Voice dialogue system based on large model

Abstract

The invention discloses a voice dialogue system based on a large model, and relates to the technical field of voice interaction, and the system comprises an intelligent scheduling module which determines a task type according to a to-be-recognized text, and dynamically adapts to a fast and slow thinking module according to the task type; the fast and slow thinking module dynamically adapts to the calculation unit according to the task type and outputs a feedback result; the voice interaction module generates a to-be-recognized text according to the voice characteristics of the user and generates a voice instruction based on a feedback result; the context management module stores the interaction log to a corresponding storage unit according to the index key value; the interaction log comprises a task type belonging to the same data flow, an adaptive path and a feedback result; and the data centralized control module determines a parameter adjustment strategy according to the performance index factors in the interaction logs, and the parameter adjustment strategy acts on the intelligent scheduling module or / and the fast and slow thinking module to optimize corresponding parameters. The intelligent balance of quick response and deep reasoning of the dialogue system under a single architecture is realized, and the real-time performance and accuracy of interaction are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of voice interaction technology, and in particular to a large model-based voice dialogue system. Background Art

[0002] With the rise of large-scale pre-trained language models (such as BERT and GPT), natural language processing (NLP) capabilities have significantly improved. While large models can handle complex language tasks, their high computational requirements and inference latency limit their application in real-time interactions. As the integrated embodiment of AI and NLP, voice dialogue systems face a difficult trade-off between speed and quality. First, dialogue systems based on lightweight models offer fast responses for simple tasks. However, due to their fixed model structure and limited computational resources, their understanding and response quality are lower for dialogue tasks with ambiguous or multi-intent input. Second, cloud-based systems based on large models, while capable of providing high-quality responses, suffer from long inference latency, making them particularly inefficient for simple tasks. This degrades user experience due to extended wait times. Furthermore, while hybrid architecture systems attempt to balance these two requirements, their simple task scheduling mechanisms (based on fixed rules rather than intelligent judgment) can lead to inconsistent response styles and fragmented contextual understanding when switching between modules. This hinders the full potential of hybrid architecture systems in balancing real-time performance and accuracy. Summary of the Invention

[0003] The purpose of the present invention is to address the technical problem that existing dialogue systems are difficult to achieve real-time and accurate interactions. A speech dialogue system based on a large model is proposed. Through the dynamic identification of task types by the intelligent scheduling module and the adaptive adaptation of the fast and slow thinking module, combined with the structured splicing of interaction logs by the context management module and the parameter collaborative optimization mechanism of the data set control module, a "fast response-deep reasoning-dynamic optimization" language dialogue closed-loop processing system is constructed; it effectively overcomes the defects of poor response quality of simple tasks, high latency of complex tasks and rigid scheduling of hybrid architectures in the existing system, and realizes the intelligent balance between fast response and deep reasoning of the dialogue system under a single architecture, significantly improving the real-time nature, fluency and accuracy of semantic understanding of interactions.

[0004] A technical solution provided in an embodiment of the present invention is: a voice dialogue system based on a large model, comprising: an intelligent scheduling module, a fast and slow thinking module, a voice interaction module, a context management module, and a data set control module; The intelligent scheduling module determines the task type based on the text to be recognized and dynamically adapts the fast and slow thinking modules according to the task type; The fast and slow thinking module dynamically adapts the computing unit according to the task type and outputs feedback results; The voice interaction module generates text to be recognized based on the user's voice characteristics and generates voice commands based on the feedback results; The context management module stores the interaction log to the corresponding storage unit according to the index key value; the interaction log includes the task type, adaptation path and feedback result belonging to the same data flow; The data set control module determines a parameter adjustment strategy based on the performance indicator factors in the interaction log, and the parameter adjustment strategy acts on the intelligent scheduling module and / or the fast and slow thinking module to optimize the corresponding parameters.

[0005] This solution achieves real-time speech-to-text conversion through the voice interaction module, accurately classifies task types through analysis of text features with the help of the intelligent scheduling module, and dynamically adapts the differentiated computing units (rapid response submodule and deep thinking submodule) in the fast and slow thinking module, thereby achieving low-latency response in simple tasks and ensuring deep reasoning capabilities in complex tasks; at the same time, the voice interaction module ensures the natural and smooth voice input and output, the context management module maintains the coherence of multiple rounds of dialogue through structured storage of interaction logs through indexed key-value pairs, and the data set control module dynamically optimizes scheduling strategies and model parameters based on performance indicators, ultimately building a closed-loop system of "real-time perception-intelligent scheduling-precise reasoning-context maintenance-continuous optimization", effectively overcoming the shortcomings of existing technologies such as the difficulty in balancing response speed and reasoning quality, broken context understanding, and insufficient system adaptability, and achieving a comprehensive improvement in the interaction efficiency, semantic understanding accuracy, and long-term operational stability of the voice dialogue system.

