An intelligent voice interconnection method for vehicle cockpit control

Abstract

The present invention discloses an intelligent voice interconnection method for vehicle cockpit control, which relates to the field of intelligent voice technology and includes: performing denoising, echo cancellation and voice enhancement processing on digital signals; extracting voice features in digital signals by using the linear predictive cepstral coefficient algorithm; decoding digital signals based on an acoustic model to map the input voice features to corresponding phonemes or words; performing lexical analysis on the text content; performing syntactic analysis and semantic analysis processing on the text content based on fuzzy instruction parsing of the dialogue history; parsing specific instructions and operations to convert the natural language understanding result into a vehicle control instruction; and feeding back relevant information or results to the user through voice synthesis technology. By performing lexical analysis, syntactic analysis and semantic analysis processing on the text content, the intention and requirements of the user are parsed out, the interaction experience between the user and the vehicle is enhanced, and the usability and reliability of the system are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent voice technology, and specifically to an intelligent voice interconnection method for vehicle cockpit control. Background Art

[0002] With the rapid development of intelligent cockpit technology, in-vehicle voice interaction systems have become the core configuration for enhancing driving safety and optimizing the human-machine interaction experience. In the complex acoustic environment of the vehicle cockpit, accurately capturing and understanding voice commands through microphone array technology is the fundamental link in building a complete voice ecosystem. With the rapid development of technology, intelligentization has become one of the core trends in the modern automotive industry. Against this background, intelligent voice interconnection methods, as an innovative interaction method, are gradually changing people's driving experiences.

[0003] In the fast-paced modern life, people are increasingly pursuing efficient and convenient travel methods. Traditional vehicle control methods, such as manual buttons and knobs, although stable and reliable, may seem cumbersome and distract the driver's attention in some cases. Especially during long-distance driving or in complex road conditions, the driver needs to focus more on the road conditions, while intelligent voice interconnection methods can effectively reduce this burden. In addition, with the continuous increase in consumers' demands for vehicle intelligence and personalization, traditional control methods are no longer able to meet market needs. Intelligent voice interconnection methods capture users' voice commands and convert them into vehicle control commands, achieving a more intuitive and natural interaction method, meeting consumers' pursuit of convenience and comfort. Summary of the Invention

[0004] To solve the above technical problems, an intelligent voice interconnection method for vehicle cockpit control is provided, and this technical solution solves the problems raised in the above background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] An intelligent voice interconnection method for vehicle cockpit control, comprising:

[0007] Capturing the user's voice signal through the vehicle cockpit microphone array technology and converting it into a digital signal, and the intelligent voice interconnection system performs denoising, echo cancellation, and voice enhancement processing on the digital signal;

[0008] The preprocessed digital signal is input into the feature extraction module, and the linear prediction cepstral coefficient algorithm is used to extract the voice features in the digital signal;

[0009] Input the extracted speech features into a speech recognition model, decode the digital signal based on an acoustic model, map the input speech features to corresponding phonemes or words, and utilize the word-to-text mapping relationship provided by a dictionary to output the corresponding text content;

[0010] Input the text content into a natural language understanding module, perform lexical analysis on the text content, split the text into words and phrases, and determine their parts of speech;

[0011] Based on the fuzzy instruction parsing of the dialogue history, perform syntactic analysis and semantic analysis on the text content, analyze the grammatical relationships between words and phrases, and parse the user's intentions and requirements;

[0012] The intelligent voice interconnection system parses specific instructions and operations based on the user's intentions and requirements, and converts the natural language understanding result into a vehicle control instruction;

[0013] During and after the execution of the instruction, the intelligent voice interconnection system feeds back relevant information or results to the user through speech synthesis technology.

