Recognition and detection system and method for AI voice dialogue scene application

Abstract

The invention belongs to the technical field, and particularly relates to a recognition and detection system and method for AI voice dialogue scene application, and the system comprises an ASR automatic voice recognition unit which employs an advanced noise elimination algorithm to improve the definition of a voice signal; the NLP natural language processing unit adopts an intelligent recommendation analysis module; a machine learning unit; a TTS text-to-speech unit; a deep learning unit; an intelligent dialogue platform unit; a multi-modal interaction unit; a personalized customization unit; a dictionary noise elimination unit for eliminating noise by analyzing dictionary composition is arranged in the noise elimination algorithm; the intelligent recommendation analysis module comprises an intention recognition unit; and a feedback circulation unit. According to the method, the function of eliminating the noise by analyzing the composition of the dictionary can be realized, the noise and useful information are effectively distinguished, the application effect of the improved LMS algorithm and the improved RLS algorithm in noise elimination is remarkably improved, and in addition, a recommendation first choice for solving a problem can be selected according to secondary inquiry of a plurality of users.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to an identification and detection system and method for AI voice dialogue scenario application. Background Art

[0002] AI scene dialogue refers to the use of artificial intelligence technology to simulate human dialogue behavior in specific application scenarios to achieve natural language communication with users. This dialogue system can understand user queries and commands and provide corresponding answers or perform tasks. Its core technology: The implementation of AI scene dialogue relies on the integration of multiple technologies, including automatic speech recognition, natural language processing, machine learning, etc. These technologies work together to enable machines to capture contextual information and generate intelligent responses.

[0003] Problems with existing technologies: Regarding the scene speech with AI, intelligent AI still has large errors and defects in the recognition and detection of user voice, which are specifically manifested in the following points: When a user is in the same room with a person who has a similar voice and tone to the user, and the user is talking to the AI ​​and the other person is also speaking, the AI ​​is very likely to treat the other person's voice as input speech, and then pile up different words, causing the AI ​​to distort the input speech content, reducing the voice interaction experience; In the existing AI dialogue model, when users consult about a problem, most of them will not directly express satisfaction or acceptance of a solution. Moreover, since AI dialogue is different from operating computer equipment, most AI scene dialogues do not have a corresponding feedback mechanism, which makes it impossible for AI scene big data to know the adoption rate of different solutions, and thus unable to provide recommended solutions for the same problem. Summary of the invention

[0004] The purpose of the present invention is to provide an AI voice dialogue scene application recognition detection system and method, which can eliminate noise by analyzing the composition of words, effectively distinguish noise from useful information, and significantly improve the application effect of the improved LMS algorithm and RLS algorithm in noise elimination. In addition, the recommended first choice for solving the problem can be selected based on secondary inquiries of multiple users.

[0005] The technical solution adopted by the present invention is as follows: An AI voice dialogue scene application recognition and detection system, comprising: ASR automatic speech recognition unit, which uses a high signal-to-noise ratio (SNR) microphone, such as a MEMS microphone; uses advanced noise cancellation algorithms to improve the clarity of speech signals; and uses deep learning technology to train models to recognize and understand a variety of different accents and dialects; NLP natural language processing unit, which adopts a large model technology application module, and uses the latest large language models such as GPT-3 or BERT to enhance the system's ability to understand complex language structures and semantics; adopts an integrated sentiment analysis module to enable the system to recognize the user's emotional changes and make corresponding emotional responses; adopts an intelligent recommendation analysis module to determine the user's preferred method through the user's secondary inquiry of one of the methods, and combines big data to recommend this method as the first choice for solving the problem; Knowledge graph unit; A machine learning unit that enables the AI ​​voice dialogue system to continuously optimize its performance through large amounts of data training, thereby improving the accuracy and naturalness of the dialogue; TTS text-to-speech unit, which uses more efficient endpoint detection (VAD) technology to accurately predict the punctuation points in speech, especially in continuous speech streams; builds a large model that can understand complex contexts and improves the system's understanding of context dependencies during real-time transcription, especially in long sequence conversations; A deep learning unit, wherein the deep learning algorithm adopts a recurrent neural network (RNN) and a long short-term memory network (LSTM); An intelligent dialogue platform unit, which uses advanced echo cancellation algorithms and multi-channel signal processing technology to reduce interference caused by sound reflections; and uses separation accuracy optimization technology to accurately distinguish the voiceprints of different speakers in a scene where multiple people are speaking at the same time; A multimodal interaction unit, which is used to combine multiple sensory information such as vision and touch; A personalized customization unit, which analyzes the user's historical data and behavior patterns through a machine learning algorithm and automatically adjusts system settings to meet the user's personalized needs; An emotion recognition and expression unit, which is used to enable the AI ​​voice system to perceive and express the user's emotions; Privacy protection unit.

