A voice interaction system and method based on artificial intelligence technology

Abstract

The application discloses a voice interaction system and method based on artificial intelligence technology, and relates to the technical field of voice interaction, comprising: double-layer collaborative recognition adopting a main channel Transformer model and an auxiliary channel CTC-Attention hybrid architecture, obtaining text with multi-dimensional feature labels, and performing scene classification and dependency syntax analysis through a bidirectional recurrent neural network, combining a dynamic knowledge graph to perform real-time retrieval of associated entities, generating structured intent data, performing three-dimensional analysis through an explainability engine, generating a multi-dimensional traceability report, and triggering an intelligent decision fusion engine through a confidence threshold to generate structured enhanced interaction data; the application performs scene classification and dependency syntax analysis through a bidirectional recurrent neural network, combines a dynamic knowledge graph to perform real-time retrieval of associated entities, and generates structured intent data, so as to accurately understand user intent and provide personalized interaction services.

Description

Technical Field

[0001] This invention relates to the field of voice interaction technology, and in particular to a voice interaction system and method based on artificial intelligence technology. Background Technology

[0002] With the continuous advancement of artificial intelligence technology, voice interaction systems have achieved significant development and application in many fields. Traditional voice interaction systems mainly rely on speech recognition technology and natural language processing technology. By converting users' voice commands into text and then performing semantic analysis on the text, interactive functions can be realized. The development of these technologies has greatly improved the convenience and naturalness of human-computer interaction, enabling users to easily control devices and obtain information to complete complex tasks through voice commands.

[0003] While some improvements have been made in existing technologies, traditional voice interaction systems have gradually revealed many shortcomings as application scenarios become more complex and user needs become more diverse. In terms of multimodal interaction, although existing technologies attempt to combine visual and tactile feedback, they lack effective synchronization mechanisms and interpretability, and cannot achieve true multimodal fusion. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a voice interaction method based on artificial intelligence technology to address the shortcomings of existing technologies, which, although attempting to combine visual and tactile feedback, lack effective synchronization mechanisms and interpretability.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a voice interaction method based on artificial intelligence technology, which includes: capturing user voice data, synchronizing a sound source focusing algorithm to lock the target location, and using a deep neural network noise reduction model to eliminate environmental noise and output voice data packets;

[0008] A two-layer collaborative recognition model using a hybrid architecture of the main channel Transformer model and the auxiliary channel CTC-Attention is employed to obtain text with multi-dimensional feature annotations.

[0009] Based on text with multi-dimensional feature annotations, a bidirectional recurrent neural network is used for scene classification and dependency parsing, and dynamic knowledge graphs are combined to retrieve related entities in real time to generate structured intent data.

[0010] Structured intent data is analyzed in three dimensions using an interpretability engine to generate a multi-dimensional source tracing report;

[0011] Based on multi-dimensional source tracing reports, an intelligent decision fusion engine is triggered by confidence thresholds to generate structured enhanced interactive data;

[0012] Based on structured enhanced interactive data, multimodal interaction is achieved through dynamic source tracing visualization interface rendering and tactile coding feedback from smart wearable devices, and a complete interaction log is generated.

[0013] As a preferred embodiment of the voice interaction method based on artificial intelligence technology described in this invention, the specific steps for eliminating environmental noise using a deep neural network noise reduction model are as follows:

[0014] A deep neural network noise reduction model is constructed based on a hybrid framework of convolutional neural networks and recurrent neural networks.

[0015] The spectral features of the beamformed speech signal are extracted by the convolutional neural network layer, and the temporal features of the speech signal are captured by the recurrent neural network layer to distinguish noise.

[0016] A deep neural network denoising model was trained using paired data of noisy and clean speech, and the difference between the denoised features and the clean speech was calculated using the mean squared error loss function.

[0017] The Adam optimizer is used to update parameters through backpropagation. After being optimized on the validation set, the trained deep neural network denoising model is deployed, and noise components are estimated and suppressed. The output layer outputs the denoised speech features, which are then encoded to form speech data packets.

[0018] As a preferred embodiment of the voice interaction method based on artificial intelligence technology described in this invention, the specific steps of employing a two-layer collaborative recognition model using a hybrid architecture of a main channel Transformer model and an auxiliary channel CTC-Attention are as follows:

[0019] After preprocessing the speech processing package, it is input into the main channel Transformer model and the auxiliary channel CTC-Attention model for feature encoding.

[0020] The Transformer encoding process generates a context representation through multi-head self-attention, then processes it through residual connections and layer normalization, and finally performs residual connections and layer normalization a second time after nonlinear transformation through a feedforward neural network, ultimately outputting an encoded feature sequence.

[0021] The interaction weights between each character in the target text and the encoded feature sequence are dynamically calculated through the attention mechanism to generate a context vector. The time-step features of the CTC-Attention model are concatenated with the attention context vector to generate an optimized feature sequence.

[0022] The main channel and auxiliary channel are reconstructed into a global semantic strategy layer and a local alignment execution layer through functional decoupling. The global semantic strategy layer focuses on cross-frame semantic association, while the local alignment execution layer strengthens the alignment confidence at the time step granularity.

[0023] A gated residual fusion strategy is adopted to dynamically fuse global semantic features and local alignment features to generate multi-dimensional feature-annotated text.

[0024] As a preferred embodiment of the voice interaction method based on artificial intelligence technology described in this invention, the specific steps for generating structured intent data by combining dynamic knowledge graphs for real-time retrieval of related entities are as follows:

[0025] Extract scene classification results and dependency relationship information from text with multi-dimensional feature annotations, dynamically combine them into query conditions, and retrieve related entities in the knowledge graph in real time;

[0026] Relevant entities are filtered through entity matching and similarity calculation, and then integrated into structured intent data in JSON format.

[0027] The process of performing three-dimensional analysis using an interpretability engine to generate a multi-dimensional source tracing report involves the following steps:

[0028] The structured intent data is input into the interpretability engine. Through hierarchical feature decoupling, the structured intent data is split into three independent encoders: intent dimension encoder channel, entity association dimension encoder channel, and time-series trajectory dimension encoder channel. The split data is processed separately.