[0014] Preferably, the step of inputting the extracted speech features into a speech recognition model, decoding the digital signal based on an acoustic model, mapping the input speech features to corresponding phonemes or words, and utilizing the word-to-text mapping relationship provided by a dictionary to output the corresponding text content specifically includes:

[0015] Input the extracted speech feature sequence into a trained acoustic model, align the feature sequence by frame, and retain the time dimension information;

[0016] Convert the frame-level output into a phoneme sequence through a blank symbol and repetition merging mechanism;

[0017] Utilize an encoder-decoder structure to align the acoustic features with the phoneme sequence;

[0018] Obtain a mapping table from phonemes to words and denote it as a dictionary, perform dictionary constraint processing on the dictionary, and only retain the phoneme combinations that can form valid words;

[0019] Calculate the matching probability between the current phoneme and the speech features and output it as an acoustic score;

[0020] Perform weighted fusion on the acoustic score and the language model score;

[0021] Obtain the globally optimal path through the A* algorithm, construct a search space with frames as nodes, phoneme transitions as edges, and the total score as the edge weight;

[0022] Design a heuristic function to estimate the minimum cost from the current node to the end point, and preferentially expand the node with the lowest sum of the minimum cost and the actual cost;

[0023] Determine whether the end of the sequence is reached. If so, end the search for the globally optimal path. If not, do not output anything.

[0024] Restore out-of-vocabulary words in the dictionary through byte pair encoding subword units.

[0025] Obtain a punctuation prediction model, input the phoneme sequence and context, and output the punctuation positions.

[0026] Insert capital letters and line breaks based on the text structure to generate formatted text, completing the conversion from speech to text.

[0027] Preferably, the fuzzy instruction parsing based on the dialogue history performs syntactic analysis and semantic analysis on the text content, analyzes the grammatical relationships between words and phrases, and parses the user's intentions and requirements, specifically including:

[0028] Maintain a dialogue state machine to record the user instructions, system responses, and slot filling results of historical rounds.

[0029] Use a neural network model to identify the pronoun references in the user's speech and construct an entity coreference chain.

[0030] Copy the unknown parameters not mentioned in the user's speech based on historical actions, where the unknown parameters include time parameters and location parameters.

[0031] Update the domain dictionary in real-time based on the user's speech and cache context-related entities.

[0032] Construct a phrase structure tree, identify complex nested structures, and perform pruning based on the dialogue history to limit the candidate structures according to historical actions.

[0033] Map the current parsing result to the dialogue state variables.

[0034] Query the knowledge graph to verify the existence of entities, filter unreasonable requests based on business rules, and ensure that the current request is consistent with the dialogue goal.

[0035] Use multi-task learning to jointly optimize the domain and sub-intentions.

[0036] Inherit the dialogue history parameters, perform requirement conflict detection, and generate corresponding guiding questions after discovering conflicting parameters.

[0037] Use the user's clarification result as new training data and optimize the parameter extraction strategy using reinforcement learning.

[0038] Optimize the slot filling model in combination with the user click log and analyze the user's speech stress features to improve intention recognition.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] Reduces the interference of environmental noise and echo, improves the signal-to-noise ratio and clarity of the voice signal, ensures the accuracy and clarity of the voice signal, provides a reliable basis for subsequent processing, extracts voice features in the digital signal using the linear prediction cepstral coefficient algorithm, which is crucial for subsequent speech recognition and natural language understanding, enhances the interaction experience between the user and the vehicle, improves the usability and reliability of the system, enhances driving safety, and improves the overall satisfaction of the user. Description of the Drawings

[0041] Figure 1 It is a flowchart of the intelligent voice interconnection method for vehicle cockpit control of the present invention;

[0042] Figure 2 It is a flowchart of the method for denoising, echo cancellation and voice enhancement processing of digital signals of the present invention;

[0043] Figure 3 It is a diagram of extracting voice features in the digital signal using the linear prediction cepstral coefficient algorithm of the present invention;

[0044] Figure 4 It is a flowchart of the method for decoding digital signals and mapping the input voice features to corresponding phonemes or words of the present invention;

[0045] Figure 5 It is a flowchart of the method for lexical analysis of text content of the present invention;

[0046] Figure 6 It is a flowchart of the method for syntactic analysis and semantic analysis processing of text content of the present invention;

[0047] Figure 7 It is a flowchart of the method for parsing specific instructions and operations and converting the natural language understanding result into vehicle control instructions of the present invention. Detailed Embodiments

[0048] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments in the following description are only examples, and those skilled in the art can think of other obvious variations.

