Conference room speech intent recognition method and system based on scenario corpus

Abstract

The present application relates to the technical field of intelligent voice interaction and recognition, and particularly relates to a conference room voice intention recognition method and system based on scene corpus, comprising: providing a data basis for multi-dimensional intention analysis by constructing a multi-modal structured conversation flow, while supporting multi-person voice separation; effectively avoiding single dimension misjudgment by concurrent analysis in three dimensions of semantics, acoustics and device cooperation and setting double-dimensional trigger threshold; solving the decision ambiguity problem when multiple intentions coexist by re-screening candidate intentions with the help of historical association patterns in conflict coupling graph, locking the unique dominant intention or generating a clarification anchor; giving consideration to response efficiency and safety by dynamic adaptation of three-level strategy of high-confidence direct execution, clarification dialogue and conservative safe operation, avoiding the risk of misoperation in uncertain scenarios; continuously optimizing with the running of the conflict coupling graph, adapting to different conference room environment changes, and gradually improving long-term accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent voice interaction and recognition technology, and in particular to a method and system for recognizing voice intent in conference rooms based on scene corpus. Background Technology

[0002] Intelligent conference room systems widely employ voice interaction technology. They use pre-set microphone arrays to collect participants' voice commands, which are then translated into control operations for equipment such as projectors, lights, and air conditioning through automatic speech recognition and natural language understanding. Existing voice intent recognition methods typically rely on a single semantic understanding model, supplemented by sound source localization and noise reduction, enabling basic equipment operation in a single-person, clear command environment.

[0003] However, in real-world multi-person conference scenarios, complex situations arise such as multiple participants speaking freely, overlapping speech, contextual jumps, and collaborative control of multiple devices, posing fundamental limitations to existing methods. On one hand, unexpected jumps or logical loops in intent association during multi-turn dialogues can lead to semantic conflicts. On the other hand, the acoustic environment within the conference room changes due to personnel movement, equipment noise, and reverberation variations, interfering with sound source localization and acoustic feature extraction, causing the acoustic field model to deviate from its historical baseline. Simultaneously, continuous control commands involving multiple devices, without verification of the timing tolerance window and state interlocking between devices, will result in collaborative deviations. These semantic and logical conflicts, acoustic field anomalies, and multi-device collaborative deviations often occur concurrently, and existing methods lack a joint detection and dynamic intent confirmation mechanism for multi-dimensional conflicts. This leads to the system either executing incorrectly, frequently initiating invalid clarifications, or directly refusing to respond when intent is ambiguous, severely restricting the accuracy and efficiency of conference room voice interaction. Therefore, this invention proposes a conference room voice intent recognition method and system based on scenario-specific corpora. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the background art, and to propose a method and system for recognizing speech intent in conference rooms based on scene corpus.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a method for recognizing speech intent in a conference room based on a contextual corpus, comprising: S1. Collect and parse the original voice stream of users in the conference room to generate a structured conversation stream containing semantic text, acoustic features and sound source localization information; S2. Perform multi-dimensional intent conflict judgment on structured conversation streams: concurrently analyze semantic logic conflicts, acoustic field anomalies, and multi-device coordination deviations in structured conversation streams. When at least two dimensions simultaneously indicate the existence of intent understanding risks, it is determined that there are potential intent conflicts in the current conversation stream. S3. When a potential intent conflict is determined, the intent anchor points of the current conversation flow are re-filtered based on the pre-constructed conflict coupling graph to obtain the re-filtering results. The conflict coupling graph reflects the historical association patterns between different conflict dimensions, and the re-filtering is used to determine the unique dominant intent from multiple candidate intents or to generate multi-turn dialogue anchor points that need to be clarified. S4. Based on the re-screening results, dynamically adapt the intent response strategy, which includes at least high-confidence direct execution, initiating targeted clarification dialogue, or performing conservative security operations. S5. Update the conflict coupling graph based on the execution feedback of the intent response strategy.

[0006] Further, S1 includes: Raw audio streams containing multiple audio signals are captured using a microphone array deployed in the conference room; Speech activity detection is performed on the original speech stream to separate human voices from non-human voices, and the separated human voice signals are then subjected to noise reduction and echo cancellation processing to obtain effective human voice segments. Automatic speech recognition is performed on valid human voice segments to generate corresponding semantic text sequences, and fundamental frequency, energy and speech rate are extracted from valid human voice segments as acoustic feature sets simultaneously; Based on the time difference of arrival of multiple signals in the microphone array, the location coordinates of the sound source of the effective human voice segment in the conference room space are obtained; The semantic text sequence, acoustic feature set, and sound source location coordinates are bound to the same time window to form a timestamped structured conversation stream unit, and multiple consecutive structured conversation stream units are organized into a structured conversation stream in chronological order.

[0007] Furthermore, the semantic logic conflict in the analysis in S2 includes: Based on a pre-defined knowledge graph of the meeting room scenario, the semantic text sequence in the structured conversation stream is parsed to identify the entities, actions, and contextual relationships involved, and the parsing results are obtained. Using the parsing results as input, calculate the correlation between the semantic elements of the current statement and the historical successful interaction patterns stored in the knowledge graph of the conference room scenario, and track the penetration and decay path of this correlation in consecutive rounds of dialogue; When an unexpected jump in the correlation degree is detected at a preset key logical node or a logical loop appears in the penetration path, it is determined that there is a semantic logical conflict, and the location of the key logical node where the conflict occurs is recorded.

[0008] Furthermore, the analyzed acoustic field anomalies in S2 include: Based on the sound source location coordinate sequence in the structured conversation flow, combined with the preset three-dimensional spatial model of the conference room, the propagation of sound waves is dynamically simulated to generate a real-time acoustic field model that reflects the spatial distribution and attenuation of sound energy. Pattern matching is performed between the real-time acoustic field model and a baseline acoustic field feature library built from historical normal conference audio. If the matching results show that there are unexplainable acoustic variations in the current acoustic field that cannot be explained by known noise sources or normal speech activity, then it is determined that there is an acoustic field anomaly.

[0009] Furthermore, the method of dynamically simulating sound wave propagation by combining a pre-set three-dimensional spatial model of the conference room to generate a real-time acoustic field model reflecting the spatial distribution and attenuation of sound energy includes: Using the sound source location coordinates as the starting point of the sound source, and based on the geometric structure and acoustic material properties of the three-dimensional spatial model of the conference room, the set of propagation paths of the sound wave from this starting point to various directions of the conference room space is obtained. Using the set of propagation paths as input, and based on the sound wave propagation equation, the physical effects of energy attenuation, reflection, diffraction, and interference of sound waves on each path are simulated sequentially to obtain the simulation results of the sound wave state on each path. By combining the simulation results of sound wave states along all propagation paths, a three-dimensional distribution field representing the sound pressure level or sound energy density at each point in the conference room space is generated as a real-time acoustic field model.

[0010] Furthermore, the analysis of multi-device collaborative deviation in S2 includes: Parse the semantic text sequence in the structured session stream, identify the set of devices to be controlled and the timing of operation instructions, and construct the expected device cooperation network graph of the current session; From the IoT platform of the conference room, retrieve the dynamic collaborative baseband formed by the device set in historical collaborative tasks. The dynamic collaborative baseband defines the timing tolerance window and state interlock relationship of the action response between devices. The expected device collaboration network graph is overlaid on the dynamic collaboration baseband to detect whether there are any device nodes or action edges that deviate from the baseband range, and all deviated nodes and edges are recorded to form a deviation set. If a non-empty deviation set is detected, the analysis will determine whether the elements in the deviation set will trigger a chain reaction or form logical islands in the expected device collaboration network graph, thereby determining whether there is a multi-device collaboration deviation.