[0006] Preferably, the intelligent scheduling module includes a natural language understanding submodule, a task classification submodule and a task allocation submodule; The natural language understanding submodule extracts sentence features, keyword features, and contextual features of the text to be recognized; The task classification submodule determines the task type based on sentence features, keyword features and contextual features; The task allocation submodule dynamically adapts the fast and slow thinking modules according to the task type.

[0007] Preferably, the task assignment submodule is provided with a path selection mechanism, which includes a condition determination logic or a routing assignment logic; the task type determines the corresponding adaptation path through the condition determination logic or the routing assignment logic.

[0008] Preferably, the fast and slow thinking module includes a quick response submodule and a deep thinking submodule; The quick response submodule and / or the deep thinking submodule performs intent analysis, word meaning analysis and context relevance analysis on the recognized text to determine a feedback sequence and a corresponding serial number, and the feedback sequence and the corresponding serial number construct the feedback result.

[0009] Preferably, the quick response submodule includes N computing units, and the deep thinking submodule includes M computing units, where M>N.

[0010] Preferably, the voice interaction module includes a voice recognition submodule and a voice synthesis submodule; The language recognition submodule receives the user's voice information and performs streaming processing on it to obtain the text to be processed; The speech synthesis submodule reorganizes the feedback results according to the feedback sequence and the corresponding sequence number to obtain the speech instruction.

[0011] Preferably, the context management module includes an API calling unit, a message queue unit and a storage unit; The API calling unit is used to obtain the identity tag information of the text to be recognized, the adaptation path of the intelligent scheduling module, the task type, and the feedback result of the fast and slow thinking module, wherein the feedback result includes a feedback sequence and a corresponding sequence number; The message queue unit generates an interaction log according to the task type, adaptation path, and feedback sequence, and performs a splicing operation on the interaction log according to the identity tag information and the corresponding sequence number to obtain a target interaction log; allocates a storage unit for the target interaction log and configures an index key value for the storage unit.

[0012] Preferably, the step of performing a splicing operation on the interaction logs according to the identity tag information and the corresponding sequence number to obtain the target interaction log comprises the following steps: Extract the voiceprint features of the user's voice information as identity tag information, and determine the data flow channel according to the transmission direction of the data flow. Use identity tag information as the authentication medium for the data stream channel, and fork the data stream channel according to the adaptation path corresponding to the task type; configure the sequence number of the current interaction log based on the time sequence of the data stream processed at the fork point; Based on the sequence number, the interaction logs belonging to the same data stream channel are spliced together to obtain the target interaction log.

[0013] Preferably, the data set control module includes a performance analysis module and a parameter correction module; The performance analysis module analyzes the performance indicator factors in the interaction log to determine the parameter adjustment strategy of each intelligent scheduling module and / or the fast and slow thinking module; the performance indicator factors include at least task response time and model reliability score; The parameter correction module maps the parameter adjustment strategy to the intelligent scheduling module and / or the fast and slow thinking module to correct the corresponding parameters.

[0014] Preferably, the parameter adjustment strategy includes adjusting the condition determination logic or routing allocation logic of the intelligent scheduling module, or / and adjusting the network level and the number of computing units of the fast and slow thinking module.

[0015] Beneficial effects of the present invention: (1) In response to the technical bottleneck that lightweight models and large models in traditional dialogue systems cannot balance response efficiency and reasoning depth, this application constructs a dynamic adaptation architecture based on intelligent scheduling and differentiated computing units. Through the real-time extraction of sentence features, keyword features and context-related features by the natural language understanding submodule, the task classification submodule is driven to divide the input into simple / complex task types. The task allocation submodule then dynamically calls the fast response submodule or the deep thinking submodule through conditional judgment logic or routing allocation logic, achieving parallel support for low-latency response of simple tasks and multi-level deep reasoning of complex tasks, breaking through the inherent contradiction in the existing technology that "a single model is difficult to balance speed and quality"; (2) In response to the problem of context understanding interruption caused by fixed rule scheduling in hybrid architecture systems, this application proposes a time-series management solution for interaction logs based on voiceprint identity tags and data stream serial numbers. The unique identity tag composed of voiceprint features is obtained through the API call unit, and the data stream channel is bifurcated in combination with the message queue unit, and a globally unique serial number is configured based on the time series. This realizes the time-series splicing of task types, adaptation paths, and feedback results in the same data stream, significantly improving the coherence of context information transmission in multiple rounds of dialogue, and solving the technical problem of semantic faults when switching modules in the existing technology. (3) In response to the problem of long-term performance degradation caused by the lack of adaptive optimization mechanism in existing systems, this application designs a closed-loop parameter optimization system based on interactive log mining. The performance analysis module conducts real-time analysis of indicator factors such as task response time and model reliability score. The parameter correction module dynamically maps the adjustment strategy to the condition judgment threshold of the intelligent scheduling module or the network layer / number of computing units of the fast and slow thinking module, thereby realizing periodic automatic optimization of the scheduling logic and model architecture, so that the system can maintain a dynamic balance between response efficiency and reasoning accuracy during long-term operation.

