Anthropomorphic AI robot intelligent conversation method and system

Abstract

This invention relates to the technical fields of machine question answering, and provides a human-like AI robot intelligent dialogue method and system. The method includes: in an idle state, calling a local lightweight large language model to determine the semantic validity of the speech-recognized text; if the determination is valid, extracting voiceprint features to establish a conversation and switching to a working state; in the working state, determining the semantic priority of the speech-recognized text, with high-priority text immediately interrupting the current task, and low-priority text inserted into a waiting queue for sequential response. This invention overcomes the shortcomings of existing technologies, such as single interaction mode, poor anti-interference ability, and inability to intelligently handle concurrent requests, achieving natural, fluent, stable, and reliable intelligent dialogue interaction in complex environments. The dual-model architecture of a local lightweight large language model and a cloud-based large language model ensures real-time semantic determination and low latency, guarantees the intelligence and richness of the generated responses, and achieves a balance between efficiency and effectiveness.

Description

Technical Field

[0001] This invention relates to the fields of service consumption robots, machine question answering, and more particularly to a human-like AI robot intelligent dialogue method and system. Background Technology

[0002] With the deep integration of artificial intelligence technology, anthropomorphic AI robot intelligent dialogue systems are being used more and more widely in fields such as culture and tourism and customer service. Existing technologies have explored several directions to achieve more natural interaction.

[0003] First, there's the forced wake-up technology. This technology requires users to wake up the robot in a specific way before each conversation, including voice wake-up (such as saying a specific wake word) or facial wake-up (such as facing the robot and looking at it). While this mechanism can effectively avoid misidentification, the user experience is poor, and the interaction process is not natural and smooth. Especially when users need to ask questions continuously or engage in multi-turn conversations, the repeated wake-up operation greatly reduces interaction efficiency and fails to meet users' expectations for natural dialogue.

[0004] Second, multimodal interaction triggering technologies. For example, determining whether the user's voice is a complete sentence to initiate a wake-word-free dialogue, or introducing multimodal information (such as changes in lip contours) to help determine the user's interaction intent, thereby improving the accuracy of wake-up and interruption. These technologies attempt to solve the inconvenience of traditional interactions that rely on fixed wake words.

[0005] Despite the progress made by the aforementioned technologies at different levels, there are still many significant technical shortcomings in complex and realistic interactive scenarios, especially in noisy public environments such as museums, science and technology museums, and libraries. These shortcomings are specifically manifested as follows.

[0006] First, the interaction mode is simplistic, lacking dynamic perception and management of the robot's working state. Most existing technologies treat the robot as a single entity always in a ready-to-respond state, failing to simulate the intelligent switching dialogue management strategies humans employ between "idle" and "busy" working states. For example, when the robot is performing a task (such as providing guidance or thinking for another user), existing systems cannot, like a human, intelligently decide whether to respond immediately, delay processing, or ignore it based on the urgency of new requests and its own current state. This results in the robot either being easily interrupted by arbitrary dialogue, causing task chaos and a fragmented user experience, or exhibiting a "rigid" single-response mode, unable to handle concurrent multi-priority requests.

[0007] Second, the intent recognition and scene adaptability are insufficient, lacking a refined priority discrimination mechanism. Although existing technologies have introduced mechanisms to initiate dialogue based on complete sentences or without wake words, and to use multimodal information for intent judgment, their discrimination criteria often remain at a binary level of "whether it is effective." When the robot is in working mode, existing technologies lack the ability to finely distinguish the semantic priority of interaction requests (for example, they cannot effectively distinguish between "emergency help" and "general chat"). This leads to high-priority urgent inquiries and low-priority chat requests being treated equally, failing to ensure that critical tasks are not interrupted by irrelevant information while also accommodating reasonable responses to non-urgent requests.

[0008] Third, the utilization of multimodal information is insufficient, and the ability to resist interference from complex environments is limited. Although some existing technical solutions integrate visual (such as lip movements and facial expressions) and auditory information, their applications are mostly concentrated on the initial judgment of interactive intent or the auxiliary enhancement of speech recognition, failing to be deeply integrated with dialogue state management and speaker recognition technologies. In complex acoustic scenarios such as "cocktail parties," the system struggles to accurately distinguish between the target user's valid commands and background noise or irrelevant conversations from others, leading to false wake-ups, misidentification, or dialogue thread chaos, affecting the system's reliability and usability. Summary of the Invention

[0009] To address the shortcomings of the existing technologies, this invention provides a human-like AI robot intelligent dialogue method and system, which enables efficient, natural, stable and reliable intelligent dialogue interaction in complex environments, effectively balancing the convenience and accuracy of the interaction with the system's anti-interference capabilities.