[0006] The noise elimination algorithm is provided with a word elimination unit for eliminating noise by analyzing word composition, and the word elimination unit adopts word analysis technology, word feature extraction, adaptive filter optimization and multi-modal noise elimination; The word analysis technology uses the NLP natural language processing unit to analyze the vocabulary and syntax in the voice signal and identify the noise components therein; uses a deep learning unit, such as BERT or GPT-3, to perform semantic understanding of the voice signal, extract key information from it and eliminate noise; and uses an integrated sentiment analysis module to detect the user's emotional changes, thereby identifying the noise caused by emotional fluctuations; The word feature extraction uses word frequency analysis to count the frequency of occurrence of each word in the speech signal, and filters out words with low occurrence frequency by setting a threshold; uses syntactic structure analysis to analyze the structure of the sentence to identify ungrammatical or illogical noise parts; uses semantic relevance analysis to calculate the semantic relevance between words, and determines whether the words with low relevance are noise; The adaptive filter optimization adopts the LMS algorithm to improve the least mean square algorithm, and adds a word analysis module to dynamically filter out noise words by adjusting the filter coefficient in real time; the RLS algorithm is applied, and the recursive least squares algorithm has a faster convergence speed; the Kalman filter optimization is adopted to suppress the noise of the dynamic system through the state space model and the observation model; The multimodal noise elimination adopts visual assistance, combined with visual sensor data, to identify the noise source in the environment through image processing technology; adopts tactile feedback, using tactile sensors to detect the user's body movements and behaviors, so as to identify possible noise sources; adopts multi-channel signal processing, combined with data from different sensors, to perform joint noise elimination.

[0007] The core of the LMS algorithm improvement is to establish a nonlinear functional relationship between the step size factor μ and the error signal e(n), and to improve it through the Sigmoid function. The specific form is: ; The Sigmoid function is an S-shaped curve, which is mathematically defined as , where EXP represents the natural exponential function. The improvement is as follows: In order to overcome the shortcomings of the Sigmoid function in the step size adjustment process in the steady-state stage, it is multiplied by an even function, such as , so that the step factor changes slowly when the error is close to zero; Based on the above improvements, the nonlinear functional relationship between the step size factor μ and the error signal e(n) can be expressed as: , where β and α are adjustable parameters. By adjusting the parameters β and α, the performance of the algorithm can be optimized according to the actual application requirements.

[0008] The core equations of the RLS algorithm include weight update formula and error calculation; Weight update formula: In the RLS algorithm, the update formula of the weight vector w(n) is as follows: , where g(n) is the gain vector and e(n) is the error signal; The calculation formula of the gain vector g(n) is: , where P(n) is the correlation matrix, x(n) is the input signal, and λ is the forgetting factor; The update formula of the correlation matrix P(n) is: ; Error calculation: The error signal e(n) is defined as the difference between the desired signal d(n) and the filter output y(n): ,in, , the expected signal d(n) is the known reference signal, and the output signal y(n) is the filter output at the current moment; Parameter adjustment: The forgetting factor λ is used to control the convergence speed and stability of the algorithm; Initial conditions: The correlation matrix P(0) is usually initialized to the identity matrix, that is , where δ is a small constant.

[0009] The improvement measure of the RLS algorithm is to introduce an additional constant multiplier for optimizing the filter performance; By adjusting the value of the constant multiplier, optimization is performed for different signal-to-noise ratios (SNRs), and when the constant multiplier takes a value of 0.65, the best SNR value can be obtained, reaching 19.38 dB.

[0010] The intelligent recommendation analysis module includes: Intent recognition unit: By recognizing the user's initial question and subsequent secondary inquiry, it determines whether the user intends to seek more information about a specific method or is interested in other methods; Context management unit: used to maintain the consistency of context in multiple rounds of dialogue and accurately provide information related to the previous discussion when the user asks a second question; Sentiment analysis unit: used to help the system determine the user's attitude towards different methods; Feedback loop unit: through the establishment of feedback mechanisms to enable users to directly express their satisfaction with a method; Machine Learning Model Unit: Using machine learning models, the system can train algorithms based on historical data to predict methods that users may be interested in; Personalized recommendation unit: Based on the user's historical behavior and preferences, the system can provide personalized recommendations.

