A Large-Model Real-Time Speech-Text Intent Recognition Method and System Based on Text Prediction

Abstract

This invention discloses a real-time speech-text intent recognition method and system based on a large-scale model using text prediction. The method includes: acquiring real-time speech data from a user; performing real-time speech recognition processing through a streaming speech recognition interface to obtain partially transcribed text; inputting the partially transcribed text into a masked language model for text prediction to generate several highly reliable complete sentence candidates; based on these candidates, inputting them in parallel into a large-scale language model for intent recognition to obtain corresponding intent results and establishing a mapping relationship between candidate sentences and intent recognition results; acquiring the user's complete sentence; calculating the similarity between the complete actual sentence and the multiple highly reliable complete sentence candidates using a multi-level text similarity algorithm; selecting the candidate sentence with the highest similarity score; and directly obtaining the corresponding final intent recognition result based on the mapping relationship. This invention aims to solve the technical problem of high response latency in existing speech intent recognition systems.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a real-time speech-text intent recognition method and system based on a large model of text prediction, which can be applied to scenarios such as intelligent customer service, voice assistants, and human-computer interaction that require real-time processing and understanding of user voice commands. Background Technology

[0002] With the rapid development of artificial intelligence technology, speech recognition and natural language processing technologies have been widely used in fields such as intelligent customer service, voice assistants, and human-computer interaction. Modern voice interaction systems typically need to convert the user's voice input into text, then use a natural language understanding module to identify the user's intent, and finally provide corresponding service responses.

[0003] Currently, typical speech intent recognition systems mainly employ two technical approaches: one is based on traditional machine learning methods, achieving intent recognition through feature engineering and classification algorithms; the other is an end-to-end approach based on deep learning, directly predicting user intent from speech signals or text input. In recent years, with breakthroughs in large language model technology, more and more systems have begun to adopt large models based on the Transformer architecture for intent understanding. These methods can better handle complex semantic information and contextual relationships.

[0004] Current state-of-the-art speech intent recognition systems typically employ a serial processing architecture: first, an automatic speech recognition (ASR) system converts the speech into complete text; then, the complete text is input into a large language model for intent analysis. This architecture works by waiting for the user to finish speaking before the ASR system outputs the complete transcribed text. The large language model then performs semantic understanding and intent classification based on this complete text, ultimately outputting a structured intent recognition result.

[0005] However, existing technologies suffer from significant latency issues: the system can only begin intent recognition after the user has finished speaking the entire sentence. Combined with the inference time of large language models, the overall response latency often reaches several seconds, severely impacting the user's interactive experience. This latency is particularly problematic in applications requiring rapid responses, such as real-time customer service conversations and emergency command processing, leading to a decline in user experience and reduced business efficiency. Summary of the Invention

[0006] The purpose of this invention is to provide a real-time speech-text intent recognition method and system based on large-scale text prediction, which aims to solve the technical problem of high response latency in existing speech intent recognition systems, and to predict and recognize the user's possible intent in advance before the user has finished speaking, thereby significantly reducing system response time and improving user interaction experience.

[0007] To achieve the above objectives, this invention provides a large-scale real-time speech-text intent recognition method based on text prediction, comprising: acquiring real-time speech data from a user; performing real-time speech recognition processing through a streaming speech recognition interface to obtain partially transcribed text; inputting the partially transcribed text into a masked language model for text prediction based on the partially transcribed text to generate several highly reliable complete sentence candidates; inputting the complete sentence candidates in parallel into a large-scale language model for intent recognition based on the complete sentence candidates to obtain corresponding intent results, and establishing a mapping relationship between candidate sentences and intent recognition results; acquiring the user's complete expressed sentence; calculating the similarity between the complete actual sentence and the multiple highly reliable complete sentence candidates using a multi-level text similarity algorithm; selecting the candidate sentence with the highest similarity score; and directly obtaining the corresponding final intent recognition result based on the mapping relationship.

[0008] As a preferred embodiment, the step of performing real-time speech recognition processing through a streaming speech recognition interface to obtain partially transcribed text includes: establishing a real-time audio stream connection and adopting a sampling strategy with fixed time intervals to collect audio segments from the real-time speech data according to a preset time window, generating a continuous audio segment sequence; based on the continuous audio data segments, performing acoustic feature extraction and temporal information decoding through a streaming ASR engine to generate real-time text output; and for the real-time text output, using a dynamic sliding window mechanism to cache the most recent text segments and perform standardization processing to obtain the partially transcribed text.

[0009] As a preferred embodiment, the acquisition of the user's real-time voice data and the real-time speech recognition processing through a streaming speech recognition interface include: based on the real-time voice data, using an encoder-decoder architecture based on an attention mechanism to perform real-time decoding processing of audio features, achieving synchronous processing of audio input and text output, and obtaining a synchronous processing result; based on the synchronous processing result, performing text cleaning and standardization processing to remove filler words and recognition errors, ensuring that the text quality meets the requirements of subsequent processing, and generating a cleaned text segment; based on the cleaned text segment, when the text quality meets the requirements and the length reaches a preset minimum threshold, automatically triggering the subsequent text prediction process.

[0010] As a preferred embodiment, the step of inputting the partially transcribed text into a masked language model for text prediction to generate several high-confidence complete sentence candidates includes: using the partially transcribed text as contextual information, adding a mask marker to the end of the text and using a bidirectional attention mechanism for prediction to generate multiple possible sentence completion schemes; calculating the confidence score of each candidate sentence based on the sentence completion scheme to ensure the quality and reliability of the generated candidates; and retaining the top N candidates with the highest confidence based on the confidence scores to obtain the complete sentence candidates.

[0011] As a preferred embodiment, the step of inputting the partially transcribed text into a masked language model for text prediction includes: based on the partially transcribed text, using a bidirectional attention mechanism to simultaneously process preceding and following inference information to perform comprehensive semantic prediction and generate text completion results; for the text completion results, using a parallel processing method to generate multiple complete sentence structure schemes to ensure the grammatical and semantic quality of candidate sentences; based on the multiple complete sentence structure schemes, comprehensively considering the model probability distribution and semantic coherence to perform screening and output the complete sentence candidates.

[0012] As a preferred embodiment, the step of inputting the partially transcribed text into a masked language model for text prediction to generate several highly reliable complete sentence candidates includes: based on the partially transcribed text, performing text prediction using a binary weighted spike neural network instead of a traditional masked language model to generate candidate sentences encoded by time-series pulse signals; for the candidate sentences encoded by time-series pulse signals, using a sign function to quantize the weights into a binary form of {-1, +1} to obtain binary weight processing results; based on the binary weight processing results, using multiple parallel BWSNN sub-networks to predict different sentence types to obtain the complete sentence candidates.

[0013] As a preferred embodiment, the step of predicting different sentence types through multiple parallel BWSNN sub-networks to obtain the complete sentence candidates includes: constructing interrogative sentence prediction sub-networks, declarative sentence prediction sub-networks, and imperative sentence prediction sub-networks based on the binary weight processing results; employing a pulse temporal coding strategy to encode semantic information using the firing time and frequency of neurons to obtain progressive text completion results; for the progressive text completion results, weighting and fusing the prediction results of multiple sub-networks through an integrated decision fusion module to generate a preliminary candidate sentence list; and based on the preliminary candidate sentence list, employing a dynamic weight allocation strategy and a confidence evaluation mechanism to obtain the complete sentence candidates, ensuring the quality and reliability of the complete sentence candidates.

[0014] As a preferred embodiment, the step of inputting the candidate complete sentences into a large language model in parallel for intent recognition to obtain the corresponding intent result includes: constructing a constrained multi-objective optimization problem based on the candidate complete sentences to simultaneously optimize recognition accuracy, response time, and resource consumption, and identifying constrained Pareto fronts and unconstrained Pareto fronts; intelligently combining a fast classifier, a medium-complexity model, and a large language model using hybrid operator technology based on the constrained Pareto fronts and the unconstrained Pareto fronts to obtain a multi-strategy fusion result; and adjusting computing resources in real time according to system load and user priority through a dynamic resource allocation mechanism based on the multi-strategy fusion result to obtain the corresponding intent result.

[0015] As a preferred embodiment, the identification of constrained Pareto fronts and unconstrained Pareto fronts includes: identifying the optimal set of intent recognition schemes and constructing a constrained Pareto front to represent feasible schemes within the time limit based on strict delay constraints, quantifying the impact of delay constraints on recognition performance; identifying the ideal set of recognition schemes and constructing an unconstrained Pareto front to pursue the highest recognition accuracy based on ideal conditions without considering delay constraints, establishing a performance benchmark reference standard; and establishing a classification-based UPF-CPF relationship strategy based on the constrained Pareto fronts and unconstrained Pareto fronts to determine the dominance relationship between schemes, ensuring the identification of the optimal solution in the multi-objective optimization process.

[0016] As a preferred embodiment, the step of calculating the similarity between the complete actual sentence and the multiple high-confidence complete sentence candidates using a multi-level text similarity algorithm, selecting the candidate sentence with the highest similarity score, and directly obtaining the corresponding final intent recognition result based on the mapping relationship includes: based on the user's complete expression sentence and the complete sentence candidates, using a multi-objective evolutionary algorithm to decompose the similarity calculation into a multi-objective optimization problem, decomposing it into four sub-objectives: lexical level, syntactic level, semantic level, and contextual relevance, obtaining four sub-objective optimization results; based on the four sub-objective optimization results, constructing a dynamic adaptive weight vector generation mechanism to adjust the sub-objective weights according to the feature difference pattern, realizing a personalized similarity evaluation strategy, obtaining a personalized similarity evaluation result; based on the personalized similarity evaluation result, selecting the candidate sentence with the highest overall similarity score that exceeds the confidence threshold, and directly obtaining the corresponding final intent recognition result.

[0017] As a preferred embodiment, the dynamic adaptive weight vector generation mechanism adjusts the sub-target weights according to the feature difference pattern, including: based on the optimization results of the four sub-targets, designing a multi-scale attention estimator to calculate attention weights in parallel at three granular levels: word level, phrase level, and sentence level, capturing semantic correspondences at different levels, avoiding semantic information missed by single-granularity analysis, and obtaining a multi-scale attention weight distribution; based on the multi-scale attention weight distribution, implementing an online weight vector learning algorithm based on historical feedback data, continuously optimizing the weight distribution strategy, and obtaining an optimized weight distribution strategy; based on the optimized weight distribution strategy, establishing a Pareto optimal solution set maintenance mechanism to store the optimal matching scheme under multiple weight configurations, supporting personalized and scenario-based matching scheme selection, and obtaining the personalized similarity evaluation result.

[0018] As a preferred embodiment, obtaining the corresponding intent result includes: identifying the user's true intent category and extracting key entity information and important parameters based on the semantic analysis of the complete sentence candidates, and outputting a structured intent recognition result; establishing a mapping relationship between candidate sentences and intent recognition results based on the structured intent recognition result and storing it in a high-speed cache to support fast retrieval and matching; and outputting a complete intent result containing intent category identifiers, key entity information and operation parameters based on the mapping relationship to provide intelligent voice interaction services.

[0019] As a preferred embodiment, the method further includes a response time optimization mechanism: based on real-time system performance monitoring data, the number of candidates for text prediction and the resource allocation for parallel inference are dynamically adjusted to ensure that the overall response latency is controlled within a preset threshold, thereby obtaining a resource allocation optimization result; based on the resource allocation optimization result, prediction caching and result reuse strategies are adopted to reduce redundant calculations, further reducing system response latency, thereby obtaining latency optimization measures; based on the latency optimization measures, intent prediction and matching preparation before the user's complete expression are realized, significantly improving the user interaction experience.

[0020] This invention also provides a large-scale real-time speech-text intent recognition system based on text prediction, comprising: a streaming speech recognition module for acquiring real-time speech data from a user, performing real-time speech recognition processing through a streaming speech recognition interface to obtain partially transcribed text; a text prediction module for inputting the partially transcribed text into a masked language model for text prediction based on the partially transcribed text, generating several highly reliable complete sentence candidates; a parallel intent recognition module for inputting the complete sentence candidates in parallel into a large-scale language model for intent recognition based on the complete sentence candidates, obtaining corresponding intent results, and establishing a mapping relationship between candidate sentences and intent recognition results; and a similarity matching module for acquiring the user's complete expressed sentence, calculating the similarity between the complete actual sentence and the multiple highly reliable complete sentence candidates through a multi-level text similarity algorithm, selecting the candidate sentence with the highest similarity score, and directly obtaining the corresponding final intent recognition result based on the mapping relationship.

[0021] The beneficial effects of this invention are: by predicting possible complete sentences based on the identified partial text before the user has completed a full speech expression, and performing intent recognition in advance, the system response latency is significantly reduced, and the user interaction experience is improved; by adopting streaming speech recognition and parallel intent recognition, efficient collaborative processing of speech recognition and intent understanding is achieved; through a multi-level text similarity matching algorithm, the accuracy and reliability of the final intent recognition result are ensured; the overall system architecture is flexible and efficient, and can adapt to different application scenarios and user needs. Attached Figure Description

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

[0023] Figure 1 This is a flowchart illustrating the real-time speech-text intent recognition method based on a large model for text prediction provided by the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of the large-scale real-time speech-text intent recognition system based on text prediction provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0026] like Figure 1 As shown, this invention provides a large-scale real-time speech-text intent recognition method based on text prediction, comprising the following steps:

[0027] S1: Acquire the user's real-time voice data, perform real-time voice recognition processing through the streaming speech recognition interface, and obtain partial transcribed text;

[0028] S2: Based on the partially transcribed text, input the partially transcribed text into the masked language model for text prediction to generate several highly reliable complete sentence candidates;

[0029] S3: Based on the complete sentence candidates, input the complete sentence candidates into the large language model in parallel for intent recognition, obtain the corresponding intent results, and establish a mapping relationship between the candidate sentences and the intent recognition results;

[0030] S4: Obtain the complete sentence expressed by the user, calculate the similarity between the complete actual sentence and the multiple highly credible complete sentence candidates through a multi-level text similarity algorithm, select the candidate sentence with the highest similarity score, and directly obtain the corresponding final intent recognition result based on the mapping relationship.

[0031] Specifically, in step S1, raw data about the enterprise is first obtained from multiple channels. This data covers various dimensions, including the enterprise's financial and operational status, R&D capabilities, and industry position. After obtaining the raw data, this module applies a series of data preprocessing techniques to clean and normalize the data. During data cleaning, the system identifies and handles outliers, missing values, and duplicate data to ensure data quality and integrity. Subsequently, through normalization techniques, indicator data with different dimensions and ranges are transformed into a unified standard range, making the data from each dimension comparable. The application of this preprocessing technique can effectively eliminate data scale differences and improve the accuracy and fairness of subsequent analysis. Finally, the data acquisition module outputs a standardized multidimensional feature dataset of the enterprise, laying the foundation for subsequent indicator system construction and evaluation calculations.

[0032] The streaming speech recognition interface is an application programming interface (API) that supports real-time audio data input and real-time text output. By establishing a persistent network connection, it enables continuous processing of speech data and instant text return. Unlike traditional batch speech recognition, streaming recognition does not wait for complete audio submission before processing begins; instead, it performs recognition immediately upon arrival of the audio data, significantly reducing recognition latency. In its implementation, a network connection supporting bidirectional data transmission, such as WebSocket or gRPC, is first established to ensure that audio data can flow into the server in real time, while processing results are returned instantly. Subsequently, the system employs a fixed-time sampling strategy, typically in units of 10-100 milliseconds, to segment the continuous audio stream into a series of short audio segments, forming processable data units. This segmented processing method ensures both real-time recognition and maintains sufficient contextual information to improve recognition accuracy.

[0033] Partially transcribed text refers to incomplete text content obtained through speech recognition or text capture technologies, containing only a portion rather than the entirety of the original communication or audio content. This partiality can stem from various reasons: recognition gaps due to technological limitations, such as background noise interference, difficulties in accent recognition, or equipment malfunction; truncation sampling, where only specific parts of the communication are intentionally recorded; real-time processing limitations, resulting in omissions due to the system's inability to process all content; or privacy protection measures, where certain content is selectively omitted to protect sensitive information. In text similarity calculation and intent matching systems, partially transcribed text constitutes a special input type, requiring the ability to make reasonable inferences and matches based on incomplete information. This necessitates algorithms that are robust to handling information gaps, capable of extracting key semantic features from limited content, and appropriately reducing reliance on completeness. When processing partially transcribed text, techniques such as context inference, keyword weight enhancement, and uncertainty modeling are typically employed to provide the best possible matching results given limited recognition capabilities.