[0010] In a first aspect, the present invention provides a human-like AI robot intelligent dialogue method, comprising: When the AI ​​robot is in an idle state, it continuously detects voice signals in the environment, performs voice recognition on the detected environmental voice signals in the idle state, and obtains the first text information. The semantic discrimination module calls the local lightweight large language model and combines the preset idle state discrimination prompt words to perform semantic validity discrimination on the first text information; if the discrimination result is valid, the session is triggered to start, the voiceprint features of the current speaker are extracted to establish the session, and the AI ​​robot state is switched from idle state to working state. When the AI ​​robot is in working mode, it continuously detects voice signals in the environment, performs voice recognition on the detected environmental voice signals in the working mode, and obtains second text information. The semantic discrimination module calls the lightweight large language model and combines preset working state discrimination prompts to determine the semantic priority of the second text information. If the discrimination result is high priority and higher than the current task priority, the current task is immediately interrupted, and the second text information is sent to the large model intelligent agent module. The large model intelligent agent module calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response through speech synthesis. If the discrimination result is low priority, the second text information and its corresponding speaker information are inserted into the waiting queue. After the current high-priority task is completed, the second text information is retrieved from the queue and sent to the large model intelligent agent module. The large model intelligent agent module calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response through speech synthesis.

[0011] Secondly, this invention provides a human-like AI robot intelligent dialogue system, applied to AI robots, comprising: The hardware layer includes a microphone array module and a main control computing unit. The microphone array module is used to collect voice signals from the environment, and the main control computing unit is used to run various software modules. The perception layer, deployed in the main control computing unit, includes a voice activity detection module, an automatic speech recognition module, and a speaker recognition module; the voice activity detection module is used to detect human voice segments from the audio stream acquired by the microphone array module; the automatic speech recognition module is used to convert the detected voice segments into text information; and the speaker recognition module is used to extract the speaker's voiceprint features. The processing layer, deployed in the main control computing unit, includes a semantic discrimination module and a speaker separation module. The semantic discrimination module calls a local lightweight large language model, loads corresponding discrimination prompts according to the current system running state, and performs semantic validity or priority discrimination on the text information. The speaker separation module separates the speech of the target speaker from the mixed speech based on the voiceprint features extracted by the speaker recognition module. The decision-making layer, deployed in the main control computing unit, includes a state management module, a session management module, a priority scheduling module, and a queue management module. The state management module maintains the system's idle and working states. The session management module creates a session and records the current speaker's voiceprint features when a session is triggered. The priority scheduling module determines the task processing order based on the semantic discrimination module's discrimination result. The queue management module maintains the queue to be processed and supports a speaker-based request-to-queue mechanism. The interaction layer, deployed in the main control computing unit, includes a large model intelligent agent module and a speech synthesis module; the large model intelligent agent module is used to receive filtered user voice text, call the cloud-based large language model for intent understanding and response generation; the speech synthesis module is used to convert the generated text response into a speech signal for output. When the state management module controls the AI ​​robot to operate in the idle state, the perception layer continuously detects environmental speech signals and generates first text information. The semantic discrimination module calls a local lightweight large language model and combines the idle state discrimination prompt words to perform semantic validity discrimination on the first text information. If the discrimination result is valid, the session management module triggers session initiation, extracts the voiceprint features of the current speaker to establish a session, and the state management module switches the AI ​​robot's state from the idle state to the working state. When the state management module controls the AI ​​robot to operate in the working state, the perception layer continuously detects environmental speech signals and generates second text information. The semantic discrimination module calls the lightweight large language model and combines the working state discrimination prompt words to perform semantic priority discrimination on the second text information. If the discrimination result is high priority, the priority scheduling module immediately interrupts the current task, sends the second text information to the large model intelligent agent module for response generation and output. If the discrimination result is low priority, the queue management module inserts the second text information and its corresponding speaker information into the waiting queue. After the current high priority task is processed, the priority scheduling module retrieves the second text information from the queue and sends it to the large model intelligent agent module for response generation and output.

[0012] Compared with the prior art, the beneficial effects of this invention are as follows: The anthropomorphic AI robot intelligent dialogue method and system provided by this invention effectively solves the technical defects of existing technologies, such as single interaction mode and lack of dynamic perception capabilities, by using a dual-state management mechanism of idle and working states. In the idle state, semantic validity judgment based on a lightweight large language model is adopted, which can naturally start the dialogue without the user having to repeatedly use a wake word, thus solving the problem of cumbersome interaction caused by traditional forced wake-up technology. When the judgment result is valid, the AI ​​robot automatically extracts voiceprint features to establish a conversation and switches to the working state. This process simulates the natural reaction of humans when they notice someone calling them and begin to listen attentively when they are idle. In the working state, a semantic priority judgment mechanism is further used to distinguish between emergency requests for help and general chatter: high-priority requests can immediately interrupt the current task, while low-priority requests are inserted into a waiting queue for sequential response. This effectively solves the defect of existing technologies that cannot intelligently handle concurrent requests, ensuring timely response to emergency matters while avoiding interference from irrelevant chatter to core tasks. The dual-model architecture, where the semantic discrimination module calls a lightweight large language model for semantic discrimination, and the large model intelligent agent module calls a cloud-based large language model for response generation, ensures both the real-time performance and low latency of semantic discrimination and the intelligence and richness of response generation, achieving a balance between efficiency and effectiveness. Simultaneously, the use of a processing queue allows the AI ​​robot to manage multiple low-priority requests in an orderly manner, preventing request loss or confusion. Attached Figure Description

[0013] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. Some specific embodiments of the invention will be described in detail below with reference to the accompanying drawings in an exemplary and non-limiting manner. The same reference numerals in the drawings designate the same or similar parts or components. It should be understood by those skilled in the art that these drawings are not necessarily drawn to scale. In the drawings: Figure 1 This is a flowchart illustrating an anthropomorphic AI robot intelligent dialogue method according to an embodiment of the present invention. Detailed Implementation