[0011] The intelligent recommendation analysis module also includes: Data collection and preprocessing unit: collects a large amount of user behavior data, cleans the data, and extracts useful features from the user behavior data; User portrait construction unit: Analyze the user's basic information and historical behavior data, identify the user's interest preferences in different methods, and then give each user a personalized label; Method evaluation and ranking unit: Evaluate the effectiveness of each method, rank the methods using ranking algorithms (such as collaborative filtering, content-based recommendation, etc.), and regularly update the ranking of the methods; Recommendation strategy formulation unit: recommend methods related to users based on their interests, preferences and needs; make recommendations based on similarities between users and methods; combine the advantages of multiple recommendation strategies to formulate more comprehensive and accurate recommendation plans; Real-time feedback and optimization unit: Encourage users to provide feedback on the recommendation method, continuously iterate and update the recommendation model based on user feedback and new user behavior data, and then integrate the real-time feedback and optimization process into daily operations to ensure the continuous improvement and optimization of the recommendation system; Interface display and interaction unit: clearly display the recommendation results and provide detailed explanations to guide user interaction; Cross-platform integration and expansion unit: supports multi-platform operation and provides API interfaces for other developers or third-party applications to call.

[0012] The knowledge graph unit provides rich background knowledge and information for the AI ​​voice dialogue system, supporting the system to perform logical reasoning and knowledge retrieval when answering user questions.

[0013] The privacy protection unit is used to protect user privacy.

[0014] A recognition and detection method for AI voice dialogue scenario application, the specific steps are as follows: Step 1, Voice Collection: The system collects the user's voice signal through microphones and other devices, which is the basis of the entire recognition process; Step 2, preprocessing: noise reduction, filtering, amplification and other processing are performed on the collected speech signal to improve the quality of the speech signal; Step 3, feature extraction: convert the preprocessed speech signal into features that can represent the essence of speech. This process involves complex signal processing and feature engineering techniques. Step 4, model matching: Match the extracted features with the pre-trained model to identify the specific content of the speech. This model is usually obtained through machine learning technology, such as deep learning training; Step 5, data upload: The method or approach determined to be adopted by the user will be uploaded to the big database as recommended content, and used as a recommended solution for other users facing the same problem.

[0015] The technical effects achieved by the present invention are: The present invention, by introducing a rhetoric silencing unit into the noise elimination algorithm, can realize the function of eliminating noise by analyzing the composition of rhetoric. The rhetoric analysis technology can not only effectively distinguish noise from useful information, but also combine multiple methods such as deep learning models, sentiment analysis, adaptive filter optimization and multimodal noise elimination to comprehensively improve the intelligence level of the system. In the future, with the continuous advancement of technology, noise elimination algorithms will play an important role in more application scenarios and provide users with a better voice interaction experience.

[0016] The present invention significantly improves the application effect of the LMS algorithm in noise elimination, which is specifically reflected in the following points: First, convergence speed: in the initial stage and the time-varying stage of the unknown system, the algorithm can automatically increase the step size and accelerate the convergence speed. This feature enables the algorithm to quickly adapt to environmental changes and reduce training time; Second, steady-state error: when the error is close to zero, the step size of the algorithm changes slowly, which effectively reduces the steady-state error. This feature ensures the stability and accuracy of the algorithm in the steady process; Third, computational complexity: although the improved LMS algorithm increases the amount of calculation in some cases, it still maintains a low computational complexity as a whole and is suitable for real-time processing; In general, by establishing a nonlinear functional relationship between the step size factor μ and the error signal e(n), the application effect of the improved LMS algorithm in noise elimination is significantly improved, especially in a complex noise environment, and useful information can be more effectively separated.

[0017] In the present invention, the RLS algorithm can adjust the filter coefficients in real time to minimize the output error. This adaptive filtering method performs well in applications such as noise elimination and echo elimination, and has significant advantages in particular in rapidly changing environments.