[0034] In step 2, based on a standardized enterprise multidimensional feature dataset, an improved WP-PVC algorithm (Weighted Principal Component-Principal Variable Combination Algorithm) is used to construct a scientifically sound indicator system. The WP-PVC algorithm is an advanced statistical method combining principal component analysis and variable clustering, effectively identifying correlations between indicators and grouping them appropriately. The improved WP-PVC algorithm, by introducing an adaptive weight adjustment mechanism, can more accurately capture the nonlinear relationships between indicators. In practical applications, the algorithm first performs correlation analysis on all indicators, clustering highly correlated indicators to reduce redundancy; then, it calculates the weight coefficient based on the importance of each indicator in its respective dimension. Through this process, the module ultimately constructs a comprehensive evaluation indicator system encompassing three dimensions: financial performance, R&D capability level, and industry position and brand effect. Each dimension has multiple sub-indicators with clearly defined weights, forming a well-structured, scientifically sound, multi-level evaluation framework.

[0035] In step S3, after obtaining the multi-dimensional evaluation index system, the dual-standard WP-PVC algorithm is used for deeper analysis and scoring calculation. The dual-standard WP-PVC algorithm is a further improvement on the traditional WP-PVC algorithm. It achieves a more comprehensive enterprise qualification assessment by introducing two evaluation standards: industry standard reference and historical data comparison. The algorithm first analyzes the correlation between various indicators, constructs an indicator correlation network, and identifies the direct and indirect influence relationships between indicators. Then, for each indicator, the system compares and analyzes it from two dimensions: the industry average level and the enterprise's historical performance, calculating a relative score. During the calculation process, the system weights and summarizes the scores of each sub-item according to the indicator weights, ultimately generating a comprehensive enterprise qualification score and multi-dimensional scoring results. This dual-standard scoring mechanism can simultaneously reflect the enterprise's relative position in the industry and its own dynamic development changes, providing a more scientific methodological foundation for the comprehensive assessment of enterprise qualifications.

[0036] In step S4, based on the comprehensive score of enterprise qualifications, and combined with the Random Online Multinomial Logarithmic Algorithm (LOLA), intelligent matching between enterprises and various technology projects is achieved. The LOLA is a highly efficient combinatorial optimization algorithm capable of quickly finding the optimal match from a large number of options. The algorithm first constructs feature vectors for enterprise qualifications and project requirements, calculating the multidimensional similarity between them. Then, it estimates the probability of success for each project using a LOLA model and sorts them according to the probability values. The algorithm's "random online" characteristic allows it to adapt to dynamic updates in the project database; when new projects are added, the system can adjust the matching results in real time without recalculating all the data. Finally, this module outputs a project matching recommendation list containing project name, matching score, and predicted application success rate, providing enterprises with accurate project application suggestions and helping them concentrate their limited application resources on projects with the highest probability of success.

[0037] In this embodiment, step S1 involves real-time speech recognition processing via a streaming speech recognition interface to obtain partially transcribed text. Specifically, this includes: establishing a real-time audio stream connection and employing a sampling strategy with fixed time intervals to collect audio segments from the real-time speech data according to a preset time window, generating a continuous audio segment sequence; based on the continuous audio data segments, performing acoustic feature extraction and temporal information decoding through a streaming ASR engine to generate real-time text output; and for the real-time text output, using a dynamic sliding window mechanism to cache the most recent text segments and perform standardization processing to obtain the partially transcribed text.

[0038] Specifically, a real-time audio stream connection is established, and a sampling strategy with fixed time intervals is adopted to collect audio segments from the real-time speech data according to a preset time window, generating a continuous audio segment sequence. In actual implementation, the system first establishes a bidirectional communication channel between the client and the server through a network protocol that supports persistent connections, such as WebSocket or HTTP / 2, to ensure that audio data can be continuously and stably transmitted from the user device to the recognition server. After the connection is established, audio sampling parameters are configured, including sampling rate (usually 16kHz or 44.1kHz), bit depth (usually 16-bit), and number of channels (usually mono) to ensure the audio quality required for speech recognition. Subsequently, the system samples the continuous audio stream in segments at fixed time intervals (usually 20ms to 100ms). Each sampling captures a preset time window (usually 0.5 seconds to 2 seconds) of audio data, forming an audio segment. These continuously generated audio segments are organized into an audio segment sequence according to time order, preserving the temporal relationship of the original speech. To ensure a smooth transition between audio segments, the system typically sets a certain overlap area between adjacent segments (usually 10% to 25% of the window length) to avoid losing crucial information at audio boundaries during segmentation. Simultaneously, the system monitors network conditions and audio quality in real time, dynamically adjusting the sampling strategy. In cases of network congestion, it may increase the sampling interval or decrease the window size to maintain the continuity of the recognition service.

[0039] Based on the continuous audio data segments, a streaming ASR engine performs acoustic feature extraction and temporal information decoding to generate real-time text output. The streaming ASR engine is a processing system specifically designed for real-time speech recognition, capable of processing audio data as it arrives continuously, without waiting for the complete speech to finish. In the acoustic feature extraction stage, the system first applies a pre-emphasis filter to each audio segment to enhance the high-frequency components, then performs frame segmentation (typically 25ms frame length, 10ms frame shift), and applies a Hamming window function to each frame to reduce spectral leakage. Subsequently, the system calculates the short-time Fourier transform (STFT) of each frame to obtain the power spectrum, then maps it to the frequency scale perceived by the human ear using a Mel filter bank, finally extracting the Mel frequency cepstral coefficients (MFCC) or filter bank energy features (FBANK) as acoustic features. In the temporal information decoding stage, the streaming ASR engine employs model architectures such as recurrent neural networks (RNN), long short-term memory networks (LSTM), or Transformers, combining acoustic and language models to convert the acoustic feature sequence into a probability distribution of phonemes or characters. To adapt to the characteristics of streaming processing, decoders typically employ forward algorithms or improved beam search strategies to make decoding decisions within a limited historical context. During decoding, the internal state is continuously updated and maintained to ensure that newly arriving audio segments can form a coherent interpretation with previous content. Furthermore, to improve real-time recognition performance, the system may employ a multi-stage decoding strategy, first generating preliminary results and then refining them as more context is accumulated, thereby minimizing output latency while maintaining accuracy.

[0040] For the real-time text output, a dynamic sliding window mechanism is used to cache the most recent text segments and perform standardization processing to obtain the partially transcribed text. The dynamic sliding window mechanism is a special text processing technique designed to balance real-time requirements and text quality requirements. A variable-size text buffer is maintained to store the most recent recognition result segments. The initial size of this window is typically set to contain the length of 3-5 complete sentences. As the recognition process progresses, the window slides forward, discarding the oldest content and incorporating newly recognized text. The window size is not fixed; the system dynamically adjusts the window range based on the semantic integrity of the speech content. For example, when a complete semantic unit (such as the end of a natural paragraph) is detected, the window may shrink to release the processed content; while when processing complex long sentences, the system may expand the window to obtain richer contextual information. For the text within the window, a series of standardization processes are performed, including the normalized insertion of punctuation marks (based on speech pauses and intonation changes), the unified representation conversion of numbers and units (such as converting "one thousand two hundred" to "1200"), the expansion processing of common abbreviations and idioms, and grammatical error correction based on a language model. Furthermore, it identifies and removes common filler words, repetitions, and incomplete expressions in the speech, ensuring that the final output partially transcribed text has a standardized format and good readability. In high-noise environments or when the user's speech rate is abnormal, the window processing strategy is adjusted accordingly, such as extending the window retention time or enhancing the error correction weights of the language model, to adapt to complex recognition scenarios. Through this dynamic sliding window mechanism, the system can extract structured and standardized text content from a continuous speech stream, providing high-quality input for subsequent semantic understanding and content processing.

[0041] In this embodiment, step S1 involves acquiring the user's real-time voice data and performing real-time voice recognition processing through a streaming speech recognition interface. Specifically, this includes: based on the real-time voice data, using an encoder-decoder architecture based on an attention mechanism to perform real-time decoding processing of audio features, achieving synchronous processing of audio input and text output, and obtaining a synchronous processing result; based on the synchronous processing result, performing text cleaning and standardization processing to remove filler words and recognition errors, ensuring that the text quality meets the requirements of subsequent processing, and generating a cleaned text segment; based on the cleaned text segment, when the text quality meets the requirements and the length reaches a preset minimum threshold, automatically triggering the subsequent text prediction process.

[0042] The system continuously receives user voice input via a streaming speech recognition interface, obtaining partial transcribed text in real time. During the audio acquisition phase, the system first establishes a stable real-time audio stream connection to ensure the continuity and integrity of the audio data. The audio acquisition module employs a fixed-time sampling strategy, typically set to a 100-millisecond time window. This ensures both real-time performance requirements and sufficient audio information for subsequent recognition processing.

[0043] In the speech recognition processing stage, the acquired audio segments are input into a specially optimized streaming ASR engine. This engine employs an encoder-decoder architecture based on an attention mechanism, enabling real-time decoding of audio features. The encoder is responsible for extracting the acoustic features and temporal information of the audio, while the decoder generates the corresponding text output based on these features. This architectural design allows the system to output text simultaneously with audio input, achieving true streaming processing.

[0044] To ensure the continuity and accuracy of text processing, the ASR engine maintains a dynamic sliding window mechanism to cache recent text segments while outputting the currently recognized partial text. This sliding window not only saves historical text information but also provides richer contextual information for subsequent text prediction. The text preprocessing module performs necessary cleaning and standardization on these partial texts, including removing filler words, colloquial expressions, and obvious recognition errors that may occur during speech recognition, ensuring the quality of the text input to subsequent modules.

[0045] When the length of a portion of the text reaches the preset minimum threshold, typically set to 3-5 effective words, the system will automatically trigger the subsequent text prediction process to prepare for the next stage of processing.

[0046] Specifically, the encoder-decoder architecture based on attention mechanisms has become a mainstream technical framework in the field of speech recognition in recent years, effectively handling the sequence conversion problem between audio and text. In this architecture, the encoder is responsible for mapping the input audio feature sequence to hidden representations, while the decoder generates the corresponding text sequence based on these hidden representations. The attention mechanism, acting as a bridge between the encoder and decoder, helps the decoder adaptively focus on relevant parts of the input audio sequence when generating each text unit, thereby effectively capturing long-range dependencies. In streaming scenarios, this architecture requires special design to support real-time processing. Common improvements include using a unidirectional attention mechanism (focusing only on what has been seen), employing a block-based processing strategy (processing small chunks of audio at a time), and introducing a context caching mechanism (saving historical information). Through these optimizations, the system can achieve low-latency synchronous processing of audio input and text output while maintaining high recognition accuracy.

[0047] Text cleaning and standardization are crucial steps in ensuring the quality of speech recognition results, especially in streaming scenarios where raw recognition results often contain various imperfections. Text cleaning primarily addresses three types of issues: filler word removal, error correction, and duplication removal. Filler words such as "um," "ah," and "that" typically have no actual semantic value in formal text and are identified and filtered using a pre-defined filler word list. Error correction relies on language models and contextual information, employing n-gram probabilistic or neural network models to re-evaluate and potentially replace low-confidence words. Duplication removal addresses the common word repetition phenomenon in streaming recognition by comparing overlapping portions of adjacent text segments to identify and remove redundant content. Standardization focuses on format consistency, including standardizing punctuation usage, normalizing number and unit representations, and converting special symbol expressions to ensure a consistent and standardized final text format for easier subsequent processing and understanding.

[0048] Setting a minimum threshold is an important quality control mechanism used to determine when to trigger subsequent processing flows, balancing real-time and accuracy requirements. This threshold typically includes two dimensions: text quality and text length. Text quality metrics are usually based on a comprehensive evaluation of multiple factors during the recognition process, such as model confidence, grammatical completeness, and semantic coherence. A minimum quality score is set; only when the quality score of the current text segment exceeds this score is it considered for proceeding to the next stage. The text length threshold ensures sufficient contextual information has been accumulated and is typically set as the smallest unit capable of expressing complete semantics, such as a complete sentence or a phrase containing key information points. When the cleaned text segment simultaneously meets both quality and length conditions, the subsequent text prediction process is automatically triggered, initiating higher-level semantic understanding and content generation. This dual-threshold triggering mechanism effectively balances user experience and system performance, avoiding premature processing of incomplete information or delayed processing of sufficient information.

[0049] In this embodiment, step S2 involves inputting the partially transcribed text into a masked language model for text prediction to generate several high-confidence complete sentence candidates. Specifically, this includes: using the partially transcribed text as contextual information, adding a mask marker to the end of the text and using a bidirectional attention mechanism for prediction to generate multiple possible sentence completion schemes; calculating the confidence score of each candidate sentence based on the sentence completion scheme to ensure the quality and reliability of the generated candidates; and retaining the top N candidates with the highest confidence based on the confidence scores to obtain the complete sentence candidates.

[0050] In the text prediction stage, the identified portion of the text is input into a specially trained masked language model, and multiple highly reliable complete sentence candidates are generated through advanced prediction algorithms. First, a portion of the text is input into the model as important contextual information, and special mask markers are added to the end of the text. These markers explicitly indicate the text portions that the model needs to predict and complete.

[0051] The masked language model employs an advanced bidirectional attention mechanism, unique in that it can simultaneously utilize existing preceding context information and inference information about potential following context for comprehensive prediction. Compared to traditional unidirectional prediction models, the bidirectional attention mechanism can better understand the overall semantic structure and logical relationships of the text, thereby generating more accurate and natural text completion results.

[0052] During the prediction generation process, the model employs parallel processing to generate multiple possible sentence completion schemes. Each generated scheme contains a complete sentence structure and rich semantic information, ensuring that the generated candidate sentences have high quality at both the grammatical and semantic levels. To guarantee the reliability of the prediction results, the system is equipped with a dedicated confidence evaluation module, which comprehensively considers the probability distribution of the model output and the semantic coherence of the generated text, calculating a detailed confidence score for each candidate sentence.

[0053] Based on the confidence assessment results, the candidate sentence selection module retains the top N candidate results with the highest confidence, typically set to 5-10 candidate sentences. This ensures both candidate diversity and controls the computational complexity of subsequent processing. Finally, the system passes the selected candidate sentence list to the parallel intent recognition module, providing high-quality input data for the next stage of processing.

[0054] Specifically, a masked language model is a pre-trained model that learns language representations by predicting masked (i.e., occluded or replaced) text segments. Unlike traditional left-to-right language models, masked language models can simultaneously utilize contextual information before and after the target word for prediction, capturing richer semantic relationships. In this invention, a masked language model is used to predict the possible continuation of speech-transcribed text. Its core idea is to take the identified portion of the transcribed text as deterministic input, add special mask markers (usually represented as [MASK] or [MASK]) to its end, and then let the model predict what these mask markers should be replaced with. This model is typically based on the Transformer architecture and pre-trained on massive amounts of text data, enabling it to learn rich language knowledge and patterns. In practical applications, appropriate pre-trained models, such as BERT, RoBERTa, or their domain-adapted versions, are selected based on the speech scene and domain characteristics, and may be fine-tuned using domain-specific data to improve prediction accuracy in the target scene. The introduction of masked language models enables the system to go beyond simple word sequence prediction, achieving a deep understanding of semantic coherence and contextual relevance.

[0055] Bidirectional attention is a core technology in masked language models, allowing the model to simultaneously focus on all other positions in the sequence while processing information at each position, achieving comprehensive context awareness. Specifically, bidirectional attention generates attention weights by calculating the similarity between the query vector and the key vectors at all positions. Then, it performs a weighted summation of the value vectors based on these weights to obtain a context-aware representation. During prediction, it considers the relationships between each word in the identified text and all other words, as well as the relationships between the masked tokens and the context. This mechanism effectively captures long-distance dependencies and complex semantic structures, making the prediction results more coherent and context-aware. To adapt to the characteristics of streaming prediction, a chunking strategy may be adopted, dividing long texts into multiple overlapping chunks for processing, and then merging the results, thereby reducing computational complexity while maintaining the advantages of bidirectional attention. Furthermore, a hierarchical attention mechanism may be introduced, focusing first on local context and then expanding to the global context, further improving prediction efficiency and quality.