[0014] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0015] Example 1 See Figure 1 This embodiment provides a human-like AI robot intelligent dialogue method, including the following steps: S101. When the AI ​​robot is in an idle state, the AI ​​robot continuously detects the voice signals in the environment, performs voice recognition on the detected environmental voice signals in the idle state, and obtains the first text information. S102. The semantic discrimination module calls the local lightweight large language model and combines the preset idle state discrimination prompt words to perform semantic validity discrimination on the first text information; if the discrimination result is valid, the session is triggered to start, the voiceprint features of the current speaker are extracted to establish the session, and the AI ​​robot state is switched from idle state to working state; wherein, the session management module can extract the voiceprint features of the current speaker to establish the session. S103. When the AI ​​robot is in working state, the AI ​​robot continuously detects the voice signals in the environment, performs voice recognition on the detected environmental voice signals in the working state, and obtains the second text information. S104. The semantic discrimination module calls the lightweight large language model and combines it with preset working state discrimination prompts to perform semantic priority discrimination on the second text information. If the discrimination result is high priority and higher than the current task priority, the current task is immediately interrupted, and the second text information is sent to the large model intelligent agent module. The large model intelligent agent module calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response through speech synthesis. If the discrimination result is low priority, the second text information and its corresponding speaker information are inserted into the waiting queue. After the current high-priority task is processed, the second text information is taken out of the queue and sent to the large model intelligent agent module. The large model intelligent agent module calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response through speech synthesis.

[0016] It should be noted that the method provided in this embodiment, by using a dual-state management mechanism of idle and working states, effectively solves the technical defects of existing technologies, such as a single interaction mode and a lack of dynamic perception capabilities. In existing technologies, robots are often regarded as a single entity that is always in a ready-to-respond state, which is either easily interrupted by arbitrary dialogue, leading to task chaos, or exhibits a rigid single response mode that cannot handle concurrent requests. However, this embodiment achieves a deep simulation of human dialogue habits by clearly distinguishing between idle and working states and designing differentiated processing strategies for different states.

[0017] Specifically, in the idle state, this embodiment employs semantic validity judgment based on a lightweight large language model, allowing for natural conversation initiation without the user repeatedly using a wake word, thus solving the cumbersome interaction problem caused by traditional forced wake-up technology. When the judgment result is valid, the AI ​​robot automatically extracts voiceprint features to establish a conversation and switches to the working state. This process simulates the natural human reaction of noticing someone calling and starting to listen attentively when idle. In the working state, this embodiment further uses a semantic priority judgment mechanism to distinguish between urgent requests for help and general casual conversation: high-priority requests can immediately interrupt the current task, while low-priority requests are inserted into a waiting queue for sequential response. This design effectively solves the deficiency of existing technologies in intelligently handling concurrent requests, ensuring timely response to urgent matters while avoiding interference from irrelevant casual conversations with core tasks.

[0018] Furthermore, in this embodiment, a dual-model architecture is employed: the semantic discrimination module calls a lightweight large language model for semantic discrimination, while the large model intelligent agent module calls a cloud-based large language model for response generation. This layered design ensures both the real-time performance and low latency of semantic discrimination (local lightweight model) and the intelligence and richness of response generation (cloud-based large model), achieving a balance between efficiency and effectiveness. Simultaneously, the use of a processing queue enables the AI ​​robot to manage multiple low-priority requests in an orderly manner, preventing request loss or confusion.

[0019] Furthermore, after inserting the second text information and its corresponding speaker information into the queue to be processed, the process also includes queue scheduling and state maintenance, specifically including: The priority scheduling module retrieves the next pending request from the queue according to the principle of prioritizing higher priority requests and consecutive requests from the same speaker. The request to retrieve the data is sent to the large model intelligent agent module, which calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response through speech synthesis. Repeat the above retrieval and processing steps until the queue is empty. While the AI ​​robot is in working state, it continuously monitors for new high-priority requests and allows high-priority requests to interrupt the queue processing flow at any time.

[0020] It should be noted that in this embodiment, the priority scheduling module retrieves the next pending request from the queue according to the principles of high priority and consecutive priority for the same speaker, sends it to the large model intelligent agent module for processing and outputs a response, and repeats this process until the queue is empty. Simultaneously, during the working state, it continuously monitors for newly arriving high-priority requests, allowing high-priority requests to interrupt the queue processing flow at any time. First, the solution provided in this embodiment defines the specific rules for queue scheduling, namely high priority and consecutive priority for the same speaker. The organic combination of these two principles ensures the priority handling of urgent matters while maintaining the continuity of multi-turn dialogues with the same user, achieving a balance between fairness and efficiency. Second, the mechanism of cyclical processing until the queue is empty ensures that all pending requests are eventually responded to, avoiding the problem of requests being indefinitely shelved. Third, continuously monitoring for newly arriving high-priority requests during the working state and allowing them to interrupt the current queue processing at any time demonstrates the AI ​​robot's sensitivity to dynamic changes. Even when dealing with backlogged tasks, it will not ignore new urgent needs and always maintains a rapid response capability to high-priority events. Fourth, the queue scheduling mechanism provided in this embodiment enables the AI ​​robot to manage multiple concurrent requests in an orderly manner, avoiding the problems of chaotic requests and disordered responses in traditional solutions. In museums with high visitor traffic during holidays, the robot may receive questions from multiple visitors simultaneously. The queue scheduling mechanism provided in this embodiment ensures that all questions are recorded and processed in order, while emergency assistance can be interrupted at any time, greatly improving the practicality of the AI ​​robot and user satisfaction.