[0016] The above content of the invention is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Other features, objects, and advantages of the present invention will become more apparent upon reading the detailed description of the non-limiting embodiments made with reference to the following drawings. The drawings are for the purpose of illustrating preferred embodiments only and are not to be construed as limiting the present invention. Like reference characters are used throughout the drawings to designate like parts.

[0018] Figure 1 This is a system structure block diagram of the large model-based voice dialogue system of the present invention.

[0019] Figure 2 This is a structural block diagram of the subsystems of the large-model-based voice dialogue system of the present invention.

[0020] Explanation of the marks in the figure: 1-intelligent scheduling module, 11-natural language understanding sub-module, 12-task classification sub-module, 13-task allocation sub-module, 2-fast and slow thinking module, 21-quick response sub-module, 22-deep thinking sub-module, 3-voice interaction module, 31-voice recognition sub-module, 32-voice synthesis sub-module, 4-context management module, 41-API call unit, 42-message queue unit, 43-storage unit, 5-dataset control module, 51-performance analysis module, 52-parameter correction module. DETAILED DESCRIPTION

[0021] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0022] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the operations (or steps) as sequential processes, many of the operations (or steps) therein can be performed in parallel, concurrently, or simultaneously. In addition, the order of the operations can be rearranged. The process can be terminated when its operations are completed, but can also have additional steps not included in the figures; the process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0023] In the fields of artificial intelligence and natural language processing, existing technical solutions for voice dialogue systems primarily include the following typical implementation approaches. The first is fast-response systems based on lightweight models, commonly found in smart speakers or mobile assistants. These systems use end-to-end speech recognition models (such as CTC-based recurrent neural networks) to convert user speech into text. They then employ small deep neural networks or support vector machines for intent classification, quickly identifying simple commands (such as "check the weather" or "set an alarm") and generating responses using pre-set templates or external APIs. Response times are typically under one second. The second is cloud-based deep dialogue systems based on large models. These employ large-scale pre-trained language models (such as the Transformer-based GPT series) to handle complex tasks (such as "plan a trip") through high-precision speech recognition and deep reasoning, generating logically coherent natural language responses. These models contain billions of parameters, support multi-turn conversations, and understand context, but rely on cloud-based GPU computing. The third is hybrid architecture systems, such as Microsoft Cortana, which combine lightweight local models with large cloud-based models. The former handles simple tasks, while the latter handles complex requirements. Tasks are scheduled using rules (such as input length) to achieve a balance between speed and quality. These solutions all rely on the core processes of speech recognition, intent understanding, and response generation, and are widely used in smart devices.

[0024] While the above-mentioned technical solutions have achieved some success in voice interaction, they suffer from several key drawbacks, which are precisely overcome by the embodiments of the present invention. First, systems based on lightweight models respond quickly for simple tasks, but due to limited model capacity, their comprehension and response quality are low. For example, they cannot accurately parse ambiguous or multi-intent input (such as "What's the weather like tomorrow? Also, tell me what to wear"), resulting in users having to repeat input or receiving unsatisfactory answers. Second, cloud-based systems based on large models, while capable of providing high-quality answers, suffer from long inference latency (typically 2-5 seconds), making them particularly inefficient for simple tasks. This degrades the user experience due to extended wait times. Furthermore, while hybrid architecture systems attempt to balance the two, their scheduling mechanisms are relatively simple (mostly based on fixed rules rather than intelligent judgment), leading to inconsistent response styles or broken contextual understanding when switching between modules. Furthermore, they fail to fully exploit the potential of large models for rapid response. These issues make it difficult for existing systems to achieve both rapid response and high intelligence within a single architecture.

[0025] In response to the above problems, the technical problem proposed by the present invention is: how to design a voice dialogue system based on a large model by optimizing the large model and intelligent scheduling mechanism, so that it can maintain low latency and fast response when processing simple tasks, while providing high-quality deep reasoning in complex tasks, thereby improving the fluency and intelligence of the overall interactive experience.

[0026] Example: Figure 1 As shown in the figure, the voice dialogue system based on the large model is specifically composed of an intelligent scheduling module 1, a fast and slow thinking module 2, a voice interaction module 3, a context management module 4, and a data set control module 5; The intelligent scheduling module determines the task type based on the text to be recognized and dynamically adapts the fast and slow thinking modules according to the task type; The fast and slow thinking module dynamically adapts the computing unit according to the task type and outputs feedback results; The voice interaction module generates text to be recognized based on the user's voice characteristics and generates voice commands based on the feedback results; The context management module stores the interaction log to the corresponding storage unit according to the index key value; the interaction log includes the task type, adaptation path and feedback result belonging to the same data flow; The data set control module determines a parameter adjustment strategy based on the performance indicator factors in the interaction log, and the parameter adjustment strategy acts on the intelligent scheduling module and / or the fast and slow thinking module to optimize the corresponding parameters.