[0056] Sentence completion schemes refer to multiple possible text continuations generated by the system based on partially transcribed text and a masked language model. During generation, techniques such as bundle search or diversity sampling are employed to extract multiple high-probability word sequences from the probability distribution predicted by the model. The bundle search strategy retains multiple most likely candidate sequences at each step and continues to expand based on cumulative probability, ultimately forming multiple complete sentence paths. Diversity sampling generates semantically similar but expressively diverse completion schemes through temperature parameter adjustment and sampling strategy optimization. To enhance the practicality of completion, grammatical rules and domain knowledge constraints are also considered, such as ensuring that the generated sentences conform to the terminology usage habits and expression norms of a specific domain. In interactive scenarios, user history and preference information may also be incorporated to generate completion schemes that better match the user's expression style. These completion schemes not only fill in the information gaps caused by the uncertainty of speech recognition but may also expand and enrich the original content, providing a more complete semantic expression.

[0057] The credibility score is a quantitative indicator that evaluates the quality and reliability of generated candidate sentences. It comprehensively considers information from multiple dimensions to determine the final score. The main dimensions include: language model probability, contextual consistency, grammatical correctness, and domain relevance. Language model probability is the most fundamental metric, directly reflecting the model's confidence in the generated sequence. It is typically calculated using normalized log-likelihood probability (i.e., the average log probability of each word) to eliminate the influence of sentence length. Contextual consistency assesses the semantic coherence of the generated content with existing transcribed text, which can be achieved by calculating the cosine similarity of semantic vectors or using a specialized coherence assessment model. Grammatical correctness is evaluated using a grammar checker or a specific grammar scoring model to ensure that the generated content conforms to language norms. Domain relevance examines whether the generated content conforms to the professional knowledge and expression habits of a specific scenario, and can be measured by domain vocabulary coverage or similarity to a domain corpus. The scores of these dimensions are synthesized into the final credibility score through a weighted average or more complex ensemble method, with weights potentially dynamically adjusted according to the application scenario and user needs. In real-time scenarios, the scoring mechanism is continuously updated and adjusted to adapt to different speech content and quality conditions.

[0058] The selection of the top N candidates with the highest confidence is a crucial step in ensuring the quality of the system output. The value of N is typically between 3 and 10, dynamically adjusted based on the complexity of the application scenario and subsequent processing requirements. The selection process first sorts all candidate sentences in descending order according to their confidence scores, then selects the top N as the final candidate set. To ensure the diversity and representativeness of the candidate set, diversity constraints may be introduced to avoid selecting overly similar candidate sentences. In specific implementations, the Maximum Marginal Relevance (MMR) algorithm can be used to penalize new candidates that are too similar to the already selected candidates while maintaining high confidence. A minimum confidence threshold is also set to ensure that all retained candidates meet basic quality standards, even under conditions of difficult prediction. When processing content in critical domains (such as medical and legal fields), this threshold may be increased, erring on providing fewer but more reliable candidates rather than generating content that could lead to misunderstandings. These carefully selected complete sentence candidates not only provide users with a variety of possible expression choices but also provide a rich information foundation for subsequent semantic understanding and automatic correction.

[0059] Step S2 involves inputting the partially transcribed text into a masked language model for text prediction, specifically including:

[0060] Based on the partially transcribed text, a bidirectional attention mechanism is used to simultaneously process preceding and following inference information for comprehensive semantic prediction, generating text completion results. This step first preprocesses the partially transcribed text, including word segmentation, cleaning, and standardization, converting it into a standardized input format. The processed text is then input into a bidirectional attention encoder, which contains a multi-layer Transformer structure, each layer consisting of a multi-head self-attention sublayer and a feedforward neural network sublayer. In the self-attention calculation, the system maps each word to three vector representations: query vector (Q), key vector (K), and value vector (V). Attention weights are obtained by calculating the dot product of Q and all K, followed by scaling and softmax normalization. These weights are then used to weighted sum V to generate a context-aware representation. To enhance adaptation to speech characteristics, a speech feature enhancement module is introduced, encoding pauses, stresses, and intonation information in the speech as additional features and incorporating them into the attention calculation. After processing, a contextualized representation of each word is obtained, containing bidirectional contextual information. Next, a predetermined number of mask markers (typically 5-15, dynamically adjusted according to prediction needs) are added to the end of the transcribed text, and a mask prediction head network is used to predict the possible lexical units at these positions. The prediction process employs a combination of autoregressive and non-autoregressive methods, first generating keyframe words and then filling in the details. To enhance prediction quality, a knowledge graph auxiliary module is also introduced to retrieve relevant information from external knowledge bases, aiding in the understanding of specific domain content. Finally, post-processing rules are applied to correct possible grammatical errors and format the output, yielding preliminary text completion results. The entire process fully utilizes bidirectional context, considering both the semantic constraints of the identified text and forward-looking reasoning based on linguistic knowledge, providing a high-quality text foundation for the next step.

[0061] For the text completion results, a parallel processing approach is used to generate multiple complete sentence structure schemes to ensure the grammatical and semantic quality of the candidate sentences. In this stage, the text completion results generated in the previous step are first subjected to structured analysis to identify their semantic core, main argument relationships, and potential sentence type categories. Based on the analysis results, a multi-channel parallel processing framework is initiated, simultaneously activating 5-8 dedicated generation channels, each optimized for a specific sentence structure or expression style. These channels include: a basic declarative sentence channel responsible for generating direct declarative expressions; a complex syntax channel handling complex sentences containing clauses and parallel structures; an interrogative sentence channel focusing on generating interrogative forms consistent with the context; a conditional reasoning channel constructing conditional expressions containing hypotheses and inferences; and a domain-specific channel handling specific industry terminology and expression norms. Each channel is equipped with a dedicated syntactic checker and semantic consistency evaluator to monitor the generation quality in real time and make necessary adjustments. To ensure sentence integrity, a structured generation strategy is adopted, first identifying the core components (subject-verb-object), then expanding modifiers and subordinate components, and finally adding polishing elements such as conjunctions and modifiers. During the generation process, the system implements syntactic constraints to ensure that various syntactic components (such as tense, voice, and noun-verb agreement) conform to grammatical rules. Simultaneously, a context adaptation mechanism is applied to ensure that the generated content maintains consistency with existing transcribed texts in style, register, and subject matter. To enhance diversity, different decoding strategies and parameter settings are employed for each channel, such as temperature sampling, kernel sampling, or bundle search. This highly parallel processing architecture enables the rapid generation of 10-20 complete sentence candidates with different structural features while maintaining basic quality, providing a rich variety of choices for the final selection.

[0062] Based on the multiple complete sentence structure schemes, the model probability distribution and semantic coherence are comprehensively considered for selection, and the complete sentence candidates are output. This step constructs a multi-dimensional evaluation system to comprehensively evaluate and prioritize all sentence schemes generated in parallel. First, model-based probability indicators are calculated, including: normalized sentence probability score, which is the logarithmic average of the conditional probabilities of each word in the sentence to eliminate the influence of length; key position fluency score, which pays special attention to the probability values ​​of the connection between the transcribed text and the generated content and the key points of sentence structure; and keyword salience, which evaluates the probability strength of core concept words in the sentence. Second, semantic coherence is evaluated, which includes multiple levels: the vector similarity between the candidate sentence and the context is calculated through a pre-trained semantic encoder; a topic consistency detection algorithm is applied to ensure that the candidate sentence continues the topic context of the existing text; a logical relation validator is used to check the rationality of logical relations such as causality, temporality, and conditionality; and referential consistency analysis is performed to ensure that the referencing of pronouns and nouns remains coherent throughout the text. Third, domain adaptability is evaluated by comparing with a domain knowledge base to verify the accuracy of the use of professional terminology and the conformity of the expression paradigm. Fourth, pragmatic appropriateness is assessed, evaluating the suitability and effectiveness of candidate sentences in the current context. These multi-dimensional scores are integrated into a final score using an adaptive weighted fusion algorithm, with weights dynamically adjusted based on the application scenario and speech type. In the final selection stage, the system not only selects the highest-scoring candidates but also considers diversity coverage, applying a deterministic maximization algorithm to ensure that the retained candidate sentences have sufficient differences in structure and expression. Finally, 3-5 of the highest-quality and stylistically diverse candidate sentences are retained, ranked according to their comprehensive scores, and provided to the user for selection or for subsequent processing. This rigorous multi-dimensional selection mechanism ensures that the final output candidate sentences possess both high probability and high coherence, while also meeting the practical needs of specific scenarios.

[0063] Specifically, bidirectional attention is a key technology in natural language processing, overcoming the limitations of traditional unidirectional language models that can only utilize preceding or following information, thus achieving comprehensive context awareness. In this system, bidirectional attention is implemented through a self-attention layer, where the representation of each word simultaneously focuses on all other words in the sequence, forming a context-aware representation. Specifically, for each word in the partially transcribed text, attention weights are calculated between it and all other word elements. These weights reflect the strength of semantic associations between different words. During the calculation, each word element is mapped to three representations: query vector, key vector, and value vector. Similarity is calculated by taking the dot product of the query vector and all key vectors, then normalized using softmax to obtain the attention weights. Finally, a weighted sum of all value vectors is performed to obtain a word representation containing global information. When processing partially transcribed text, the system particularly strengthens the attention weights for sentence boundary words and key information words, ensuring that this important contextual information can effectively guide the prediction of subsequent text. Meanwhile, a multi-head attention mechanism is introduced to compute attention in parallel in different representation subspaces, capturing different types of semantic dependencies, such as close syntactic relations and long-distance logical associations, thereby improving the accuracy and richness of predictions.

[0064] Comprehensive semantic prediction is an advanced text generation process that combines bidirectional attention processing with linguistic knowledge and contextual understanding. The process first maps partially transcribed text into a high-dimensional semantic representation using an encoder network, capturing information such as the text's topic, sentiment, and intent. Then, based on these semantic representations and linguistic knowledge learned from a pre-trained language model, the system predicts the most likely text continuation. The prediction process employs an autoregressive strategy, generating words one by one, immediately incorporating each generated word into the context as a condition for predicting the next word. To improve the accuracy and stability of the prediction, multiple linguistic features are comprehensively considered, including word frequency statistics, grammatical rules, semantic associations, and domain knowledge. Furthermore, an adaptive language model calibration mechanism is introduced to dynamically adjust the prediction probability distribution according to different topics and styles, making the generated text more consistent with the expression habits of specific scenarios. When processing professional domain content, the corresponding domain knowledge module is activated to ensure that the generated professional terminology and expressions conform to industry standards. Through this comprehensive semantic prediction method, coherent, natural, and context-aware text completion results can be generated, effectively compensating for the uncertainties and information gaps that may exist in speech recognition.

[0065] Parallel processing is a highly efficient technique for generating diverse sentence structures. It involves simultaneously launching multiple generation threads, each employing different generation parameters or constraints to produce complete sentence structure schemes with varying characteristics. At the implementation level, firstly, based on the semantic and structural features of the partially transcribed text, potentially applicable sentence templates, such as declarative, interrogative, and conditional sentences, are determined. Then, a dedicated generation thread is assigned to each sentence type, with each thread generating based on a shared semantic representation but following different syntactic rules. To ensure generation quality, a syntactic checker and semantic consistency evaluation module are embedded in each thread to monitor the generation process in real time and make necessary adjustments. Simultaneously, a diversity control mechanism is designed, adjusting sampling temperature, kernel sampling parameters, or using bundle diversity techniques to ensure that the content generated by different threads conforms to language norms while possessing sufficient diversity. When handling complex content, a hierarchical generation strategy may be introduced, first determining the overall framework of the sentence and then filling in specific details to ensure the structural integrity and logical coherence of long sentences. This parallel processing approach can generate multiple possible complete sentence structure schemes in a short time, providing a rich candidate pool for subsequent screening.

[0066] The model probability distribution is a quantitative assessment of the likelihood of each candidate sentence by the language model, reflecting the reasonableness of each candidate from a statistical and linguistic perspective. It calculates the overall probability score for each candidate sentence, typically using normalized log-likelihood probability (the average log probability of all words in the sentence) as the basic metric to eliminate the influence of sentence length. During the calculation, special attention is paid to the probability values ​​at the junctions between the transcribed text and the generated content to ensure a natural and smooth transition. Furthermore, the conditional probability distribution of keywords in the candidate sentences is analyzed; high-probability keywords usually indicate that the candidate sentence is more consistent with the contextual theme. Semantic coherence assesses the degree of matching between the candidate sentence and the existing text from a content understanding perspective, mainly measured by calculating the similarity of semantic vectors, topic consistency, and logical relationships. A pre-trained semantic encoder is used to map the candidate sentence and the existing text to the same semantic space, and then their cosine similarity or other relevance metrics are calculated. Simultaneously, it checks whether the candidate sentence continues the thematic thread of the existing text and whether there are obvious topic jumps or content breaks. By comprehensively considering the two dimensions of model probability distribution and semantic coherence, the quality of each candidate sentence can be fully evaluated, and complete sentence candidates that conform to language rules and are highly relevant to the context can be selected and provided to users or used for subsequent processing.

[0067] In a preferred embodiment, step S2 involves inputting the partially transcribed text into a masked language model for text prediction to generate several highly reliable complete sentence candidates. Specifically, this includes: using a binary weighted spike neural network to replace the traditional masked language model for text prediction based on the partially transcribed text, generating candidate sentences encoded by time-series pulse signals; quantizing the weights of the time-series pulse signal-encoded candidate sentences into a binary form of {-1, +1} using a sign function to obtain binary weight processing results; and based on the binary weight processing results, predicting different sentence types using multiple parallel BWSNN sub-networks to obtain the complete sentence candidates.

[0068] A more efficient text prediction method is achieved by replacing the traditional masked language model with a Binary Weighted Spiking Neural Network (BWSNN). In the text encoding stage, a text encoder based on spike neurons is constructed to convert a portion of the input text into temporal pulse signals. This encoder employs a hierarchical encoding strategy: first, each word in the text is mapped to a corresponding word vector representation; then, a temporal encoding module converts the word vectors into a pulse sequence with temporal features. The intensity and frequency of the pulse signals correspond to the semantic importance of the words and their positional information within the sentence, respectively. This encoding method better preserves the temporal dependencies of the text.

[0069] In terms of network architecture design, a binary weight matrix is ​​designed, quantizing continuous weights into a binary form of {-1, +1}, which significantly reduces computational complexity and storage requirements. The binary quantization process uses a sign function to discretize the weights, while introducing a scaling factor to compensate for quantization errors. This design simplifies the network's multiplication operations to sign judgment and addition operations, significantly improving inference speed, while compressing the model's storage space to 1 / 32 of its original size.

[0070] Specifically, the Binary-Weight Spiking Neural Network (BWSNN) is a highly efficient neural network model inspired by biological neurons. It combines the temporal processing capabilities of spike neural networks with the computational efficiency of binary weight networks. Unlike traditional deep learning models that use continuous values ​​for information transmission and processing, BWSNN uses discrete neural spikes (i.e., "spikes") as information carriers, simulating the firing mechanism of biological neurons. In this example, BWSNN receives the encoded representation of partially transcribed text as input and then converts it into a spike sequence through temporal encoding. Specifically, each word in the text is encoded into a spike pattern with different firing times and frequencies based on its importance and semantic features. These spikes propagate through artificial neurons in the network. Each neuron accumulates an internal potential based on the spatiotemporal pattern of the input spikes and its connection weights, firing a spike when the potential exceeds a threshold. The core advantage of BWSNN lies in the fact that its network connection weights are restricted to binary values ​​(+1 or -1), which significantly simplifies computational complexity while maintaining sufficient model expressive power. This design enables the network to operate efficiently in resource-constrained environments (such as mobile devices or embedded systems), making it particularly suitable for real-time voice interaction scenarios requiring low-latency responses. During training, BWSNN learns the mapping relationship from text features to impulse patterns through special gradient approximation techniques (such as pass-through estimators) and binarization constraints. This allows the network to effectively capture the temporal characteristics and structural patterns of language, providing a unique processing perspective for text prediction.