[0021] Preferably, the step of calling a local lightweight large language model by the semantic discrimination module and combining it with preset idle state discrimination prompts to perform semantic validity discrimination on the first text information specifically includes: Load the idle state discrimination prompt, which includes typical feature examples of valid semantics and typical feature examples of invalid speech; The first text information is combined with the idle state discrimination prompt word and then input into the lightweight large language model. The model is required to output a valid or invalid binary structure discrimination result. If the output result is valid, the semantics of the first text information are determined to be valid, and the session initiation process is executed; if the output result is invalid, the current speech recognition result is discarded, and idle state monitoring continues.

[0022] It should be noted that in this embodiment, the large language model is guided to make judgments using preset typical feature examples, enabling the model to quickly and accurately identify whether a user intends to engage in dialogue with the robot. This avoids the problems of misjudgment or missed judgment that are easily encountered by traditional methods based on keyword matching or rule templates. Requiring the model to output structured binary results simplifies subsequent processing logic and improves system response speed. For speech judged as invalid (such as conversations with others, meaningless noise, etc.), it is directly discarded while idle state monitoring continues, effectively avoiding false wake-ups and invalid processing, and improving the AI ​​robot's anti-interference capability. The semantic discrimination method based on the large language model provided in this embodiment has good generalization ability and can adapt to the diverse expression styles of different users. It does not require preset rules for each expression style, greatly reducing the maintenance cost of the robot.

[0023] Preferably, the step of the semantic discrimination module calling the lightweight large language model and combining preset working state discrimination prompts to perform semantic priority discrimination on the second text information specifically includes: Load the working state discrimination prompt words, which predefine typical categories of high-priority semantics and low-priority semantics using a few-sample example approach; The second text information is combined with the working state discrimination prompt word and then input into the lightweight large language model. The model is required to output a valid high priority, a valid low priority, or an invalid ternary structured discrimination result. If the output result is valid and of high priority, the interruption mechanism of the current task is triggered; if the output result is valid and of low priority, it is transferred to the queue processing flow; if the output result is invalid, the current speech recognition result is discarded and the working state monitoring continues.

[0024] It should be noted that in this embodiment, priority categories are defined using a few-sample example approach, enabling the large language model to accurately understand the differences in the importance of different semantics. For example, it can clearly distinguish between "I'm not feeling well and need help" (high priority) and "How are you feeling today?" (low priority), solving the deficiency of existing technologies in their inability to finely differentiate the urgency of requests. The design of the ternary output provides the robot with a clear decision-making basis: high-priority requests trigger interruptions to ensure that urgent matters are handled immediately, low-priority requests are queued to prevent core tasks from being interrupted by irrelevant information, and invalid requests are directly discarded to reduce the robot's burden. This dynamic priority discrimination mechanism based on semantic content provided in this embodiment is more intelligent and flexible than traditional priority allocation methods based on request source or fixed rules, and can adapt to various emergencies and diverse user needs. Moreover, echoing the binary discrimination in the idle state, the ternary discrimination in the working state further reflects the robot's differentiated processing strategy in different states, enabling the robot to "selectively listen" like a human when busy, responding immediately only to truly important matters, and "noting down" other matters for later processing, greatly improving the anthropomorphism of the interaction and the user experience.

[0025] Preferably, after the session is initiated, the process further includes speaker separation, specifically including: Extract and register the voiceprint features of the current session initiator as a reference template; During the session, voiceprint features are extracted for each subsequently received speech segment and compared with the reference template; Only voiceprints that pass the voiceprint comparison are sent to the speech recognition and subsequent processing flow, while speech interference from speakers other than those in the current session is filtered out. The subsequent processing flow may include semantic discrimination and response generation.

[0026] It should be noted that the solution provided in this embodiment uses voiceprint recognition to lock onto the current conversation subject, allowing the robot to focus on handling consecutive questions from the same user, ensuring the coherence and contextual consistency of the dialogue. This voiceprint-based speaker separation mechanism is more reliable than methods based on sound source localization (such as based on sound direction). Even if the user moves during the conversation, the robot can still continuously track the same user through voiceprint recognition, and it can also filter out speech interference from speakers other than the current conversation speaker, avoiding the problem of dialogue thread chaos caused by simultaneous interaction by multiple users. For example, when the robot is providing a guided tour for a tourist, the idle chatter or questions from other tourists around will not interfere with the ongoing guided tour service, ensuring the smooth completion of the core task.

[0027] Preferably, after inserting the second text information and its corresponding speaker information into the queue to be processed, the process further includes speaker-based queue management, specifically including: Extract the speaker's voiceprint features for the current low-priority request; Traverse the queue of pending requests and determine whether there are any unprocessed requests from the same speaker in the queue. If other requests from the same speaker exist, the current request is inserted after the last request from that speaker to ensure continuous processing of requests from the same speaker; otherwise, the current request is appended to the end of the queue.