[0049] Referring to Figure 1 As shown, an intelligent voice interconnection method for vehicle cockpit control includes:

[0050] Capturing the user's voice signal through the vehicle cockpit microphone array technology and converting it into a digital signal, and the intelligent voice interconnection system performs denoising, echo cancellation and voice enhancement processing on the digital signal;

[0051] The preprocessed digital signal is input into the feature extraction module, and the linear predictive cepstral coefficient algorithm is used to extract the speech features in the digital signal;

[0052] The extracted speech features are input into the speech recognition model. Based on the acoustic model, the digital signal is decoded, the input speech features are mapped to the corresponding phonemes or words, and using the mapping relationship from words to text provided by the dictionary, the corresponding text content is output;

[0053] The text content is input into the natural language understanding module, and lexical analysis is performed on the text content. The text is split into words and phrases, and their parts of speech are determined;

[0054] Based on the fuzzy instruction parsing of the dialogue history, syntactic analysis and semantic analysis processing are performed on the text content to analyze the grammatical relationships between words and phrases and parse the user's intentions and requirements;

[0055] Based on the user's intentions and requirements, the intelligent voice interconnection system parses specific instructions and operations, and converts the natural language understanding result into a vehicle control instruction;

[0056] During and after the execution of the instruction, the intelligent voice interconnection system feeds back relevant information or results to the user through speech synthesis technology.

[0057] Refer to Figure 2 As shown, the intelligent voice interconnection system captures the user's speech signal through the automotive cockpit microphone array technology and converts it into a digital signal. The specific processing of denoising, echo cancellation, and speech enhancement for the digital signal includes:

[0058] Adopt a distributed microphone array, and based on the matching of the sound source frequency and the array aperture, design the spacing of the microphones;

[0059] Suppress low-frequency road noise and engine harmonic interference through a hardware high-pass filter, and dynamically adjust the microphone gain based on the signal saturation;

[0060] Realize the time-domain alignment of at least one microphone channel through a field-programmable gate array, transmit the data to the main processor, and manage the real-time data stream using a circular buffer;

[0061] Use the navigation audio as a reference signal to construct an echo path model, and perform echo cancellation processing through an adaptive algorithm;

[0062] Based on the noise spectrum estimation, distinguish the speech segment from the non-speech segment, and use Wiener filtering to attenuate the noise component in the frequency domain and retain the harmonic structure of the speech segment;

[0063] Calculate the time delay difference between each microphone, align the signals, perform weighted superposition, and enhance the target direction signal. Combined with the direction of arrival estimation, locate the speaker position and optimize the beam pointing;

[0064] Detect the fundamental frequency of speech and enhance the energy of formant to improve the clarity of digital signals.

[0065] The system calculates the time delay difference between each microphone, aligns the signals and performs weighted superposition to enhance the target direction signal. Combined with the direction of arrival estimation technology, it locates the speaker's position and optimizes the beam pointing based on the speaker's position to ensure accurate capture of the voice signal. It detects the voice fundamental frequency, that is, the basic frequency component in the voice signal, and enhances the resonance peak energy, that is, the energy of the harmonic component in the voice signal, to improve the clarity of the digital signal.

[0066] Reference Figure 3 As shown, the pre-processed digital signal is input into the feature extraction module, and the speech features in the digital signal are extracted using the linear prediction cepstral coefficient algorithm, which specifically includes:

[0067] The preprocessed speech signal is cut into at least one short time frame, and a Hamming window function is applied to each frame signal to smooth the signal edge and reduce spectral leakage;

[0068] Based on the approximate estimation of the vocal tract response, the linear prediction coefficients of each frame signal are calculated using covariance;

[0069] Convert the linear prediction coefficients into linear prediction cepstral coefficients, map the prediction coefficients to the cepstral domain through recursive operation, and generate a linear prediction cepstral vector;

