuring no request is missed. This improves the reliability of the hearing aid decision service under dynamically changing user needs.

[0077] Furthermore, the candidate matching scheme generation module 201 includes a vector extraction unit 301, a similarity calculation unit 302, a comprehensive matching score calculation unit 303, and a matching scheme filtering unit 304; wherein: the vector extraction unit 301 is used to encode the pre-acquired natural language text into query vectors, and extract the trigger word vectors and applicable scenario vectors corresponding to each initial adjustment scheme from the adjustment scheme library according to the pre-acquired hearing aid model parameters; the similarity calculation unit 302 is used to calculate the trigger similarity between each query vector and each trigger word vector, and the scenario similarity between each query vector and each applicable scenario vector; the comprehensive matching score calculation unit 303 is used to perform weighted fusion of the trigger similarity and the scenario similarity to obtain the comprehensive matching score of each initial adjustment scheme; the matching scheme filtering unit 304 is used to sort the comprehensive matching scores according to preset sorting requirements to obtain a comprehensive matching score sorting result, and select the initial adjustment schemes whose comprehensive matching scores meet the preset filtering requirements as candidate matching schemes according to the comprehensive matching score sorting result to obtain several candidate matching schemes.

[0078] It is understood that the above-described device embodiments correspond to the method embodiments of the present invention, and can implement a multi-node collaborative hearing aid decision service method provided by any of the above-described method embodiments of the present invention.

[0079] It should be noted that the system embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0080] Based on the above-described embodiment of a multi-node collaborative hearing aid decision service method, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a multi-node collaborative hearing aid decision service method according to any embodiment of the present invention.

[0081] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0082] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0083] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0084] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute a multi-node collaborative hearing aid decision service method as described in any of the above-described method embodiments of the present invention.

[0085] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0086] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A multi-node collaborative hearing aid decision service method, characterized in that, This is applied to hearing aids, which include an audio adjustment node and an intelligent customer service node; the audio adjustment node and the intelligent customer service node operate in the same shared state stream. The multi-node collaborative hearing aid decision service method is executed by the audio adjustment node; The multi-node collaborative hearing aid decision service method includes: Receive pre-acquired natural language text and hearing aid model parameters, perform scheme matching on the natural language text from the adjustment scheme library corresponding to the hearing aid model parameters, and obtain several candidate matching schemes; The target matching scheme is selected from the candidate matching schemes according to the preset large language model, and a status marker of the user intent is generated according to the target matching scheme. If the status marker is the first intention status marker, then a structured instruction is generated according to the target matching scheme, and the structured instruction is sent to the hearing aid to complete the hearing aid decision service; If the state marker is the second intent state marker, the shared state in the shared state stream will be transferred to the intelligent customer service node, which will then generate a customer service response decision using a preset retrieval enhancement generation technology. This will enable the intelligent customer service node to send the customer service response decision to the hearing aid, thus completing the hearing aid decision service.

2. The hearing aid decision service method with multi-node collaboration as described in claim 1, characterized in that, The system receives pre-acquired natural language text and hearing aid model parameters, performs scheme matching on the natural language text from the adjustment scheme library corresponding to the hearing aid model parameters, and obtains several candidate matching schemes, including: The pre-acquired natural language text is encoded into a query vector, and the trigger word vector and applicable scenario vector corresponding to each initial adjustment scheme are extracted from the adjustment scheme library according to the pre-acquired hearing aid model parameters. Calculate the trigger similarity between each query vector and each trigger word vector, and the scenario similarity between each query vector and each applicable scenario vector. The trigger similarity and the scene similarity are weighted and fused to obtain the comprehensive matching score of each initial adjustment scheme; The comprehensive matching scores are sorted according to preset sorting requirements to obtain comprehensive matching score sorting results. Based on the comprehensive matching score sorting results, initial adjustment schemes that meet preset screening requirements are selected as candidate matching schemes to obtain several candidate matching schemes.

3. The hearing aid decision service method with multi-node collaboration as described in claim 1, characterized in that, Based on a preset large language model, a target matching scheme is selected from the candidate matching schemes, and a status marker of the user intent is generated based on the target matching scheme, including: The execution result and boundary conclusion are calculated based on the adjustment method in the candidate matching scheme, and the execution result and the boundary conclusion are written into the candidate matching scheme to obtain the updated candidate matching scheme. The updated candidate matching scheme, the natural language text, and the historical dialogue data in the shared state stream are concatenated to form a decision prompt input into the preset large language model; The target matching scheme is selected from the candidate matching schemes by the preset large language model, and the intent classification result is output; wherein, there is one and only one target matching scheme; When the intent classification result is an adjustment intent and the parameters of the target matching scheme are within the safety boundary corresponding to the hearing aid model parameters, the first intent state label is generated; When the intent classification result is a non-adjustment intent or the parameters of the target matching scheme exceed the safety boundary, a second intent state label is generated.

4. The multi-node collaborative hearing aid decision service method as described in claim 3, characterized in that, Also includes: When the intent classification result is incomplete information or vague request, a third intent status marker is generated; wherein, based on the third intent status marker, follow-up questions are generated but no structured instructions are generated. When the intent classification result is a consultation intent, a fourth intent status marker is generated; wherein, based on the fourth intent status marker, a response script is generated through a candidate matching scheme without generating a structured instruction.