[0071] Temporal pulse signal encoding is the core mechanism of BWSNN for information processing. It utilizes the temporal dimension of neural pulses to represent complex linguistic information. In this encoding scheme, each word in a candidate sentence is not only represented by the presence or absence of a pulse (0 or 1), but also by the precise time point of the pulse occurrence and the temporal pattern of the pulse sequence. This temporal encoding method can express rich semantic information with a limited number of neurons, significantly improving information density. In specific implementations, multiple temporal encoding strategies are employed: rate encoding adjusts the pulse firing frequency according to the importance of words, with neurons corresponding to keywords firing pulses at a higher frequency; time encoding maps the semantic features of words to the precise time point of pulse firing, with semantically similar words triggering pulses within similar time windows; phase encoding uses the relative temporal relationship between pulses to express the syntactic and semantic dependencies between words, with pulses corresponding to words with strong master-slave relationships exhibiting a stable phase difference. These temporal encoding strategies collectively construct a multi-dimensional information representation space, enabling the network to simultaneously process multi-level linguistic features such as vocabulary, syntax, and semantics. During the decoding stage, the dominant neuron activity within each time window is extracted by analyzing the spiking patterns of the output neuron layer and mapped back to the lexical space, thereby recovering the complete candidate sentence text. This time-series-based information processing approach is particularly suitable for capturing the sequential relationships and long-distance dependencies in language, providing a new technical path for high-quality text prediction.

[0072] Sign function weight quantization is a key technique for precisely restricting the connection weights in BWSNN to binary values. A sign function is a mathematical operation that outputs +1 when x > 0, -1 when x < 0, and can be defined as 0 when x = 0, or +1 or -1 depending on the application requirements. In this paper, the sign function is applied to the real-valued weights generated during network training, quantizing them into a strict {-1, +1} binary form. Specifically, the gradient of the weights is first calculated using the standard backpropagation algorithm. Then, a special optimizer (such as a binarized perceptron learning rule or a gradient-based binarization method) is used to update the real-valued weights. Finally, the sign function is applied to binarize the weights during the forward propagation stage. To address the issue of the sign function being non-differentiable at zero, a straight-through estimator is used to approximate the gradient during backpropagation, maintaining gradient flow. Furthermore, a scaling factor is introduced to preserve weight magnitude information. The mean or standard deviation of the weights in each layer is calculated as a scaling parameter, participating in the calculation along with the binary weights to improve the model's expressive power. In the later stages of training, the strength of the binarization constraints is gradually increased, allowing the network to adapt to the limitations of binary weights while maintaining predictive performance. This weight quantization process not only significantly reduces the model's storage and computational requirements but also improves inference speed, enabling complex language models to run efficiently in computationally limited environments and providing strong technical support for real-time voice interaction.

[0073] The BWSNN sub-networks are specialized neural network modules designed for different sentence structure characteristics, with each sub-network focusing on processing specific types of sentence structures and expressions. In this study, multiple BWSNN sub-networks work in parallel to form a comprehensive text prediction system. Each sub-network inherits the basic architectural features of BWSNN, employing binary weights and pulse coding mechanisms, but undergoes specific optimizations in network structure, neuron parameters, and training objectives. Typical sub-network types include: the declarative sentence network focuses on generating logically clear and informationally complete declarative expressions; the interrogative sentence network optimizes the processing of interrogative words and tone, generating natural interrogative structures; the conditional sentence network excels at handling hypotheses, inferences, and causal relationships, generating complex conditional expressions; and the imperative sentence network is optimized for command and request types of statements. Each sub-network is trained using a large amount of corresponding type corpus during the training phase, forming a specialized processing capability for specific sentence types. During the prediction process, the contextual and intent features of a portion of the transcribed text are first analyzed to preliminarily determine the most likely sentence type category. Then, the corresponding main sub-network and several auxiliary sub-networks are activated to generate multiple candidate results in parallel. The outputs of each sub-network are integrated and scored by a post-processing module to select the most relevant complete sentence candidates. This distributed parallel processing architecture not only improves the diversity of predictions but also enhances its adaptability to complex linguistic phenomena, enabling it to provide customized prediction services for different speech scenarios and expression needs.

[0074] In step S2, different sentence types are predicted by multiple parallel BWSNN sub-networks to obtain the complete sentence candidates. Specifically, this includes: constructing interrogative sentence prediction sub-networks, declarative sentence prediction sub-networks, and imperative sentence prediction sub-networks based on the binary weight processing results; using a pulse temporal coding strategy to encode semantic information using the firing time and frequency of neurons to obtain progressive text completion results; for the progressive text completion results, weighted fusion of the prediction results of multiple sub-networks is performed through an integrated decision fusion module to generate a preliminary candidate sentence list; based on the preliminary candidate sentence list, a dynamic weight allocation strategy and a confidence evaluation mechanism are used to obtain the complete sentence candidates, ensuring the quality and reliability of the complete sentence candidates.

[0075] In terms of prediction mechanism implementation, a self-integration heuristic mechanism is adopted, which uses multiple parallel BWSNN sub-networks to predict different types of text completion, with each sub-network focusing on a specific semantic pattern. Specifically, these include interrogative sentence prediction sub-networks, declarative sentence prediction sub-networks, and imperative sentence prediction sub-networks, etc. Each sub-network is optimized with a dedicated training dataset, enabling it to more accurately capture the linguistic features and expression patterns of the corresponding sentence type.

[0076] Regarding temporal coding strategies, a pulse temporal coding strategy is adopted, utilizing the firing time and frequency of neurons to encode semantic information and improve the time efficiency of prediction. Early firing pulses correspond to high-frequency words and common grammatical structures, while late firing pulses correspond to low-frequency words and complex semantic relationships. This temporal coding mechanism enables the network to progressively refine the prediction results during processing, achieving incremental text completion.

[0077] In the results fusion stage, an integrated decision fusion module is established to weight and fuse the prediction results of multiple sub-networks to generate a final list of candidate sentences. The fusion module employs a dynamic weight allocation strategy, automatically adjusting the contribution weights of each sub-network based on the features of the input text. Simultaneously, a confidence evaluation mechanism is introduced to calculate a comprehensive confidence score for each candidate sentence, ensuring the quality and reliability of the output results.

[0078] Specifically, constructing interrogative sentence prediction subnetworks, declarative sentence prediction subnetworks, and imperative sentence prediction subnetworks is a key step in achieving diversified text prediction. The interrogative sentence prediction subnetwork specifically handles sentences containing interrogative mood and structure. Its network architecture includes a special interrogative word recognition module and a mood detection unit. This subnetwork adopts a three-layer structure: the input layer receives pulse signals processed by binary weights; the middle layer contains multiple clusters of interrogative pattern recognition neurons, each corresponding to different types of interrogative structures (such as yes / no questions, specific questions, alternative questions, etc.); the output layer generates the corresponding interrogative sentence based on the activation pattern. The declarative sentence prediction subnetwork is the most basic component, handling sentences that directly state facts, opinions, or descriptions. This subnetwork adopts a relatively balanced structure. The input layer is the same as the interrogative sentence network, but the middle layer is optimized for declarative mood, including a subject-predicate structure recognition unit, an adverbial processing unit, and a complex sentence pattern unit. Its characteristic is its ability to handle logically complex declarative structures containing multiple layers of clauses. The imperative sentence prediction subnetwork focuses on handling sentences expressing commands, requests, or suggestions. Its unique feature lies in the inclusion of an imperative tone recognition module and a verb command form processing unit within the network. The neurons in the intermediate layers are primarily divided into three categories: direct command mode, euphemistic request mode, and suggestive expression mode, capable of generating imperative expressions of varying intensities based on the context. While these three sub-networks each focus on different sentence structures, they share a basic semantic understanding layer, ensuring that the generated content remains semantically consistent with some transcribed texts. Through this specialized network structure, appropriate sentence structures can be generated for different contextual needs, significantly improving the diversity and adaptability of text prediction.

[0079] The spiking temporal coding strategy is the core mechanism for information processing and transmission in the BWSNN sub-network. It draws inspiration from the working principles of biological neurons, utilizing the temporal patterns and frequencies of neuron firing to encode complex semantic information. In traditional deep learning models, information is typically represented and transmitted through continuous real-valued signals. However, in spiking neural networks, information is encoded as discrete spiking events, more closely resembling the actual workings of biological neurons. Specifically, the spiking temporal coding strategy comprises three key layers: First, rate coding, which represents information strength by adjusting the frequency of neuron firing spiking signals; neurons corresponding to semantically important words will fire spiking signals at a higher frequency. Second, temporal coding, which encodes information using the precise timing of spiking; words with similar semantics will trigger spiking signals within similar time windows. Third, phase coding, which expresses syntactic and semantic dependencies between words through the relative temporal relationships between spiking sequences; grammatically closely related words will exhibit stable phase relationships. This multi-dimensional coding strategy enables the network to express rich linguistic information within a limited number of neurons. In the processing, each input word is first mapped to a specific initial impulse pattern, then propagates and transforms through synaptic connections in the network, ultimately forming a temporal impulse pattern representing the prediction result at the output layer. An adaptive thresholding mechanism is also implemented, where the firing threshold of neurons is dynamically adjusted based on historical activity states to prevent overactivity or underactivity, ensuring the stability and efficiency of information transmission. This time-based information processing approach is particularly suitable for capturing sequential relationships and long-distance dependencies in language, providing a new technical path for high-quality text prediction.

[0080] Progressive text completion is a dynamic, iterative text generation method. Unlike traditional methods that generate all content at once, it gradually builds and refines the text through multiple progressive steps. In this paper, progressive text completion combines the spurious characteristics of the BWSNN subnetwork and consists of four main stages: the skeleton construction stage first generates the core structure of the sentence, including the main components (subject-verb-object) and key conjunctions, establishing the basic framework of the sentence; the content filling stage adds secondary components such as modifiers, adverbs, and attributives to enrich the semantic content based on the skeleton; the detail optimization stage adjusts word order, adds transition words and modifiers to improve fluency and naturalness; and finally, the consistency check stage ensures that the generated content is consistent with the existing transcribed text in terms of syntax, semantics, and style. The core advantage of this progressive method is that it can continuously integrate feedback information and adjust the generation strategy during the generation process. For example, when it is found that the skeleton structure may lead to difficulties in subsequent filling, the skeleton can be backtracked and adjusted in time; or when potential semantic conflicts are detected, specific parts can be modified in a targeted manner without affecting the overall structure. In implementation, a multi-level spiking neuron group is employed. Lower-level neurons are primarily responsible for skeleton construction, mid-level neurons handle content filling, and higher-level neurons focus on detail optimization and consistency checks. Neurons at different levels form a cascaded activation pattern through specific time delays and threshold settings, ensuring that each stage proceeds in an orderly manner. Furthermore, a dynamic time window mechanism is implemented, automatically adjusting the processing time of each stage based on sentence complexity, ensuring that simple sentence structures are generated quickly, while complex sentence structures have sufficient time for thorough reasoning and optimization. This progressive text completion method generates well-structured, content-rich text that is highly consistent with its context.

[0081] The integrated decision fusion module is the core component that integrates the outputs of multiple specialized sub-networks into a coherent and consistent prediction result. This module adopts a hierarchical fusion architecture, comprising three processing layers: a signal-level fusion layer directly processes the raw impulse patterns output by each sub-network, identifying consensus patterns across networks through impulse synchronization analysis and temporal correlation calculation; a feature-level fusion layer decodes the impulse patterns into intermediate semantic features, then integrates them through feature mapping and alignment to generate a unified semantic representation; and a decision-level fusion layer, based on the integrated semantic features, applies a multi-strategy combination algorithm to generate the final text prediction. Several advanced technologies are employed during the fusion process: adaptive weight calculation dynamically adjusts the influence of each sub-network in the fusion process based on its historical performance and current context, with better-performing sub-networks receiving higher weights; a complementarity enhancement mechanism emphasizes complementary information from different sub-networks, prioritizing prediction results that can compensate for the weaknesses of other networks; a consistency score quantitatively evaluates the internal and mutual consistency of predictions from each sub-network, serving as an important reference for weight calculation; and a conflict resolution algorithm handles contradictory predictions from different sub-networks, determining the most reasonable solution through contextual correlation analysis and probabilistic reasoning. Furthermore, an information gain evaluation mechanism was implemented to calculate the unique information provided by each sub-network, avoiding excessive influence caused by redundant predictions. In real-time processing, the fusion module adopts a pipelined architecture, allowing parallel operation at each level and significantly improving processing efficiency. Through this complex and sophisticated fusion mechanism, the advantages of each specialized sub-network can be fully utilized, while mitigating the potential biases and limitations of individual sub-networks, ultimately generating a preliminary candidate sentence list that reflects consensus while preserving diversity.

[0082] The preliminary candidate sentence list is a set of potential text completion schemes output by the integrated decision fusion module, containing multiple candidate sentences with varying structures and styles but inconsistent quality. These candidate sentences retain the predictive characteristics of each sub-network, including interrogative structures generated by the interrogative sentence network, declarative expressions output by the declarative sentence network, and imperative statements provided by the imperative sentence network. Each candidate sentence contains four key attributes: original text content, sub-network source identifier, preliminary confidence score, and sentence type category label. The preliminary candidate sentence list typically contains 15-30 candidates, providing ample selection space for subsequent fine-tuning. These candidate sentences may vary significantly in structural completeness and semantic accuracy; some may require only minor adjustments for usability, while others may require substantial reconstruction or be discarded outright. A unique identifier is assigned to each candidate sentence for easy reference and tracking in subsequent processing. Furthermore, a similarity matrix is ​​calculated between candidate sentences to identify duplicate or highly similar candidates, providing a reference for subsequent diversity control. The generation of the preliminary candidate sentence list marks the shift from the multi-network parallel prediction stage to the unified evaluation and optimization stage, and is a key intermediate product connecting the individual predictions of sub-networks and the final high-quality output.

[0083] The confidence assessment mechanism is a systematic method for measuring the reliability of candidate sentences, comprehensively evaluating the certainty and credibility of the prediction results through multiple complementary indicators. In this paper, the confidence assessment is constructed based on five main indicators: First, the model consistency indicator calculates the consensus among different sub-networks on the same predicted content; predictions agreed upon by multiple sub-networks generally have higher credibility. Second, the prediction stability indicator tests the stability of the prediction results by adding slight perturbations during the prediction process (such as adjusting temperature parameters or randomly discarding some inputs); stable predictions are generally more reliable. Third, the contextual support indicator assesses the semantic association between the predicted content and the existing transcribed text; strong associations usually imply higher confidence. Fourth, the lexical certainty indicator is calculated based on the probability distribution of the model output; if the predicted word at a certain position has a much higher probability than other candidates, the prediction at that position is more certain. Finally, the grammatical correctness indicator evaluates the grammatical correctness of the predicted content through a grammar checker; correct grammar usually indicates higher generation quality. These five indicators are integrated into a comprehensive confidence score between 0 and 1 through a non-linear mapping function, and each candidate sentence is assigned a confidence level (e.g., high, medium, low). In practical applications, confidence scores are used not only to filter candidate sentences but also to determine actions. For example, high-confidence predictions may be provided directly to the user, while medium-confidence predictions may require multiple alternatives, and low-confidence predictions may trigger additional verification steps or prompt the user for confirmation. Furthermore, a confidence score calibration mechanism is implemented, adjusting the original confidence score estimate through historical data analysis to ensure that the confidence score truly reflects the actual reliability of the prediction. This sophisticated confidence score assessment mechanism enables an accurate understanding of the reliability of the predictions, providing users with transparent and reliable decision support.

[0084] Complete sentence candidates are the final prediction results after comprehensive screening and optimization, achieving the highest standards in structural completeness, semantic coherence, and naturalness of expression. These candidates are a select few (typically 3-5) elite sentences chosen from the initial candidate sentence list after dynamic weighting and confidence assessment. They represent the most reasonable completion schemes for partially transcribed text. Each candidate sentence contains rich metadata: original text content, overall quality score, confidence level, sentence type, potential application scenario, and language style tags. This ensures sufficient diversity in style and structure, providing users with meaningful choices. Furthermore, the candidate sentences undergo final optimization, including punctuation normalization, special terminology correction, and formatting adjustments, ensuring the output conforms to standard text formatting requirements. In multi-round interactive scenarios, a historical record of candidate sentences is maintained for subsequent learning and improvement. The quality of complete sentence candidates directly impacts the user experience and effectiveness; therefore, a rigorous quality assurance mechanism is applied during the generation process, including multi-layered filtering, cross-validation, and human-machine collaborative evaluation. These high-quality complete sentence candidates will eventually be presented to the user or passed to the next processing module for further applications, such as automatic document generation, meeting minutes organization, or real-time dialogue assistance.

[0085] The parallel intent recognition process is described in step S3. In this stage, the predicted complete sentence candidates are input in parallel into a large language model for deep intent recognition analysis, thereby obtaining the corresponding intent results in advance. To achieve efficient parallel processing, an independent inference task is first created for each candidate sentence, ensuring that the tasks do not interfere with each other. The task scheduler employs an intelligent allocation strategy, rationally distributing multiple inference tasks across different GPU devices or inference instances to fully utilize the parallel computing capabilities of the hardware resources.