[0011] Further, S3 includes: Extract the identifiers of each specific conflict dimension that triggers the determination of potential intention conflicts, and their corresponding quantitative evaluation results; The conflict dimension identifier and the quantitative evaluation result are used as query vectors and input into the conflict coupling graph for matching query; the conflict coupling graph stores the true intentions and dialogue clarification patterns that correspond to different combinations of conflict dimensions in history in a graph structure. Based on the matching query results, calculate the conflict resolution confidence of each candidate intent in the current structured session flow, where the conflict resolution confidence reflects the historical probability that the candidate intent is the correct intent under the combination of conflict dimensions; If the confidence level of conflict resolution for any candidate intent is higher than the preset first intent analysis threshold, then that candidate intent is determined as the unique dominant intent. If the confidence level of conflict resolution for all candidate intents is lower than the preset first intent analysis threshold but higher than the preset second intent analysis threshold, then based on the clarification patterns associated in the conflict coupling graph, multi-turn dialogue anchors to be initiated to the user are generated.

[0012] Furthermore, the construction process of the conflict coupling map includes: During initialization or early operation, historical session data is collected, and the combination of conflict dimensions determined in each session and the true intent of the final verification are extracted. The association rules between the combination of conflict dimensions and the true intent are analyzed, and the association relationship with the confidence of conflict resolution higher than the preset rule is initialized as a conflict coupling graph in the form of nodes and weighted edges. The nodes represent conflict dimensions or the final intent, the edges represent their co-occurrence relationship, and the weights represent the co-occurrence strength. After each S1 to S4 identification process is completed, record the specific combination of conflict dimensions output by S2 in this process, the re-screening process and results of S3, and the true intent finally confirmed by the user or verified by the system. Based on the actual execution effect feedback of the intent response strategy finally adopted in this identification process, the weights of the relevant edges in the conflict coupling graph are dynamically adjusted: if the execution is successful, the weight of the corresponding edge is increased; if the execution fails or triggers user correction, the weight of the corresponding edge is decreased.

[0013] Furthermore, the dynamic adaptation intent response strategy in S4 includes: S41. Determine the type of re-filtered result: If the type of re-filtered result is a single dominant intent, proceed to S42; if the type of re-filtered result is a multi-turn dialogue anchor, proceed to S43. S42. Implement a high-confidence direct execution strategy: Transform the unique dominant intent into structured device control commands and send them to the conference room IoT platform for execution; S43. Execute a targeted clarification dialogue strategy: Based on the multi-round dialogue anchor points, retrieve the corresponding interaction template from the predefined clarification script template library, and initiate an interactive clarification inquiry to the user through the speech synthesis module; S44. If S3 does not output a clear result or the confidence level of conflict resolution for all candidate intentions is lower than the safety threshold, then a conservative safety operation strategy is executed: under this strategy, only the default safety operation corresponding to the current conference room scenario is executed, or only the receipt of the instruction is confirmed through voice feedback without any device control action is executed.

[0014] A second aspect of the present invention provides a conference room speech intent recognition system based on scene corpus, comprising: The speech stream acquisition, parsing, and structured module acquires and parses the raw speech streams of users in the conference room, generating a structured conversation stream that includes semantic text, acoustic features, and sound source location information. Multi-dimensional Intent Conflict Detection and Judgment Module: Performs multi-dimensional intent conflict judgment on structured conversation streams: Concurrently analyzes semantic logic conflicts, acoustic field anomalies, and multi-device coordination deviations in structured conversation streams. When at least two dimensions simultaneously indicate the existence of intent understanding risks, the current conversation stream is judged to have potential intent conflicts. Conversation Flow Intent Anchor Filtering Module: When a potential intent conflict is determined, the intent anchors of the current conversation flow are re-filtered based on a pre-built conflict coupling graph to obtain the re-filtering results. The conflict coupling graph reflects the historical association patterns between different conflict dimensions, and the re-filtering is used to determine the unique dominant intent from multiple candidate intents or to generate multi-turn dialogue anchors that need to be clarified. Dynamic Adaptation Module for Intent Response Strategy: Based on the re-filtering results, dynamically adapt the intent response strategy, which includes at least high-confidence direct execution, initiating a targeted clarification dialogue, or performing a conservative security operation.

[0015] Compared with existing technologies, the beneficial effects of the meeting room speech intent recognition method and system based on scene corpus provided by this invention are as follows: 1) By synchronously collecting semantic text, acoustic features, and sound source localization information through a microphone array, a multimodal structured conversation stream is constructed, providing a unified data foundation for subsequent multi-dimensional analysis. At the same time, sound source localization can distinguish the voice input of different participants. By introducing three dimensions of concurrent analysis—semantic logic, acoustic field anomalies, and multi-device collaboration—and setting a judgment threshold that indicates at least two dimensions simultaneously, conflicts and false triggers caused by misjudgments of a single dimension are effectively avoided, improving the accuracy of intent conflict detection and reducing the false alarm rate in complex meeting scenarios. 2) By leveraging historical association patterns in the conflict coupling graph, anchor points are re-filtered for multiple candidate intentions in conflict states, resolving the issue of decision ambiguity when multiple intentions coexist. This allows for the identification of a single dominant intention or the generation of targeted clarification anchor points from multiple candidate intentions, avoiding the output of erroneous instructions in conflict scenarios. Through dynamic adaptation of a three-tiered strategy—high-confidence direct execution, initiating clarification dialogue, and conservative security operation—response efficiency is ensured in high-deterministic scenarios, proactive clarification is provided to avoid misoperation in uncertain scenarios, and a safety net is provided in low-confidence scenarios, balancing user experience and system security. The conflict coupling graph is continuously optimized with actual operation, adapting to changes in different meeting room environments and user habits. Over long-term operation, the accuracy of intention conflict discrimination and the adaptability of response strategies gradually improve. Attached Figure Description

[0016] Figure 1 This is a flowchart of the conference room speech intent recognition method based on scene corpus proposed in this invention.

[0017] Figure 2 This is a block diagram of the conference room speech intent recognition system based on scene corpus proposed in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only 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 are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a method for recognizing speech intent in a conference room based on a contextual corpus, comprising: S1. Collect and parse the original voice stream of users in the conference room to generate a structured conversation stream containing semantic text, acoustic features and sound source localization information; S2. Perform multi-dimensional intent conflict judgment on structured conversation streams: concurrently analyze semantic logic conflicts, acoustic field anomalies, and multi-device coordination deviations in structured conversation streams. When at least two dimensions simultaneously indicate the existence of intent understanding risks, it is determined that there are potential intent conflicts in the current conversation stream. S3. When a potential intent conflict is determined, the intent anchor points of the current session flow are re-filtered based on the pre-built conflict coupling graph to obtain the re-filtering results. In the initialization or early operation phase, historical session data is collected, and the combination of conflict dimensions determined in each session and the true intent of the final verification are extracted. The association rules between the combination of conflict dimensions and the true intent are analyzed, and the association relationship with the confidence of conflict resolution higher than the preset rule is initialized as a conflict coupling graph in the form of nodes and weighted edges. The nodes represent conflict dimensions or the final intent, the edges represent their co-occurrence relationship, and the weights represent the co-occurrence strength. Specifically, the process involves loading anonymized historical meeting session data from the storage system; for each complete session record, extracting the output conflict dimension determination result: the specific activated conflict dimension and its quantified value, which serves as the condition part; simultaneously, extracting the true intent of the session confirmed by user feedback or marked by the administrator, which serves as the conclusion part; applying an association rule mining algorithm to calculate the support and conflict resolution confidence of each condition part to the conclusion part; pre-setting a minimum confidence rule, for example, 0.7; filtering out all association rules that meet the conditions; in the conflict coupling graph, creating a conflict pattern node for each unique combination of conflict dimensions and an intent node for each true intent; for each rule, if its conflict resolution confidence is higher than 0.7, creating a weighted edge between its corresponding conflict pattern node and intent node, with the initial weight of the edge equal to the conflict resolution confidence of the rule; simultaneously, creating a clarification pattern node for the session segment describing the clarification process, and connecting it to the relevant conflict pattern node with an edge; thus, forming an initial conflict coupling graph containing three types of nodes—conflict pattern, intent, and clarification pattern—and the weighted edges between them; S4. Based on the re-screening results, dynamically adapt the intent response strategy, which includes at least high-confidence direct execution, initiating targeted clarification dialogue, or performing conservative security operations. S5. Update the conflict coupling graph based on the execution feedback of the intent response strategy; Understandably, after each identification process from S1 to S4 is completed, the specific combination of conflicting dimensions output by S2 in this process, the re-screening process and results of S3, and the true intent finally confirmed by the user or verified by the system are recorded. Based on the actual execution effect feedback of the intent response strategy finally adopted in this identification process, the weights of the relevant edges in the conflict coupling graph are dynamically adjusted: if the execution is successful, the weight of the corresponding edge is increased; if the execution fails or triggers user correction, the weight of the corresponding edge is decreased. After each execution of the intent response strategy in S4, continuously monitor the effect feedback from the user or environment. Successful effect feedback includes: the device action is completed without error and no user immediately interrupts or corrects it, or the user gives positive confirmation after the clarification dialogue. Failed effect feedback includes: the user immediately issues a cancellation or correction command, or the user expresses denial and rephrases the intent after the clarification dialogue. Locate the conflict coupling graph edge on which the S3 step in this process is based, that is, the edge pointing from the conflict mode node matched at that time to the intent that is finally selected as the sole dominant intent or initiates the clarification mode. If the feedback is successful, the weight of that edge is increased. The weight is increased incrementally: the new weight equals the old weight plus a positive learning rate parameter multiplied by a success reward value. If the feedback is unsuccessful, the weight of that edge is decreased: the new weight equals the old weight minus a positive learning rate parameter multiplied by a failure penalty value, ensuring that the weight is not lower than zero. The learning rate parameter is set after balancing convergence speed and stability through offline simulation, and the example value is 0.1. The example success reward value is 1.0, and the example failure penalty value is 0.8. In addition, if the current attempt fails but a new true intent is eventually confirmed, it is checked whether there is an edge between the new intent node and the current conflict mode node. If it exists, it is increased; if it does not exist, a new edge is created and assigned an initial weight.