[0018] The present invention, through the operation of the intelligent recommendation analysis module, the AI ​​system can not only provide users with multiple solutions to problems, but also effectively determine the user's preferred method based on secondary inquiries of multiple users, and use this method as the recommended first choice for solving the problem. This system not only improves user experience and satisfaction, but also brings higher conversion rates and commercial value to enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a schematic diagram of the composition of a detection system provided by an embodiment of the present invention; Figure 2It is a schematic diagram of the composition of the NLP natural language processing unit provided in an embodiment of the present invention; Figure 3 It is a flow chart of the identification method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0021] like Figure 1-Figure 2 As shown, a recognition and detection system for AI voice dialogue scenario application includes: ASR automatic speech recognition unit, which uses high signal-to-noise ratio (SNR) microphones, such as MEMS micro-electromechanical system microphones, to reduce interference from ambient noise; uses advanced noise cancellation algorithms to improve the clarity of voice signals, ensuring that the user's voice can be accurately captured even in noisy environments; uses deep learning technology to train models to recognize and understand a variety of different accents and dialects, improving the universality and accuracy of the system; NLP natural language processing unit: The NLP natural language processing unit adopts a large model technology application module. With the help of the latest large language models such as GPT-3 or BERT, the system's ability to understand complex language structures and semantics is enhanced. These models can reason and generate coherent dialogue content in a wider context; the integrated sentiment analysis module enables the system to identify the user's emotional changes and make corresponding emotional responses, thereby providing a more humane interactive experience; the intelligent recommendation analysis module is used to determine the user's preferred method through the user's secondary inquiry of one of the methods, and combined with big data, the method is recommended as the first choice for solving the problem, which improves user experience and satisfaction, and also brings higher conversion rates and commercial value to the enterprise; Knowledge graph unit: The knowledge graph unit provides rich background knowledge and information for the AI ​​voice dialogue system, supporting the system to perform logical reasoning and knowledge retrieval when answering user questions; The machine learning unit enables the AI ​​voice dialogue system to continuously optimize its performance through large amounts of data training, improving the accuracy and naturalness of the dialogue; TTS text-to-speech unit: The TTS text-to-speech unit uses more efficient endpoint detection (VAD) technology to accurately predict the punctuation points in speech, especially in continuous speech streams, which helps improve the clarity and accuracy of real-time transcription; build a large model that can understand complex contexts to improve the system's understanding of context dependencies during real-time transcription, especially in long sequence conversations; Deep learning unit, deep learning algorithms use recurrent neural networks (RNN) and long short-term memory networks (LSTM), which play a key role in speech recognition and natural language processing, improving the accuracy of recognition and understanding; Intelligent dialogue platform unit: The intelligent dialogue platform unit adopts advanced echo cancellation algorithm and multi-channel signal processing technology to reduce the interference caused by sound reflection. In far-field interaction, effective echo cancellation technology is the key. It adopts separation accuracy optimization technology to accurately distinguish the voiceprints of different speakers in the scene where multiple people speak at the same time, realize effective sound source separation, and improve recognition accuracy. Multimodal interaction unit, which is used to combine multiple sensory information such as vision and touch; Personalization and customization unit: The personalization and customization unit analyzes the user's historical data and behavior patterns through machine learning algorithms, and automatically adjusts system settings to meet the user's personalized needs; Emotion recognition and expression unit, which is used to enable the AI ​​voice system to perceive and express the user's emotions; Privacy protection unit,The privacy protection unit is used to protect user privacy.

[0022] Embodiment 1: The noise elimination algorithm is provided with a word elimination unit that eliminates noise by analyzing word composition. The word elimination unit adopts word analysis technology, word feature extraction, adaptive filter optimization and multi-modal noise elimination; The word analysis technology uses the NLP natural language processing unit to analyze the vocabulary and syntax in the voice signal and identify the noise components. For example, through part-of-speech tagging and dependency syntax analysis, the content words and function words in the sentence are distinguished, thereby filtering out some insignificant words; using deep learning units such as BERT or GPT-3, the voice signal is semantically understood, key information is extracted and noise is excluded. These models can capture the semantic relationship in the context and effectively distinguish noise from useful information; using the integrated sentiment analysis module, the user's emotional changes are detected, thereby identifying the noise caused by emotional fluctuations. This function can help the system better understand the user's true intentions and improve the accuracy of the conversation; Lexical feature extraction: lexical frequency analysis is used to count the frequency of occurrence of each word in the speech signal. By setting a threshold, the words with low frequency of occurrence are filtered out. These low-frequency words often contain more noise components. Syntactic structure analysis is used to analyze the structure of the sentence to identify ungrammatical or illogical noise parts. For example, through dependency syntax analysis, isolated components in the sentence or phrases that do not conform to grammatical rules can be found. Semantic relevance analysis is used to calculate the semantic relevance between words. Words with low relevance are judged as noise. This method is based on the assumption that words with high semantic relevance are more likely to be key information in the conversation. Adaptive filter optimization, using LMS algorithm improvement, based on the least mean square algorithm, adding a word analysis module, by real-time adjustment of the filter coefficient, dynamically filter out noise words, variable step size LMS algorithm can automatically adjust the step size according to the change of noise, improve the noise suppression effect; using RLS algorithm application, recursive least squares algorithm has a faster convergence speed, combined with word analysis technology, can achieve effective noise suppression in a shorter time; using Kalman filter optimization, through the state space model and observation model, the dynamic system is subjected to noise suppression, combined with word analysis, it can more accurately estimate the signal state and improve the noise elimination effect; Multimodal noise elimination uses visual assistance and combines visual sensor data to identify noise sources in the environment through image processing technology, such as noise emitted by fans, air conditioners and other equipment. With the assistance of visual information, the system can more accurately locate and eliminate noise; tactile feedback is used, using tactile sensors to detect the user's body movements and behaviors, so as to identify possible noise sources. For example, by detecting the user's gestures, it can be determined whether it is an unintentional touch or friction sound; multi-channel signal processing is used to combine data from different sensors for joint noise elimination. Multi-channel signal processing can improve the accuracy and robustness of noise elimination.