[0086] In the core processing of intent recognition, the large language model performs in-depth semantic analysis on each candidate sentence. This involves not only identifying the user's true intent category but also extracting key entity information and important parameters contained within the sentence. This deep analysis ensures a more accurate and comprehensive understanding of user needs. After completing the analysis, the model outputs structured intent recognition results, which include detailed intent category labels, corresponding confidence scores, and extracted key parameter information.

[0087] To support the subsequent rapid matching process, the result caching module establishes a clear mapping relationship between each candidate sentence and its corresponding intent recognition result, and securely stores this mapping information in a high-speed cache. Simultaneously, a real-time monitoring module continuously tracks the execution status and progress of each inference task, ensuring that all prediction tasks are successfully completed before the user finishes full voice input.

[0088] This parallel processing design significantly improves overall efficiency and ultimately passes the complete candidate sentence-intent result mapping table to the final matching module, providing users with fast and accurate intent recognition services.

[0089] In step S3, the complete sentence candidates are input into a large language model in parallel for intent recognition to obtain the corresponding intent result. Specifically, this includes: constructing a constrained multi-objective optimization problem based on the complete sentence candidates to simultaneously optimize recognition accuracy, response time, and resource consumption, and identifying constrained Pareto fronts and unconstrained Pareto fronts; using hybrid operator technology to intelligently combine a fast classifier, a medium-complexity model, and a large language model based on the constrained Pareto fronts and the unconstrained Pareto fronts to obtain a multi-strategy fusion result; and adjusting computing resources in real time according to system load and user priority through a dynamic resource allocation mechanism to obtain the corresponding intent result.

[0090] Specifically, constrained multi-objective optimization is a mathematical optimization framework used to find the optimal balance among multiple competing objectives while satisfying specific constraints. In intent recognition, this problem involves three core objectives: recognition accuracy (the proportion of correctly recognized user intents), response time (the delay from input to output), and resource consumption (the amount of computational resources used in the processing). These three objectives are often inherently conflicting—improving accuracy usually requires more complex models, leading to increased response time and resource consumption. First, each objective is formally modeled: recognition accuracy is defined as the ratio of correctly recognized samples to the total number of samples, denoted by the function facc(x); response time is defined as the total time from input processing to output generation, denoted by the function ftime(x); and resource consumption is quantified as the total amount of computational resources used in the processing, including memory usage, processor time, and energy consumption, denoted by the function fres(x). Here, x represents a configuration vector containing decision variables such as model selection, parameter settings, and resource allocation strategies. A series of constraints are also defined, such as the maximum acceptable response time Tmax, the maximum available memory Mmax, and the maximum available number of processor cores Cmax. Thus, the constrained multi-objective optimization problem can be formally represented as: maximizing facc(x), minimizing ftime(x), and minimizing fres(x), while satisfying constraints such as ftime(x) ≤ Tmax, mem(x) ≤ Mmax, and cpu(x) ≤ Cmax. To solve this problem, an improved NSGA-II (Non-dominated sorting genetic algorithm II) is employed. This algorithm simulates an evolutionary process, generating and evaluating candidate solutions generation by generation, and retaining elite solutions based on non-dominated sorting and crowding distance calculations. Compared to the traditional NSGA-II, a constraint handling mechanism is introduced, employing a feasibility-first principle to handle constraint violations and using an adaptive mutation operator to enhance search capabilities. Furthermore, a solution space partitioning strategy is implemented, maintaining dedicated solution sets for different operating conditions (such as high load / low load, high priority / low priority) to ensure suitable solutions can be found in various real-world scenarios. Through this carefully designed optimization framework, an optimal balance can be found between accuracy, response time, and resource consumption, providing a solid foundation for subsequent intent recognition.

[0091] A Constrained Pareto Front (CPF) is the set of non-dominated solutions that satisfy all constraints in a multi-objective optimization problem, representing the optimal performance trade-off achievable under given constraints. In intent recognition, the CPF includes all solutions that satisfy constraints such as upper limits on response time, memory limits, and processor core count limits, and cannot simultaneously improve all objectives without compromising at least one objective. The CPF is constructed through the following steps: First, starting with an initial population, a multi-generation evolutionary search is performed using the constraint-handled NSGA-II algorithm; then, in each generation, a two-stage sorting is performed based on the degree of constraint violation and the objective function value, prioritizing solutions that satisfy the constraints and sorting them by non-dominated relation; next, crowding calculation is used to maintain solution diversity, ensuring the CPF covers the entire feasible solution space; finally, after a sufficient number of generations of evolution (typically 100-200 generations), all non-dominated solutions that satisfy the constraints are collected to form the CPF. In practical implementation, an adaptive mesh generation technique is used to maintain the CPF (Continuous Performance Factor), an archive update strategy is used to save historical optimal solutions, and the CPF's characteristic indicators, such as hypervolume, distribution uniformity, and extensibility, are calculated through a solution analysis module to evaluate the CPF's quality. The CPF construction process is a computationally intensive task, and parallel computing and incremental update techniques are employed to improve efficiency. Once constructed, the CPF provides the optimal performance combination achievable under constraints, serving as an important reference for configuration decisions.

[0092] An Unconstrained Pareto Front (UPF) is a set of non-dominated solutions obtained solely based on trade-offs between multiple optimization objectives, without considering any constraints. It represents the optimal performance boundary achievable under ideal conditions. The construction of a UPF is similar to that of a Common Processing Front (CPF), but it removes the constraint handling mechanism, focusing entirely on optimizing the objective function. In intent recognition, the UPF is generated through the following steps: First, using the standard NSGA-II algorithm, only three objective functions are considered: recognition accuracy, response time, and resource consumption, without imposing any constraints. Then, during the evolutionary process, the ranking of solutions is determined entirely based on non-dominated sorting, with crowding calculations used to maintain solution diversity. Next, non-dominated solutions are retained in each generation of selection, gradually approximating the true unconstrained Pareto Front. Finally, after sufficient evolution, all non-dominated solutions are collected to form the UPF. The construction of a UPF typically requires more computational resources and more iterations (usually 200-300 generations) due to the wider search space. Subpopulation techniques are employed to accelerate UPF exploration, while adaptive mutation strategies are used to enhance the search capability in different regions. Once built, the UPF can serve as an ideal performance benchmark to assess the impact of constraints on performance and guide decisions regarding architecture and resource allocation. The UPF will be updated regularly (e.g., weekly or monthly) in an offline environment to reflect the impact of technological advancements and new algorithms.

[0093] The difference analysis between constrained and unconstrained Pareto fronts is a crucial step in understanding the impact of constraints and guiding design. The analysis is conducted as follows: First, the geometric distance between the two fronts is calculated, including Euclidean, Manhattan, and directional distances, quantifying the overall impact of constraints on performance. Then, a regional impact analysis is performed to identify the uneven impact of constraints on different performance regions; for example, high-accuracy regions may be more significantly affected by constraints. Next, a sensitivity analysis is performed, observing changes in the Pareto Filter Factor (CPF) by gradually relaxing or tightening specific constraints to identify key constraint factors. Finally, a performance gap analysis is conducted to identify the regions with the largest gap between the Unconstrained Pareto Filter Factor (UPF) and the Pareto Filter Factor (CPF), representing the most significant performance losses caused by constraints. The analysis results are visualized as an impact heatmap, visually displaying the distribution of constraint impacts. Furthermore, a constraint impact model is built to predict the potential performance improvement from relaxing specific constraints, aiding resource allocation decisions. The analysis results are not only used for current configurations but also provide a basis for long-term upgrades and resource planning.

[0094] Hybrid operator technology is an innovative method for intelligently combining models of different complexities. It dynamically selects and fuses various models based on specific task requirements and the operating environment. In intent recognition, hybrid operator technology involves three main types of models: fast classifiers (such as decision trees, Naive Bayes, or lightweight neural networks) offer high processing speed and low resource consumption, but relatively low accuracy; medium-complexity models (such as LSTM, CNN, or medium-sized Transformers) provide a good balance between accuracy and response time; and large language models (such as pre-trained Transformer-based models) offer the highest recognition accuracy but require longer processing time and more computational resources. Hybrid operator technology achieves intelligent combination through the following components: a model feature analyzer evaluates the performance characteristics of each model on different types of intents and establishes a model-intent fit matrix; a task complexity estimator analyzes the complexity of the input sentence, including dimensions such as grammatical complexity, domain specialization, and intent ambiguity; a runtime environment monitor tracks resource status and load levels, providing resource constraint information; and a combination policy generator, based on the above information, uses reinforcement learning methods to learn the optimal model combination policy, forming a decision policy network. In practical operation, the hybrid operator operates in three basic modes: cascade mode uses a fast model for initial screening, then applies a more complex model to samples with insufficient confidence; parallel mode simultaneously launches multiple models of varying complexity, employing an early stopping mechanism to terminate other models when a satisfactory result is obtained; and ensemble mode combines the outputs of multiple models, generating the final result through weighted voting or stacked ensemble. An adaptive learning mechanism is also implemented to continuously optimize the combination strategy based on historical performance. Through hybrid operator technology, while maintaining high recognition accuracy, average response time and resource consumption can be significantly reduced, achieving an efficient balance between multiple objectives.

[0095] Multi-strategy fusion results are comprehensive intent recognition outputs from hybrid operator techniques. They integrate the predictive advantages of models with varying complexity, providing more accurate and reliable intent judgments than single models. The multi-strategy fusion result comprises the following core elements: first, a primary intent label representing the most likely user intent; second, an intent confidence score quantifying the degree of certainty regarding the intent judgment; third, a list of secondary intents containing other highly probable alternative intents and their confidence scores; fourth, intent parameter extraction results, including specific parameters related to the intent (such as time, location, and object); and finally, a fusion justification description, recording the formation process of the final result, including the models involved in the fusion and their respective contributions. During the fusion process, several advanced techniques are employed: confidence calibration ensures the comparability of confidence scores from different models, typically achieved through temperature scaling or equal-bin calibration; evidence-theory fusion uses Dempster-Shafer theory or Bayesian networks to integrate evidence from different models; uncertainty propagation tracks and quantifies uncertainty throughout the fusion process, providing reliability guarantees for high-risk scenarios; and a conflict resolution mechanism addresses conflicting predictions from different models through a comprehensive assessment of historical performance, current confidence, and domain rules. Furthermore, an intent transition graph is maintained to capture the sequential relationships and transition probabilities between different intents, thereby improving the accuracy of intent recognition in multi-turn dialogues. The multi-strategy fusion results not only provide the current recognition results but also contain rich metadata, providing decision support for downstream applications and feedback data for continuous optimization.

[0096] Dynamic resource allocation is an efficient and adaptive mechanism for managing computing resources. It adjusts the allocation of computing resources to the intent recognition process in real time based on current task demands, load, and user priorities. This mechanism consists of four main components: a resource monitoring system that continuously tracks the status of available computing resources, including processor utilization, memory usage, network bandwidth, and energy consumption; a workload analyzer that assesses current and expected load levels, including request queue length, average processing time, and peak forecasts; a user hierarchy manager that maintains a user priority system, determining the priority of different users based on service level agreements, historical usage patterns, or real-time payment status; and a resource allocation optimizer that combines the above information and uses mixed-integer programming or constraint satisfaction optimization algorithms to calculate the optimal resource allocation scheme. In actual operation, dynamic resource allocation follows these principles: elastic scaling automatically adjusts the amount of allocated resources according to load fluctuations, increasing resources during peak periods and releasing resources during idle periods; priority guarantee ensures that high-priority users receive sufficient resources even under high loads, ensuring service quality; resource isolation prevents a single user or task from consuming excessive shared resources through containerization or virtualization technologies; and cost-effectiveness balance minimizes overall resource costs, including computing and energy costs, while meeting performance requirements. It also implements a predictive resource allocation strategy, preparing resources in advance for potential load peaks based on historical patterns and time series forecasts. In abnormal situations, a resource degradation strategy is activated, gradually reducing resource consumption for non-critical functions according to a pre-set service quality degradation plan, ensuring the availability of core functions. Through this intelligent resource management, stable intent recognition services can be provided under various load conditions, while optimizing resource utilization efficiency and operating costs.

[0097] The intent result is the final output of the intent recognition process, accurately capturing the user's basic intent and detailed needs expressed in a complete sentence. In this paper, the intent result adopts a structured format, containing multiple levels of information: the intent graph category (e.g., query, command, information provision) defines the basic interaction type of the sentence; sub-intent categories (e.g., weather query, schedule, product search) further refine the specific domain and function; the intent parameter set contains specific information related to intent execution, such as time, location, quantity, and object identifier; modifier information captures non-functional characteristics of the intent, such as its emotional tone, urgency, or politeness level; and context dependency markers indicate whether the intent depends on previous dialogue history. In addition to the intent itself, rich metadata is provided: a confidence index quantifies the degree of certainty about the recognition result; a list of alternative intents provides other possible interpretations and their confidence levels; the recognition path records the model and decision-making steps used to obtain the result; and the processing time statistics show the total time spent completing the recognition and the breakdown of each stage. To support high-quality downstream applications, post-processing enhancements are performed on intent results: Normalization unifies the same intent expressed in different ways, improving consistency; intent expansion uses knowledge graph reasoning to supplement implicit but unexpressed related intent information; intent conflict detection identifies and resolves potentially contradictory intent expressions within the same sentence. Intent result logs are also maintained for subsequent analysis and improvement. This structured intent representation provides downstream applications (such as dialogues, intelligent assistants, or automated processes) with a clear, accurate, and information-rich understanding of user intent, significantly improving the interactive experience and task completion efficiency.

[0098] In step S3, identifying the constrained Pareto front and the unconstrained Pareto front specifically includes: identifying the optimal set of intent recognition schemes and constructing a constrained Pareto front to represent feasible schemes within the time limit based on strict delay constraints, quantifying the impact of delay constraints on recognition performance; identifying the ideal set of recognition schemes and constructing an unconstrained Pareto front to pursue the highest recognition accuracy based on ideal conditions without considering delay constraints, and establishing a performance benchmark reference standard; and establishing a classification-based UPF-CPF relationship strategy based on the constrained Pareto front and the unconstrained Pareto front to determine the dominance relationship between schemes, ensuring the identification of the optimal solution in the multi-objective optimization process.

[0099] First, we need to identify the Constrained Pareto Front (CPF), which represents the set of optimal intent recognition schemes achievable under strict delay constraints. These schemes must complete the intent recognition task within a preset time limit. Simultaneously, we also need to identify the Unconstrained Pareto Front (UPF), which represents the set of ideal recognition schemes without considering delay constraints. These schemes aim for the highest recognition accuracy and the most comprehensive semantic understanding. By comparing the differences between the CPF and UPF, we can quantify the impact of delay constraints on recognition performance.

[0100] Employing Hybrid Operators (HOps) technology, this approach intelligently combines various intent recognition strategies of varying complexity. These strategies include fast classifiers with extremely high response speeds but relatively low accuracy, medium-complexity models that strike a balance between speed and accuracy, and large language models that offer the highest precision despite longer processing times. Based on an epsilon constraint processing method, the weight allocation of different strategies is dynamically adjusted according to current performance requirements and constraints.

[0101] Regarding the scheme selection mechanism, a classification-based UPF-CPF relationship strategy was established. This strategy establishes a dominance relationship judgment mechanism between different identification schemes, ensuring accurate identification of the true optimal solution during multi-objective optimization. The Dynamic Resource Allocation (DRA) mechanism, as the core scheduling module, can adjust the computing resources allocated to various identification strategies in real time based on the current computing load, network latency, and user priority, ensuring that high-priority users receive better service quality.