5. The hearing aid decision service method with multi-node collaboration as described in claim 1, characterized in that, If the status marker is a first intent status marker, then a structured instruction is generated according to the target matching scheme, and the structured instruction is sent to the hearing aid to complete the hearing aid decision service, including: The adjustment parameters in the target matching scheme are mapped to the instruction format supported by the corresponding hearing aid model to generate the structured instruction; The structured instructions are written into the instruction queue of the shared state stream so that the hearing aid can read and execute the corresponding parameter adjustments; The target matching scheme and the comparison information before and after adjustment are written into the shared state stream to update the historical dialogue data of the shared state stream and complete the hearing aid decision service.

6. The multi-node collaborative hearing aid decision service method as described in claim 5, characterized in that, Also includes: Obtain the target matching scheme and the comparison information before and after adjustment written into the shared state stream after each parameter adjustment is completed, and generate a visualization card based on the target matching scheme and the comparison information before and after adjustment; Receive user interaction commands triggered by the visualization card; wherein the interaction commands include undo adjustment commands or readjustment commands; In response to the undo adjustment command, the audio parameters of the hearing aid are restored to their pre-adjustment state and the shared state stream is updated; In response to the readjustment instruction, the system receives the user's re-entered natural language text and the hearing aid model parameters, and re-executes the hearing aid decision service process.

7. The multi-node collaborative hearing aid decision service method as described in claim 5, characterized in that, If the state marker is the second intent state marker, the shared state in the shared state stream will be transferred to the intelligent customer service node. The intelligent customer service node will then generate a customer service response decision using a preset retrieval enhancement generation technology. This decision will be sent to the hearing aid to complete the hearing aid decision service, including: The historical dialogue data in the shared state stream is transferred to the intelligent customer service node; The intelligent customer service node is instructed to perform hearing aid decision service actions: obtain the corresponding natural language text from the historical dialogue data, rewrite the natural language text, and generate a search query; Based on the current query and the retrieval query, a candidate knowledge set is obtained by performing multi-path parallel recall from the pre-built knowledge base; The candidate knowledge set is reordered by a cross encoder to obtain a reordering result, and knowledge items that meet the preset screening requirements are selected from the reordering result as reference knowledge. The reference knowledge and the natural language text are submitted to the preset large language model to generate customer service response decisions constrained by the reference knowledge; The customer service response decision is written into the shared state stream and sent to the hearing aid for output, thus ending the hearing aid decision service action.

8. The multi-node collaborative hearing aid decision service method as described in claim 7, characterized in that, In the step of performing multi-path parallel retrieval from a pre-built knowledge base based on the current query and the retrieval query to obtain a candidate knowledge set, the pre-building process of the knowledge base includes: Obtain pre-written question-and-answer pairs and product documentation, as well as historical dialogue data; The historical dialogue data is anonymized, and structured question-answer pairs are extracted using the preset large language model; The structured question-and-answer pairs are deduplicated based on similarity and classified for redundancy, and then vectorized together with the pre-written question-and-answer pairs and product documents and stored in the knowledge base.

9. A multi-node collaborative hearing aid decision service system, characterized in that, This is applied to hearing aids, which include an audio adjustment node and an intelligent customer service node; the audio adjustment node and the intelligent customer service node operate in the same shared state stream. The multi-node collaborative hearing aid decision service system is controlled by the audio adjustment node; The multi-node collaborative hearing aid decision service system includes a candidate matching scheme generation module, a status marker generation module, a first hearing aid decision service module, and a second hearing aid decision service module; wherein: The candidate matching scheme generation module is used to receive pre-acquired natural language text and hearing aid model parameters, and perform scheme matching on the natural language text from the adjustment scheme library corresponding to the hearing aid model parameters to obtain several candidate matching schemes. The status marker generation module is used to select a target matching scheme from the candidate matching schemes according to a preset large language model, and generate a status marker of the user intent according to the target matching scheme; The first hearing aid decision service module is used to generate a structured instruction according to the target matching scheme if the status mark is a first intention status mark, and send the structured instruction to the hearing aid to complete the hearing aid decision service; The second hearing aid decision service module is used to transfer the shared state in the shared state stream to the intelligent customer service node if the state marker is the second intention state marker. The intelligent customer service node then generates a customer service response decision through a preset retrieval enhancement generation technology, so that the intelligent customer service node sends the customer service response decision to the hearing aid to complete the hearing aid decision service.

10. The multi-node collaborative hearing aid decision service system as described in claim 9, characterized in that, The candidate matching scheme generation module includes a vector extraction unit, a similarity calculation unit, a comprehensive matching score calculation unit, and a matching scheme filtering unit; wherein: The vector extraction unit is used to encode the pre-acquired natural language text into a query vector, and extract the trigger word vector and applicable scenario vector corresponding to each initial adjustment scheme from the adjustment scheme library according to the pre-acquired hearing aid model parameters; The similarity calculation unit is used to calculate the trigger similarity between each query vector and each trigger word vector, and the scene similarity between each query vector and each applicable scene vector. The comprehensive matching score calculation unit is used to perform weighted fusion of the trigger similarity and the scene similarity to obtain the comprehensive matching score of each initial adjustment scheme; The matching scheme filtering unit is used to sort the comprehensive matching scores based on preset sorting requirements, obtain the comprehensive matching score sorting result, and select the initial adjustment scheme that meets the preset screening requirements as the candidate matching scheme according to the comprehensive matching score sorting result, thereby obtaining several candidate matching schemes.

Images