[0029] The semantic confidence vector, entity association strength vector, and temporal feature vector output by the three independent encoders are enhanced by residual connection and then input into the Gram-Schmidt orthogonalization process to eliminate the linear correlation between dimensions and generate mutually independent orthogonal basis vector sets. A multi-dimensional source tracing report is generated based on the orthogonal basis vector sets.

[0030] As a preferred embodiment of the voice interaction method based on artificial intelligence technology described in this invention, the specific steps for generating structured enhanced interaction data by triggering an intelligent decision fusion engine through a confidence threshold are as follows:

[0031] Set thresholds for semantic dimension, context dimension, and intent dimension, and compare them with the confidence level of each dimension to dynamically trigger the intelligent decision fusion engine;

[0032] Feature fusion is performed on dependency relationship annotations and associated entity annotations in the semantic dimension to enhance the semantic representation vector; context integration is performed on scene classification results and context labels in the context dimension to enhance the context representation vector; and intent labels and intent verification results in the intent dimension are merged through weighted fusion operation to optimize the intent representation vector.

[0033] The data is aggregated to integrate the semantic, contextual, and intent dimensions, and then structured, enhanced interactive data is generated using JSON encoding.

[0034] As a preferred embodiment of the voice interaction method based on artificial intelligence technology described in this invention, the steps of performing multimodal interaction through dynamic source-tracing visualization interface rendering and tactile coding feedback from smart wearable devices, and generating a complete interaction log, are as follows:

[0035] The dynamic traceability visualization interface is based on enhanced interactive data rendering of a three-layer interactive view. The smart wearable device tactile coding engine converts the entity association strength and intent confidence into dual-channel tactile signals.

[0036] The tactile pulses and key frames of the visual animation are linked by time stamp synchronization, and the pressure sensor data is used to adjust the interface mechanical parameters in real time to form a closed-loop interaction.

[0037] The timeline-aligned voice data stream, visual operation path, and haptic parameter triplet are encoded according to the W3CActivityStreams 2.0 specification and verified with SHA-256 to generate a complete interaction log that can be restored across devices.

[0038] Secondly, the present invention provides a voice interaction system based on artificial intelligence technology, comprising: a data acquisition module, an annotation module, a parsing module, an analysis module, a fusion module, and an interaction module.

[0039] The acquisition module captures user voice data, simultaneously uses a sound source focusing algorithm to lock the target location, and employs a deep neural network noise reduction model to eliminate environmental noise and output voice data packets.

[0040] The annotation module employs a two-layer collaborative recognition approach combining the main channel Transformer model and the auxiliary channel CTC-Attention hybrid architecture to obtain text with multi-dimensional feature annotations.

[0041] The parsing module, based on text with multi-dimensional feature annotations, performs scene classification and dependency parsing through a bidirectional recurrent neural network, and combines dynamic knowledge graphs to retrieve related entities in real time, generating structured intent data;

[0042] The analysis module performs three-dimensional analysis on structured intent data through an interpretability engine, generating a multi-dimensional source tracing report.

[0043] The fusion module, based on multi-dimensional source tracing reports, triggers an intelligent decision fusion engine through confidence thresholds to generate structured, enhanced interactive data;

[0044] The interaction module, based on structured enhanced interactive data, performs multimodal interaction through dynamic traceability visualization interface rendering and tactile coding feedback from smart wearable devices, and generates a complete interaction log.

[0045] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the voice interaction method based on artificial intelligence technology as described in the first aspect of the present invention.

[0046] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the voice interaction method based on artificial intelligence technology as described in the first aspect of the present invention.

[0047] The beneficial effects of this invention are as follows: environmental noise is eliminated by a deep neural network noise reduction model, and a two-layer collaborative recognition is achieved by using a hybrid architecture of the main channel Transformer model and the auxiliary channel CTC-Attention. This enables high-precision recognition and multi-dimensional feature annotation of speech signals. Scene classification and dependency parsing are performed by a bidirectional recurrent neural network, and related entities are retrieved in real time by combining a dynamic knowledge graph to generate structured intent data, thereby accurately understanding user intent and providing personalized interactive services. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a voice interaction method based on artificial intelligence technology.

[0050] Figure 2 This is a flowchart of a deep neural network noise reduction model.

[0051] Figure 3 Flowchart of collaborative identification between main channel and auxiliary channel.

[0052] Figure 4 This is the core flowchart for multimodal interaction. Detailed Implementation

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0055] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0056] Reference Figures 1 to 4 This is one embodiment of the present invention, which provides a voice interaction method based on artificial intelligence technology, including the following steps:

[0057] S1. Capture user voice data, synchronize the sound source focusing algorithm to lock the target location, and use a deep neural network noise reduction model to eliminate environmental noise and output voice data packets;

[0058] User voice data is collected using a multi-microphone array arranged in a ring. This ring arrangement provides even coverage in all directions, making it suitable for omnidirectional sound acquisition. The signal from each microphone is stored in digital audio format.

[0059] Beamforming technology is used to preprocess speech signals acquired by multiple microphones. By calculating the time delay and phase difference of each microphone signal, the speech signal is focused in a specific direction, locking the location of the target sound source. In a multi-microphone array, based on the speed of sound in air, the time difference between the signals received by any two microphones is calculated. The time difference reflects the difference in the path length of the sound wave from the sound source to the two microphones divided by the speed of sound. Based on the calculated arrival time difference and the microphone's position information in the array, a weighting coefficient for each microphone signal is determined. The weighting coefficient is dynamically adjusted according to the microphone's position and the direction of the target sound source, so that the signal is enhanced in the direction of the target sound source and attenuated in other directions. The coefficients weight the digital audio signals collected by each microphone, adjusting the amplitude of each microphone signal so that the signal in the direction of the target sound source is highlighted, while the signal in other directions is suppressed. The weighted signal retains the characteristics of the target sound source while reducing the influence of background noise and other interference sources. All the weighted microphone signals are superimposed. Since the weighting process has adjusted the amplitude and phase of each signal, the signals from the direction of the target sound source can enhance each other when superimposed, forming a beam pointing to the target sound source. The beam pointing to the target sound source has directionality and can be focused on the target sound source, thereby locking the location of the target sound source. Through the directionality of the beam, the position of the target sound source in space can be determined.