[0102] Specifically, strict latency constraints are the time performance requirements that real-time or near-real-time intent recognition must meet. They specify the maximum allowed time from receiving input to generating the intent result. In this paper, strict latency constraints are formalized as a series of time thresholds and service level agreements (SLAs): the maximum response time T_max defines the absolute upper limit that processing must be completed within this time, typically set to 100-500 milliseconds; the service level target specifies the proportion of requests that must be processed within a specific time, such as "95% of requests must be completed within 200 milliseconds"; the quantile latency requirement specifies the upper limit of response time for different percentages of requests, such as "50% of requests within 100 milliseconds, 90% within 200 milliseconds, and 99% within 300 milliseconds"; and jitter control limits the variability of response time to ensure a stable user experience. Several factors were considered to determine these constraints: user perception studies showed that latency exceeding 300 milliseconds was noticeably perceived by users, impacting interaction smoothness; application scenarios required different latency tolerances, such as voice assistants needing faster responses while document analysis could tolerate slightly longer processing times; device capabilities limited processing power on resource-constrained devices; and network conditions considered transmission latency under different network environments. Latency performance was evaluated through precise time measurement and statistical analysis: latency distribution analysis captured the probability distribution characteristics of response times; long-tail latency analysis focused on performance under extreme conditions; and time series analysis tracked latency trends over time. To meet strict latency constraints, multiple optimization techniques were employed: request priority queues scheduled requests based on urgency and user priority; pre-calculation of common intents during idle periods; incremental processing allowing for early return of partial results; and time budget allocation assigning time quotas to each stage of the processing flow. These mechanisms enabled the provision of high-quality intent recognition services within strict time constraints.

[0103] The optimal intent recognition scheme set is the collection of configurations and algorithm choices that offer the best performance trade-offs given latency constraints. In constructing this set, a scheme space is first defined, encompassing multiple dimensions: model selection includes a variety of choices from lightweight models to large pre-trained models; batching strategy defines how input data is organized and processed; quantization level involves different options for reducing model accuracy to improve speed; hardware acceleration configuration includes strategies for using accelerators such as GPUs, TPUs, or FPGAs; and caching strategy defines how to leverage historical results to improve response speed. Optimal schemes are identified through a two-stage search process: the initial exploration phase uses Latin hypercube sampling to uniformly select sample points across the entire scheme space and evaluate their performance; the refined search phase uses Bayesian optimization or evolutionary algorithms to explore potential regions in depth. Each candidate scheme undergoes a comprehensive evaluation: accuracy testing assesses recognition accuracy on a standard test set; latency testing measures response time under different load conditions; resource consumption testing monitors resource usage during processing; and stability testing evaluates performance under long-term operation and edge conditions. The Pareto optimization principle is used to identify non-dominated schemes, i.e., those schemes that cannot simultaneously improve all objectives without sacrificing at least one objective. Finally, a subset of these non-dominated solutions that satisfy the strict delay constraint is selected to form the Constrained Pareto Front (CPF). Each solution in the CPF comes with detailed configuration parameters, performance characteristics, and applicable scenario descriptions, making it easy to select the most suitable solution based on actual needs.

[0104] The construction process of a constrained Pareto front involves organizing optimal solutions that satisfy delay constraints into a structured representation, which intuitively shows the performance trade-offs achievable within the time limit. The construction process includes five key steps: First, the data collection phase obtains performance data of candidate solutions under different conditions through experiments or simulations, including accuracy, response time, and resource consumption. Second, the constraint filtering phase applies strict delay constraints to eliminate all solutions that do not meet the time requirements, ensuring that all solutions on the front are feasible. Third, the non-dominated ranking phase applies the Pareto optimization principle to identify non-dominated solutions among the remaining solutions—those not completely superior to any other solution. Fourth, the front representation phase organizes the non-dominated solutions into a structured dataset, typically represented in a multi-dimensional space, with each dimension corresponding to an optimization objective. Fifth, the density evaluation phase analyzes the distribution characteristics of solutions on the front, identifies sparse and dense regions, and evaluates the coverage quality of the front. In practical implementation, a grid-based representation method is used, dividing the target space into grid cells and recording the optimal solution in each cell. This method is particularly efficient when handling high-dimensional spaces and a large number of solutions. To improve the accuracy and robustness of the representation, smoothing techniques are applied to reduce noise, and interpolation techniques are used to fill in blank areas on the front. Once the front is constructed, a series of descriptive metrics are calculated: hypervolume measures the volume of the target space covered by the front, reflecting the overall performance level; uniformity of distribution assesses the distribution of solutions on the front; extension measures the range of the target space covered by the front; and reference point difference compares the distance between the front and the ideal reference point. These metrics are used not only to evaluate the quality of the current front but also to compare it with historical fronts or unconstrained fronts, quantifying performance improvements or the impact of constraints.

[0105] The ideal set of recognition schemes is the set of optimal solutions based solely on recognition accuracy and resource efficiency, without considering latency constraints. These schemes represent the theoretically best performance achievable under no time pressure. The recognition process is similar to the optimal set of schemes, but all time-related constraints are removed, allowing the exploration of solutions that may have longer response times but extremely high accuracy. The construction of the ideal set of schemes focuses on the following aspects: full-precision versions of large pre-trained models, which typically have the highest accuracy but require longer processing times; ensemble learning techniques, combining predictions from multiple models to improve accuracy; deep inference chains, allowing models to perform multi-step inference to generate more accurate results; knowledge augmentation, integrating external knowledge bases to improve recognition quality; and contextual extension analysis, considering a broader range of dialogue history and background information. These schemes are evaluated through offline batch processing, unrestricted by real-time response requirements, comprehensively testing their performance on a variety of complex inputs. The construction of the ideal set of schemes also applies the Pareto optimization principle, but only considers accuracy and resource efficiency, without using response time as a selection criterion. The resulting set forms the unconstrained Pareto front (UPF), representing the upper limit of performance achievable under ideal conditions. While the set of ideal solutions may not be directly used in real-time applications, it provides important reference value: a performance ceiling benchmark, quantifying the gap between current real-time performance and theoretical optimality; a technology potential assessment, anticipating potential performance improvements brought about by future technological advancements; an algorithm research direction, guiding R&D teams to optimize existing algorithms and narrow the gap with ideal performance; and a basis for hardware upgrade decisions, evaluating the potential performance return on investment in new hardware.

[0106] The construction of unconstrained Pareto fronts is similar to the constrained version, but focuses on achieving the highest recognition accuracy under ideal conditions without time constraints. The construction process employs a broader search strategy, allowing for the exploration of computationally intensive solutions. First, sampling is performed across a larger solution space, including complex model architectures, deep ensemble strategies, and computationally intensive processing flows. Then, these candidate solutions are comprehensively evaluated, focusing on accuracy metrics while recording resource consumption, but disregarding response time constraints. The evaluation process is typically conducted in a high-performance computing environment, using batch processing to handle a large number of test samples. Next, a non-dominated ranking algorithm is applied to identify non-dominated solutions in terms of accuracy and resource efficiency, forming the initial unconstrained Pareto front. To improve the quality and coverage of the front, various optimization techniques are employed: front density enhancement algorithms generate new candidate solutions in the sparse regions of the front; extrapolation techniques predict the extension trend of the front, guiding the generation of new solutions; and ensemble optimization focuses on designing new model combinations, attempting to overcome the performance limitations of single models. After construction, the quality of the UPF is evaluated using the same metrics as the CPF, but it is generally expected that the UPF will exhibit higher performance levels and broader coverage. Once built, the UPF becomes an ideal benchmark for performance, used to evaluate the gap between any practical solution and the theoretical optimum. The UPF is updated regularly, typically whenever a new algorithm or model is released, to ensure it always represents the best possible performance with current technology.

[0107] The classification-based UPF-CPF relationship strategy is a method used to analyze and utilize the relationship between constrained and unconstrained Pareto fronts to assist in making optimal allocation decisions. The core of this strategy is to establish a classification framework that compares solutions in the UPF with those in the CPF, identifying their dominance relationships and performance gaps. First, four basic relationship categories are defined: Completely substitutable, indicating that the performance of the solution in the CPF is almost identical to its corresponding solution in the UPF, suggesting that the delay constraint has minimal impact on the solution; Partially degraded, indicating that the CPF solution performs slightly worse than its corresponding UPF solution, but the degradation is acceptable; Significantly degraded, indicating a significant performance gap between the CPF and UPF solutions, suggesting that the delay constraint has a significant impact; and Incomparable, indicating that the CPF solution has no direct counterpart in the UPF, typically representing a specific optimization result for particular constraints. Based on these fundamental relationships, a decision support framework was constructed: for solutions with completely substitutable relationships, they are adopted directly because they achieve near-ideal performance while satisfying latency constraints; for solutions with partially degraded relationships, the trade-off between performance degradation and response time improvement is evaluated, and latency constraints may be relaxed in critical scenarios; for solutions with significantly degraded relationships, alternative designs or resource enhancements are explored to attempt to narrow the gap with ideal performance; for solutions with incomparable relationships, specialized analysis is conducted to evaluate their applicability in specific application scenarios. To support this strategy, automated analysis tools were implemented: a gap quantification module calculates the Euclidean distance and performance ratio between the CPF solution and the nearest UPF solution; an equivalent solution identifier finds the most similar counterpart to the CPF solution in the UPF; and a performance trade-off visualization tool visually displays the trade-offs of different solutions for various objectives. Furthermore, a historical database is maintained to record the evolution of solution relationships over time and with technological advancements, providing a reference for long-term optimization strategies. Through this structured relationship analysis strategy, the true optimal solution can be identified under various operating conditions, maximizing performance while meeting practical time constraints.

[0108] Step S3 obtains the corresponding intent result, specifically including: based on the semantic analysis of the complete sentence candidate, identifying the user's true intent category and extracting key entity information and important parameters, and outputting a structured intent recognition result; based on the structured intent recognition result, establishing a mapping relationship between candidate sentences and intent recognition results and storing it in a high-speed cache to support fast retrieval and matching; based on the mapping relationship, outputting a complete intent result containing intent category identifiers, key entity information and operation parameters, and providing intelligent voice interaction services.

[0109] Specifically, semantic analysis is a natural language processing technique used to understand the meaning and internal structural relationships of text. In this paper, semantic analysis first performs multi-level processing on candidate complete sentences: lexical analysis decomposes the sentence into lexical units and labels them with parts of speech; syntactic analysis identifies the grammatical structure of the sentence and constructs a dependency tree; semantic role labeling identifies the function of each component in the sentence, such as agent, patient, and instrument; coreference resolution determines the referents of pronouns and demonstratives; and discourse intent analysis identifies the communicative function of the sentence, such as inquiry, command, or statement. A hybrid semantic model is used for analysis, combining statistical learning methods and neural network techniques, especially a bidirectional Transformer architecture enhanced with attention mechanisms, to capture context-dependent semantic representations. Furthermore, a domain knowledge graph is integrated to improve the understanding of domain-specific expressions. The semantic analysis process is adaptive, dynamically adjusting the analysis depth according to the complexity and ambiguity of the sentence, using lightweight analysis for simple and clear expressions and deep semantic reasoning for complex and ambiguous expressions. Through this comprehensive analysis, the true meaning of the sentence can be accurately understood, laying the foundation for subsequent intent recognition.

[0110] True intent category identification is the process of determining a user's actual purpose, mapping semantic analysis results to a predefined intent category system. A hierarchical intent classification framework is employed, comprising three levels: the top-level intent categories distinguish basic interaction purposes, such as querying information, performing actions, and expressing emotions; the middle-level intent categories subdivide specific domains and functions, such as weather queries, scheduling, and device control; and the bottom-level intent categories further distinguish specific operations, such as checking tomorrow's weather, adding a reminder, and adjusting volume. The identification process combines multiple technologies: rule-based intent matching uses predefined patterns to identify explicit expressions; a statistical classifier processes standard expressions, classifying them through feature extraction and machine learning models; deep learning classification uses neural networks to process complex expressions, capturing implicit relationships; and an intent reasoning engine processes implicit expressions, determining potential intents through logical reasoning. An ensemble approach is used to integrate the results of multiple models, improving identification accuracy. Simultaneously, the certainty of the intent is evaluated, assigning a reliability score to each identification result. This multi-level, multi-strategy identification method accurately captures the user's true intent, even in cases of ambiguous or indirect expression.

[0111] Key entity information and important parameter extraction is the process of identifying specific data points in a sentence that are relevant to the execution of the intent. Key entities refer to noun phrases in a sentence, such as names, locations, organizations, and products; important parameters are attribute values ​​related to the execution of the intent, such as time, quantity, and status. Named entity recognition technology is used to identify standard entity types, and sequence labeling models (such as BiLSTM-CRF or BERT-based models) are used to process the input text. For domain-specific entities, a professional dictionary and a customized model are combined for identification. Parameter extraction involves value normalization, converting attribute values ​​expressed in natural language (such as "next Tuesday afternoon") into a standard format (such as "2023-10-10 14:00:00"). Relationship extraction is also implemented to identify the semantic relationships between entities and intents, as well as between entities, determining the role of each entity in the execution of the intent. For incomplete expressions, a context completion mechanism is applied to supplement missing entities and parameters from dialogue history or environmental information. Through this comprehensive information extraction, all specific data points required to execute the intent can be obtained.

[0112] Structured intent recognition results convert natural language understanding into a standard machine-processable output format. It uses a JSON or similar structure and includes several key components: an intent identifier indicating the identified intent category; a confidence score quantifying the certainty of the identification; an entity list containing all extracted entities, their types, and values; a parameter map recording all operation parameters and their values; a context dependency flag indicating whether the result depends on the dialogue history; and a slot fill status recording the completeness of necessary information. The results are standardized to ensure that the same intent expressed in different ways produces a consistent structured representation. The structured results also include the correspondence between the original text and structured elements, facilitating interpretation and error analysis. This standardized structure allows downstream components to directly process intent information without further natural language understanding.

[0113] Mapping relationships and high-speed caching are key mechanisms for improving response speed and efficiency. Mapping relationships associate complete sentence candidates with their corresponding structured intent results, forming <sentence, intent result> pairs. A two-level mapping structure is employed: precise mapping stores perfectly matching sentence-intent pairs; fuzzy mapping stores the association between sentence templates and intent templates, supporting variable substitution. The mapping construction process includes sentence normalization, feature extraction, and index building. The high-speed cache employs a multi-layered design: L1 cache stores the most recently and frequently used mappings, residing in memory; L2 cache stores frequently used mappings, possibly in memory or fast storage devices; persistent storage holds the complete mapping library for cold starts and recovery. Cache management uses a strategy combining the LRU-K algorithm and frequency analysis, prioritizing the retention of high-value mappings. An incremental update mechanism is also implemented, continuously enriching and optimizing the mapping library based on new identification results. Through this efficient caching, results can be retrieved directly for repetitive or similar expressions without re-executing the entire analysis process, significantly improving response speed.

[0114] The complete intent result is the final output of user intent understanding. It integrates the results of all analysis and processing, providing accurate and comprehensive intent representations for downstream applications. Building upon structured intent recognition, the complete intent result adds extra information: execution priority indicates the urgency of the intent; intent dependencies record the preceding or following relationships with other intents; execution status tracking indicates the processing stage of the intent; and the applicable service list indicates the components that can handle the intent. A consistency check is performed on the complete intent result to ensure there are no logical conflicts between its components. Finally, these intent results are passed to downstream components of the intelligent voice interaction service, such as dialogue managers, task execution engines, or content recommendations, achieving a complete closed loop from natural language understanding to actual function execution.

[0115] In step S4, after the user finishes expressing the sentence, the system uses an efficient and accurate text similarity algorithm to match the most similar predicted sentence and directly obtains the corresponding intent result. In the initial stage of this phase, the speech recognition system outputs the user's complete actual sentence, while the sentence integrity detection module uses advanced grammatical analysis technology and speech pause detection algorithms to confirm the complete end of the sentence, ensuring that the obtained result represents the user's complete expressed intent.

[0116] During the text preprocessing stage, the text standardization module performs uniform standardization on the user's actual sentence and all pre-generated candidate sentences. This process includes precise word segmentation, detailed part-of-speech tagging analysis, and intelligent synonym normalization, ensuring that all texts involved in the comparison are processed under the same standardized format, thereby improving the accuracy and reliability of similarity calculation.

[0117] In the core stage of similarity calculation, an advanced multi-level similarity algorithm is employed, which can comprehensively evaluate text from multiple dimensions. Specifically, this includes lexical similarity analysis, primarily focusing on word overlap and frequency distribution; semantic similarity calculation, utilizing deep learning models to understand the deeper meaning of the text; and structural similarity assessment, analyzing sentence grammatical structure and expression patterns. This multi-level comprehensive evaluation method can more comprehensively capture the similarity features between texts.

[0118] During the matching decision phase, the matching algorithm comprehensively considers similarity scores at various levels and selects the candidate sentence with the highest overall similarity score as the best matching result. Simultaneously, it verifies whether the highest similarity score exceeds a preset credibility threshold to ensure the reliability of the matching result. Once the best matching candidate sentence is determined, the corresponding intent recognition result can be quickly retrieved directly from the previously established cache, completely avoiding the time delay caused by re-performing large-scale model inference.