[0028] It should be noted that the solution provided in this embodiment enables continuous processing of requests from the same speaker, avoiding interruptions caused by other users' requests from the same user, thus maintaining the contextual coherence of multi-turn dialogues. For example, when a user asks multiple questions consecutively, such as "What is this exhibit?", "What era was it made in?", and "Where was it unearthed?", the robot will process these requests consecutively and respond in sequence, providing the user with a smooth experience as if having a continuous conversation with a real person. This queue-jumping mechanism in this embodiment optimizes the processing order of low-priority requests as much as possible while ensuring that high-priority requests are processed first, reflecting a balance between fairness and efficiency. Compared to the simple first-in-first-out queue in traditional technologies, the speaker-based queue-jumping mechanism can significantly improve the user experience and reduce the awkward situation where users have to wait for multiple irrelevant requests to be processed before receiving a response.

[0029] Preferably, if a user-initiated wake-up signal is detected in the working state, a wake-up strong interruption process is executed as a parallel approach to voice processing in the working state, specifically including: Immediately interrupt the currently executing task, including stopping voice output and canceling ongoing large model inference requests; Extract the speaker's voiceprint features corresponding to the wake-up voice and compare them with the voiceprint template of the current session; If the voiceprint matches, the current session will continue to process the voice after wake-up; if it does not match, a new session will be created, and the original session will be suspended or destroyed.

[0030] It should be noted that this embodiment provides a stronger forced interruption mechanism than semantic priority interruption. Regardless of the current task priority, user-initiated wake-up immediately interrupts the current task, ensuring the user's absolute control over the robot and conforming to the basic safety principles of human-computer interaction. By identifying the waker through voiceprint comparison, the robot can intelligently decide whether to continue the current session or create a new one: if the waker is the current session participant (e.g., the user wants to ask additional questions), the original session is continued to maintain contextual coherence; if the waker is another user (e.g., another visitor with urgent needs), a new session is created to handle the new request, while the original session is suspended rather than directly destroyed, preserving the possibility of later recovery. This design in this embodiment avoids the problem of losing original session information due to simple and abrupt interruptions, reflecting the robot's refined design for multi-tasking. Immediate interruption actions such as stopping voice output and canceling large model inference requests can quickly release system resources to respond to new high-priority requests.

[0031] Preferably, after the queue to be processed is empty in the working state, a state transition process is executed, specifically including: Monitor the current task processing status to determine whether all tasks have been completed and the queue is empty; If the state transition conditions are met, the AI ​​robot's state will be switched from the working state back to the idle state. In idle state, reset the context information of the current session, release the voiceprint template and session identifier resources, and prepare to receive new interaction requests.

[0032] It should be noted that this embodiment provides the triggering conditions and specific actions for the robot to return from the working state to the idle state, making the state transition logic a complete closed loop and avoiding the problem of the system remaining in the working state for a long time, thus preventing it from receiving new user requests. Resetting the session context and releasing resources when transitioning back to the idle state can promptly clean up unused session data, avoiding resource leaks and excessive memory consumption, and ensuring the long-term stability and reliability of the robot. This automated state transition mechanism in this embodiment simulates the natural state transition of humans after completing a task—"relaxing and preparing for a new task"—making the robot's behavior more in line with user expectations. For example, after the robot has answered all the queued questions, it will automatically return to the "standby listening" state, and users can initiate new conversations at any time without additional operation. Furthermore, releasing the voiceprint template and session identifier also clears the way for the establishment of subsequent new sessions, ensuring that each new session has an independent context space and avoiding data confusion between old and new sessions.

[0033] Preferably, if a user-initiated wake-up signal is detected in the idle state, a wake-up session initiation process is executed as a parallel method to trigger session initiation in the idle state, specifically including: After detecting a preset voice wake-up word or touch wake-up signal, confirm the validity of the wake-up signal; Extract the speaker's voiceprint features corresponding to the wake-up signal, create a new session, and assign a session identifier; The voice recognition text after wake-up is sent to the large model intelligent agent module, which calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response through speech synthesis. At the same time, the AI ​​robot's state is switched from idle state to working state.

[0034] The solution provided in this embodiment complements the semantic wake-up mechanism, offering users multiple interaction options. Users can either speak naturally without a wake word (semantic wake-up) or actively wake the robot using a wake word in specific situations (such as at a distance or in a noisy environment), balancing convenience and reliability. Through voiceprint feature extraction and session creation, the robot can identify the current user from the outset, laying the foundation for subsequent speaker separation and continuous dialogue. The step of verifying the validity of the wake-up signal (such as verifying the wake word confidence threshold) avoids false wake-ups and improves the robot's anti-interference capability.

[0035] Preferably, when performing high-priority semantic interruption in the working state, it further includes resource release and task switching, specifically including: Obtain the speaker's voiceprint features corresponding to high-priority semantics and verify whether they are the current session object or other legitimate users; Release the computing resources and voice output channels currently occupied by the interrupted task; The speech recognition text with high-priority semantics is sent to the large model intelligent agent module, which calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response first through speech synthesis.