[0027] It can be understood that this embodiment uses a voice interaction module to convert the user's voice characteristics into text to be recognized in real time and generates voice instructions based on the feedback results. With the help of the intelligent scheduling module, the text characteristics are analyzed to determine the task type and dynamically adapt the fast and slow thinking module, so that the fast and slow thinking module can flexibly adapt the computing unit according to the task type and output the feedback result. At the same time, the context management module stores the interaction log containing the task type, adaptation path and feedback result of the same data stream to the corresponding storage unit according to the index key value, and then uses the data set control module to determine the parameter adjustment strategy according to the performance indicator factor in the interaction log to optimize the intelligent scheduling module or / and the fast and slow thinking module. The corresponding parameters of the thinking module realize real-time conversion of speech to text, accurate classification of task types and dynamic adaptation of computing units, ensuring low-latency response for simple tasks and deep reasoning capabilities for complex tasks, maintaining the continuity of multiple rounds of dialogue, and improving the system's adaptive optimization capabilities. It builds a closed-loop system of "real-time perception-intelligent scheduling-precise reasoning-context maintenance-continuous optimization", effectively overcoming the defects of existing technologies such as the difficulty in balancing response speed and reasoning quality, the interruption of context understanding and the insufficient system adaptability, and realizes a comprehensive improvement in the interaction efficiency, semantic understanding accuracy and long-term operation stability of the voice dialogue system.

[0028] As an optional embodiment, Figure 2 As shown, the intelligent scheduling module includes a natural language understanding submodule 11, a task classification submodule 12 and a task allocation submodule 13; The natural language understanding submodule extracts sentence features, keyword features, and contextual features of the text to be recognized; The task classification submodule determines the task type based on sentence features, keyword features and contextual features; The task allocation submodule dynamically adapts the fast and slow thinking modules according to the task type.

[0029] It can be understood that this embodiment uses the natural language understanding sub-module to extract the sentence features, keyword features and context-related features of the text to be recognized, and determines the task type based on the above features through the task classification sub-module. The task allocation sub-module then dynamically adapts the fast and slow thinking modules according to the task type, thereby realizing multi-dimensional feature analysis of the input text and accurate determination of the task type, thereby driving the differentiated computing resource allocation of the fast and slow thinking modules, effectively solving the problems of rigid task scheduling and inability to balance response efficiency and reasoning depth in traditional systems, and establishing an intelligent switching mechanism for the voice dialogue system between low-latency processing of simple tasks and deep reasoning of complex tasks, thereby improving the system's adaptive processing capabilities for different types of tasks.

[0030] For example, when a user says, "What's the weather like tomorrow? And remind me to bring an umbrella," the natural language understanding submodule first performs syntactic analysis on the sentence to extract sentence features, identifying the first half of the sentence, "What's the weather like tomorrow?", as a typical interrogative sentence, and the second half, "And remind me to bring an umbrella," as an imperative sentence. Simultaneously, a keyword extraction algorithm captures words such as "tomorrow," "weather," "reminder," and "bring an umbrella," forming a keyword feature set. If the user previously recorded a conversation about "going hiking the day after tomorrow," the context management module retrieves the interaction log using the index key value, enabling the system to understand the association between "reminder to bring an umbrella" and the "hiking" scenario (constructing contextual association features). Based on these features, the task classification submodule first determines that the first half of the sentence matches the sentence structure and keyword pattern of the "weather query" task. Then, combining the "reminder" keyword and the contextual hiking scenario in the second half of the sentence, it determines that the overall task type is a composite task of "weather query + life reminder" (complex task). The task allocation submodule then dynamically adapts the computational units of the fast and slow thinking modules based on this composite task type.

[0031] As an optional embodiment, the task assignment submodule is provided with a path selection mechanism, which includes condition determination logic or route assignment logic; the task type determines the corresponding adaptation path through the condition determination logic or route assignment logic.

[0032] It can be understood that this embodiment adopts a path selection mechanism that includes conditional judgment logic or routing allocation logic in the task allocation sub-module, which realizes the dynamic determination of the corresponding adaptation path according to the task type, and then drives the intelligent calling of differentiated computing units in the fast and slow thinking module, effectively solving the problems of fixed task scheduling rules and inconsistent response styles when switching modules in traditional hybrid architecture systems, and improving the system's dynamic balance ability between low-latency processing of simple tasks and deep reasoning of complex tasks.

[0033] For example, when a user asks, "Please check the weather for next week and plan some attractions suitable for my children," the task classification submodule first determines that the task belongs to the combined task type of "weather query + itinerary planning." The conditional decision logic within the path selection mechanism then detects that the keyword "attraction planning" corresponds to complex task characteristics (e.g., requiring multiple rounds of reasoning and knowledge integration). Simultaneously, the routing logic, based on a pre-set task type-path mapping table, maps "weather query" to the N computational units of the fast response submodule (e.g., a lightweight convolutional neural network) and "itinerary planning" to the M computational units of the deep thinking submodule (e.g., a multi-layer Transformer architecture). After the task allocation submodule confirms that the proportion of complex tasks exceeds a threshold through the conditional decision logic, it combines this with the routing logic to generate parallel processing paths: a fast path processes the "weather query" to ensure low-latency response, while a deep path processes the "itinerary planning" to leverage a large model for semantic understanding and knowledge reasoning. Ultimately, adaptive paths are dynamically generated for the different subtasks within the complex task.