[0060] A hybrid architecture of convolutional neural networks and recurrent neural networks is adopted. The beamforming processed speech signal is input into the deep neural network noise reduction model. The convolutional neural network is used to extract features from the input speech signal. The convolutional layer in the convolutional neural network extracts local features through a sliding window, and the pooling layer is used to reduce the dimensionality of the features while retaining important information. The spectral features in the speech signal are automatically learned through the convolutional and pooling layers in the convolutional neural network. The spectral features include information on the frequency distribution and energy distribution of the speech signal, and can identify the differences between the speech signal and the noise signal in the spectrum.

[0061] The features extracted by the convolutional neural network are input into the recurrent neural network. The recurrent neural network can capture the temporal features of the continuity and change pattern of the speech signal in time. The recurrent neural network can remember the information of the previous moment and use it for the processing of the current moment, better understand the dynamic changes of the speech signal, and further distinguish noise and speech signal.

[0062] A large number of noisy speech samples and corresponding clean speech samples are collected as training datasets. Noisy speech samples refer to speech signals recorded in real-world environments containing background noise, while clean speech samples refer to pure speech signals recorded in noise-free environments. Long speech signals are segmented into shorter frames, and spectral features are extracted from each frame. The training dataset is further expanded by adding different types of noise and adjusting the signal-to-noise ratio. A deep neural network denoising model based on a hybrid architecture of convolutional neural networks and recurrent neural networks is constructed. The input layer receives preprocessed noisy speech features, the convolutional neural network layer extracts the spectral features of the speech signal, the recurrent neural network layer captures the temporal features of the speech signal, and the output layer outputs the denoised speech features. A loss function is defined to measure the difference between the denoised speech features output by the model and the clean speech features. Commonly used loss functions include mean squared error loss. The squared error B between the model output and the target clean speech features is calculated using the following formula:

[0063]

[0064] Where N is the number of samples, y i Features of clean speech denoised speech features are output by the model, and i is the index of the i-th sample in the training dataset;

[0065] The deep neural network denoising model is trained using a training dataset. Noisy speech features are input into the model, and denoised speech features are output using spectral and temporal features. A loss function is used to calculate the error between the model output and the clean speech features. Based on the calculated loss, the parameters of the deep neural network denoising model are updated using backpropagation to minimize the loss function. The Adam optimization algorithm is selected to adjust the model parameters. The Adam optimization algorithm gradually adjusts the model weights based on the gradient of the loss function to achieve the optimal denoising effect. During training, the deep neural network denoising model is validated using a validation dataset to evaluate its performance. Based on the validation results, the model's hyperparameters are adjusted to optimize its performance, and the data is saved. The validation dataset is a subset of data from the training dataset used to simulate the model's performance on unknown data. The trained deep neural network denoising model is then output.

[0066] By using a trained deep neural network denoising model, noise components in speech signals are estimated and removed. A large number of noisy speech samples and corresponding clean speech samples are used to optimize the parameters of the deep neural network denoising model by minimizing the noise estimation error. This enables the deep neural network denoising model to accurately identify and remove noise components. In practical applications, the deep neural network denoising model outputs a denoised speech signal based on the input noisy speech signal. The speech signal processed by the deep neural network denoising model is then encoded and encapsulated to form a speech data packet. The speech data packet contains metadata information such as the sampling rate and quantization bit depth of the speech signal and is stored in a digital audio format.

[0067] S2. A two-layer collaborative recognition model using a hybrid architecture of the main channel Transformer model and the auxiliary channel CTC-Attention is adopted to obtain text with multi-dimensional feature annotations.

[0068] The voice data packets are preprocessed to adapt to the input requirements of the main channel Transformer model. The voice data packets are segmented into fixed-length frame sequences, and Fourier transform is performed on each frame signal to obtain spectral features. The spectral features are filtered by a Mel filter bank and the energy value is calculated. The energy value is then subjected to discrete cosine transform to extract Mel frequency cepstral coefficient features. The extracted spectral features and Mel frequency cepstral coefficient features are normalized to obtain Mel frequency cepstral feature values. The Mel frequency cepstral feature values ​​are scaled to a fixed range, such as between 0 and 1. The normalized feature values ​​are input into the main channel Transformer model. The normalized feature values ​​of each frame are embedded and mapped through a learnable linear transformation, which improves the low-dimensional normalized features to a high-dimensional space and generates the embedded feature vector for each frame.

[0069] During the model training phase, the preprocessed speech feature sequences are jointly trained end-to-end. The main channel Transformer model encodes the embedded feature vectors through a multi-head self-attention mechanism. The entire embedded vector sequence is batch-projected into a query vector matrix, a key vector matrix, and a value vector matrix through three sets of independent and learnable weight matrices. Attention scores are calculated by the dot product of the query vector matrix and the key vector matrix. After scaling, the scores are normalized to attention weights using the Softmax function. These weighted sums are then added to the value vector matrix to generate an intermediate context representation. After the multi-head self-attention mechanism, the Transformer model adds the residuals of the input and output of the multi-head self-attention layer. Layer normalization is applied to the sum to adjust the mean and variance of the features. The Transformer model performs a nonlinear transformation on the normalized features through a feedforward neural network. The feedforward neural network processes the features through two fully connected layers. The ReLU activation function is applied in the intermediate layer. The residuals of the input and output of the feedforward neural network are added again. Layer normalization is repeated on the sum, and finally, the encoded feature sequence is output.

[0070] Simultaneously, the encoded feature sequence is input into the auxiliary channel's CTC-Attention model. The shape and length of the encoded feature sequence are adjusted to match the input requirements of the CTC-Attention model. The encoded feature sequence is then propagated forward through the CTC-Attention model to generate the output probability distribution at each time step, forming an output probability matrix. Each element represents the prediction probability of the CTC-Attention model for a specific character. The target text is converted into a character sequence, and all possible alignment paths are enumerated using a dynamic programming algorithm. The probability P(π|X) of each path is calculated using the following formula:

[0071]

[0072] Where X represents the encoded feature sequence input into the CTC-Attention model, T represents the length of the input sequence, and π t This indicates the character corresponding to the alignment path π at time step t. This indicates that at time step t, the CTC-Attention model performs a certain action on character π. t The predicted probability;

[0073] Calculate the total probability value P(Y|X) for each alignment path to obtain the alignment error, using the following formula:

[0074] P(Y|X)=∑ π∈Π P(π|X);

[0075] A = 1 - P(Y|X);

[0076] Where Y represents the character sequence of the target text, Π represents the set of all possible alignment paths, and A represents the alignment error;

[0077] The total probability value is converted into a CTC loss value by taking the negative log-likelihood. The smaller the CTC loss value, the higher the degree of matching between the model output and the target text. The formula is as follows:

[0078] CTCL = -log P(Y|X);

[0079] Wherein, CTCL represents the CTC loss value;

[0080] Building upon the CTC-Attention model, an attention mechanism is introduced to further optimize feature alignment. This involves calculating the attention weight between each character in the target text and the feature vector at each time step of the input encoded feature sequence, and then normalizing this weight to represent the contribution of the input encoded feature sequence to the target text. The calculated attention weights are then used to perform a weighted summation of the input encoded feature sequence to generate a context vector.