[0119] Ultimately, it will output complete and detailed intent recognition results, which include accurate intent category identifiers, extracted key entity information, and related operation parameters, providing users with accurate and efficient intelligent voice interaction services.

[0120] In step S4, a multi-level text similarity algorithm is used to calculate the similarity between the complete actual sentence and the multiple high-confidence complete sentence candidates. The candidate sentence with the highest similarity score is selected, and the corresponding final intent recognition result is directly obtained based on the mapping relationship. Specifically, this includes: based on the user's complete sentence and the complete sentence candidates, a multi-objective evolutionary algorithm is used to decompose the similarity calculation into a multi-objective optimization problem, which is decomposed into four sub-objectives: lexical level, syntactic level, semantic level, and contextual relevance, resulting in four sub-objective optimization results; based on the four sub-objective optimization results, a dynamic adaptive weight vector generation mechanism is constructed to adjust the sub-objective weights according to the feature difference pattern, realizing a personalized similarity evaluation strategy, resulting in a personalized similarity evaluation result; based on the personalized similarity evaluation result, the candidate sentence with the highest overall similarity score and exceeding the confidence threshold is selected, and the corresponding final intent recognition result is directly obtained.

[0121] By introducing the decomposition concept of multi-objective evolutionary algorithms (MOEA / D), text similarity calculation is transformed into a multi-objective optimization problem, and adaptive weight vectors are used to improve the multi-scale attention (MSA) estimation mechanism. The traditional single text similarity calculation task is decomposed into four interrelated sub-objective optimization problems. Lexical similarity primarily evaluates the matching degree between two sentences in terms of word selection and word frequency distribution, quantified by calculating the coverage of common words and TF-IDF weight similarity. Syntactic similarity focuses on the grammatical structure and sentence patterns of sentences, evaluated using the structural similarity of dependency parsing trees and the matching degree of part-of-speech tagging sequences. Semantic similarity employs a pre-trained semantic vector model to calculate the cosine similarity and semantic angle bias of sentences in a high-dimensional semantic space. The contextual relevance sub-objective considers the adaptability of sentences in specific dialogue scenarios and application domains, measured through the relevance analysis of context embedding vectors.

[0122] Specifically, the multi-level text similarity algorithm is a method for comprehensively evaluating the similarity between two texts at different levels. This algorithm overcomes the limitations of traditional single-dimensional similarity calculations by considering text features at multiple levels simultaneously, achieving more accurate similarity assessment. Specifically, the algorithm calculates similarity at four key levels: at the lexical level, it analyzes the shared lexical units and their distribution between the two texts, applying an improved Jaccard coefficient and cosine similarity to calculate the overlap of common words; at the syntactic level, it evaluates the syntactic structural similarity between the two texts, measuring structural similarity through syntactic tree edit distance and dependency relation matching; at the semantic level, it explores the semantic closeness between the two texts, using word embedding techniques and context-dependent semantic representations to calculate the similarity of semantic vectors; and at the context-dependent level, it analyzes the coherence of the text with the preceding and following dialogue history, evaluating the naturalness of each candidate in the current dialogue flow. The algorithm uses a sliding window technique to process long texts, ensuring that local similarity is not diluted by overall averaging, and employs a hierarchical weighted mechanism to integrate the similarity scores at each level, generating a final comprehensive similarity score. It also implements parallel processing of similarity calculation, improves computational efficiency through GPU acceleration and batch processing technology, and ensures that real-time response can still be maintained even with a large number of candidate sentences.

[0123] Multi-objective evolutionary decomposition is a methodology that breaks down complex problems into multiple simpler sub-problems and solves them through evolutionary computation. In similarity calculation, this approach decomposes the overall similarity assessment into multiple independent but related sub-objectives, each focusing on a different aspect of text similarity. First, a specific objective function is defined for each sub-objective: at the lexical level, TF-IDF-weighted n-gram matching rate is used; at the syntactic level, a tree kernel function is applied to calculate syntactic structural similarity; at the semantic level, a pre-trained language model is used to extract semantic vectors and calculate cosine similarity; contextual relevance is quantified using conditional probability and coherence scoring. Then, a dedicated optimizer is built for each sub-objective, using evolutionary computational methods such as genetic algorithms or particle swarm optimization to search for optimal parameter configurations. These optimizers run in parallel, evolve independently, and periodically exchange information to leverage cross-objective synergies. A non-dominated ranking mechanism is used to evaluate the quality of solutions, identifying the Pareto optimal set of solutions—those solutions that cannot be simultaneously surpassed by other solutions on at least one sub-objective. This decomposition and evolutionary approach effectively addresses the multi-dimensional trade-offs in similarity calculation, yielding more comprehensive and accurate similarity assessment results.

[0124] The four sub-objective optimization results are the outputs of independent optimization of each sub-objective after decomposition by the multi-objective evolutionary algorithm, each reflecting a specific aspect of text similarity. Lexical optimization results include three indicators: word overlap, keyword matching rate, and word frequency distribution similarity, reflecting the degree of sharing in surface text. Syntactic optimization results include three indicators: syntactic tree similarity, dependency relation matching rate, and syntactic pattern overlap rate, quantifying structural similarity. Semantic optimization results include three indicators: semantic vector similarity, topic distribution similarity, and entity relation consistency, representing the degree of semantic proximity. Contextual relevance optimization results include historical coherence score, logical transition naturalness, and dialogue fluency indicators, assessing adaptability in dialogue. Each sub-objective optimization result includes not only numerical scores but also detailed contribution analysis, identifying which specific features contribute most to the similarity score, providing a basis for subsequent weight adjustments. These optimization results collectively constitute a complete profile of similarity assessment, providing a comprehensive information foundation for the final decision.

[0125] The dynamic adaptive weight vector generation mechanism is an intelligent approach that dynamically adjusts the importance of each sub-objective based on specific input text features. Unlike traditional fixed-weight methods, this mechanism can identify text feature patterns and adaptively generate the most suitable weight allocation scheme for the current comparison task. First, it extracts the feature differences between text pairs, such as differences in lexical complexity, syntactic complexity, and semantic richness. Then, it maps these features into weight vectors using a trained neural network model trained on a large amount of labeled data, learning the most effective weight configurations under different text feature patterns. Next, it applies a context-sensitive adjustment mechanism to fine-tune the initial weights based on the current dialogue history and domain context. Finally, it performs real-time feedback optimization, continuously adjusting and improving the weight generation model based on the accuracy of historical similarity calculations. This dynamic mechanism can adapt to different types of text comparison tasks. For example, for texts dense with technical terms, it automatically increases the weights at the lexical level; for texts with implicit expressions, it enhances the weights at the semantic level. Through this personalized weight allocation, it achieves more accurate and adaptive similarity assessment.

[0126] The personalized similarity assessment result is a comprehensive score generated after applying a dynamically adaptive weight vector to the optimization results of four sub-objectives. It reflects the overall similarity between two texts under a specific weight configuration. The scores of each sub-objective are integrated using a weighted summation method, and the confidence interval of the comprehensive score is calculated to quantify the uncertainty of the assessment result. In addition to the overall similarity score, the personalized assessment result also includes sub-level contribution analysis, feature matching visualization, and a similarity explanation summary to help understand the composition and basis of the similarity score. A similarity tier classification is also generated, mapping scores to intuitive levels such as high, medium, and low for quick decision-making. For candidate sentences with close similarity, a difference analysis is performed to highlight subtle but potentially crucial differences, aiding in the final selection. This personalized assessment result not only provides numerical scores but also rich contextual information, making the similarity assessment process more transparent and interpretable.

[0127] The credibility threshold is a dynamic decision criterion used to filter high-quality candidate sentences. It defines the minimum similarity requirement for accepting a candidate sentence as a matching result. The threshold is dynamically set based on the application scenario and historical performance: higher thresholds are used for critical operations (such as financial transactions or control) to ensure high accuracy; lower thresholds are used for low-risk scenarios such as information queries to improve recall; and personalized thresholds are adaptively adjusted based on user feedback history. Threshold calculation considers multiple factors: a global baseline threshold sets the basic standard; task complexity adjustment adjusts the threshold according to the current task difficulty; error cost analysis assesses the potential impact of mismatches; and a user preference model reflects individual error tolerance. A progressive application strategy for the threshold is implemented: a standard threshold is used for the initial matching; if no candidates meet the conditions, the threshold is gradually lowered and additional verification steps are introduced; when the threshold is lowered to the minimum and there is still no match, a backup processing flow is triggered, such as prompting the user to confirm or rephrase. Through this flexible threshold mechanism, the probability of successful matching is maximized while ensuring matching quality.

[0128] The final intent recognition result is the output of the final understanding of the user's intent. It is obtained by directly querying the mapping relationship in the cache, avoiding the repeated execution of complex intent analysis processes. This result adopts a standardized format and contains five core components: an intent type identifier that clearly defines the user's basic intent category; an entity list that records all key entities identified and their attributes in detail; an operation parameter set containing all parameters and their values ​​required to execute the intent; context dependency information that marks the association between the intent and the dialogue history; and metadata that adds auxiliary information such as processing time, confidence level, and source. A consistency check is performed on the final result to ensure that the logic of all components is coordinated, and normalization processing is applied to unify the representation of the same intent in different expressions. In addition, an intent result summary is generated to describe the user's intent in concise natural language, facilitating review and confirmation. The final intent recognition result is passed to downstream modules, such as the dialogue manager or task execution engine, to achieve a seamless connection from natural language understanding to function execution, completing the entire intelligent voice interaction process.

[0129] Step S4 involves constructing a dynamic adaptive weight vector generation mechanism to adjust the sub-target weights based on feature difference patterns. Specifically, this includes: designing a multi-scale attention estimator based on the optimization results of the four sub-targets to compute attention weights in parallel at three granular levels: word, phrase, and sentence, capturing semantic correspondences at different levels, avoiding semantic information missed by single-granularity analysis, and obtaining a multi-scale attention weight distribution; implementing an online weight vector learning algorithm based on historical feedback data based on the multi-scale attention weight distribution, continuously optimizing the weight distribution strategy, and obtaining an optimized weight distribution strategy; and establishing a Pareto optimal solution set maintenance mechanism based on the optimized weight distribution strategy to store the optimal matching scheme under multiple weight configurations, supporting personalized and scenario-based matching scheme selection, and obtaining the personalized similarity evaluation result.

[0130] A dynamic and adaptive weight vector generation mechanism was constructed, which can adjust the importance weight distribution of the four sub-objectives in real time based on the feature difference patterns between candidate sentences and actual sentences. When a significant difference in word selection is detected between the candidate sentence and the actual sentence, the weight of semantic similarity is automatically increased, while the influence of lexical similarity is reduced; when there are significant differences in sentence length and structure, the weight contribution of syntactic similarity is enhanced. This adaptive adjustment mechanism is implemented based on statistical analysis of feature vectors and machine learning prediction models, and can effectively cope with changes in the expression habits and language styles of different users.

[0131] A multi-scale attention estimator was designed, which computes attention weight distributions in parallel at three different linguistic granularities. At the word level, the estimator focuses on the correspondence and semantic relevance between individual words, identifying keyword matching patterns. At the phrase level, the estimator analyzes the phrase structure and semantic units formed by word combinations, capturing local semantic correspondences. At the sentence level, the estimator evaluates the similarity between sentences from the perspective of overall semantics and expressive intent. This multi-scale attention mechanism can more comprehensively capture semantic correspondences at different levels, avoiding important semantic information that may be missed by single-granularity analysis.

[0132] An online weight vector learning algorithm based on historical data feedback was implemented. This algorithm continuously collects implicit and explicit user feedback on matching results, including subsequent user actions, error correction information, and satisfaction ratings. By analyzing this feedback data, the algorithm can identify the most effective weight configuration patterns in different application scenarios and continuously optimize the weight distribution strategy using incremental learning techniques. The learning process employs a combination of gradient descent and reinforcement learning, ensuring both learning efficiency and the stability and convergence of weight adjustments.

[0133] A robust Pareto-optimal solution set maintenance mechanism has been established, capable of simultaneously storing multiple optimal matching schemes under different weight configurations. A dynamically updated solution set storage structure is maintained, containing non-dominated solutions obtained under different combinations of optimization objective weights. When faced with a new matching task, the most suitable matching strategy for the current scenario characteristics can be selected from this diverse solution set, supporting personalized and scenario-based matching scheme selection.

[0134] Specifically, the multi-scale attention estimator is a comprehensive text analysis tool capable of evaluating correspondences between texts at multiple linguistic granular levels. At the word level, the estimator uses a bidirectional attention flow network to compute a similarity matrix between words, with each word represented as a context-sensitive embedding vector, considering positional information and contextual factors. At the phrase level, a sliding window technique is employed to identify n-gram phrase units, using combinatorial semantic representation and tree-structured encoding to capture correspondences between phrases. At the sentence level, the estimator applies a global encoder to generate an overall semantic representation, evaluating the consistency of sentence topic and structure. Multi-scale processing integrates information from each layer through residual connections and attention fusion mechanisms to generate a three-dimensional semantic mapping. It also implements adaptive granularity adjustment, dynamically determining the most suitable analysis granularity weight based on text characteristics; for example, it enhances word-level analysis weights for texts dense with technical terminology, and enhances sentence-level analysis weights for complex expressions.

[0135] The multi-scale attention weight distribution is a three-dimensional text correspondence mapping generated by the estimator, containing three levels of weight data. A word-level weight matrix records the similarity score between each pair of words, highlighting keyword matches; a phrase-level weight tensor captures the semantic correspondence between phrases of different lengths; and a sentence-level weight vector represents the overall matching degree and the importance distribution of key structures. These three weight levels are normalized to ensure numerical comparability, and the concentration and dispersion indices of each layer's distribution are calculated to quantify the weight pattern features. The weight distribution also includes key matching point identifiers, identifying the most significant semantic connection nodes at each level. This multi-level weight distribution provides a comprehensive feature map for subsequent weight vector adjustments, enabling the customization of optimal weight configurations based on the specific characteristics of the text.

[0136] The online weight vector learning algorithm based on historical feedback data is an adaptive optimization that continuously improves the weight allocation strategy by analyzing user interaction feedback. The algorithm collects various feedback signals, including direct user confirmation, task completion status, and internal evaluation results; it constructs training samples, storing input text features, application weight configurations, and feedback results in a correlated manner; and it applies online learning techniques, including incremental gradient descent, Bayesian updates, and an exploration-utilization balancing strategy, to continuously optimize model parameters. The learning adopts a hierarchical architecture, comprising three levels: general pattern learning, domain-specific optimization, and personalized adjustment, ensuring that the weight configuration is both universal and adaptable to specific scenarios and users. The algorithm also implements a rapid adaptation mechanism triggered by key events, accelerating the learning process and improving responsiveness when significant errors or pattern changes are identified.

[0137] The optimized weight distribution strategy is a mature decision-making model generated by an online learning algorithm, capable of intelligently generating the most suitable weight configuration based on text features. The core of the strategy is a feature-weight mapping model, combined with deep neural networks to handle complex nonlinear relationships and decision trees to provide interpretable adjustment paths. The strategy includes a context-aware adjuster that fine-tunes the base weights based on the dialogue state; a domain adapter that applies domain-specific rules; and a personalized component that integrates user preferences and historical patterns. To ensure stability, the strategy sets weight change limits and a gradual adjustment mechanism to prevent instability caused by a single large adjustment; it also maintains a weight version library to support rollback to a reliable configuration when the new strategy performs poorly.

[0138] The Pareto optimal solution set maintenance mechanism is a multi-objective optimization management approach that saves and maintains multiple non-dominated optimal solutions, providing diverse optimal choices for different scenarios. The mechanism performs multi-dimensional evaluation of matching schemes generated for each weight configuration, applies a non-dominated ranking algorithm to identify Pareto optimal solutions, which represent the best trade-off points under their respective weight configurations; uses crowding calculation to maintain solution set diversity, ensuring coverage of the entire decision space; and maintains a dynamic solution set, continuously updating it as new solutions are generated. It employs grid indexing technology to accelerate solution lookup and uses incremental update algorithms to reduce the need for complete reordering. The application process includes three steps: scenario matching, personalized filtering, and final selection. Based on the characteristics of the current task and user preferences, it selects the most suitable solution, achieving truly personalized similarity assessment.