[0070] The periodic peak intervals are detected in the cepstral domain, which correspond to the glottal vibration period;

[0071] By obtaining the peak points in the cepstrum domain, the distance between adjacent peak points is calculated and output as the fundamental frequency of the digital signal;

[0072] Use the linear prediction coefficient to construct the transfer function, calculate its frequency response curve, obtain the first three local maximum points in the amplitude spectrum, and output the resonance peaks of the vocal tract corresponding to the first three local maximum points in the amplitude spectrum;

[0073] The linear prediction cepstral coefficients, fundamental frequency and formant parameters are combined into a feature vector to form a complete speech feature representation. The feature vector includes spectrum envelope information, excitation source characteristics and vocal tract resonance characteristics.

[0074] The preprocessed speech signal is cut into multiple short time frames. The length of each frame is usually selected to be 10-30 milliseconds. The frame shift (the overlapping part between adjacent frames) is usually 1 / 2 to 1 / 3 of the frame length. A Hamming window function is applied to each frame signal to reduce spectral leakage and smooth the signal edges.

[0075] The Hamming window function is as follows: ,

[0076] wherein, is the Hamming window function, is the sample index within the frame, is the frame length.

[0077] Referring to Figure 4 as shown, the extracted speech features are input into a speech recognition model, the digital signal is decoded based on an acoustic model, the input speech features are mapped to corresponding phonemes or words, and using the mapping relationship from words to text provided by a dictionary, the corresponding text content is output, specifically including:

[0078] Input the extracted speech feature sequence into the trained acoustic model, the feature sequence is frame-aligned, and the time dimension information is retained;

[0079] Convert the frame-level output into a phoneme sequence through a blank symbol and repetition merging mechanism;

[0080] Use an encoder-decoder structure to align the acoustic features with the phoneme sequence;

[0081] Obtain a mapping table from phonemes to words and denote it as a dictionary, perform dictionary constraint processing on the dictionary, and only retain the phoneme combinations that can form valid words;

[0082] Calculate the matching probability between the current phoneme and the speech features and output it as an acoustic score;

[0083] Perform weighted fusion on the acoustic score and the language model score;

[0084] Obtain the globally optimal path through the A* algorithm, use frames as nodes, phoneme transitions as edges, and the total score as the edge weight to construct a search space;

[0085] Design a heuristic function to estimate the minimum cost from the current node to the end point, and preferentially expand the node with the lowest sum of the minimum cost and the actual cost;

[0086] Judge whether the end point of the sequence is reached. If so, end the search for the globally optimal path. If not, do not make an output;

[0087] Restore out-of-vocabulary words in the dictionary through subword units encoded by byte pair encoding;

[0088] Obtain a punctuation prediction model, input the phoneme sequence and context, and output the punctuation positions;

[0089] Insert capital letters and line breaks based on the text structure to generate formatted text, and complete the conversion from speech to text.

[0090] Using the encoder-decoder structure, the acoustic feature sequence is encoded into a hidden state and then decoded into a phoneme sequence. The encoder part is usually used to extract the deep representation of the acoustic features, while the decoder part is used to generate the phoneme sequence, obtain the mapping table (dictionary) of phonemes to words, and perform dictionary constraint processing to retain only the phoneme combinations that can constitute valid words to reduce the search space and improve decoding efficiency. The phonemes corresponding to each frame feature are predicted, and the probability of matching with the real phonemes is calculated and output as an acoustic score to obtain the language model score, which reflects the probability that the phoneme sequence constitutes a legal word and sentence. The acoustic score and the language model score are weighted and fused to obtain the total score.

[0091] Reference Figure 5 As shown, the text content is input into the natural language understanding module, and the text content is lexically analyzed to split the text into words and phrases, and its parts of speech are determined, including:

[0092] Splitting the text into the smallest semantic units, preserving the linguistic structure, wherein the smallest semantic units include words, phrases and symbols;

[0093] For each minimum semantic unit, use the rules in the rule base based on the context of the current word to match;

[0094] By searching the rule in the rule base that best matches the tag, a part-of-speech tag is assigned to each minimum semantic unit, where the part-of-speech tags include nouns, verbs, adjectives, and adverbs.