[0023] The working principle of the present invention is: In summary, by introducing a rhetoric silencing unit into the noise elimination algorithm, the function of eliminating noise by analyzing the composition of rhetoric can be realized. The rhetoric analysis technology can not only effectively distinguish noise from useful information, but also combine deep learning models, sentiment analysis, adaptive filter optimization and multimodal noise elimination and other methods to comprehensively improve the intelligence level of the system. In the future, with the continuous advancement of technology, noise elimination algorithms will play an important role in more application scenarios and provide users with a better voice interaction experience.

[0024] Embodiment 2: The core of the LMS algorithm improvement is to establish a nonlinear functional relationship between the step size factor μ and the error signal e(n), and improve it through the Sigmoid function. The specific form is: ; Among them, the Sigmoid function is an S-shaped curve, which is mathematically defined as , where EXP represents the natural exponential function. The improvement is as follows: In order to overcome the shortcomings of the Sigmoid function in the step size adjustment process in the steady-state stage, it is multiplied by an even function, such as , so that the step factor changes slowly when the error is close to zero; Based on the above improvements, the nonlinear functional relationship between the step size factor μ and the error signal e(n) can be expressed as: , where β and α are adjustable parameters. By adjusting the parameters β and α, the performance of the algorithm can be optimized according to the actual application requirements. For example, increasing β can increase the convergence speed, while increasing α can make the step size factor more sensitive to errors.

[0025] The working principle of the present invention is as follows: Through the above-mentioned improvement measures, the application effect of the LMS algorithm in noise elimination has been significantly improved, which is specifically reflected in the following points: 1. Convergence speed: In the initial stage and the time-varying stage of the unknown system, the algorithm can automatically increase the step size and speed up the convergence speed. This feature enables the algorithm to quickly adapt to environmental changes and reduce training time; 2. Steady-state error: When the error is close to zero, the step size of the algorithm changes slowly, which effectively reduces the steady-state error. This feature ensures the stability and accuracy of the algorithm in the steady process; 3. Computational complexity: Although the improved LMS algorithm increases the amount of calculation in some cases, it still maintains a low computational complexity overall and is suitable for real-time processing; In general, by establishing a nonlinear functional relationship between the step size factor μ and the error signal e(n), the application effect of the improved LMS algorithm in noise elimination has been significantly improved, especially in complex noise environments, and useful information can be separated more effectively.

[0026] Embodiment three: The core equations of the RLS algorithm include the weight update formula and error calculation; Weight update formula: In the RLS algorithm, the update formula of the weight vector w(n) is as follows: , where g(n) is the gain vector and e(n) is the error signal; The calculation formula of the gain vector g(n) is: , where P(n) is the correlation matrix, x(n) is the input signal, and λ is the forgetting factor; The update formula of the correlation matrix P(n) is: ; Error calculation: The error signal e(n) is defined as the difference between the desired signal d(n) and the filter output y(n): ,in, , the expected signal d(n) is the known reference signal, and the output signal y(n) is the filter output at the current moment; Parameter adjustment: The forgetting factor λ is used to control the convergence speed and stability of the algorithm. A smaller λ value can speed up the convergence speed, but may lead to an increase in the steady-state error; a larger λ value can improve the steady-state performance, but the convergence speed is slower. Initial conditions: The correlation matrix P(0) is usually initialized to the identity matrix, that is , where δ is a small constant.

[0027] The improvement measure of RLS algorithm is to introduce an additional constant multiplier for optimizing the filter performance; By adjusting the value of the constant multiplier, it is used to optimize for different signal-to-noise ratios (SNRs). When the constant multiplier value is 0.65, the best SNR value of 19.38 dB can be obtained. This optimization significantly improves the performance stability of the algorithm.