[0020] In this embodiment of the invention, the specific implementation of S1, which collects and parses the user's original voice stream in the conference room to generate a structured conversation stream containing semantic text, acoustic features, and sound source localization information, is as follows: S11. Acquire raw audio streams containing multiple audio signals using a microphone array deployed in the conference room, including: A microphone array consisting of multiple omnidirectional or directional microphone units is deployed at a predetermined location in the conference room. After startup, all microphone units in the array synchronously enter the sound pickup state, continuously collecting sound wave vibrations in the environment and converting them into independent analog electrical signals. Each analog electrical signal is quantized into a digital audio stream according to a preset sampling rate and bit depth via its respective channel's analog-to-digital converter. All channels' digital audio streams are marked with a unified timestamp at the acquisition end and synchronously transmitted to the processing unit through a multi-channel audio interface. The processing unit aligns and encapsulates the received digital audio streams according to the channel number and timestamp, forming a raw voice stream data packet containing multiple synchronous audio signals, providing a multi-view, spatially redundant acoustic information foundation for subsequent processing.

[0021] S12. Perform speech activity detection on the original speech stream to separate human voices from non-human voices, and then perform noise reduction and echo cancellation processing on the separated human voice signals to obtain effective human voice segments, specifically: For each synchronized audio signal of the original speech stream, calculate its short-time energy and zero-crossing rate; preliminarily mark continuous audio frames whose short-time energy exceeds the preset energy threshold and whose zero-crossing rate is within the human voice frequency range as human voice candidate segments; conversely, mark continuous frames whose energy is below the energy threshold or whose zero-crossing rate characteristics do not match as non-human voice silence segments. A speech endpoint detection unit is applied to the human voice candidate segments. This unit outputs the posterior probability of speech presence frame by frame. By smoothing the decision through Viterbi decoding, the start and end points of each voice event are determined, achieving fine separation of human voice and non-human voice. For the separated human voice signal, an adaptive beamforming algorithm is adopted. This algorithm uses a microphone array to receive signals and dynamically adjusts the complex weighting coefficients of each channel according to the real-time estimated sound source location. This maximizes the response amplitude of the array in the direction of the desired sound source and forms concave null points in other interference directions, thereby enhancing the target speech and suppressing environmental noise. The beamformed human voice signal and the reference signal are fed into the echo canceller together. The reference signal is the playback signal directly derived from the output port of the conference room speaker, representing the far-end audio content currently being played by the speaker. The echo canceller uses an adaptive filter to identify and simulate the echo path from the speaker to the microphone, subtracts the estimated echo component from the human voice signal, and finally outputs a clean, echo-free effective human voice segment.

[0022] S13. The method for automatically recognizing valid human voice segments, generating corresponding semantic text sequences, and simultaneously extracting fundamental frequency, energy, and speech rate from valid human voice segments as an acoustic feature set is as follows: For each valid human voice segment, two parallel processing pipelines are simultaneously launched. In the speech recognition pipeline, acoustic features are extracted from the valid human voice segments, and Mel-frequency cepstral coefficients or filter bank features are calculated, resulting in a feature vector for each frame. The feature vectors from each frame are concatenated in chronological order to form a feature vector sequence. This feature vector sequence serves as input to the acoustic model and is fed into an end-to-end automatic speech recognition engine composed of acoustic units, language units, and a dictionary. The acoustic unit is responsible for establishing the mapping relationship between feature vectors and phonemes or words, while the language unit provides the prior probability of the word sequence. The most probable word sequence is obtained through beam search decoding. The beam search retains the current cumulative value at each step. The candidate hypotheses with the highest probabilities are continuously expanded based on the prior probabilities of language units until a sentence end marker is encountered. The final output is the complete word sequence with the highest probability as the semantic text sequence. In the acoustic feature extraction pipeline, based on the time-domain waveform of the effective human voice segment, the fundamental frequency value of each frame is calculated using the autocorrelation function method or the cepstral method to obtain the fundamental frequency trajectory. Simultaneously, the mean of the sum of squares of the signal in each frame is calculated and the logarithm is taken to obtain the short-time energy trajectory. The speech rate is obtained by identifying syllable or word boundaries in the text sequence and calculating the number of syllables or words per unit time. Finally, the fundamental frequency trajectory, energy trajectory, and speech rate value together constitute the acoustic feature set bound to the human voice segment.

[0023] S14. The steps for obtaining the location coordinates of the sound source of the effective human voice segment within the conference room space based on the arrival time difference of multiple signals in the microphone array are as follows: For a specific valid human voice segment, at least three pairs of non-collinear microphone units with known spatial positions are selected from the microphone array. For each pair of microphone units, two signals are extracted during the valid human voice segment, and the generalized cross-correlation function of the two signals is calculated using one of them as a reference. In the generalized cross-correlation function curve, the time index corresponding to the highest peak is the time difference of the signal arriving at the pair of microphones. The subsampling accuracy can be obtained through quadratic interpolation. The arrival time differences of at least three non-coplanar signals are converted into distance differences. A hyperboloid equation system is established with the corresponding microphone positions as the focus, and linearized least squares iteration is used. The method directly solves for the spatial coordinates of the sound source. During the iteration process, the Euclidean distance between the sound source positions obtained from two adjacent iterations is calculated as the residual change. When the residual change is lower than the preset convergence threshold, the iteration is terminated and the final sound source coordinates are output. The convergence threshold is set based on the minimum spatial resolution required to distinguish the speaking positions of adjacent participants in a conference room scenario: the distance between adjacent seats is usually about 0.8 meters to 1.2 meters. In order to ensure that the intent attribution analysis can accurately distinguish the speaker, the positioning error is required to be less than half of this distance. Therefore, the convergence threshold is set to 0.1 meters. The example value is 0.1 meters.