[0095] In practical applications, the rule base needs to be continuously updated and optimized to adapt to text content in different fields and contexts. At the same time, machine learning or deep learning technology can be combined to improve the accuracy and work efficiency of part-of-speech tagging.

[0096] Reference Figure 6 As shown, fuzzy instruction parsing based on conversation history performs syntactic and semantic analysis on text content, analyzes the grammatical relationship between words and phrases, and parses user intentions and needs. Specifically, it includes:

[0097] Maintain the dialogue state machine, recording historical rounds of user commands, system responses, and slot filling results;

[0098] Use a neural network model to identify pronoun references in user speech and construct entity coreference chains;

[0099] Copy unknown parameters not mentioned in the user's voice based on historical actions, wherein the unknown parameters include time parameters and location parameters;

[0100] Real-time updates to domain dictionaries based on user voice, caching context-related entities;

[0101] Construct a phrase structure tree, identify complex nested structures, and guide pruning based on the dialogue history, restricting candidate structures according to historical actions;

[0102] Map the current parsing result to the dialogue state variables;

[0103] Query the knowledge graph to verify the existence of entities, filter unreasonable requests based on business rules, and ensure that the current request is consistent with the dialogue goal;

[0104] Use multi-task learning to jointly optimize the domain and sub-intentions;

[0105] Inherit dialogue history parameters, perform requirement conflict detection, and generate corresponding guiding questions after discovering conflicting parameters;

[0106] Take the user's clarification result as new training data and use reinforcement learning to optimize the parameter extraction strategy;

[0107] Optimize the slot filling model in combination with the user's click logs, and analyze the user's speech stress features to improve intention recognition.

[0108] According to the pronoun resolution result, construct an entity co-reference chain, connecting pronouns and nouns that refer to the same entity. The entity co-reference chain helps to accurately identify the user's intention and slot filling in subsequent processing. Use a multi-task learning framework to jointly optimize the domain recognition and sub-intention recognition tasks, improving the generalization ability and recognition accuracy of the model by sharing neural network layers or joint loss functions.

[0109] Refer to Figure 7 As shown, the intelligent voice interconnection system analyzes specific instructions and operations based on the user's intention and requirements, and converts the natural language understanding result into vehicle control instructions, specifically including:

[0110] Analyze the semantic content of the text based on the intention recognition model, extract intention-related information from the text, and determine the user's intention;

[0111] Track and manage the interaction history between the user and the system, and maintain the current dialogue state;

[0112] Handle the clarification requests, correction requests, and multi-round interactions proposed by the user to ensure that the system understands the user's complete requirements;

[0113] Based on the user's intention and context information, decide the next operation and response;

[0114] Generate specific vehicle control instructions based on the decision result of the dialogue management;

[0115] Parse the generated instructions into a format that the vehicle control system can understand and encapsulate them into corresponding control signals.

[0116] Parse the generated vehicle control instructions into a format that can be understood by the vehicle control system, which involves converting natural language instructions into specific command codes, protocols, or signals, and encapsulating the parsed instructions into corresponding control signals for sending to the vehicle control system for execution.

[0117] Furthermore, this solution also proposes a computer-readable storage medium, on which a computer-readable program is stored. When the computer-readable program is called, it executes the above-mentioned intelligent voice interconnection method for vehicle cockpit control.

[0118] It can be understood that the storage medium can be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid-state drive (SSD).

[0119] In summary, the advantages of the present invention are as follows: reducing the interference of environmental noise and echo, improving the signal-to-noise ratio and clarity of the voice signal, ensuring the accuracy and clarity of the voice signal, providing a reliable basis for subsequent processing, extracting voice features in the digital signal using the linear predictive cepstral coefficient algorithm, which is crucial for subsequent speech recognition and natural language understanding, enhancing the interaction experience between the user and the vehicle, improving the usability and reliability of the system, improving driving safety, and enhancing the overall satisfaction of the user.