[0081] After completing the CTC loss calculation and attention mechanism optimization, the feature vector of each time step is merged with the corresponding context vector by concatenation to generate the optimized feature vector. The integrated feature vector is then arranged in the order of time steps to generate the optimized feature sequence.

[0082] The main channel Transformer model and the auxiliary channel CTC-Attention model are reconstructed into a global semantic strategy layer and a local alignment execution layer through functional decoupling. The global semantic strategy layer focuses on cross-frame semantic association, while the local alignment execution layer strengthens the alignment confidence at the time step granularity.

[0083] The global semantic strategy layer processes the input encoded feature sequence through a multi-head self-attention mechanism and a feedforward neural network to generate global semantic features containing long-distance dependencies between phonemes. The local alignment execution layer optimizes the alignment feature sequence based on CTC path probability and attention weights to generate local phoneme boundary features. The confidence of the two layers is quantified by calculating the self-attention weight entropy of the global semantic strategy layer and the CTC loss back-mapping value of the local alignment execution layer. A gated residual fusion strategy is adopted to dynamically fuse global semantic features and local alignment features. Finally, the fused features are annotated in multiple dimensions (phonemes, semantics, and context) through a multi-task joint framework to generate multi-dimensional feature-annotated text with cross-granularity alignment capabilities.

[0084] S3. Based on text with multi-dimensional feature annotations, scene classification and dependency parsing are performed through a bidirectional recurrent neural network, and related entities are retrieved in real time using a dynamic knowledge graph to generate structured intent data.

[0085] The phoneme features, semantic features, and contextual features of each time step are extracted from the text with multi-dimensional feature annotations. These features are then fused and organized into a multi-dimensional feature vector sequence. The extracted multi-dimensional feature vector sequence is input into the feedforward recurrent neural network of a bidirectional recurrent network to process the contextual information from left to right. The hidden state of the feedforward recurrent neural network is initialized, usually set to zero vector. The multi-dimensional feature vectors of each time step are then processed sequentially from front to back, and the hidden state of the feedforward recurrent neural network is updated. The hidden state of each time step depends on the input of the current time step and the hidden state of the previous time step. After processing by the feedforward recurrent neural network, a feedforward contextual information feature sequence is generated. The multi-dimensional feature vector of each time step contains the contextual information from the beginning of the sequence to the current time step.

[0086] The extracted multidimensional feature vector sequence is input into the feedforward recurrent neural network of the bidirectional recurrent network to process the context information from right to left. The hidden state of the feedforward recurrent neural network is initialized, usually set to zero vector. Then, the multidimensional feature vector of each time step is processed sequentially from back to front, and the hidden state of the feedforward recurrent neural network is updated. The hidden state of each time step depends on the input of the current time step and the hidden state of the next time step. After processing by the feedforward recurrent neural network, a feature sequence of backward context information is generated. The feature vector of each time step contains the context information from the current time step to the end of the sequence.

[0087] For each time step, the multidimensional feature vectors generated in the forward recurrent neural network and the backward recurrent neural network are concatenated to form a comprehensive multidimensional feature vector, which is then arranged in chronological order to form a bidirectional feature sequence.

[0088] The system receives the bidirectional feature sequence processed by a bidirectional recurrent network. The bidirectional feature vector at each time step is input into a fully connected layer. The output dimension of the fully connected layer is consistent with the number of scene categories. The fully connected layer maps the bidirectional feature vectors to the scene category space, generating a scene category score for each time step. A Softmax function is applied to the scene category scores output by the fully connected layer, converting the scores into probability values. These probability values ​​represent the predicted probability of each scene category at the corresponding time step. The predicted probabilities of all time steps are arranged in time step order to form a probability matrix, where each row corresponds to a time step and each column corresponds to a scene category. Each element in the probability matrix is ​​the predicted probability of the corresponding scene category at the given time step. The generated prediction probability matrix is ​​analyzed. For all time step prediction probability matrices, the average probability of each scene category is calculated along the time dimension. The scene category with the highest average probability is selected as the scene classification result.

[0089] The text with multi-dimensional feature annotations is preprocessed, segmented into individual words, and each word is tagged with part-of-speech tags to determine its grammatical role in the sentence. These tags are then converted into vector forms suitable for dependency parsing models. A rule-based approach and machine learning model are used to load a pre-trained dependency parsing model. Large-scale corpora with dependency relation annotations are collected through data preparation and annotation. The corpora are then segmented, tagged with part-of-speech tags, and have dependency relation triplets labeled. Finally, the dependency parsing model is loaded and initialized. The rule-based approach requires defining a grammar rule base. Dependency matching is achieved through hard coding, while machine learning methods load a pre-trained dependency parsing model. The pre-training method loads configuration files and weights, and uses domain data to fine-tune the pre-trained dependency parsing model to adapt to the task. Then, it enters the feature input and parsing stage, where vectorized text is input into the dependency parsing model. The classifier predicts the head word index and dependency relation type, and the prediction results are post-processed in combination with the rule base to correct conflicting dependency edges. Finally, the output and integration are completed, outputting the prediction result for each word, and the prediction result is back-annotated to the original text to generate enhanced dependency relation annotation text.