[0024] S15. Bind the semantic text sequence, acoustic feature set, and sound source location coordinates to the same time window to form a timestamped structured conversation stream unit, and organize multiple consecutive structured conversation stream units into a structured conversation stream in chronological order.

[0025] In this embodiment of the invention, the detailed implementation steps of S2 for multi-dimensional intent conflict judgment of the structured session stream, including concurrent analysis of semantic logic conflicts, acoustic field anomalies, and multi-device coordination deviations in the structured session stream, include: S21. Based on the pre-defined knowledge graph of the conference room scenario, parse the semantic text sequence in the structured conversation flow, identify the entities, actions and contextual relationships involved, and obtain the parsing results; Specifically, the core conceptual ontology related to conference room operation is defined, including five types of nodes: equipment entities, spatial areas, personnel roles, operational actions, and scene states. Equipment entity nodes include projectors, lights, air conditioners, curtains, etc., and are associated with attributes such as model, supported functions, and control interfaces. Spatial area nodes include the main conference table area, whiteboard area, projection area, etc. Operational action nodes include play, pause, dim, open, etc. Scene state nodes include presentation mode, discussion mode, departure mode, etc., which are used to represent the overall operating status of the conference room. From historically successful and manually verified meeting interaction logs, five types of node instances and their relationships were extracted using natural language processing and information extraction techniques. For example, from log fragments, the following relationships were extracted: user statements containing action nodes and device entity nodes controlled by them constitute a control relationship; actions initiated in a specific spatial area constitute a spatial relationship; changes in device operating status constitute a state transition relationship. Five types of nodes and relationships are stored in a graph database in the form of a directed graph and assigned confidence scores to form an initial knowledge graph of the meeting room scenario. Based on newly confirmed meeting interaction data, the knowledge graph of the meeting room scenario is continuously updated dynamically by adding and deleting nodes and adjusting relationship weights. The current semantic text sequence in the structured conversation stream is input into a pre-defined semantic role labeling unit. The semantic role labeling unit first performs word segmentation and part-of-speech tagging on the text, and then identifies the predicate that serves as the core of the action. For each predicate, within its syntactic scope, the semantic roles associated with the predicate are identified and mapped to predefined entities, actions, and attribute tags in the knowledge graph. For example, the agent role is mapped to the operator entity, the patient role is mapped to the controlled equipment entity, and the action predicate itself is mapped to the operation action node. Simultaneously, a conversation context stack is set up to store entities and actions parsed in the past. When parsing the current statement, for pronouns or omitted components, the most matching antecedent is searched in the conversation context stack using backreference parsing technology. This antecedent is then parsed as a specific entity. Specifically, the matching score between the current pronoun and each candidate entity in the stack is calculated based on features such as gender, number, and semantic role. The entity with the highest matching score and the most recently mentioned is selected as the antecedent to complete the referential resolution, thereby establishing the contextual association between the current statement and the historical conversation. The final output of this process is a structured data object, which contains the action node of the current statement, all involved entity nodes and their semantic relationships, and contextual links with entities confirmed in the historical dialogue.

[0026] S22. Taking the parsing results as input, the specific process of calculating the correlation between the semantic elements of the current statement and the historical successful interaction patterns stored in the conference room scenario knowledge graph, and tracing the penetration and decay path of this correlation in consecutive multi-turn dialogues includes: The combination of action nodes and entity nodes contained in the parsing results is used as the query intent pattern. The knowledge graph of the conference room scene is searched to see if there is a subgraph of historical successful interaction patterns that exactly or partially matches the query intent pattern. The matching process adopts graph isomorphism or subgraph matching algorithm to calculate the structural similarity and semantic similarity between the query pattern and the candidate subgraph. The structural similarity is evaluated based on the degree of matching of node and edge labels and the consistency of topological connections. The semantic similarity is calculated by calculating the cosine distance of the corresponding node labels through pre-trained word vectors. The two are weighted and summed to obtain the relevance value, which represents the degree to which the current operation intent has corresponding support in historical successful cases. At the start of the session, a global trajectory is initialized for the overall intent relevance of the session. After each round of new utterance parsing, the calculated relevance value becomes a new penetration point, attempting to propagate its influence into subsequent trajectories. The penetration path is characterized by a decay function: if the relevance of subsequent utterances remains high, the penetration continues and decays slowly; if the relevance of a certain round of utterances drops to a low level, the penetration path experiences a decay breakpoint at that point. The example decay function can adopt an exponential decay form: New relevance influence = current round relevance value × exp(-α × round interval), where α is the decay coefficient, determined by fitting the development dataset, with a typical value of 0.3. By continuously tracking the trajectory of this correlation value on the conference round coordinate axis, the trajectory represents the penetration and decay path, reflecting the process of consensus building or disintegration.

[0027] S23. When an unexpected jump in the correlation degree is detected at a preset key logical node or a logical loop appears in the penetration path, it is determined that there is a semantic logical conflict, and the location of the key logical node where the conflict occurs is recorded. Specifically, a series of key logical nodes are preset in the knowledge graph of the conference room scenario. These key logical nodes correspond to decision points in the meeting process where user intent needs to be clarified, such as pre-start confirmation for device control and critical points for mode switching. The intersection of the penetration path and key logical nodes is continuously monitored. When the penetration path passes through a key logical node, the rate of change of the correlation value is checked. If the correlation value is detected to drop or rise sharply in adjacent rounds, exceeding the preset jump threshold, an unexpected jump is considered to have occurred. The jump threshold is based on the minimum value when intent is successfully switched in historical sessions. The correlation change magnitude is statistically obtained, and the example value can be set to 0.4, that is, the absolute value of the correlation change between adjacent rounds exceeds 0.4 and is regarded as an unexpected jump; at the same time, the shape of the penetration path is analyzed. If a loop pointing back and forth is found on the time axis, that is, the intent feature of the current round traces back and repeatedly points to the key logical node that has been previously confirmed, it is identified as a logical loop. If any of the above conditions are met, it is determined that there is a semantic logical conflict, and the position of the key logical node where the first jump or loop occurs on the penetration path is recorded as the key logical node position where the conflict occurs.

[0028] S24. Based on the sound source location coordinate sequence in the structured conversation flow, combined with the preset three-dimensional spatial model of the conference room, dynamically simulate sound wave propagation and generate a real-time acoustic field model that reflects the spatial distribution and attenuation of sound energy. Specifically, S241, taking the sound source location coordinates as the starting point of the sound source, and based on the geometric structure and acoustic material property parameters of the three-dimensional spatial model of the conference room, obtain the set of propagation paths of the sound wave from this starting point to various directions of the conference room space, including: A pre-built 3D spatial model of the conference room is provided. This model uses boundary representation to accurately describe the room's size, shape, and polygonal meshes of all major obstacles such as walls, ceilings, floors, conference tables, and large equipment. Simultaneously, each polygonal facet of the model is associated with its corresponding acoustic material properties, including absorption coefficient, scattering coefficient, and transmission coefficient. Using the sound source's location coordinates as the starting point, a ray tracing algorithm is employed to emit virtual sound rays from this starting point in a large number of uniformly distributed directions on a unit sphere centered at that point, ensuring uniform directional coverage. Each sound ray carries initial energy and travels along its... The sound propagates in a straight line. When a sound ray hits a polygonal facet in the three-dimensional spatial model of the conference room, a new reflected ray is calculated and generated based on the material properties of the facet. Its direction follows the specular reflection or Lambertian scattering model, and its energy is attenuated accordingly based on the sound absorption and scattering coefficients. At the same time, a transmitted ray may be generated. Each derivative ray directly emitted by the sound source or generated through reflection and transmission is tracked until its energy attenuates below a preset threshold or its propagation time exceeds the set reverberation time. All the tracked rays that start from the sound source and reach the receiving area, along with their propagation history, together constitute the propagation path set.