[0036] It should be noted that in this embodiment, the step of verifying the speaker's identity enhances the robot's security, ensuring that only high-priority requests from legitimate users can interrupt the current task, thus preventing malicious interference or accidental interruptions. Furthermore, releasing the computing resources and voice output channels occupied by the interrupted task allows robot resources to be quickly reallocated to handle high-priority requests, improving resource utilization efficiency and response speed. For example, when the robot is answering a low-priority casual question and suddenly detects a user's high-priority request for help, the robot will immediately stop its current answer, release the voice channel and model inference resources, and prioritize processing the request for help.

[0037] Example 2 See Figure 1 This embodiment provides a human-like AI robot intelligent dialogue system, applied to AI robots, including: The hardware layer includes a microphone array module and a main control computing unit. The microphone array module is used to collect voice signals from the environment, and the main control computing unit is used to run various software modules. The perception layer, deployed in the main control computing unit, includes a voice activity detection module, an automatic speech recognition module, and a speaker recognition module; the voice activity detection module is used to detect human voice segments from the audio stream acquired by the microphone array module; the automatic speech recognition module is used to convert the detected voice segments into text information; and the speaker recognition module is used to extract the speaker's voiceprint features. The processing layer, deployed in the main control computing unit, includes a semantic discrimination module and a speaker separation module. The semantic discrimination module calls a local lightweight large language model, loads corresponding discrimination prompts according to the current system running state, and performs semantic validity or priority discrimination on the text information. The speaker separation module separates the speech of the target speaker from the mixed speech based on the voiceprint features extracted by the speaker recognition module. The decision-making layer, deployed in the main control computing unit, includes a state management module, a session management module, a priority scheduling module, and a queue management module. The state management module maintains the system's idle and working states. The session management module creates a session and records the current speaker's voiceprint features when a session is triggered. The priority scheduling module determines the task processing order based on the semantic discrimination module's discrimination result. The queue management module maintains the queue to be processed and supports a speaker-based request-to-queue mechanism. The interaction layer, deployed in the main control computing unit, includes a large model intelligent agent module and a speech synthesis module; the large model intelligent agent module is used to receive filtered user voice text, call the cloud-based large language model for intent understanding and response generation; the speech synthesis module is used to convert the generated text response into a speech signal for output. When the state management module controls the AI ​​robot to operate in the idle state, the perception layer continuously detects environmental speech signals and generates first text information. The semantic discrimination module calls a local lightweight large language model and combines the idle state discrimination prompt words to perform semantic validity discrimination on the first text information. If the discrimination result is valid, the session management module triggers session initiation, extracts the voiceprint features of the current speaker to establish a session, and the state management module switches the AI ​​robot's state from the idle state to the working state. When the state management module controls the AI ​​robot to operate in the working state, the perception layer continuously detects environmental speech signals and generates second text information. The semantic discrimination module calls the lightweight large language model and combines the working state discrimination prompt words to perform semantic priority discrimination on the second text information. If the discrimination result is high priority, the priority scheduling module immediately interrupts the current task, sends the second text information to the large model intelligent agent module for response generation and output. If the discrimination result is low priority, the queue management module inserts the second text information and its corresponding speaker information into the waiting queue. After the current high priority task is processed, the priority scheduling module retrieves the second text information from the queue and sends it to the large model intelligent agent module for response generation and output.

[0038] It should be noted that in this embodiment, the system is arranged from bottom to top as a hardware layer, a perception layer, a processing layer, a decision layer, and an interaction layer, with clear responsibilities and division of labor for each layer. The hardware layer is responsible for the acquisition of voice signals and basic computing support, providing high-quality audio input for upper-layer processing; the perception layer completes the conversion from raw audio to text information and voiceprint features, transforming physical signals into data that can be processed by algorithms; the processing layer focuses on semantic understanding and speaker separation, and is the core of the system's intelligent discrimination; the decision layer is responsible for state management and task scheduling, and is the control center of the system; the interaction layer finally completes response generation and voice output, realizing natural dialogue with the user. This layered design allows each module to be optimized and iterated independently. For example, the perception layer can be upgraded to a more advanced speech recognition model without affecting other layers, and the processing layer can be replaced with a lightweight model of different scales without changing the overall architecture, greatly improving the maintainability and scalability of the system.

[0039] Furthermore, the dual-state management mechanism achieves a human-like interactive behavior pattern. By maintaining idle and working states through the state management module, the system can simulate human conversational habits in different states. In the idle state, the system continuously listens to ambient speech and performs semantic validity judgment, allowing the user to naturally initiate a conversation without repeatedly using a wake word, thus solving the cumbersome interaction problem caused by traditional forced wake-up technology. In the working state, the system intelligently decides whether to respond immediately, queue, or ignore new requests based on semantic priority judgment results, reflecting the human behavior of "selectively listening" when busy. When a high-priority request arrives, the priority scheduling module can immediately interrupt the current task to release resources and prioritize urgent matters; when a low-priority request arrives, the queue management module inserts it into the waiting queue to ensure that no user requests are lost. This state-aware dynamic management mechanism in this embodiment makes the robot's behavior more flexible, natural, and reliable in complex environments.