[0139] In a preferred embodiment, the method further includes a response time optimization mechanism: based on real-time system performance monitoring data, the number of candidates for text prediction and the resource allocation for parallel inference are dynamically adjusted to ensure that the overall response latency is controlled within a preset threshold, thereby obtaining a resource allocation optimization result; based on the resource allocation optimization result, a prediction caching and result reuse strategy is adopted to reduce redundant calculations, further reducing the system response latency, thereby obtaining latency optimization measures; based on the latency optimization measures, intent prediction and matching preparation before the user's complete expression are realized, significantly improving the user interaction experience.

[0140] Specifically, the response time optimization mechanism is a performance management system that ensures text similarity calculations are completed within a specified time, maintaining a good interactive experience. It achieves comprehensive performance monitoring and continuously collects key indicators: processing queue length reflects the number of pending requests; component response time tracks the time consumed at each processing stage; resource utilization monitors CPU, memory, and network usage; and terminal performance data collects the processing capabilities of different user devices. The system employs an adaptive scheduling algorithm, dynamically adjusting processing parameters based on monitoring data and response time targets: candidate number control dynamically limits the number of candidate sentences generated and evaluated based on the current system load, reducing the number of candidates under high load to ensure response speed; parallelism adjustment adjusts the number of tasks processed in parallel based on available computing resources; and batch processing optimization merges multiple requests for processing, improving throughput. It also implements intelligent resource allocation strategies: priority queues ensure critical requests are processed first; resource pool management dynamically allocates and reclaims computing resources; and dedicated hardware allocation assigns GPUs or dedicated accelerators to computationally intensive tasks. When the load exceeds processing capacity, a service degradation strategy is automatically activated, systematically reducing service complexity and ensuring the availability of core functions.

[0141] The resource allocation optimization result is the output of the response time optimization mechanism, allocating the most suitable computing resources to each component of the system. This result includes a detailed resource configuration scheme: processing unit allocation indicates the number of CPU cores and GPU units used by each component; memory quota sets memory limits for each processing stage; batch size determines the optimal batch processing scale; and the number of parallel tasks limits the number of tasks executed simultaneously. The optimization result adopts a layered design: a guarantee layer ensures basic resources for core functions; an elastic layer dynamically adjusts according to load fluctuations; and a reservation layer reserves emergency resources for sudden demands. Resource allocation is continuously adjusted through a continuous feedback loop: performance metrics monitoring tracks the latency and throughput of each component in real time; resource utilization analysis identifies system bottlenecks; and a load prediction model predicts recent demand changes based on historical patterns. The optimization result also defines boundary conditions for resource usage: a maximum resource limit prevents a single request from consuming too many shared resources; a minimum resource guarantee ensures basic service quality; and a fair usage strategy balances resource allocation among different users.

[0142] Predictive caching and result reuse strategies are a set of specialized techniques that further reduce system response time by minimizing redundant computation. The predictive caching mechanism pre-calculates possible inputs and their results based on user behavior patterns and language habits: frequency analysis identifies high-frequency query patterns and prioritizes caching the processing results of common expressions; context prediction predicts possible user input for the next round based on dialogue history and current state; user model caching builds personalized caches based on specific users' historical interaction patterns. Result reuse technology reduces redundant processing by identifying repetitive parts of the computation: computation result caching stores intermediate computation results to avoid repeated evaluation; partial result sharing allows different requests to share common computation steps; incremental computation recalculates only the changed parts for slightly modified inputs. An intelligent cache management strategy is employed: time-to-live (TTL) control sets cache expiration times based on data timeliness; a priority replacement algorithm prioritizes high-value results when cache space is limited; and a cache preheating mechanism proactively fills the cache when system load is low. An adaptive control mechanism is also implemented to continuously evaluate the effectiveness of each optimization strategy and dynamically activate the most effective measures.

[0143] Intent prediction and matching preparation is a proactive processing technique that begins analysis and preparation before the user completes their expression, significantly reducing perceived latency. It achieves real-time incremental processing: continuously updating analysis results as user input progresses, rather than waiting for complete input; employing a progressive recognition strategy, gradually refining from initial guesses to precise understanding. Prediction technology is the core mechanism: input prediction predicts possible complete expressions based on partial input and a language model; pre-identification of intent begins identifying possible intent categories before input is complete; parallel hypothesis processing considers multiple possible input completion paths simultaneously. Proactive resource scheduling is implemented: pre-loading potentially needed models and data; pre-allocating computing resources to prepare for upcoming processing demands; and pre-generating responses to high-probability predicted paths. To balance prediction benefits and costs, an intelligent control mechanism is implemented: value assessment models calculate potential time savings in prediction processing; risk assessment prevents resource waste due to prediction errors; and adaptive thresholds adjust the aggressiveness of predictions based on historical accuracy. This proactive processing approach significantly reduces user-perceived response latency, creating a more natural and fluid interactive experience, particularly effective in voice interaction scenarios.

[0144] like Figure 2 As shown, the present invention also provides a large-scale real-time speech-text intent recognition system based on text prediction, comprising:

[0145] The streaming speech recognition module 10 is used to acquire the user's real-time speech data, perform real-time speech recognition processing through the streaming speech recognition interface, and obtain partial transcribed text.

[0146] The text prediction module 20 is used to input the partially transcribed text into a masked language model for text prediction based on the partially transcribed text, and generate several highly reliable complete sentence candidates.

[0147] The parallel intent recognition module 30 is used to input the complete sentence candidates into the large language model in parallel for intent recognition based on the complete sentence candidates, obtain the corresponding intent results, and establish a mapping relationship between the candidate sentences and the intent recognition results;

[0148] The similarity matching module 40 is used to obtain the user's complete expression of the sentence, calculate the similarity between the complete actual sentence and the multiple high-confidence complete sentence candidates through a multi-level text similarity algorithm, select the candidate sentence with the highest similarity score, and directly obtain the corresponding final intent recognition result based on the mapping relationship.

[0149] The advantages and effects of this invention are significant. By predicting possible complete sentences based on identified partial text before the user has completed their speech expression, and performing intent recognition in advance, the system response latency is significantly reduced, improving the user interaction experience. The use of streaming speech recognition and parallel intent recognition achieves efficient collaborative processing of speech recognition and intent understanding. A multi-level text similarity matching algorithm ensures the accuracy and reliability of the final intent recognition result. The overall system architecture is flexible and efficient, adaptable to different application scenarios and user needs.

[0150] Although the embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A real-time speech-text intent recognition method based on a large model of text prediction, characterized in that, include: The system acquires the user's real-time voice data, performs real-time speech recognition processing through a streaming speech recognition interface, and obtains a partial transcribed text. Based on the partially transcribed text, the partially transcribed text is input into a masked language model for text prediction, generating several highly reliable complete sentence candidates. Based on the complete sentence candidates, the complete sentence candidates are input into the large language model in parallel for intent recognition to obtain the corresponding intent results, and a mapping relationship between the candidate sentences and the intent recognition results is established. The user's complete sentence is obtained, and the similarity between the complete actual sentence and the multiple highly credible complete sentence candidates is calculated using a multi-level text similarity algorithm. The candidate sentence with the highest similarity score is selected, and the corresponding final intent recognition result is directly obtained based on the mapping relationship. The step of inputting the candidate complete sentences into a large language model in parallel for intent recognition to obtain the corresponding intent result includes: constructing a constrained multi-objective optimization problem based on the candidate complete sentences to simultaneously optimize recognition accuracy, response time, and resource consumption, and identifying constrained Pareto fronts and unconstrained Pareto fronts; intelligently combining a fast classifier, a medium-complexity model, and a large language model using hybrid operator technology based on the constrained Pareto fronts and the unconstrained Pareto fronts to obtain a multi-strategy fusion result; and adjusting computing resources in real time according to system load and user priority based on the multi-strategy fusion result through a dynamic resource allocation mechanism to obtain the corresponding intent result.

2. The method according to claim 1, characterized in that, The real-time speech recognition processing via the streaming speech recognition interface yields partially transcribed text, including: A real-time audio stream connection is established and a sampling strategy with a fixed time interval is adopted to collect audio segments from the real-time voice data according to a preset time window, thereby generating a continuous audio segment sequence. Based on the continuous audio data segments, acoustic feature extraction and temporal information decoding are performed using a streaming ASR engine to generate real-time text output; For the real-time text output, a dynamic sliding window mechanism is used to cache the most recent text fragments and perform standardization processing to obtain the partially transcribed text.

3. The method according to claim 1, characterized in that, The process of acquiring the user's real-time voice data and performing real-time voice recognition processing through a streaming speech recognition interface includes: Based on the real-time speech data, an encoder-decoder architecture based on an attention mechanism is used to perform real-time decoding of audio features, thereby achieving synchronous processing of audio input and text output and obtaining synchronous processing results. Based on the synchronous processing results, text cleaning and standardization are performed to remove filler words and recognition errors, ensuring that the text quality meets the requirements of subsequent processing, and generating cleaned text fragments. Based on the cleaned text fragments, when the text quality meets the requirements and the length reaches the preset minimum threshold, the subsequent text prediction process is automatically triggered.

4. The method according to claim 1, characterized in that, The process involves inputting the partially transcribed text into a masked language model for text prediction, generating several highly reliable candidate complete sentences, including: Using the transcribed text as contextual information, a mask marker is added to the end of the text, and a bidirectional attention mechanism is used for prediction to generate multiple possible sentence completion schemes. Based on the sentence completion scheme, the credibility score of each candidate sentence is calculated to ensure the quality and reliability of the generated candidates; Based on the confidence scores, the top N candidate results with the highest confidence are retained to obtain the candidate complete sentences.

5. The method according to claim 1, characterized in that, The step of inputting the partially transcribed text into a masked language model for text prediction based on the partially transcribed text includes: Based on the partially transcribed text, a bidirectional attention mechanism is used to simultaneously process the preceding and following inference information to perform comprehensive semantic prediction and generate text completion results. For the text completion results, multiple complete sentence structure schemes are generated using parallel processing to ensure the grammatical and semantic quality of the candidate sentences; Based on the multiple complete sentence structure schemes, the model probability distribution and semantic coherence are comprehensively considered for selection, and the candidate complete sentences are output.

6. The method according to claim 1, characterized in that, The process involves inputting the partially transcribed text into a masked language model for text prediction, generating several highly reliable candidate complete sentences, including: Based on the partially transcribed text, text prediction is performed by replacing the traditional masked language model with a binary weighted spike neural network to generate candidate sentences encoded by time-series pulse signals. For the candidate sentences encoded by the time-series pulse signal, the weights are quantized into a binary form of {-1, +1} using a sign function to obtain the binary weight processing result; Based on the binary weight processing results, different sentence types are predicted by multiple parallel BWSNN sub-networks to obtain the complete sentence candidates.

7. The method according to claim 6, characterized in that, The process of predicting different sentence types through multiple parallel BWSNN sub-networks to obtain the complete sentence candidate includes: Based on the binary weight processing results, interrogative sentence prediction subnetwork, declarative sentence prediction subnetwork and imperative sentence prediction subnetwork are constructed. The pulse timing coding strategy is used to encode semantic information by utilizing the firing time and frequency of neurons to obtain progressive text completion results. Based on the progressive text completion results, the prediction results of multiple sub-networks are weighted and fused by an integrated decision fusion module to generate a preliminary candidate sentence list; Based on the preliminary candidate sentence list, a dynamic weight allocation strategy and confidence evaluation mechanism are used to obtain the complete sentence candidates, ensuring the quality and reliability of the complete sentence candidates.

8. The method according to claim 1, characterized in that, The identification of constrained Pareto fronts and unconstrained Pareto fronts includes: Based on strict delay constraints, we identify the set of optimal intent recognition schemes and construct a constrained Pareto front to represent feasible schemes within the time limit, thus quantifying the impact of delay constraints on recognition performance. Based on ideal conditions without considering delay constraints, we identify a set of ideal recognition schemes and construct an unconstrained Pareto front to pursue the highest recognition accuracy, and establish a performance benchmark reference standard. Based on the constrained Pareto front and the unconstrained Pareto front, a classification-based UPF-CPF relationship strategy is established to determine the dominance relationship between schemes, ensuring the identification of the optimal solution in the multi-objective optimization process.

9. The method according to claim 1, characterized in that, The process involves calculating the similarity between the complete actual sentence and the multiple highly reliable candidate complete sentences using a multi-level text similarity algorithm, selecting the candidate sentence with the highest similarity score, and directly obtaining the corresponding final intent recognition result based on the mapping relationship. This includes: Based on the user's complete sentence and the candidate complete sentences, the similarity calculation is transformed into a multi-objective optimization problem using the idea of ​​multi-objective evolutionary algorithm decomposition. It is decomposed into four sub-objectives: lexical level, syntactic level, semantic level, and contextual relevance, and the optimization results of the four sub-objectives are obtained. Based on the optimization results of the four sub-objectives, a dynamic adaptive weight vector generation mechanism is constructed to adjust the weights of the sub-objectives according to the feature difference pattern, thereby realizing a personalized similarity evaluation strategy and obtaining personalized similarity evaluation results. Based on the personalized similarity evaluation results, the candidate sentences with the highest overall similarity scores that exceed the credibility threshold are selected, and the corresponding final intent recognition results are directly obtained.

10. The method according to claim 9, characterized in that, The mechanism for constructing a dynamic adaptive weight vector generation method adjusts the weights of sub-targets based on feature difference patterns, including: Based on the optimization results of the four sub-objectives, a multi-scale attention estimator is designed to compute attention weights in parallel at three granularity levels: word level, phrase level, and sentence level. This captures the semantic correspondences at different levels, avoids missing semantic information in single-granularity analysis, and yields a multi-scale attention weight distribution. Based on the multi-scale attention weight distribution, an online learning algorithm for weight vectors based on historical feedback data is implemented to continuously optimize the weight distribution strategy and obtain an optimized weight distribution strategy. Based on the optimized weight distribution strategy, a Pareto optimal solution set maintenance mechanism is established to store the optimal matching scheme under multiple weight configurations, supporting personalized and scenario-based matching scheme selection, and obtaining the personalized similarity evaluation result.

11. The method according to claim 1, characterized in that, Obtaining the corresponding intent result includes: Based on the semantic analysis of the complete sentence candidates, the user's true intent category is identified and key entity information and important parameters are extracted, and a structured intent recognition result is output. Based on the structured intent recognition results, a mapping relationship between candidate sentences and intent recognition results is established and stored in a cache to support fast retrieval and matching; Based on the mapping relationship, a complete intent result containing intent category identifier, key entity information and operation parameters is output, providing intelligent voice interaction services.

12. The method according to claim 1, characterized in that, It also includes response time optimization mechanisms: Based on real-time system performance monitoring data, the number of candidates for text prediction and the resource allocation for parallel inference are dynamically adjusted to ensure that the overall response latency is controlled within a preset threshold, thus obtaining optimized resource allocation results. Based on the resource allocation optimization results, predictive caching and result reuse strategies are adopted to reduce redundant calculations and further reduce system response latency, resulting in latency optimization measures. Based on the aforementioned delay optimization measures, intent prediction and matching preparation can be achieved before the user fully expresses their opinion, significantly improving the user interaction experience.

13. A large-scale real-time speech-text intent recognition system based on text prediction, characterized in that, include: The streaming speech recognition module is used to acquire the user's real-time speech data, perform real-time speech recognition processing through the streaming speech recognition interface, and obtain partial transcribed text. The text prediction module is used to input the partially transcribed text into a masked language model for text prediction based on the partially transcribed text, and generate several highly reliable complete sentence candidates. A parallel intent recognition module is used to input the complete sentence candidates in parallel into a large language model for intent recognition, obtain the corresponding intent results, and establish a mapping relationship between candidate sentences and intent recognition results. The step of inputting the complete sentence candidates in parallel into the large language model for intent recognition to obtain the corresponding intent results includes: constructing a constrained multi-objective optimization problem based on the complete sentence candidates to simultaneously optimize recognition accuracy, response time, and resource consumption, and identifying constrained Pareto fronts and unconstrained Pareto fronts; intelligently combining a fast classifier, a medium-complexity model, and a large language model using hybrid operator technology based on the constrained Pareto fronts and the unconstrained Pareto fronts to obtain a multi-strategy fusion result; and adjusting computing resources in real time according to system load and user priority based on the multi-strategy fusion result through a dynamic resource allocation mechanism to obtain the corresponding intent results. The similarity matching module is used to obtain the user's complete sentence expression, calculate the similarity between the complete actual sentence and the multiple highly credible complete sentence candidates through a multi-level text similarity algorithm, select the candidate sentence with the highest similarity score, and directly obtain the corresponding final intent recognition result based on the mapping relationship.

Images