[0029] S242. Using the set of propagation paths as input, and based on the sound wave propagation equation, simulate the physical effects of energy attenuation, reflection, diffraction, and interference of the sound wave on each path in sequence to obtain the simulation results of the sound wave state on each path. Specifically, for each propagation path in the set of propagation paths, the corresponding sound wave state is simulated; the sound source signal is modeled as a multi-band signal covering the frequency range of human voice; for each band, the energy attenuation is calculated along the propagation path, and the attenuation consists of two parts: one part is the spherical wave spread loss proportional to the propagation distance; the other part is the air absorption loss related to the propagation distance and frequency. At each reflection intersection point of the propagation path and the scene model, the absorbed portion is subtracted from the energy of that frequency band based on the absorption coefficient of the corresponding surface at that point. For cases where the path passes through the edge of an obstacle, the diffraction effect is calculated based on geometric diffraction theory, resulting in additional energy attenuation behind the obstacle that increases with the diffraction angle. When sound waves from different propagation paths converge at the receiving point, the relative time difference and phase difference of the signals arriving from each path are considered, and the signals are superimposed in the complex domain to simulate constructive or destructive interference effects. After the above processing, the sound wave state simulation results for each propagation path, including arrival time, arrival direction, and complex sound pressure or energy attenuation in each frequency band, are obtained.

[0030] S243. By integrating the simulation results of sound wave states along all propagation paths, a three-dimensional distribution field representing the sound pressure level or sound energy density at each point in the conference room space is generated as a real-time acoustic field model. Understandably, the conference room space is discretized into a three-dimensional voxel grid, with each voxel representing a tiny cubic region within the conference room space. For each microphone position in the receiver array and each voxel center point in the voxel grid, the acoustic state simulation results for all propagation paths are iterated. For each propagation path reaching the target point, a vector superposition and summation in the complex domain is performed at the target point based on its direction of arrival and complex sound pressure, yielding the total complex sound pressure at that point. The acoustic energy density at the target point is calculated based on the ratio of the square of the total complex sound pressure to the characteristic impedance of the air. After performing this operation on all voxel grid points, the acoustic energy density value for each voxel center point is obtained. The coordinates of these voxel center points are then... Combined with its corresponding sound energy density value, a three-dimensional scalar field is formed, which is the real-time acoustic field model. Inside the model, each voxel is connected to its 26 neighboring voxels through its spatial index, forming a data hierarchy relationship of the field. When training or calibrating the model, a known sound source is used to play standard sound signals at different standard positions, and the sound pressure level at each point of the array is measured as the true value. By adjusting the material parameters such as the sound absorption coefficient in the model, the root mean square error between the simulated field and the measured field is minimized. This process is the parameter fitting step of the model. When the model is combined with the scene, the input is the sound source location coordinates and the three-dimensional spatial model of the conference room, and the output is the three-dimensional sound energy distribution data corresponding to the real physical sound field.

[0031] S25. Perform pattern matching between the real-time acoustic field model and the baseline acoustic field feature library established based on historical normal conference audio, including: The real-time acoustic field model is decomposed into features to express the acoustic energy distribution in its three-dimensional space as a set of spatial statistical feature vectors. The feature dimensions include at least: acoustic energy centroid coordinates, eigenvalues ​​of the energy distribution covariance matrix, and energy gradient histograms between regions. The benchmark acoustic field feature library is a set of feature vectors formed during the initialization phase by playing standard sound sources at different typical locations, collecting signals from various microphone arrays, constructing an acoustic field model, and extracting spatial statistical feature vectors of the same dimension. After clustering and denoising, each vector is accompanied by a corresponding meeting scene label. During pattern matching, the Mahalanobis distance between the feature vector of the real-time acoustic field model and each feature vector in the reference acoustic field feature library is calculated. A preset matching threshold is set, which is determined by analyzing the intra-class distance distribution of the feature vectors of a normal conference acoustic field. The upper quartile of the intra-class distance is taken as the threshold, and the example value is 2.5. If the minimum Mahalanobis distance is less than the preset matching threshold, the current acoustic field pattern is considered to have successfully matched a certain historical normal conference acoustic field pattern, indicating that the current sound field conforms to the norm. If the minimum Mahalanobis distance is greater than the matching threshold, it indicates that the real-time acoustic field model does not belong to any known normal conference acoustic field pattern in the feature space.

[0032] S26. If the matching results show that there are unexplainable acoustic anomalies in the current acoustic field that cannot be explained by known noise sources or normal speech activities, then it is determined that there is an acoustic field anomaly; among them, unexplainable acoustic anomalies include energy vortices, missing reflection paths, or abrupt changes in reverberation characteristics. When pattern matching shows that the minimum Mahalanobis distance exceeds the matching threshold, the operating status and location information of all registered devices in the conference room IoT platform are retrieved, and the distribution map of common environmental noise sources recorded in the historical normal conference acoustic field is queried. The real-time acoustic field model is compared with the theoretical sound field model of these known noise sources region by region. Specifically, the theoretical sound field of the matched known noise sources is differentially eliminated from the real-time acoustic field model to obtain the residual acoustic field. The residual acoustic field is analyzed to detect whether there are three types of unexplained acoustic anomalies: energy vortices, discontinuous missing sound reflection paths in the model, or abrupt changes in the reverberation decay time constant of a certain region. Among them, energy vortex refers to the local area where the sound energy flux density forms a closed loop; missing reflection path refers to the abnormally low sound energy in the model at the corresponding position of the wall where reflected waves should be superimposed; abrupt change in reverberation characteristics refers to the reverberation time of a certain area being seriously inconsistent with the inherent characteristics of the conference room. If any of the above-mentioned aberration characteristics are detected in the residual acoustic field, it is determined that there is an acoustic field anomaly.

[0033] S27. Parse the semantic text sequence in the structured session stream, identify the set of devices to be controlled and the timing of operation instructions, and construct the expected device collaboration network diagram of the current session; Specifically, the semantic text sequence in the structured conversation stream is parsed to identify the control intent contained therein; for each identified control intent, its operation object is extracted as the device node to be controlled, and its action and / or target state is extracted as the operation instruction node; at the same time, based on the order of temporal conjunctions, action predicates or their positions in the dialogue rounds, the temporal relationship between the operation instruction nodes is determined and classified as sequential execution, parallel execution or conditional triggering. Treat the controlled device as one type of node and the operation instructions as another type of node to construct a bipartite graph or hybrid network graph; mark the temporal relationship attributes on the directed edges; for example, the directed edge from operation instruction node P to operation instruction node Q is marked as precedent, indicating that operation instruction node Q can only start after operation instruction node P has been executed; if there are consecutive instructions for the same controlled device, a conditional dependency is generated, and subsequent operation instruction nodes depend on the conditions after the previous instruction has been executed and returned a status; All extracted device nodes to be controlled, operation instruction nodes, and edges with timing and condition labels from this round of dialogue are combined into a graph structure. The directed edges in this graph structure are action edges, which are directly derived from the instruction timing relationships extracted during semantic parsing. This graph structure is the expected device cooperation network graph of the current session.

[0034] S28. Retrieve the dynamic collaborative baseband formed by the device set in historical collaborative tasks from the conference room IoT platform. The dynamic collaborative baseband defines the timing tolerance window and state interlocking relationship of the action response between devices. Understandably, a query request is sent to the conference room IoT platform, containing all device identifiers involved in the current expected device collaboration network graph; the IoT platform retrieves its historical collaborative task execution logs, filtering out collaborative task instances jointly completed by these devices; the collaborative task instances are analyzed, and machine learning algorithms are used to statistically analyze the time interval distribution between the completion time of device X's action and the time when device Y should reach the target state for any two devices X and Y with a direct collaborative relationship; specifically, the machine learning algorithms (such as statistical modeling or sequence pattern mining) are used as follows: for the extraction of the time series tolerance window, a Gaussian mixture model is used to statistically model the historical time interval distribution between device pairs to fit its multimodal characteristics, and a specific confidence interval (such as 95%) is calculated based on the fitted distribution model as the dynamic tolerance window boundary; for the mining of state interlocking relationships, an association rule learning algorithm (such as the Apriori algorithm) is applied. The upper and lower quartiles of this distribution are taken as the upper and lower limits of the collaborative timing tolerance window, respectively, thus defining the normal timing range of the action response between the devices. At the same time, the state dependency records between the devices are analyzed. If the function of device Y depends on device X being in a specific state, the logical condition that device Y must check the state of device X before performing an action is extracted, forming a state interlock relationship. The timing tolerance window set and the state interlock relationship set of all involved device pairs are structured and encapsulated to form a dynamic collaborative baseband data, which is returned to the intent recognition process.