[0120] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection required by the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent voice interconnection method for vehicle cockpit control, characterized in that, Including: Capturing the user's voice signal through the automotive cockpit microphone array technology and converting it into a digital signal, and the intelligent voice interconnection system performs denoising, echo cancellation, and voice enhancement processing on the digital signal; The preprocessed digital signal is input into the feature extraction module, and the linear predictive cepstral coefficient algorithm is used to extract the voice features in the digital signal; The extracted voice features are input into the speech recognition model, decoding the digital signal based on the acoustic model, mapping the input voice features to the corresponding phonemes or words, and using the mapping relationship from words to text provided by the dictionary to output the corresponding text content; The text content is input into the natural language understanding module, and lexical analysis is performed on the text content, splitting the text into words and phrases and determining their part of speech; Based on the fuzzy instruction parsing of the dialogue history, syntactic analysis and semantic analysis processing are performed on the text content, analyzing the grammatical relationships between words and phrases and parsing the user's intentions and requirements; The intelligent voice interconnection system parses specific instructions and operations based on the user's intentions and requirements, and converts the natural language understanding result into a vehicle control instruction; During and after the execution of the instruction, the intelligent voice interconnection system feeds back relevant information or results to the user through speech synthesis technology; The capturing of the user's voice signal through the automotive cockpit microphone array technology and converting it into a digital signal, and the intelligent voice interconnection system performing denoising, echo cancellation, and voice enhancement processing on the digital signal specifically includes: Adopting a distributed microphone array and designing the microphone spacing based on the matching of the sound source frequency and the array aperture; Suppressing low-frequency road noise and engine harmonic interference through a hardware high-pass filter, and dynamically adjusting the microphone gain based on the signal saturation; Realizing the time-domain alignment of at least one microphone channel through a field programmable gate array, transmitting the data to the main processor, and using a circular buffer to manage the real-time data stream; Using the navigation audio as a reference signal to construct an echo path model and performing echo cancellation processing through an adaptive algorithm; Based on the noise spectrum estimation, distinguishing the speech segment from the non-speech segment, and using Wiener filtering to attenuate the noise component in the frequency domain and retain the harmonic structure of the speech segment; Calculating the time delay difference between each microphone, performing weighted superposition after aligning the signals, enhancing the target direction signal, combining the direction of arrival estimation to locate the speaker's position, and optimizing the beam pointing; Detecting the fundamental frequency of the speech and enhancing the formant energy to improve the clarity of the digital signal; The preprocessed digital signal is input into the feature extraction module, and the linear predictive cepstral coefficient algorithm is used to extract the voice features in the digital signal specifically includes: Cutting the preprocessed voice signal into at least one short-time frame, applying a Hamming window function to each frame of the signal to smooth the signal edge and reduce the spectral leakage; Based on the approximate estimation of the vocal tract response, using covariance to calculate the linear prediction coefficients of each frame of the signal; Converting the linear prediction coefficients into linear predictive cepstral coefficients, mapping the prediction coefficients to the cepstral domain through recursive operations to generate a linear predictive cepstral vector; Detecting the periodic peak interval in the cepstral domain, and this interval corresponds to the glottal vibration period; By obtaining the peak points in the cepstral domain, calculating the distances between adjacent peak points, and outputting the fundamental frequency of the digital signal; Constructing a transfer function using the linear prediction coefficients, calculating its frequency response curve, obtaining the first three local maximum points in the amplitude spectrum, and outputting the formants of the corresponding channels of the first three local maximum points in the amplitude spectrum; Combining the linear prediction cepstral coefficients, fundamental frequency, and formant parameters into a feature vector to form a complete speech feature representation, where the feature vector includes spectral envelope information, excitation source characteristics, and vocal tract resonance characteristics; Inputting the extracted speech features into a speech recognition model, decoding the digital signal based on the acoustic model, mapping the input speech features to the corresponding phonemes or words, and using the mapping relationship from words to text provided by the dictionary to output the corresponding text content, which specifically includes: Inputting the extracted speech feature sequence into the trained acoustic model, aligning the feature sequence frame by frame, and retaining the time dimension information; Converting the frame-level output into a phoneme sequence through a blank and repetition merging mechanism; Using an encoder-decoder structure to align the acoustic features with the phoneme sequence; Obtaining the mapping table from phonemes to words and denoting it as a dictionary, performing dictionary constraint processing on the dictionary, and only retaining the phoneme combinations that can form valid words; Calculating the matching probability between the current phoneme and the speech features and outputting it as an acoustic score; Performing weighted fusion of the acoustic score and the language model score; Obtaining the globally optimal path through the A* algorithm, using frames as nodes, phoneme transitions as edges, and the total score as the edge weights to construct a search space; Designing a heuristic function to estimate the minimum cost from the current node to the end point, and preferentially expanding the node with the lowest sum of the minimum cost and the actual cost; Judging whether the end of the sequence is reached. If so, end the search for the globally optimal path. If not, do not make an output; Restoring the out-of-vocabulary words in the dictionary through byte pair encoding subword units; Obtaining a punctuation prediction model, inputting the phoneme sequence and context, and outputting the punctuation positions; Inserting capital letters and line breaks based on the text structure to generate formatted text and complete the speech-to-text conversion.