[0090] Based on the prediction results, a dependency relation list containing dependent word indexes, head word indexes, and dependency relation types is extracted. The dependency relations are then sorted according to the dependent word indexes to ensure their processing order matches the word order in the sentence. During dependency tree construction, the root node of the sentence is identified as the starting point of the dependency tree. The root node has no head word and represents the main structural center of the sentence. Based on the root node, the dependency tree structure is initialized to store nodes and edges. The root node is then added to the initialized dependency tree structure, becoming the first node of the dependency tree. The previously sorted dependencies are then processed... For each dependent word in the dependency list, the corresponding head word node is found in the dependency tree based on the head word index. A node is added to the dependency tree for the dependent word, and an edge is created from the head word node to the dependent word node. Each node represents a word, and the edge represents a dependency relationship. The entire dependency tree is gradually constructed. After adding each dependent word node, the structure of the dependency tree is updated so that the hierarchical relationship of the dependency tree can accurately reflect the grammatical and semantic levels of the sentence. On the basis of text with multi-dimensional feature annotation, dependency relationship annotation information is added as the result of dependency parsing.

[0091] Extract the scene category to which the text belongs from the scene classification results, extract semantic entity identifiers and relation types from the dependency parsing results, combine the scene category and semantic information into query conditions, initialize dynamic knowledge graph query based on the query conditions, and in the dynamic knowledge graph, query related entities and relations according to scene category and dependency relation information, and generate query results;

[0092] Extract entities related to the text content from the query results, match the extracted entities with the text content to associate the entities with the words in the text, further filter out entities that are highly related to the text content based on dependency information, evaluate the relevance of each entity by using similarity calculation to determine the degree of matching with the text content, organize the identifier, type and related attributes of each filtered entity into entity information, and combine all the organized entity information to generate associated entity annotations.

[0093] The scene classification results, dependency relationship annotation text, and associated entity annotation are integrated into a unified data structure. The scene classification results, dependency relationship annotation text, and associated entity annotation in the unified data structure are parsed to extract key information about scene categories, main dependencies, and key entities. The extracted key information is organized in JSON structured data format to form structured intent data.

[0094] S4. Perform three-dimensional analysis on the structured intent data through the interpretability engine to generate a multi-dimensional source tracing report;

[0095] Structured intent data is input into the interpretability engine, hierarchical feature decoupling is performed, and three independent encoders extract intent dimension features, entity association dimension features and temporal trajectory dimension features respectively;

[0096] The intent dimension encoder channel loads a pre-trained self-attention mechanism layer, performs multi-head attention computation on the input intent label sequence, and generates a semantic confidence vector. The number of attention heads strictly corresponds to the number of intent label categories, and the attention weight matrix is ​​normalized by the Softmax function.

[0097] The entity association dimension encoder channel activation graph convolutional network layer constructs a graph structure with knowledge graph entity IDs as nodes and dependency paths as edges, performs two-hop neighborhood information aggregation, and uses mean pooling to aggregate neighbor information for node feature updates, outputting an entity association strength vector.

[0098] The temporal trajectory dimension encoder channel deploys a bidirectional long short-term memory network layer. The sequence of event continuity indicators of the temporal trajectory dimension is input in the order of timestamps. The forward and backward hidden states are concatenated to generate temporal feature vectors. The output vector of each encoder channel is enhanced by residual connection and then input into the Gram-Schmidt orthogonalization process to eliminate the linear correlation between dimensions and generate mutually independent orthogonal basis vector sets.

[0099] The orthogonalized feature vectors are entered into dynamic weight allocation. Monte Carlo sampling is used to simulate feature missing scenarios. The difference in the impact of each dimension on the decision result is statistically analyzed. After normalization, the intention dimension weight, entity association dimension weight, and time-series trajectory dimension weight are generated. When the weight coefficient is lower than the dynamic average threshold, the feature is marked as low contribution and the orthogonalized copy is retained. The high-weight features are concatenated with the orthogonalized copy to generate a decoupled feature vector group.

[0100] The decoupled feature vector group is input into a two-layer rule engine. A rule base is constructed using the C4.5 algorithm and DBSCAN clustering. The rule base is sorted by feature weights to match tasks and perform temporal continuity, entity association, and semantic matching. Figure 1 The three-stage verification of consistency ultimately generates a structured decision path, while anomalies are fed back to the feature decoupling layer to dynamically adjust weights to achieve closed-loop optimization.

[0101] After the validated structured decision path enters the dynamic tracing stage, the tracing path is constructed. The intent dimension tracing generates the node trigger chain in chronological order by using the unique identifier and confidence change trajectory of the trigger node in the rule base path.

[0102] Entity-dimensional tracing extracts the diffusion paths of core and peripheral entities in the knowledge graph, marks the decay trend of association strength during the diffusion process, and forms a curve of change in entity relationship strength.

[0103] Time-series tracing involves constructing an event chain causal relationship graph and marking the timestamp offsets of key event nodes.

[0104] Multimodal reports are synthesized based on the source tracing path results. The text report generates natural language descriptions according to the structure of decision conclusion - rule basis - anomaly warning. The visualization map uses a force-directed graph layout to show the relationship between entities and events. The node size represents the feature weight, the edge transparency reflects the temporal correlation strength, and the overlaid time axis heatmap uses color gradient to represent the confidence change. The interactive log fully records the application interface call sequence and intermediate state snapshots during the decision-making process, and supports going back along the time axis to view the data status at any historical moment.

[0105] The node trajectories, entity decay curves, and timestamp offsets output during the source tracing path construction phase are directly used as input data for the multimodal report. The visualized graph and interactive log are linked through a unified timeline. Clicking on a graph node can locate the corresponding operation record in the log, ultimately generating a multidimensional source tracing report. The multidimensional source tracing report consists of three parts: a structured text report, a visualized force-oriented graph, and an interactive log. These three parts are linked through a unified timeline to achieve collaborative analysis and backtracking of text descriptions, graph displays, and operation records.

[0106] S5. Based on multi-dimensional source tracing reports, the intelligent decision fusion engine is triggered by confidence thresholds to generate structured enhanced interactive data;

[0107] The sliding window algorithm is used to process the distribution data of historical semantic validation confidence and semantic consistency confidence. The sliding window algorithm traverses the historical data in a fixed time window, calculates the mean of semantic validation confidence and semantic consistency confidence within each window, and uses the mean minus one standard deviation as a dynamic baseline value to generate the semantic dimension threshold;

[0108] Historical scene classification confidence data is initialized based on semantic dimension thresholds, and the probability density distribution of these data is fitted using kernel density estimation. Kernel density estimation smooths the data using a Gaussian kernel function, generates a probability density curve, and identifies the inflection point value of the probability density decrease in the curve to generate a context dimension threshold.