[0028] The working principle of the present invention is: through the above-mentioned improvement measures, the RLS algorithm can adjust the filter coefficients in real time to minimize the output error. This adaptive filtering method performs well in applications such as noise elimination and echo elimination, especially in rapidly changing environments.

[0029] Embodiment 4: The intelligent recommendation analysis module includes: Intent recognition unit: By identifying the user's initial question and subsequent secondary inquiry, it determines whether the user's intention is to seek more information about a specific method or is interested in other methods. This step is crucial because it determines how the system responds to the user's further needs; Context management unit: It is used to maintain the consistency of context in multiple rounds of dialogue and accurately provide information related to the previous discussion when the user asks a second question. This requires the NLP system to be able to track the history of the dialogue. The system has a strong memory capacity and can store and process a large amount of dialogue data. Sentiment analysis unit: used to help the system determine the user's attitude towards different methods. By analyzing the user's tone, word choice, etc., the system can perceive the user's preference or aversion to a certain method. This emotional nuance is very important for understanding the user's real needs and helps the system provide more personalized services. Feedback loop unit: A feedback mechanism is established to enable users to directly express their satisfaction with a method, not only through explicit affirmation or denial, but also by collecting implicit feedback from users, such as whether they are willing to continue to explore the details of a method. Such a feedback loop helps to continuously optimize the system's recommendation algorithm to make it more in line with user expectations; Machine Learning Model Unit: With machine learning models, the system can train algorithms based on historical data to predict methods that users may be interested in. These models become more and more accurate over time because they are able to learn patterns and trends from a large number of user interactions. This self-optimization ability is key to improving system performance; Personalized recommendation unit: Based on the user's historical behavior and preferences, the system can provide personalized suggestions. This customized service can improve user engagement and satisfaction because it shows the system's understanding of user needs. Personalized recommendation is one of the important means to improve user experience.

[0030] The intelligent recommendation analysis module also includes: Data collection and preprocessing unit: collects a large amount of user behavior data, cleans the data, and extracts useful features from the user behavior data; User portrait construction unit: Analyze the user's basic information and historical behavior data, identify the user's interest preferences in different methods, and then give each user a personalized label; Method evaluation and ranking unit: Evaluate the effectiveness of each method, rank the methods using ranking algorithms (such as collaborative filtering, content-based recommendation, etc.), and regularly update the ranking of the methods; Recommendation strategy formulation unit: recommend methods related to users based on their interests, preferences and needs; make recommendations based on similarities between users and methods; combine the advantages of multiple recommendation strategies to formulate more comprehensive and accurate recommendation plans; Real-time feedback and optimization unit: Encourage users to provide feedback on the recommendation method, continuously iterate and update the recommendation model based on user feedback and new user behavior data, and then integrate the real-time feedback and optimization process into daily operations to ensure the continuous improvement and optimization of the recommendation system; Interface display and interaction unit: clearly display the recommendation results and provide detailed explanations to guide user interaction; Cross-platform integration and expansion unit: supports multi-platform operation and provides API interfaces for other developers or third-party applications to call.

[0031] The working principle of the present invention is: through the operation of the intelligent recommendation analysis module, the AI ​​system can not only provide users with multiple solutions to problems, but also effectively determine the user's preferred method based on secondary inquiries from multiple users, and use this method as the recommended first choice for solving the problem. This system not only improves user experience and satisfaction, but also brings higher conversion rates and commercial value to enterprises.

[0032] like Figure 3 As shown, a recognition and detection method for AI voice dialogue scene application, the specific steps are as follows: Step 1, Voice Collection: The system collects the user's voice signal through microphones and other devices, which is the basis of the entire recognition process; Step 2, preprocessing: noise reduction, filtering, amplification and other processing are performed on the collected speech signal to improve the quality of the speech signal; Step 3, feature extraction: convert the preprocessed speech signal into features that can represent the essence of speech. This process involves complex signal processing and feature engineering techniques. Step 4, model matching: Match the extracted features with the pre-trained model to identify the specific content of the speech. This model is usually obtained through machine learning technology, such as deep learning training; Step 5, data upload: The method or approach determined to be adopted by the user will be uploaded to the big database as recommended content, and used as a recommended solution for other users facing the same problem.

[0033] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.