[0035] S29. Overlay the expected device collaboration network graph onto the dynamic collaboration baseband, detect whether any device nodes or action edges deviate from the baseband range, and record all deviating nodes and edges to form a deviation set, including:

[0036] Perform a graph overlay operation, comparing each action edge in the expected device cooperative network graph with the timing tolerance window between the corresponding device pairs in the dynamic cooperative baseband. Specifically, if the action from device P to device Q in the expected network is marked as a sequential timing edge, but no timing tolerance window has been defined for this pair of devices in the dynamic cooperative baseband, then the action edge is determined to be a deviation edge; if a tolerance window is defined, but the implicit timing requirements in the expected device cooperative network graph exceed the upper or lower limit of the tolerance window, then the edge is also determined to be a deviation edge. For a device node, check whether the node's state requirements in the expected device collaboration network graph conflict with the state interlocking relationship between the node and other devices in the dynamic collaboration baseband. For example, the expected device collaboration network graph requires device C to directly perform a certain action, but the baseband state interlocking requires device D to be in a specific state before the action occurs, and the current session flow does not give the intention to operate device D. If such a conflict exists, then device C is determined to be a deviating node. All detected deviating edges and deviating nodes are aggregated to form the deviation set detected in this case.

[0037] S210. If the deviation set is detected to be non-empty, further analysis is conducted to determine whether the elements in the deviation set will trigger a chain reaction or form logical islands in the expected device collaboration network graph, thereby determining whether there is a multi-device collaboration deviation. The specific implementation method is as follows: When the deviation set is not empty, impact analysis is initiated. In the expected device collaboration network graph, nodes and edges in the deviation set are removed or modified sequentially to simulate the impact of the absence of these deviation elements on the connectivity and reachability of the expected device collaboration network graph. Chain reaction analysis is performed by checking whether the topological sorting of the expected device collaboration network graph is still feasible. If the deviation edge is an edge on a critical path, the critical path refers to the longest dependent path from the initial instruction to the final device state in the expected device collaboration network graph. Any delay or removal on this path will cause the overall task to fail to complete on time, and in this case, it is determined that a chain reaction will be triggered. Logical island analysis checks whether the expected device collaboration network graph splits into an isolated subgraph that has lost all connection edges with other parts after the removal of deviation elements. The device group corresponding to this isolated subgraph cannot be reached from the initial instruction node or cannot feed back its execution results to the overall task, thus forming a logical island. If the analysis results show that the elements deviating from the set are expected to trigger a chain reaction or form logical islands, then it is formally determined that there is a multi-device collaborative deviation; otherwise, if the elements deviating from the set are marginal redundant nodes or optional edges that do not affect the overall collaborative logic, then it does not constitute a multi-device collaborative deviation.

[0038] In this embodiment of the invention, when a potential intent conflict is determined to exist, step S3 performs intent anchor point re-filtering on the current session flow based on a pre-constructed conflict coupling graph to obtain the re-filtering result. The specific implementation of this re-filtering method includes: S31. Extract the identifiers of each specific conflict dimension that triggers the determination of potential intention conflicts, and their corresponding quantitative evaluation results; In step S2, when at least two of the three dimensions of analysis units—semantic logic conflict, acoustic field anomaly, and multi-device collaborative deviation—simultaneously output results indicating a risk of intent understanding, it is determined that there is a potential intent conflict in the current session flow. At this time, all dimension analysis units are traversed, and for each analysis unit that triggers the risk indication, its unique conflict dimension identifier is extracted, such as string constant semantic logic conflict, acoustic field anomaly, or multi-device collaborative deviation. At the same time, the quantitative evaluation result output by the analysis unit is queried. The quantitative evaluation result is a scalar value related to the degree of risk. For example, the semantic logic conflict unit outputs the normalized value of the correlation jump amplitude, the acoustic field anomaly unit outputs the significance score of the variant features in the residual acoustic field, and the multi-device collaborative deviation unit outputs the severity score of the impact of the deviation set elements on the network. Each triggered dimension identifier is paired with its corresponding quantitative evaluation result to form a key-value pair. All key-value pairs are combined into an information structure, namely the identifier of each specific conflict dimension and its corresponding quantitative evaluation result.

[0039] S32. Use the conflict dimension identifier and quantitative evaluation result as query vectors and input them into the conflict coupling graph for matching query; wherein, the conflict coupling graph stores the true intentions and dialogue clarification patterns that were finally corresponding to different combinations of conflict dimensions in history in a graph structure. Understandably, the extracted conflict dimension identifiers and quantitative evaluation results are encapsulated into a query vector with a fixed format. Each dimension of the query vector corresponds to a possible conflict dimension identifier. If a conflict dimension is triggered, its quantitative evaluation result value is filled in the corresponding dimension position. If it is not triggered, the dimension position is set to zero. The constructed query vector is input into the query interface of the conflict coupling graph. In the graph structure of the conflict coupling graph, each historical conflict pattern is represented by a virtual node. The attribute vector of this virtual node stores the historical conflict dimension combinations and their normalized strengths. The matching query process calculates the similarity between the query vector and the attribute vectors of all historical conflict pattern nodes in the conflict coupling graph. The calculation method can be cosine similarity. Based on the cosine similarity from high to low, the K most similar historical conflict pattern nodes are retrieved as the matching query results. The matching query results include historical conflict pattern nodes, as well as true intent nodes and clarification pattern nodes connected by weighted edges from these historical conflict pattern nodes.

[0040] S33. Based on the matching query results, calculate the conflict resolution confidence of each candidate intent in the current structured session flow, where the conflict resolution confidence reflects the historical probability that the candidate intent is the correct intent under the combination of conflict dimensions. During the intent understanding process in step S2, a candidate intent list has been generated for the current session stream. This list contains several possible interpretation intents. For each candidate intent in the candidate intent list, the following calculations are performed: traverse the top K historical conflict pattern nodes matched in the matching query results. If a historical pattern node is connected to the candidate intent node through an edge, obtain the weight of that edge. This weight represents the co-occurrence strength of the intent being the correct intent under that historical conflict pattern. Use the cosine similarity of each matching historical pattern as the contribution coefficient of that matching historical pattern, and perform a weighted sum with the edge weights pointing to the candidate intent under that matching historical pattern. If multiple historical matching patterns all point to the same candidate intent, the contributions of these historical matching patterns will be accumulated. The accumulated result is mapped to the interval between 0 and 1 through a normalization function. This value is the confidence level of the conflict resolution of the candidate intent under the current specific combination of conflict dimensions. This confidence level reflects the probability that the current candidate intent is the user's true intent based on experience in similar conflict situations in the past.

[0041] S34. If the conflict resolution confidence of any candidate intent is higher than the preset first intent analysis threshold, then that candidate intent is determined as the unique dominant intent, specifically: A preset first intent analysis threshold is established. This threshold is set based on: collecting a large number of conflict session samples during the development and testing phase, and calculating the conflict resolution confidence distribution obtained from the correct intent confirmed by the user under various conflict combinations; selecting the lower tenths of this distribution as a reference for the first intent analysis threshold, with an example value of 0.85; its significance is that when the conflict resolution confidence is higher than the first intent analysis threshold, the historical accuracy of intent determination under this type of conflict mode exceeds 90%; when processing the current session flow, the conflict resolution confidence of each candidate intent is checked one by one; if any candidate intent is found to have a conflict resolution confidence value greater than 0.85, the check is immediately stopped, and the candidate intent is promoted to the sole dominant intent, indicating that after conflict analysis, the certainty of the intent is high enough, and no further clarification with the user is needed, so it can be directly used as the basis for the response.