[0034] As an optional embodiment, the fast and slow thinking module includes a quick response submodule 21 and a deep thinking submodule 22; the quick response submodule and / or the deep thinking submodule performs intent analysis, word meaning analysis and context relevance analysis on the text to be recognized to determine a feedback sequence and a corresponding serial number, and the feedback sequence and the corresponding serial number construct the feedback result.

[0035] As an optional embodiment, the quick response submodule includes N computing units, and the deep thinking submodule includes M computing units, where M>N.

[0036] It is understandable that this embodiment employs a technique that subdivides the fast and slow thinking modules into a quick response submodule and a deep thinking submodule. These modules determine feedback sequences and serial numbers through intent analysis, word meaning analysis, and contextual relevance analysis. Combined with the N computing units of the quick response submodule and the M computing units of the deep thinking submodule (M>N), this allows for flexible allocation of computing resources based on task complexity. Simple tasks are rapidly processed by the quick response submodule using lightweight computing units, ensuring low latency. Complex tasks are further reasoned using the deeper computing units of the deep thinking submodule. The feedback sequences and serial numbers, constructed from these units, maintain semantic coherence across multiple rounds of conversation, balancing response speed and reasoning depth within the same architecture.

[0037] For example, when a user asks "Help me check the weather next week, and then plan some attractions suitable for taking children to visit", the quick response sub-module first identifies the weather query intent through intent analysis for the "check the weather next week" part, uses N computing units to perform word meaning analysis on keywords such as "next week" and "weather", and combines the contextual relevance of past weather queries (such as the user's permanent residence, historical query preferences, etc.) to quickly generate a feedback sequence "Next week's weather is [specific weather conditions]" and assign it a sequence number 001; and when the deep thinking sub-module processes "Plan attractions suitable for taking children to visit", it uses M computing units to first perform intent analysis to clarify that it is a parent-child travel attraction planning task, deeply analyzes the meanings of words such as "children" and "attractions", associates it with contextual information such as the user's previous travel purpose and the child's age, and generates a feedback sequence containing the attraction name, features, and travel suggestions through knowledge graph and deep reasoning, "The attractions recommended for you to take children to are [attraction 1] and [attraction 2]. They have [feature 1] and [feature 2] respectively, and the travel suggestions are [suggested content]", and assigns it a sequence number 002. Finally, the two feedback sequences are integrated into a complete feedback result according to the serial number, which not only enables a quick response to weather queries, but also completes scenic spot planning through the complex calculations of the deep thinking sub-module. The serial number also facilitates subsequent interaction management and context tracing.

[0038] As an optional embodiment, the voice interaction module includes a voice recognition submodule 31 and a voice synthesis submodule 32; the voice recognition submodule receives the user's voice information and performs streaming processing on it to obtain a text to be processed; The speech synthesis submodule reorganizes the feedback results according to the feedback sequence and the corresponding sequence number to obtain the speech instruction.

[0039] It can be understood that this embodiment uses a voice recognition submodule to perform streaming processing on the received user voice information to generate a text to be processed, and at the same time uses a voice synthesis submodule to reorganize the feedback results according to the feedback sequence and its corresponding serial number to generate voice instructions. This technical means realizes real-time two-way conversion between voice and text, ensures the natural fluency of voice input and output, avoids the problems of high voice interaction delay and abrupt semantic punctuation in traditional systems, and maintains the logical coherence of voice responses in multiple rounds of conversations through the orderly reorganization of feedback results by serial numbers, thereby improving the fluency and naturalness of the user interaction experience.

[0040] For example, when a user says, "Please check the weather for next week and then plan some attractions suitable for taking the kids," the speech recognition submodule captures the speech signal in real time using streaming processing, converting it into text segments. It first recognizes "Please check the weather for me," then "Next week's weather," and then "Then plan some attractions suitable for taking the kids," and integrates them into the complete text to be processed. After the fast and slow thinking module generates the feedback sequence, the speech synthesis submodule receives the weather query feedback (sequence number 001) and the attraction planning feedback (sequence number 002). It reorganizes the feedback results according to the sequence number order, first converting "Next week's weather is [specific weather conditions]" into natural speech with a steady intonation to convey the information. It then converts "Recommended attractions suitable for taking the kids include [attraction X], [attraction Y]..." into speech, adjusting the tone based on the relationship between the characters, for example, making the tone slightly more lively to suit a parent-child setting. Finally, it generates a coherent speech command in the order of 001, 002, ensuring the user hears a clear and natural flow of content.

[0041] As an optional embodiment, the context management module includes an API calling unit 41, a message queue unit 42, and a storage unit 43; the API calling unit is used to obtain identity tag information of the text to be recognized, the adaptation path of the intelligent scheduling module, the task type, and the feedback result of the fast and slow thinking module, wherein the feedback result includes a feedback sequence and a corresponding sequence number; The message queue unit generates an interaction log according to the task type, adaptation path, and feedback sequence, and performs a splicing operation on the interaction log according to the identity tag information and the corresponding sequence number to obtain a target interaction log; allocates a storage unit for the target interaction log and configures an index key value for the storage unit.