Claims

1. A recognition and detection system for AI voice dialogue scenarios, It is characterized in that include: An ASR automatic speech recognition unit, wherein the ASR automatic speech recognition unit uses a high signal-to-noise ratio (SNR) microphone, such as a MEMS micro-electromechanical system microphone; Advanced noise cancellation algorithms are used to improve the clarity of speech signals. Deep learning technology is used to train models to recognize and understand a variety of different accents and dialects. NLP natural language processing unit, which adopts a large model technology application module, and uses the latest large language models such as GPT-3 or BERT to enhance the system's ability to understand complex language structures and semantics; adopts an integrated sentiment analysis module to enable the system to recognize the user's emotional changes and make corresponding emotional responses; adopts an intelligent recommendation analysis module to determine the user's preferred method through the user's secondary inquiry of one of the methods, and combines big data to recommend this method as the first choice for solving the problem; Knowledge graph unit; A machine learning unit that enables the AI ​​voice dialogue system to continuously optimize its performance through large amounts of data training, thereby improving the accuracy and naturalness of the dialogue; TTS text-to-speech unit, which uses more efficient endpoint detection (VAD) technology to accurately predict the punctuation points in speech, especially in continuous speech streams; builds a large model that can understand complex contexts and improves the system's understanding of context dependencies during real-time transcription, especially in long sequence conversations; A deep learning unit, wherein the deep learning algorithm adopts a recurrent neural network (RNN) and a long short-term memory network (LSTM); An intelligent dialogue platform unit, which uses advanced echo cancellation algorithms and multi-channel signal processing technology to reduce interference caused by sound reflections; Adopt separation accuracy optimization technology to accurately distinguish the voiceprints of different speakers in scenarios where multiple people are speaking at the same time; A multimodal interaction unit, which is used to combine multiple sensory information such as vision and touch; A personalized customization unit, which analyzes the user's historical data and behavior patterns through a machine learning algorithm and automatically adjusts system settings to meet the user's personalized needs; An emotion recognition and expression unit, which is used to enable the AI ​​voice system to perceive and express the user's emotions; Privacy protection unit.

2. According to claim 1, a recognition and detection system for AI voice dialogue scene application, Features: The noise elimination algorithm is provided with a word elimination unit for eliminating noise by analyzing word composition, and the word elimination unit adopts word analysis technology, word feature extraction, adaptive filter optimization and multi-modal noise elimination; The word analysis technology uses the NLP natural language processing unit to analyze the vocabulary and syntax in the voice signal and identify the noise components therein; uses a deep learning unit, such as BERT or GPT-3, to perform semantic understanding of the voice signal, extract key information from it and eliminate noise; and uses an integrated sentiment analysis module to detect the user's emotional changes, thereby identifying the noise caused by emotional fluctuations; The word feature extraction uses word frequency analysis to count the frequency of occurrence of each word in the speech signal, and filters out words with low frequency of occurrence by setting a threshold; Using syntactic analysis, the structure of the sentence is analyzed to identify the ungrammatical or illogical noise parts; using semantic relevance analysis, the semantic relevance between words is calculated, and the words with low relevance are judged as noise. The adaptive filter optimization adopts the LMS algorithm to improve the least mean square algorithm, and adds a word analysis module to dynamically filter out noise words by adjusting the filter coefficient in real time; the RLS algorithm is applied, and the recursive least squares algorithm has a faster convergence speed; the Kalman filter optimization is adopted to suppress the noise of the dynamic system through the state space model and the observation model; The multimodal noise elimination uses visual assistance, combined with visual sensor data, to identify noise sources in the environment through image processing technology; By using tactile feedback and tactile sensors, the user's body movements and behaviors are detected to identify possible noise sources; multi-channel signal processing is used to combine data from different sensors for joint noise elimination.

3. According to claim 2, a recognition and detection system for AI voice dialogue scene application, Features: The core of the LMS algorithm improvement is to establish a nonlinear functional relationship between the step size factor μ and the error signal e(n), and to improve it through the Sigmoid function. The specific form is: ; The Sigmoid function is an S-shaped curve, which is mathematically defined as , where EXP represents the natural exponential function. The improvement is as follows: In order to overcome the shortcomings of the Sigmoid function in the step size adjustment process in the steady-state stage, it is multiplied by an even function, such as , so that the step factor changes slowly when the error is close to zero; Based on the above improvements, the nonlinear functional relationship between the step size factor μ and the error signal e(n) can be expressed as: , where β and α are adjustable parameters. By adjusting the parameters β and α, the performance of the algorithm can be optimized according to the actual application requirements.