[0040] Furthermore, in this embodiment, the dual-model architecture balances real-time performance and intelligence. The system employs a local lightweight large language model in the semantic discrimination module, while calling a cloud-based large language model in the large model intelligent agent module of the interaction layer, forming a unique architecture of dual-model collaborative work. The lightweight model, deployed locally, offers fast response speed, requires no network connection, and can perform semantic validity and priority judgment in real time, ensuring the system's rapid response capability in noisy environments. The large language model used for response generation can be deployed in the cloud according to scenario requirements: cloud deployment can leverage the powerful generation capabilities of ultra-large-scale models to provide richer and more intelligent dialogue content. This division of labor design in this embodiment ensures both low latency and reliability in semantic discrimination and high quality and richness in response generation, achieving an effective balance between efficiency and effectiveness.

[0041] Furthermore, in this embodiment, the speaker recognition and separation technology significantly enhances anti-interference capabilities. The system extracts voiceprint features through the speaker recognition module and separates the target speaker's voice from mixed speech through the speaker separation module. Working together, these two technologies solve the "cocktail party problem." In noisy environments with multiple people speaking simultaneously, the system can accurately lock onto the voice of the current conversation participant, filtering out background noise and irrelevant conversations from others, avoiding false wake-ups, misidentification, and dialogue thread chaos. This mechanism, combined with the speaker-based queue-jumping function of the queue management module, enables the system to maintain the dialogue context of multiple users and achieve continuous processing of the same user's request, greatly improving its practicality and reliability in public scenarios such as museums, shopping malls, and libraries.

[0042] Furthermore, in this embodiment, queue management and priority scheduling achieve refined task processing. The queue management module maintains a queue of pending tasks and supports a speaker-based request queue-jumping mechanism, while the priority scheduling module dynamically adjusts the task processing order based on semantic discrimination results. Together, they constitute the intelligent task scheduling hub of the system. High-priority requests can immediately interrupt the current task to obtain an immediate response, reflecting the principle of prioritizing urgent matters; low-priority requests are queued in order, ensuring that no user request is missed; multiple requests from the same speaker can be processed continuously, maintaining the continuity of multi-turn dialogues. This refined scheduling mechanism enables the system to handle scenarios with concurrent requests from multiple users with ease, ensuring timely response to critical tasks while also taking into account the reasonable processing of non-urgent requests, significantly improving user experience and system throughput.

[0043] It should be noted that the above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention, and the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A human-like AI robot intelligent dialogue method, characterized in that, include: When the AI ​​robot is in an idle state, it continuously detects voice signals in the environment, performs voice recognition on the detected environmental voice signals in the idle state, and obtains the first text information. The semantic discrimination module calls a local lightweight large language model and combines it with preset idle state discrimination prompts to perform semantic validity discrimination on the first text information; If the judgment result is valid, the session is initiated, the voiceprint features of the current speaker are extracted to establish the session, and the AI ​​robot's state is switched from idle state to working state. When the AI ​​robot is in working mode, it continuously detects voice signals in the environment, performs voice recognition on the detected environmental voice signals in the working mode, and obtains second text information. The semantic discrimination module calls the lightweight large language model and combines preset working state discrimination prompts to determine the semantic priority of the second text information. If the discrimination result is high priority and higher than the current task priority, the current task is immediately interrupted, and the second text information is sent to the large model intelligent agent module. The large model intelligent agent module calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response through speech synthesis. If the discrimination result is low priority, the second text information and its corresponding speaker information are inserted into the waiting queue. After the current high-priority task is completed, the second text information is retrieved from the queue and sent to the large model intelligent agent module. The large model intelligent agent module calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response through speech synthesis.

2. The method according to claim 1, characterized in that, The step of calling a local lightweight large language model by the semantic discrimination module and combining it with preset idle state discrimination prompts to perform semantic validity discrimination on the first text information specifically includes: Load the idle state discrimination prompt, which includes typical feature examples of valid semantics and typical feature examples of invalid speech; The first text information is combined with the idle state discrimination prompt word and then input into the lightweight large language model. The model is required to output a valid or invalid binary structure discrimination result. If the output result is valid, the semantics of the first text information are determined to be valid, and the session initiation process is executed; if the output result is invalid, the current speech recognition result is discarded, and idle state monitoring continues.

3. The method according to claim 1, characterized in that, The step of the semantic discrimination module calling the lightweight large language model and combining preset working state discrimination prompts to perform semantic priority discrimination on the second text information specifically includes: Load the working state discrimination prompt words, which predefine typical categories of high-priority semantics and low-priority semantics using a few-sample example approach; The second text information is combined with the working state discrimination prompt word and then input into the lightweight large language model. The model is required to output a valid high priority, a valid low priority, or an invalid ternary structured discrimination result. If the output result is valid and of high priority, the interruption mechanism of the current task is triggered; if the output result is valid and of low priority, it is transferred to the queue processing flow; if the output result is invalid, the current speech recognition result is discarded and the working state monitoring continues.

4. The method according to claim 1, characterized in that, After the session is initiated, speaker separation processing is also included, specifically: Extract and register the voiceprint features of the current session initiator as a reference template; During the session, voiceprint features are extracted for each subsequently received speech segment and compared with the reference template; Only voiceprints that pass the voiceprint comparison are sent into the speech recognition and subsequent processing flow, while speech interference from speakers who are not currently in the conversation is blocked.