[0042] S35. If the confidence level of conflict resolution for all candidate intents is lower than the preset first intent analysis threshold but higher than the preset second intent analysis threshold, then generate multi-turn dialogue anchors to be initiated to the user based on the clarification patterns associated in the conflict coupling graph. Understandably, the second intent analysis threshold is set based on the following: analyzing data from the development and testing phases to identify the minimum confidence level required for the intent to be correctly understood. When the conflict resolution confidence level falls below this minimum confidence level, the failure rate increases sharply even with clarification dialogue. The upper quartile of this minimum confidence level can be used as a reference for the second intent analysis threshold; an example value of 0.50 can be used. When the conflict resolution confidence levels of all candidate intents are not higher than the first intent analysis threshold, but at least one candidate intent has a conflict resolution confidence level higher than the second intent analysis threshold, then clarification mode is entered. In this case, uniqueness is not directly determined. Instead of focusing on intent, the analysis extracts the clarification pattern node most closely connected to the current conflict pattern from the matching query results of the conflict coupling graph. The clarification pattern node contains the definition of clarification strategy and issue slot. The pattern is parsed to determine the key information points that need to be clarified, such as confirming the device name, operation action, or coordination sequence. These information points are packaged into a dialogue anchor data structure to generate one or more targeted clarification questions and set the state machine expectation for multi-turn dialogue. The dialogue anchor will drive one or more subsequent rounds of clarification dialogue with the user to obtain enough information to push the conflict resolution confidence level above the first intent analysis threshold.

[0043] In this embodiment of the invention, the detailed implementation steps of S4, which dynamically adapts the intent response strategy based on the re-filtering results, include: S41. Determine the type of re-filtered result: If the type of re-filtered result is a single dominant intent, proceed to S42; if the type of re-filtered result is a multi-turn dialogue anchor, proceed to S43.

[0044] S42. Implement a high-confidence direct execution strategy: Transform the unique dominant intent into structured device control commands and send them to the conference room IoT platform for execution; The process of transforming a unique dominant intent into structured device control instructions involves the following steps: A pre-defined mapping table between intents and instructions is established, defining the conversion rules between each possible atomic operation intent and its corresponding device application programming interface (API) call format. Once the unique dominant intent is determined, its semantic slots are parsed, including the target device identifier, the action to be performed, and parameters such as target percentage or mode value. This information is then populated into a predefined instruction template, which uses a data exchange format parsable by the IoT platform. After template population, a structured device control instruction object is generated, containing fields such as target device ID, action code, parameter list, priority, and timeout. This instruction object is serialized and sent via a secure protocol to the command execution interface of the conference room IoT platform to drive the actions of the physical devices.

[0045] S43. Execute a targeted clarification dialogue strategy: Based on the multi-round dialogue anchor points, retrieve the corresponding interaction template from the predefined clarification script template library, and initiate an interactive clarification inquiry to the user through the speech synthesis module; Specifically, when the output is a multi-turn dialogue anchor point, the anchor point object contains a type identifier of the clarification intent and missing semantic slot information. This type identifier is used as an index key to query a predefined clarification dialogue template library. The template library stores various interactive templates, each defining the structure of the clarification dialogue. For example, a confirmation template asks whether an operation should be performed, and a parameter-based question template asks what level of adjustment the device would like to be made. After retrieving the matching template, the specific device name, candidate actions, and other entity information from the anchor point are filled into the corresponding placeholders in the template to generate a complete natural language clarification question text. Then, the speech synthesis unit is invoked to convert the text into a natural and fluent speech waveform. The synthesized speech signal is played through the conference room's speaker system to initiate an interactive clarification inquiry to the user. Subsequently, a listening window is started to wait for user voice input to receive the next round of dialogue.

[0046] S44. If S3 does not output a clear result or the confidence level of conflict resolution for all candidate intentions is lower than the safety threshold, then a conservative safety operation strategy is executed: under this strategy, only the default safety operation corresponding to the current conference room scenario is executed, or only the instruction is confirmed by voice feedback without any device control action is executed. Understandably, the security threshold is a critical value used to determine whether to implement a conservative security operation strategy. It is set based on the analysis of receiver operation characteristic curves of a large number of conflict session samples in an offline environment. Specifically, under the premise of a fixed overall system failure tolerance, a threshold point is found so that when a conservative strategy is adopted at or below this point, the risk of irreversible consequences caused by misoperation can be reduced to the greatest extent. This threshold takes into account both system security and availability and is set relatively low. As an example, the security threshold can be set to 0.40.

[0047] Please see Figure 2 This invention provides a conference room speech intent recognition system based on scene corpus, comprising: The speech stream acquisition, parsing, and structured module acquires and parses the raw speech streams of users in the conference room, generating a structured conversation stream that includes semantic text, acoustic features, and sound source location information. Multi-dimensional Intent Conflict Detection and Judgment Module: Performs multi-dimensional intent conflict judgment on structured conversation streams: Concurrently analyzes semantic logic conflicts, acoustic field anomalies, and multi-device coordination deviations in structured conversation streams. When at least two dimensions simultaneously indicate the existence of intent understanding risks, the current conversation stream is judged to have potential intent conflicts. Conversation Flow Intent Anchor Filtering Module: When a potential intent conflict is determined, the intent anchors of the current conversation flow are re-filtered based on a pre-built conflict coupling graph to obtain the re-filtering results. The conflict coupling graph reflects the historical association patterns between different conflict dimensions, and the re-filtering is used to determine the unique dominant intent from multiple candidate intents or to generate multi-turn dialogue anchors that need to be clarified. Dynamic Adaptation Module for Intent Response Strategy: Based on the re-filtering results, dynamically adapt the intent response strategy, which includes at least high-confidence direct execution, initiating a targeted clarification dialogue, or performing a conservative security operation. Conflict Coupling Graph Feedback Update Module: Updates the conflict coupling graph based on the execution feedback of the intent response strategy.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0049] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0050] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for conference room speech intent recognition based on a scenario corpus, characterized in that, The method includes: S1. Collect and parse the original voice stream of users in the conference room to generate a structured conversation stream containing semantic text, acoustic features and sound source localization information; S2. Perform multi-dimensional intent conflict judgment on structured conversation streams: concurrently analyze semantic logic conflicts, acoustic field anomalies, and multi-device coordination deviations in structured conversation streams. When at least two dimensions simultaneously indicate the existence of intent understanding risks, it is determined that there are potential intent conflicts in the current conversation stream. S3. When a potential intent conflict is determined, the intent anchor points of the current conversation flow are re-filtered based on the pre-constructed conflict coupling graph to obtain the re-filtering results. The conflict coupling graph reflects the historical association patterns between different conflict dimensions, and the re-filtering is used to determine the unique dominant intent from multiple candidate intents or to generate multi-turn dialogue anchor points that need to be clarified. S4. Based on the re-screening results, dynamically adapt the intent response strategy, which includes at least high-confidence direct execution, initiating targeted clarification dialogue, or performing conservative security operations. S5. Update the conflict coupling graph based on the execution feedback of the intent response strategy. 2.The method according to claim 1, wherein, S1 includes: Raw audio streams containing multiple audio signals are captured using a microphone array deployed in the conference room; Speech activity detection is performed on the original speech stream to separate human voices from non-human voices, and the separated human voice signals are then subjected to noise reduction and echo cancellation processing to obtain effective human voice segments. Automatic speech recognition is performed on valid human voice segments to generate corresponding semantic text sequences, and fundamental frequency, energy and speech rate are extracted from valid human voice segments as acoustic feature sets simultaneously; Based on the time difference of arrival of multiple signals in the microphone array, the location coordinates of the sound source of the effective human voice segment in the conference room space are obtained; The semantic text sequence, acoustic feature set, and sound source location coordinates are bound to the same time window to form a timestamped structured conversation stream unit, and multiple consecutive structured conversation stream units are organized into a structured conversation stream in chronological order.