4. According to claim 2, a recognition and detection system for AI voice dialogue scene application, Features: The core equations of the RLS algorithm include weight update formula and error calculation; Weight update formula: In the RLS algorithm, the update formula of the weight vector w(n) is as follows: , where g(n) is the gain vector and e(n) is the error signal; The calculation formula of the gain vector g(n) is: , where P(n) is the correlation matrix, x(n) is the input signal, and λ is the forgetting factor; The update formula of the correlation matrix P(n) is: ; Error calculation: The error signal e(n) is defined as the difference between the desired signal d(n) and the filter output y(n): ,in, , the expected signal d(n) is the known reference signal, and the output signal y(n) is the filter output at the current moment; Parameter adjustment: The forgetting factor λ is used to control the convergence speed and stability of the algorithm; Initial conditions: The correlation matrix P(0) is usually initialized to the identity matrix, that is , where δ is a small constant.

5. According to claim 4, a recognition and detection system for AI voice dialogue scene application, Features: The improvement measure of the RLS algorithm is to introduce an additional constant multiplier for optimizing the filter performance; By adjusting the value of the constant multiplier, optimization is performed for different signal-to-noise ratios (SNRs), and when the constant multiplier takes a value of 0.65, the best SNR value can be obtained, reaching 19.38 dB.

6. The recognition and detection system for AI voice dialogue scene application according to claim 1, Features: The intelligent recommendation analysis module includes: Intent recognition unit: By recognizing the user's initial question and subsequent secondary inquiry, it determines whether the user intends to seek more information about a specific method or is interested in other methods; Context management unit: used to maintain the consistency of context in multiple rounds of dialogue and accurately provide information related to the previous discussion when the user asks a second question; Sentiment analysis unit: used to help the system determine the user's attitude towards different methods; Feedback loop unit: through the establishment of feedback mechanisms to enable users to directly express their satisfaction with a method; Machine Learning Model Unit: Using machine learning models, the system can train algorithms based on historical data to predict methods that users may be interested in; Personalized recommendation unit: Based on the user's historical behavior and preferences, the system can provide personalized recommendations.

7. The recognition and detection system for AI voice dialogue scene application according to claim 6, Features: The intelligent recommendation analysis module also includes: Data collection and preprocessing unit: collects a large amount of user behavior data, cleans the data, and extracts useful features from the user behavior data; User portrait construction unit: Analyze the user's basic information and historical behavior data, identify the user's interest preferences in different methods, and then give each user a personalized label; Method evaluation and ranking unit: Evaluate the effectiveness of each method, rank the methods using ranking algorithms (such as collaborative filtering, content-based recommendation, etc.), and regularly update the ranking of the methods; Recommendation strategy formulation unit: recommend methods related to users based on their interests, preferences and needs; make recommendations based on similarities between users and methods; combine the advantages of multiple recommendation strategies to formulate more comprehensive and accurate recommendation plans; Real-time feedback and optimization unit: Encourage users to provide feedback on the recommendation method, continuously iterate and update the recommendation model based on user feedback and new user behavior data, and then integrate the real-time feedback and optimization process into daily operations to ensure the continuous improvement and optimization of the recommendation system; Interface display and interaction unit: clearly display the recommendation results and provide detailed explanations to guide user interaction; Cross-platform integration and expansion unit: supports multi-platform operation and provides API interfaces for other developers or third-party applications to call.

8. The recognition and detection system for AI voice dialogue scene application according to claim 1, Features: The knowledge graph unit provides rich background knowledge and information for the AI ​​voice dialogue system, supporting the system to perform logical reasoning and knowledge retrieval when answering user questions.

9. The recognition and detection system for AI voice dialogue scene application according to claim 1, Features: The privacy protection unit is used to protect user privacy.

10. A recognition and detection method for AI voice dialogue scenario applications, used to use the recognition and detection system for AI voice dialogue scenario applications as described in any one of claims 1 to 9, It is characterized in that The specific steps are as follows: Step 1, Voice Collection: The system collects the user's voice signal through microphones and other devices, which is the basis of the entire recognition process; Step 2, preprocessing: noise reduction, filtering, amplification and other processing are performed on the collected speech signal to improve the quality of the speech signal; Step 3, feature extraction: convert the preprocessed speech signal into features that can represent the essence of speech. This process involves complex signal processing and feature engineering techniques. Step 4, model matching: Match the extracted features with the pre-trained model to identify the specific content of the speech. This model is usually obtained through machine learning technology, such as deep learning training; Step 5, data upload: The method or approach determined to be adopted by the user will be uploaded to the big database as recommended content, and used as a recommended solution for other users facing the same problem.

Images