[0042] It can be understood that this embodiment uses the API call unit in the context management module to obtain the identity tag information of the text to be identified, the adaptation path of the intelligent scheduling module, the task type and the feedback results of the fast and slow thinking module, and generates an interaction log according to the task type, adaptation path and feedback sequence through the message queue unit, and splices and aligns the interaction logs based on the identity tag information and serial number to generate the target interaction log, and then allocates a storage unit to it and configures the index key value. This technical means realizes the structured storage and time-series management of the interaction log, effectively solves the problem of context understanding interruption when switching modules in traditional hybrid architecture systems, ensures the consistency and traceability of semantic information in multiple rounds of dialogues, and improves the system's context maintenance capabilities for long dialogue scenarios.

[0043] For example, when a user says, "Help me check next week's weather and plan some attractions suitable for taking my kids," the API call unit first extracts the voiceprint features of the user's speech as identity tag information. It also obtains the intelligent scheduling module's adaptation path for the task (a quick-response submodule for weather query adaptation and a deep-thinking submodule for attraction planning adaptation), the "complex task" type, and the feedback sequence generated by the fast-slow thinking module (sequence number 001: next week's weather; sequence number 002: recommended attractions for parents and children). Based on this information, the message queue unit generates an initial interaction log, using the voiceprint identity tag as the data stream channel authentication medium. The data stream channel is bifurcated into two sub-channels according to the adaptation path corresponding to the task type. At the bifurcation point, the current interaction log is assigned a sequence number (001 and 002) based on the time sequence. Based on the sequence number, the interaction logs of the two sub-channels are concatenated in the chronological order of "weather query → attraction planning" to form the target interaction log. Finally, the message queue unit allocates a storage unit for the target interaction log, and generates an index key value (such as "SV1234_20250627_001-002") using "voiceprint label + timestamp + serial number range", and stores it in the corresponding storage unit so that context information can be quickly retrieved through the index later to maintain the semantic coherence of multiple rounds of conversations.

[0044] As an optional embodiment, the step of performing a splicing operation on the interaction logs according to the identity tag information and the corresponding sequence number to obtain the target interaction log includes the following steps: Extract the voiceprint features of the user's voice information as identity tag information, and determine the data flow channel according to the transmission direction of the data flow. Use identity tag information as the authentication medium for the data stream channel, and fork the data stream channel according to the adaptation path corresponding to the task type; configure the sequence number of the current interaction log based on the time sequence of the data stream processed at the fork point; Based on the sequence number, the interaction logs belonging to the same data stream channel are spliced together to obtain the target interaction log.

[0045] It can be understood that this embodiment adopts the technical means of extracting user voiceprint features as identity tag information, determining the channel based on the data stream transmission direction, using the identity tag as the channel authentication medium and bifurcating the data stream according to the task adaptation path, and then configuring the serial number according to the bifurcation point time series. Finally, based on the serial number, the same channel interaction log is spliced together to achieve the time-series structured management of the interaction log, effectively solving the problem of context information fragmentation of multiple rounds of conversations in the traditional system, ensuring the uniqueness and relevance of the multi-task interaction logs of the same user through voiceprint authentication and data stream bifurcation mechanism, and the time series configuration of the serial number ensures the temporal consistency of the interaction log, providing a complete context traceability basis for the semantic understanding of multiple rounds of conversations.

[0046] For example, when a user says, "Please help me check the weather for next week and plan some attractions suitable for my children," the system first extracts the voiceprint features (e.g., a combination of spectral feature parameters) from the speech as identity tag information and determines the main data stream channel based on the transmission direction from voice input to text output. Next, using the voiceprint tag as an authentication medium, and because the task type is a composite "weather query + attraction planning" task, the system forks the main data stream channel according to the adaptive path of the intelligent scheduling module (weather query uses the quick response submodule, and attraction planning uses the deep thinking submodule), generating two subchannels. At the bifurcation point, the system assigns sequence number 001 to the weather query task and sequence number 002 to the attraction planning task, based on the chronological order of data stream processing. Finally, based on the serial number, the system concatenates the interaction log of the weather query sub-channel (including voiceprint labels, adaptation paths, and the feedback sequence "next week's weather will be [specific conditions]") and the interaction log of the scenic spot planning sub-channel (including voiceprint labels, adaptation paths, and the feedback sequence "recommended scenic spots [X], [Y]") in the order of 001→002, forming a target interaction log that fully records the user's complex task interaction process.

[0047] As an optional embodiment, the data set control module includes a performance analysis module 51 and a parameter correction module 52; the performance analysis module analyzes the performance indicator factors in the interaction log to determine the parameter adjustment strategy of each intelligent scheduling module and / or the fast and slow thinking module; the performance indicator factors include at least task response time and model reliability score; The parameter correction module maps the parameter adjustment strategy to the intelligent scheduling module and / or the fast and slow thinking module to correct the corresponding parameters.