[0109] The momentum gradient descent algorithm is used to process time-series data containing the product of intent integrity confidence and accuracy confidence. The algorithm iteratively calculates the gradient of the product value, tracks its fluctuation boundaries, and locks in a lower bound to generate an intent dimension threshold.

[0110] The system reads semantic confidence, contextual confidence, and intent confidence from the multi-dimensional source tracing report, and compares the semantic confidence with the semantic dimension threshold, the contextual confidence with the contextual dimension threshold, and the intent confidence with the intent dimension threshold. When the confidence value of any dimension is less than the corresponding threshold, a signal is triggered to activate the intelligent decision fusion engine.

[0111] Dependency parsing is performed on the original text. Dependency parsing based on deep learning is used for automatic annotation to generate a set of syntactic dependencies as dependency annotations. Entity attributes related to dependency annotations are extracted from the knowledge graph to generate structured entity attribute vectors. Dependency annotations and associated entity annotations are concatenated using a feature dimension concatenation method to obtain an enhanced semantic representation vector.

[0112] The scene classification results are mapped from discrete categories to integer index sequences through ordered label encoding, and then transformed into category encoding vectors. The context labels are concatenated into the original context vector according to the feature dimension. The scene category encoding vector and the original context vector are weighted and summed to obtain the context enhancement matrix.

[0113] The intent label and intent verification result are concatenated into the original intent vector using a feature dimension concatenation method, and then the intent label and verification result are merged through a weighted fusion operation to obtain the optimized intent representation vector.

[0114] The semantic enhancement vector, context enhancement matrix, and intent enhancement vector are decomposed into a unified field structure according to the feature dimension using a standardized feature dimension decomposition method. The entity attributes of the semantic enhancement vector and the environmental parameters of the context enhancement matrix are stored independently, and the intent enhancement vector is aligned with the scene classification probability and used as a confidence field.

[0115] The aligned field structure is concatenated using a feature dimension key-value pair mapping method, and the triples of the semantic enhancement vector are transformed into key-value pairs; the numerical fields of the context enhancement matrix are stored independently without merging semantic entity attributes; the intent enhancement vector is inserted as an independent field, and the concatenated fields are integrated using a nested JSON construction method to generate a nested JSON structure. The industrial IoT data standard is adopted to generate structured enhanced interactive data.

[0116] S6. Based on structured enhanced interactive data, multimodal interaction is achieved through dynamic traceability visualization interface rendering and tactile coding feedback from smart wearable devices, and a complete interaction log is generated.

[0117] The enhanced interaction data is input into the multimodal distributor, and the data field decoupling is performed. The distributor receives the semantic enhancement field, timestamp correction mark, and intent metadata field from the enhanced interaction data. The semantic enhancement field generates a priority queue through the entity association strength weight sorting algorithm. Each element in the queue consists of a binding pair between the dependency path and the knowledge graph entity ID. The priority weight is determined by the reciprocal of the entity association strength value and the square of the dependency path attention weight. The timestamp correction mark is aligned with the time axis of the original voice data using a sliding window mechanism. The window size is dynamically adjusted according to the average response delay of the most recent interactions. The timestamps within the window are reassigned to coordinates through linear interpolation to eliminate temporal deviations between multimodal data. The intent metadata field is parsed by the rule engine to generate an intent state transition matrix. The rows and columns of the intent state transition matrix correspond to the current intent label and the next scene classification label, respectively. The matrix element values ​​are statistical values ​​of the number of rule triggers, which are normalized and mapped to the 0-1 interval.

[0118] A three-layer visualization rendering engine is constructed. The entity relationship layer inputs a priority queue into the force-directed graph generator. The repulsion coefficient between nodes is dynamically calculated based on the reciprocal of the entity association strength value. The edge weight is determined by the normalized result of the dependency path attention weight. When the change in entity association strength exceeds the historical standard deviation, a node position interpolation animation is triggered. The animation duration is inversely proportional to the change amplitude, making the visual transition smooth. The temporal trajectory layer inputs the timestamp sequence aligned and corrected by the time axis sliding window into the heatmap generator. A Gaussian kernel density estimation algorithm is used to estimate the context coverage of each time slot. The heat value density calculation formula is:

[0119] D = ln(C+1) × k;

[0120] Where D represents thermal density, C represents context coverage, and k represents dynamic scaling factor;

[0121] After receiving the thermal density, the Canvas2D rendering engine is used to draw the timeline baseline. The transparency decreases linearly with the distance from the current time point. The intention projection layer maps the intention state transition matrix to three-dimensional space through a dimensionality reduction algorithm. The gravity constraint mechanism calculates the spatial distribution based on the intention transition probability and the surface area of ​​the sphere. When the user drags the sphere, the neighboring spheres are rearranged based on the dynamic equilibrium of gravity. The sphere radius is adjusted in real time and is positively correlated with the confidence of the intention dimension. The sphere corresponding to the high confidence intention appears to be larger in size.

[0122] The intent dimension confidence and context coverage are extracted from the enhanced interaction data. JSON data standardization is used to eliminate the difference in units. The parsed data is stored in a dual-channel buffer. The intent data is allocated to the fast access area and the context data is allocated to the large-capacity storage area. The improved similarity algorithm for haptic pattern matching matches the waveform templates in the pre-set haptic library. The dynamic adjustment mechanism automatically corrects the matching weight based on the context coverage and outputs the top three candidate patterns.

[0123] The base frequency is automatically adapted to the preset range based on the confidence level of the intent dimension. The amplitude intensity is dynamically adjusted by combining the exponential function with the scene label. The pulse duration is set according to the average value of the sliding window of the recent interaction response time. The generated parameter group is converted into a standard instruction format after verification. The standard instruction format after verification of the parameter group is processed by the segmentation and encapsulation strategy. Data packets are transmitted in order of intent priority. High-priority instructions can preempt the transmission channel.