2. The intelligent voice interconnection method for vehicle cockpit control according to claim 1, characterized in that Inputting the text content into the natural language understanding module, performing lexical analysis on the text content, splitting the text into words and phrases, and determining their part of speech, which specifically includes: Splitting the text into the smallest semantic units, retaining the linguistic structure, where the smallest semantic units include words, phrases, and symbols; For each smallest semantic unit, using the rules in the rule base based on the context of the current word for matching; Assigning a part-of-speech label to each smallest semantic unit by finding the rule in the rule base that best matches the token, where the part-of-speech labels include nouns, verbs, adjectives, and adverbs.

3. An intelligent voice interconnection method for vehicle cockpit control according to claim 2, characterized in that, Performing syntactic analysis and semantic analysis processing on the text content based on the fuzzy instruction parsing of the dialogue history, analyzing the grammatical relationships between words and phrases, and parsing the user's intentions and requirements, which specifically includes: Maintaining a dialogue state machine to record the user instructions, system responses, and slot filling results of previous rounds; Using a neural network model to identify the pronoun references in the user's speech and constructing an entity co-reference chain; Copy unknown parameters not mentioned in the user's speech based on historical actions, where the unknown parameters include time parameters and location parameters; Update the domain dictionary in real time based on the user's speech and cache context-related entities; Construct a phrase structure tree, identify complex nested structures, and perform pruning guided by the conversation history, restricting candidate structures based on historical actions; Map the current parsing result to the dialogue state variables; Query the knowledge graph to verify the existence of entities, filter unreasonable requests based on business rules, and ensure that the current request is consistent with the dialogue goal; Use multi-task learning to jointly optimize the domain and sub-intents; Inherit the dialogue history parameters, perform requirement conflict detection, and generate corresponding guiding questions after discovering conflicting parameters; Use the user's clarification result as new training data and optimize the parameter extraction strategy using reinforcement learning; Optimize the slot filling model in combination with the user's click log and analyze the user's speech stress features to improve intent recognition.

4. An intelligent voice interconnection method for vehicle cockpit control according to claim 3, characterized in that, The intelligent voice interconnection system analyzes specific instructions and operations based on the user's intent and requirements, and converts the natural language understanding result into vehicle control instructions, specifically including: Analyze the semantic content of the text based on the intent recognition model, extract intent-related information from the text, and determine the user's intent; Track and manage the interaction history between the user and the system, and maintain the current dialogue state; Handle the user's clarification requests, correction requests, and multi-round interactions to ensure that the system understands the user's complete requirements; Decide on the next actions and responses based on the user's intent and context information; Generate specific vehicle control instructions based on the decision result of the dialogue management; Parse the generated instructions into a format that the vehicle control system can understand and encapsulate them into corresponding control signals.

Images