[0048] It can be understood that this embodiment adopts a technical means of setting up a performance analysis module and a parameter correction module in the data set control module, and analyzing the performance indicator factors such as task response time and model reliability score in the interaction log through the performance analysis module to determine the parameter adjustment strategy, and then the parameter correction module maps the strategy to the intelligent scheduling module or the fast and slow thinking module to correct the corresponding parameters, thereby achieving a closed-loop adaptive optimization effect of the system. This mechanism can mine the performance data in the interaction log in real time, dynamically adjust the scheduling logic and model architecture parameters, and solve the problem of performance degradation of the existing system after long-term operation, so that the system can maintain a dynamic balance between the response efficiency of simple tasks and the reasoning accuracy of complex tasks, thereby improving the long-term operation stability and intelligent evolution capability of the system.

[0049] For example, when a user performs the task "Check next week's weather for me and plan some attractions suitable for my children," the performance analysis module extracts from the interaction log that the response time for the weather query task is 1.2 seconds (exceeding the preset threshold by 1 second), resulting in a model reliability score of 85 out of 100. The response time for the attraction planning task is 3.5 seconds, resulting in a reliability score of 92. The model reliability score is based on the data dimensions of task response accuracy and semantic understanding coherence, and the reliability score is derived based on these two factors and their manually set weighting factors. Analysis revealed that the rapid response submodule experienced excessive response delays when handling simple tasks, likely due to insufficient N computing units. Based on this, the performance analysis module generated a parameter adjustment strategy: lowering the threshold for the conditional decision logic of the rapid response submodule in the intelligent scheduling module from "response time > 1.5 seconds" to ">1 second" and increasing the number of computing units in the rapid response submodule to N+2. After receiving the strategy, the parameter modification module immediately mapped the adjustment instructions to the conditional judgment logic parameters of the intelligent scheduling module and modified the computing unit configuration of the rapid response submodule. After the adjustment, the system's response time for subsequent similar weather query tasks was reduced to 0.8 seconds, and the reliability score increased to 90 points, achieving dynamic optimization of system parameters based on actual operating data.

[0050] Furthermore, task response accuracy can be calculated by extracting feedback from interaction logs and comparing it against pre-set accuracy standards (such as weather data interfaces, scenic spot knowledge bases, and task templates). This allows for the accuracy of intent parsing and information correctness to be calculated. For example, if a user queries "next week's weather," and the feedback is consistent with the meteorological bureau's data and unambiguous, the accuracy score can reach 95-100 points. However, if some information is incorrect or the intent parsing deviates, points will be deducted based on the degree of error.

[0051] Furthermore, semantic understanding coherence is calculated based on the interaction logs of the context management module, evaluating the logic of context splicing across multiple conversations (e.g., whether historical conversations are linked by sequence numbers and whether semantic consistency is maintained). For example, if the model can associate the context of "family travel" (e.g., mentioning "bringing children" in the historical conversation) with a user's consecutive inquiries about "weather" and "attractions"), the coherence score can reach 95-100 points. If there is a contextual disconnect (e.g., ignoring the historical travel purpose), points will be deducted based on the degree of disconnection.

[0052] As an optional embodiment, the parameter adjustment strategy includes adjusting the condition determination logic or routing allocation logic of the intelligent scheduling module, or / and adjusting the network level and the number of computing units of the fast and slow thinking module.

[0053] As can be understood, this embodiment utilizes a technical approach to dynamically optimize the scheduling logic and model architecture based on real-time performance data by using the data set control module to generate parameter adjustment strategies, adjusting the intelligent scheduling module's conditional judgment logic, routing allocation logic, or the network hierarchy and number of computing units in the fast and slow thinking module. This mechanism addresses issues such as response delays for simple tasks or insufficient reasoning for complex tasks by intelligently adjusting parameters to balance efficiency and accuracy. This addresses the long-term performance degradation of existing systems, resulting in comprehensive improvements in the interaction efficiency, semantic understanding accuracy, and long-term operational stability of voice dialogue systems.

[0054] For example, when the system processes the task of "Help me check the weather for next week, and then plan some attractions suitable for taking my children to," the performance analysis module extracts from the interaction log that the response time of the weather query task is 1.2 seconds (exceeding the preset threshold by 1 second), and the model reliability score is 85 points. In contrast, although the response time of the attraction planning task is 3.5 seconds, the reliability score is 92 points. After analysis, it was found that the conditional judgment logic threshold of the rapid response submodule in the intelligent scheduling module was set too high (for example, the original threshold was "response time>1.5 seconds" to trigger the adjustment), or the rapid response submodule task had a high response delay and the number of computing units N was insufficient. At this point, a parameter adjustment strategy is generated: the conditional judgment logic threshold of the intelligent scheduling module is lowered to "response time>1 second", and the 4-layer CNN architecture of the rapid response submodule is adjusted to 3 layers, and the number of computing units N is increased to N+2. The parameter correction module maps this strategy to the conditional judgment logic parameters of the intelligent scheduling module and the computing unit configuration of the fast and slow thinking module. After the adjustment, the system's response time in subsequent weather query tasks was shortened to 0.8 seconds, and the reliability score was increased to 90 points, realizing dynamic optimization of scheduling logic and computing resources.