5. The method according to claim 4, characterized in that, After inserting the second text information and its corresponding speaker information into the queue to be processed, the process also includes speaker-based queue management, specifically including: Extract the speaker's voiceprint features for the current low-priority request; Traverse the queue of pending requests and determine whether there are any unprocessed requests from the same speaker in the queue. If other requests from the same speaker exist, the current request is inserted after the last request from that speaker to ensure continuous processing of requests from the same speaker; otherwise, the current request is appended to the end of the queue.

6. The method according to claim 4, characterized in that, If a user-initiated wake-up signal is detected during the operating state, a wake-up strong interruption process is executed as a parallel approach to voice processing during the operating state, specifically including: Immediately interrupt the currently executing task, including stopping voice output and canceling ongoing large model inference requests; Extract the speaker's voiceprint features corresponding to the wake-up voice and compare them with the voiceprint template of the current session; If the voiceprint matches, the current session will continue to process the voice after wake-up; if it does not match, a new session will be created, and the original session will be suspended or destroyed.

7. The method according to claim 4, characterized in that, When the queue to be processed is empty in the working state, a state transition process is executed, specifically including: Monitor the current task processing status to determine whether all tasks have been completed and the queue is empty; If the state transition conditions are met, the AI ​​robot's state will be switched from the working state back to the idle state. In idle state, reset the context information of the current session, release the voiceprint template and session identifier resources, and prepare to receive new interaction requests.

8. The method according to claim 4, characterized in that, If a user-initiated wake-up signal is detected during the idle state, the wake-up to start a session procedure is executed. This is a parallel approach to triggering session initiation during the idle state, specifically including: After detecting a preset voice wake-up word or touch wake-up signal, confirm the validity of the wake-up signal; Extract the speaker's voiceprint features corresponding to the wake-up signal, create a new session, and assign a session identifier; The voice recognition text after wake-up is sent to the large model intelligent agent module, which calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response through speech synthesis. At the same time, the AI ​​robot's state is switched from idle state to working state.

9. The method according to claim 4, characterized in that, When performing high-priority semantic interruption in the working state, it also includes resource release and task switching, specifically including: Obtain the speaker's voiceprint features corresponding to high-priority semantics and verify whether they are the current session object or other legitimate users; Release the computing resources and voice output channels currently occupied by the interrupted task; The speech recognition text with high-priority semantics is sent to the large model intelligent agent module, which calls the cloud-based large language model to perform intent understanding and response generation, and outputs the response first through speech synthesis.

10. A human-like AI robot intelligent dialogue system, applied to AI robots, characterized in that, include: The hardware layer includes a microphone array module and a main control computing unit. The microphone array module is used to collect voice signals from the environment, and the main control computing unit is used to run various software modules. The perception layer, deployed in the main control computing unit, includes a voice activity detection module, an automatic speech recognition module, and a speaker recognition module; the voice activity detection module is used to detect human voice segments from the audio stream acquired by the microphone array module; the automatic speech recognition module is used to convert the detected voice segments into text information; and the speaker recognition module is used to extract the speaker's voiceprint features. The processing layer, deployed in the main control computing unit, includes a semantic discrimination module and a speaker separation module. The semantic discrimination module calls a local lightweight large language model, loads corresponding discrimination prompts according to the current system running state, and performs semantic validity or priority discrimination on the text information. The speaker separation module separates the speech of the target speaker from the mixed speech based on the voiceprint features extracted by the speaker recognition module. The decision-making layer, deployed in the main control computing unit, includes a state management module, a session management module, a priority scheduling module, and a queue management module; The state management module is used to maintain the idle and working states of the system; the session management module is used to create a session and record the voiceprint characteristics of the current speaker when a session is triggered. The priority scheduling module determines the task processing order based on the discrimination result of the semantic discrimination module; The queue management module is used to maintain the queue to be processed and supports a speaker-based request queue-jumping mechanism; The interaction layer, deployed in the main control computing unit, includes a large model intelligent agent module and a speech synthesis module; the large model intelligent agent module is used to receive filtered user voice text, call the cloud-based large language model for intent understanding and response generation; the speech synthesis module is used to convert the generated text response into a speech signal for output. When the state management module controls the AI ​​robot to operate in the idle state, the perception layer continuously detects environmental speech signals and generates first text information. The semantic discrimination module calls a local lightweight large language model and combines the idle state discrimination prompt words to perform semantic validity discrimination on the first text information. If the discrimination result is valid, the session management module triggers session initiation, extracts the voiceprint features of the current speaker to establish a session, and the state management module switches the AI ​​robot's state from the idle state to the working state. When the state management module controls the AI ​​robot to operate in the working state, the perception layer continuously detects environmental speech signals and generates second text information. The semantic discrimination module calls the lightweight large language model and combines the working state discrimination prompt words to perform semantic priority discrimination on the second text information. If the discrimination result is high priority, the priority scheduling module immediately interrupts the current task, sends the second text information to the large model intelligent agent module for response generation and output. If the discrimination result is low priority, the queue management module inserts the second text information and its corresponding speaker information into the waiting queue. After the current high priority task is processed, the priority scheduling module retrieves the second text information from the queue and sends it to the large model intelligent agent module for response generation and output.

Images