[0124] After receiving the parsed command parameters, the linear resonant motor loads the resonant frequency matching table during the initialization phase and establishes the motor coil impedance characteristic curve. During the amplitude control phase, it adjusts the pulse width modulation duty cycle and outputs the target vibration intensity. The timing controller synchronously coordinates the superposition of multi-channel tactile effects and uses a hardware-accelerated interpolation algorithm to perform phase alignment and smooth transition processing on the waveforms of different channels. During this process, the pressure sensor array collects the deformation distribution data of the contact surface in real time and feeds the pressure gradient information back to the calibration loop for dynamic parameter compensation. The calibrated data is synchronously updated to the force feedback rendering engine of the dynamic traceability visualization interface, forming a closed-loop control that links tactile and visual sensations.

[0125] Establish a collaborative control mechanism for tactile feedback and dynamic traceability visualization interface. Use timestamp alignment technology to achieve synchronous triggering of the tactile pulse start edge and the key frame of the visualized particle animation. The contact surface deformation data collected by the pressure sensor array is processed by signal filtering and then input back to the force-guided graph rendering engine of the visualization interface. Dynamically adjust the edge weight parameters of the knowledge graph entity nodes. The calibration command data achieves multi-system collaboration through shared memory.

[0126] The embedded coprocessor monitors coil temperature, current peak and skin impedance fluctuations in real time. When an abnormality occurs, it sequentially performs amplitude attenuation, backup waveform switching and power cut-off operations. The monitoring threshold is dynamically calculated based on historical impedance data and the safety status indicator on the visualization interface is updated synchronously to form a closed-loop control of device operating parameters and user protection.

[0127] The dynamic traceability visualization interface event timestamps and smart wearable device haptic feedback timestamps are synchronized and calibrated via an improved NTP protocol to generate a unified time reference axis. Voice data streams are segmented into independent interaction units via endpoint detection. Each unit is bound to the visual operation path coordinates and haptic feedback parameters within a corresponding time period, forming a time-aligned multimodal data set. The haptic feedback parameters include frequency, amplitude, and duration. Voice units are converted to UTF-8 text, visual operation records are extracted as path coordinate sequences, and haptic feedback parameters are parsed into structured triplet data. These structured triplet data are indexed and associated using shared timestamps, constructing time-stamped heterogeneous data blocks. Based on the W3C ActivityStreams 2.0 specification, voice interaction entities, visual operation entities, and haptic feedback entities are defined and connected through "trigger" and "containment" semantic relationships, enabling data fusion. The system detects timestamp overlaps and conflicts, prioritizing the preservation of voice data time boundaries. The conflict resolution algorithm dynamically adjusts entity relationships based on interaction type weights. The fused data is encoded according to JSON specifications, with core fields including semantic context, device identifier, interaction data entities, and validity evaluation indicators. After attaching SHA-256 checksums, it implements hierarchical storage. Recent logs are stored in an in-memory database, while historical logs are compressed and written to distributed storage. Access policy metadata is attached during encrypted storage, and access requires verification of both device fingerprint and cloud token credentials. Abnormal requests return de-identified digests with differential privacy processing. A standardized query interface is generated through OpenAPI 3.0, supporting natural language time expression retrieval. The query results return replayable data packets, which include the original voice waveform, a visual operation animation timeline, and a haptic parameter instruction set, supporting the reconstruction of the complete interaction process across devices.

[0128] This embodiment also provides a voice interaction system based on artificial intelligence technology, including: a data acquisition module, an annotation module, a parsing module, an analysis module, a fusion module, and an interaction module.

[0129] The acquisition module captures user voice data, simultaneously uses a sound source focusing algorithm to lock the target location, and employs a deep neural network noise reduction model to eliminate environmental noise and output voice data packets.

[0130] The annotation module employs a two-layer collaborative recognition approach combining the main channel Transformer model and the auxiliary channel CTC-Attention hybrid architecture to obtain text with multi-dimensional feature annotations.

[0131] The parsing module, based on text with multi-dimensional feature annotations, performs scene classification and dependency parsing through a bidirectional recurrent neural network, and combines dynamic knowledge graphs to retrieve related entities in real time, generating structured intent data;

[0132] The analysis module performs three-dimensional analysis on structured intent data through an interpretability engine, generating a multi-dimensional source tracing report.

[0133] The fusion module, based on multi-dimensional source tracing reports, triggers an intelligent decision fusion engine through confidence thresholds to generate structured, enhanced interactive data;

[0134] The interaction module, based on structured enhanced interactive data, performs multimodal interaction through dynamic traceability visualization interface rendering and tactile coding feedback from smart wearable devices, and generates a complete interaction log.

[0135] This embodiment also provides a computer device applicable to voice interaction methods based on artificial intelligence technology, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the voice interaction method based on artificial intelligence technology as proposed in the above embodiment.

[0136] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0137] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the voice interaction method based on artificial intelligence technology as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0138] In summary, this invention achieves high-precision recognition and multi-dimensional feature annotation of speech signals by: eliminating environmental noise through a deep neural network noise reduction model; employing a two-layer collaborative recognition architecture combining a main channel Transformer model and an auxiliary channel CTC-Attention; performing scene classification and dependency parsing through a bidirectional recurrent neural network; and combining dynamic knowledge graphs to retrieve related entities in real time, thereby generating structured intent data, thus accurately understanding user intent and providing personalized interactive services.

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A voice interaction method based on artificial intelligence technology, characterized in that: include: Capture user voice data, synchronize sound source focusing algorithm to lock the target location, and use deep neural network noise reduction model to eliminate environmental noise and output voice data packets; A two-layer collaborative recognition model using a hybrid architecture of the main channel Transformer model and the auxiliary channel CTC-Attention is employed to obtain text with multi-dimensional feature annotations. Based on text with multi-dimensional feature annotations, a bidirectional recurrent neural network is used for scene classification and dependency parsing, and dynamic knowledge graphs are combined to retrieve related entities in real time to generate structured intent data. The structured intent data is analyzed in three dimensions using an interpretability engine to generate a multi-dimensional source tracing report. The specific steps are as follows: The structured intent data is input into the interpretability engine. Through hierarchical feature decoupling, the structured intent data is split into three independent encoders: intent dimension encoder channel, entity association dimension encoder channel, and time-series trajectory dimension encoder channel. The split data is processed separately. The semantic confidence vector, entity association strength vector and temporal feature vector output by the three independent encoders are enhanced by residual connection and then input into the Gram-Schmidt orthogonalization process to eliminate the linear correlation between dimensions and generate mutually independent orthogonal basis vector sets. A multi-dimensional source tracing report is generated based on the orthogonal basis vector sets. Based on multi-dimensional source tracing reports, an intelligent decision fusion engine is triggered by confidence thresholds to generate structured enhanced interactive data; Based on structured enhanced interactive data, multimodal interaction is achieved through dynamic source tracing visualization interface rendering and tactile coding feedback from smart wearable devices, and a complete interaction log is generated.