[0055] The specific implementation described above is a preferred implementation of the large-scale model-based voice dialogue system of the present invention, and is not intended to limit the specific implementation scope of the present invention. The scope of the present invention includes but is not limited to this specific implementation. Any equivalent changes made in accordance with the shape and structure of the present invention are within the scope of protection of the present invention.

Claims

1. A large-model-based voice dialogue system, characterized by: include: Intelligent scheduling module, fast and slow thinking module, voice interaction module, context management module and data set control module; The intelligent scheduling module determines the task type based on the text to be recognized and dynamically adapts the fast and slow thinking modules according to the task type; The fast and slow thinking module dynamically adapts the computing unit according to the task type and outputs feedback results; The voice interaction module generates a text to be recognized according to the user's voice characteristics and generates a voice instruction based on the feedback result; The context management module stores the interaction log to the corresponding storage unit according to the index key value; the interaction log includes the task type, adaptation path and feedback result belonging to the same data flow; The data set control module determines a parameter adjustment strategy based on the performance indicator factors in the interaction log, and the parameter adjustment strategy acts on the intelligent scheduling module and / or the fast and slow thinking module to optimize the corresponding parameters.

2. The large model-based speech dialogue system according to claim 1, characterized in that: The intelligent scheduling module includes a natural language understanding submodule, a task classification submodule and a task allocation submodule; The natural language understanding submodule extracts sentence features, keyword features, and contextual features of the text to be recognized; The task classification submodule determines the task type based on sentence features, keyword features and contextual features; The task allocation submodule dynamically adapts the fast and slow thinking modules according to the task type.

3. The large model-based speech dialogue system according to claim 2, characterized in that: The task assignment submodule is provided with a path selection mechanism, which includes a condition determination logic or a routing assignment logic; the task type determines the corresponding adaptation path through the condition determination logic or the routing assignment logic.

4. The large model-based speech dialogue system according to claim 1, characterized in that: The fast and slow thinking module includes a quick response submodule and a deep thinking submodule; The quick response submodule and / or the deep thinking submodule performs intent analysis, word meaning analysis and context relevance analysis on the recognized text to determine a feedback sequence and a corresponding serial number, and the feedback sequence and the corresponding serial number construct the feedback result.

5. The large model-based speech dialogue system according to claim 1, 3 or 4, characterized in that: The quick response submodule includes N computing units, and the deep thinking submodule includes M computing units, where M>N.

6. The large model-based speech dialogue system according to claim 4, characterized in that: The speech interaction module includes a speech recognition submodule and a speech synthesis submodule; The language recognition submodule receives the user's voice information and performs streaming processing on it to obtain the text to be processed; The speech synthesis submodule reorganizes the feedback results according to the feedback sequence and the corresponding sequence number to obtain the speech instruction.

7. The large model-based speech dialogue system according to claim 4, characterized in that: The context management module includes an API calling unit, a message queue unit and a storage unit; The API calling unit is used to obtain the identity tag information of the text to be recognized, the adaptation path of the intelligent scheduling module, the task type, and the feedback result of the fast and slow thinking module, wherein the feedback result includes a feedback sequence and a corresponding sequence number; The message queue unit generates an interaction log according to the task type, adaptation path, and feedback sequence, and performs a splicing operation on the interaction log according to the identity tag information and the corresponding sequence number to obtain a target interaction log; allocates a storage unit for the target interaction log and configures an index key value for the storage unit.

8. The large model-based speech dialogue system according to claim 7, characterized in that: The step of performing a splicing operation on the interaction logs according to the identity tag information and the corresponding sequence number to obtain the target interaction log includes the following steps: Extract the voiceprint features of the user's voice information as identity tag information, and determine the data flow channel according to the transmission direction of the data flow. Use identity tag information as the authentication medium for the data stream channel, and fork the data stream channel according to the adaptation path corresponding to the task type; configure the sequence number of the current interaction log based on the time sequence of the data stream processed at the fork point; Based on the sequence number, the interaction logs belonging to the same data stream channel are spliced together to obtain the target interaction log.

9. The large model-based speech dialogue system according to claim 1, characterized in that: The data set control module includes a performance analysis module and a parameter correction module; The performance analysis module analyzes the performance indicator factors in the interaction log and determines the parameter adjustment strategy of each intelligent scheduling module and / or fast and slow thinking module; The performance indicator factors include at least task response time and model reliability score; The parameter correction module maps the parameter adjustment strategy to the intelligent scheduling module and / or the fast and slow thinking module to correct the corresponding parameters.

10. The large model-based speech dialogue system according to claim 9, characterized in that: The parameter adjustment strategy includes adjusting the condition determination logic or routing allocation logic of the intelligent scheduling module, or / and, Adjust the network layers and number of computing units of the fast and slow thinking modules.

Images