3. The conference room speech intent recognition method based on scene corpus according to claim 2, characterized in that, The analytical semantic logic conflict in S2 includes: Based on a pre-defined knowledge graph of the meeting room scenario, the semantic text sequence in the structured conversation stream is parsed to identify the entities, actions, and contextual relationships involved, and the parsing results are obtained. Using the parsing results as input, calculate the correlation between the semantic elements of the current statement and the historical successful interaction patterns stored in the knowledge graph of the conference room scenario, and track the penetration and decay path of this correlation in consecutive rounds of dialogue; When an unexpected jump in the correlation degree is detected at a preset key logical node or a logical loop appears in the penetration path, it is determined that there is a semantic logical conflict, and the location of the key logical node where the conflict occurs is recorded.

4. The conference room speech intent recognition method based on scene corpus according to claim 3, characterized in that, The acoustic field anomalies analyzed in S2 include: Based on the sound source location coordinate sequence in the structured conversation flow, combined with the preset three-dimensional spatial model of the conference room, the propagation of sound waves is dynamically simulated to generate a real-time acoustic field model that reflects the spatial distribution and attenuation of sound energy. Pattern matching is performed between the real-time acoustic field model and a baseline acoustic field feature library built from historical normal conference audio. If the matching results show that there are unexplainable acoustic variations in the current acoustic field that cannot be explained by known noise sources or normal speech activity, then it is determined that there is an acoustic field anomaly.

5. The conference room speech intent recognition method based on scene corpus according to claim 4, characterized in that, The method combines a pre-set 3D spatial model of the conference room to dynamically simulate sound wave propagation and generate a real-time acoustic field model reflecting the spatial distribution and attenuation of sound energy, including: Using the sound source location coordinates as the starting point of the sound source, and based on the geometric structure and acoustic material properties of the three-dimensional spatial model of the conference room, the set of propagation paths of the sound wave from this starting point to various directions of the conference room space is obtained. Using the set of propagation paths as input, and based on the sound wave propagation equation, the physical effects of energy attenuation, reflection, diffraction, and interference of sound waves on each path are simulated sequentially to obtain the simulation results of the sound wave state on each path. By combining the simulation results of sound wave states along all propagation paths, a three-dimensional distribution field representing the sound pressure level or sound energy density at each point in the conference room space is generated as a real-time acoustic field model.

6. The conference room speech intent recognition method based on scene corpus according to claim 4, characterized in that, The analysis of multi-device collaborative deviation in S2 includes: Parse the semantic text sequence in the structured session stream, identify the set of devices to be controlled and the timing of operation instructions, and construct the expected device cooperation network graph of the current session; From the IoT platform of the conference room, retrieve the dynamic collaborative baseband formed by the device set in historical collaborative tasks. The dynamic collaborative baseband defines the timing tolerance window and state interlock relationship of the action response between devices. The expected device collaboration network graph is overlaid on the dynamic collaboration baseband to detect whether there are any device nodes or action edges that deviate from the baseband range, and all deviated nodes and edges are recorded to form a deviation set. If a non-empty deviation set is detected, the analysis will determine whether the elements in the deviation set will trigger a chain reaction or form logical islands in the expected device collaboration network graph, thereby determining whether there is a multi-device collaboration deviation.

7. The conference room speech intent recognition method based on scene corpus according to claim 1, characterized in that, S3 includes: Extract the identifiers of each specific conflict dimension that triggers the determination of potential intention conflicts, and their corresponding quantitative evaluation results; The conflict dimension identifier and the quantitative evaluation result are used as query vectors and input into the conflict coupling graph for matching query; the conflict coupling graph stores the true intentions and dialogue clarification patterns that correspond to different combinations of conflict dimensions in history in a graph structure. Based on the matching query results, calculate the conflict resolution confidence of each candidate intent in the current structured session flow, where the conflict resolution confidence reflects the historical probability that the candidate intent is the correct intent under the combination of conflict dimensions; If the confidence level of conflict resolution for any candidate intent is higher than the preset first intent analysis threshold, then that candidate intent is determined as the unique dominant intent. If the confidence level of conflict resolution for all candidate intents is lower than the preset first intent analysis threshold but higher than the preset second intent analysis threshold, then based on the clarification patterns associated in the conflict coupling graph, multi-turn dialogue anchors to be initiated to the user are generated.

8. The conference room speech intent recognition method based on scene corpus according to claim 7, characterized in that, The process of constructing the conflict coupling map includes: During initialization or early operation, historical session data is collected, and the combination of conflict dimensions determined in each session and the true intent of the final verification are extracted. The association rules between the combination of conflict dimensions and the true intent are analyzed, and the association relationship with the confidence of conflict resolution higher than the preset rule is initialized as a conflict coupling graph in the form of nodes and weighted edges. The nodes represent conflict dimensions or the final intent, the edges represent their co-occurrence relationship, and the weights represent the co-occurrence strength. After each S1 to S4 identification process is completed, record the specific combination of conflict dimensions output by S2 in this process, the re-screening process and results of S3, and the true intent finally confirmed by the user or verified by the system. Based on the actual execution effect feedback of the intent response strategy finally adopted in this identification process, the weights of the relevant edges in the conflict coupling graph are dynamically adjusted: if the execution is successful, the weight of the corresponding edge is increased; if the execution fails or triggers user correction, the weight of the corresponding edge is decreased.

9. The conference room speech intent recognition method based on scene corpus according to claim 1, characterized in that, The dynamic adaptation intent response strategy in S4 includes: S41. Determine the type of re-filtered result: If the type of re-filtered result is a single dominant intent, proceed to S42; if the type of re-filtered result is a multi-turn dialogue anchor, proceed to S43. S42. Implement a high-confidence direct execution strategy: Transform the unique dominant intent into structured device control commands and send them to the conference room IoT platform for execution; S43. Execute a targeted clarification dialogue strategy: Based on the multi-round dialogue anchor points, retrieve the corresponding interaction template from the predefined clarification script template library, and initiate an interactive clarification inquiry to the user through the speech synthesis module; S44. If S3 does not output a clear result or the confidence level of conflict resolution for all candidate intentions is lower than the safety threshold, then a conservative safety operation strategy is executed: under this strategy, only the default safety operation corresponding to the current conference room scenario is executed, or only the receipt of the instruction is confirmed through voice feedback without any device control action is executed.

10. A conference room speech intent recognition system based on scene corpus, characterized in that, The system, applied to the conference room speech intent recognition method based on scene corpus as described in any one of claims 1-9, comprises: The speech stream acquisition, parsing, and structured module acquires and parses the raw speech streams of users in the conference room, generating a structured conversation stream that includes semantic text, acoustic features, and sound source location information. Multi-dimensional Intent Conflict Detection and Judgment Module: Performs multi-dimensional intent conflict judgment on structured conversation streams: Concurrently analyzes semantic logic conflicts, acoustic field anomalies, and multi-device coordination deviations in structured conversation streams. When at least two dimensions simultaneously indicate the existence of intent understanding risks, the current conversation stream is judged to have potential intent conflicts. Conversation Flow Intent Anchor Filtering Module: When a potential intent conflict is determined, the intent anchors of the current conversation flow are re-filtered based on a pre-built conflict coupling graph to obtain the re-filtering results. The conflict coupling graph reflects the historical association patterns between different conflict dimensions, and the re-filtering is used to determine the unique dominant intent from multiple candidate intents or to generate multi-turn dialogue anchors that need to be clarified. Dynamic Adaptation Module for Intent Response Strategy: Based on the re-filtering results, dynamically adapt the intent response strategy, which includes at least high-confidence direct execution, initiating a targeted clarification dialogue, or performing a conservative security operation. Conflict Coupling Graph Feedback Update Module: Updates the conflict coupling graph based on the execution feedback of the intent response strategy.

Images