2. The voice interaction method based on artificial intelligence technology as described in claim 1, characterized in that: The specific steps for eliminating environmental noise using a deep neural network noise reduction model are as follows: A deep neural network noise reduction model is constructed based on a hybrid framework of convolutional neural networks and recurrent neural networks. The spectral features of the beamformed speech signal are extracted by the convolutional neural network layer, and the temporal features of the speech signal are captured by the recurrent neural network layer to distinguish noise. A deep neural network denoising model was trained using paired data of noisy and clean speech, and the difference between the denoised features and the clean speech was calculated using the mean squared error loss function. The Adam optimizer is used to update parameters through backpropagation. After being optimized on the validation set, the trained deep neural network denoising model is deployed, and noise components are estimated and suppressed. The output layer outputs the denoised speech features, which are then encoded to form speech data packets.

3. The voice interaction method based on artificial intelligence technology as described in claim 1, characterized in that: The specific steps of the two-layer collaborative recognition using a hybrid architecture of the main channel Transformer model and the auxiliary channel CTC-Attention are as follows: After preprocessing the speech processing package, it is input into the main channel Transformer model and the auxiliary channel CTC-Attention model for feature encoding. The Transformer encoding process generates a context representation through multi-head self-attention, then processes it through residual connections and layer normalization, and finally performs residual connections and layer normalization a second time after nonlinear transformation through a feedforward neural network, ultimately outputting an encoded feature sequence. The interaction weights between each character in the target text and the encoded feature sequence are dynamically calculated through the attention mechanism to generate a context vector. The time-step features of the CTC-Attention model are concatenated with the attention context vector to generate an optimized feature sequence. The main channel and auxiliary channel are reconstructed into a global semantic strategy layer and a local alignment execution layer through functional decoupling. The global semantic strategy layer focuses on cross-frame semantic association, while the local alignment execution layer strengthens the alignment confidence at the time step granularity. A gated residual fusion strategy is adopted to dynamically fuse global semantic features and local alignment features to generate multi-dimensional feature-annotated text.

4. The voice interaction method based on artificial intelligence technology as described in claim 1, characterized in that: The specific steps for combining dynamic knowledge graphs with real-time retrieval of related entities to generate structured intent data are as follows: Extract scene classification results and dependency relationship information from text with multi-dimensional feature annotations, dynamically combine them into query conditions, and retrieve related entities in the knowledge graph in real time; Relevant entities are filtered through entity matching and similarity calculation, and then integrated into structured intent data in JSON format.

5. The voice interaction method based on artificial intelligence technology as described in claim 1, characterized in that: The specific steps for triggering the intelligent decision fusion engine through a confidence threshold to generate structured enhanced interactive data are as follows: Set thresholds for semantic dimension, context dimension, and intent dimension, and compare them with the confidence level of each dimension to dynamically trigger the intelligent decision fusion engine; Feature fusion is performed on dependency relationship annotations and associated entity annotations in the semantic dimension to enhance the semantic representation vector; context integration is performed on scene classification results and context labels in the context dimension to enhance the context representation vector; and intent labels and intent verification results in the intent dimension are merged through weighted fusion operation to optimize the intent representation vector. The data is aggregated to integrate the semantic, contextual, and intent dimensions, and then structured, enhanced interactive data is generated using JSON encoding.

6. The voice interaction method based on artificial intelligence technology as described in claim 1, characterized in that: The process involves multimodal interaction through dynamic source tracing visualization interface rendering and tactile coding feedback from smart wearable devices, generating a complete interaction log. The specific steps are as follows: The dynamic traceability visualization interface is based on enhanced interactive data rendering of a three-layer interactive view. The smart wearable device tactile coding engine converts the entity association strength and intent confidence into dual-channel tactile signals. The tactile pulses and key frames of the visual animation are linked by time stamp synchronization, and the pressure sensor data is used to adjust the interface mechanical parameters in real time to form a closed-loop interaction. The timeline-aligned voice data stream, visual operation path, and haptic parameter triplet are encoded according to the W3CActivityStreams 2.0 specification and verified with SHA-256 to generate a complete interaction log that can be restored across devices.

7. A voice interaction system based on artificial intelligence technology, based on the voice interaction method based on artificial intelligence technology according to any one of claims 1 to 6, characterized in that: include: The module includes a data acquisition module, a labeling module, a parsing module, an analysis module, a fusion module, and an interaction module. The acquisition module captures user voice data, simultaneously uses a sound source focusing algorithm to lock the target location, and employs a deep neural network noise reduction model to eliminate environmental noise and output voice data packets. The annotation module employs a two-layer collaborative recognition approach combining the main channel Transformer model and the auxiliary channel CTC-Attention hybrid architecture to obtain text with multi-dimensional feature annotations. The parsing module, based on text with multi-dimensional feature annotations, performs scene classification and dependency parsing through a bidirectional recurrent neural network, and combines dynamic knowledge graphs to retrieve related entities in real time, generating structured intent data; The analysis module performs three-dimensional analysis on structured intent data through an interpretability engine, generating a multi-dimensional source tracing report. The fusion module, based on multi-dimensional source tracing reports, triggers an intelligent decision fusion engine through confidence thresholds to generate structured, enhanced interactive data; The interaction module, based on structured enhanced interactive data, performs multimodal interaction through dynamic traceability visualization interface rendering and tactile coding feedback from smart wearable devices, and generates a complete interaction log.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the voice interaction method based on artificial intelligence technology as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the voice interaction method based on artificial intelligence technology as described in any one of claims 1 to 6.

Images