Anti-fraud communication method based on voice change recognition

Abstract

The invention relates to the technical field of voiceprint recognition, in particular to an anti-fraud communication method based on voice change recognition. The method comprises the following steps: acquiring multi-dimensional voice feature data of a user, wherein the multi-dimensional voice feature data comprises timbre, frequency features, language rhythm, emotion features and pause modes; performing voice change possibility prediction according to the multi-dimensional voice feature data to obtain voiceprint feature prediction data; performing deep learning model training according to the voiceprint feature prediction data, and mapping voiceprint states and features under different call conditions to obtain voiceprint feature map data; and according to the voiceprint feature map data, carrying out simulation analysis on voice change recognition behaviors in different call states so as to obtain voice change recognition simulation data. According to the method, the call environment noise data and the network quality data are obtained in real time, the call quality model is constructed, cross-scene optimization and real-time adjustment are carried out, and the influence of the environment on recognition is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of voiceprint recognition, and in particular to an anti-fraud communication method based on voice change recognition. Background Art

[0002] Voice recognition is a technology designed to identify and detect disguised or impersonated behavior in voice. This technology is particularly important in the field of voice recognition and security, because advanced voice-changing technology can mimic the voice characteristics of a specific individual, making traditional voiceprint recognition systems unreliable. With the rapid development of communication technology and artificial intelligence, voice communication has become more and more common in daily life and business activities. However, advanced voice-changing technology is also becoming increasingly popular, making it increasingly difficult to rely solely on traditional voiceprint recognition technology for identity authentication and fraud prevention.

[0003] Traditional anti-fraud communication methods often have the following problems: Modern voice-changing technology is becoming more and more advanced and can produce very realistic voices, which makes it more difficult to identify fraud based on voice characteristics alone; for example, some advanced voice-changing software can imitate the voice of a specific individual, including details such as timbre, intonation, and accent. Fluctuations in the quality of communication networks can cause voice distortion, increasing the difficulty of identification. For example, in areas with poor signals, the decline in call quality may affect the accuracy of voiceprint analysis. Setting the appropriate recognition threshold to balance false positives (misidentifying normal voices as voice changes) and false negatives (failing to identify voice changes) is a difficult problem; for example, the system may misidentify voice changes caused by a cold as voice changes, or fail to recognize highly realistic voice-changing technology. Summary of the invention

[0004] Based on this, it is necessary for the present invention to provide an anti-fraud communication method based on voice change recognition to solve at least one of the above technical problems.

[0005] To achieve the above purpose, a fraud prevention communication method based on voice change recognition includes the following steps:

[0006] Step S1: obtaining multi-dimensional voice feature data of the user, including timbre, frequency features, language rhythm, emotional features, and pause patterns; predicting the possibility of voice change based on the multi-dimensional voice feature data to obtain voiceprint feature prediction data; training a deep learning model based on the voiceprint feature prediction data, and mapping the voiceprint states and features under different call conditions to obtain voiceprint feature graph data;

[0007] Step S2: simulating and analyzing the voice change recognition behavior under different call states according to the voiceprint feature graph data, thereby obtaining voice change recognition simulation data; establishing a dynamic recognition threshold model according to the environmental parameters corresponding to the different call states and the voice change recognition simulation data;

[0008] Step S3: adjusting the control parameters in real time according to the preset target recognition accuracy and the dynamic recognition threshold model, thereby obtaining control parameter adjustment data; constructing a multi-level feedback control system according to the control parameter adjustment data, thereby obtaining an adaptive recognition model;

[0009] Step S4: Obtain the user's historical call data, perform personalized analysis on the user's historical call data, and perform feature extraction to obtain user portrait data; perform abnormal behavior prediction based on the user portrait data and the adaptive recognition model, and obtain the optimal recognition threshold through machine learning technology to obtain personalized recognition solution data;

[0010] Step S5: Acquire call environment noise data and network quality data, and build a call quality model based on the call environment noise data and network quality data; optimize the call quality model across scenarios, and adjust communication parameters in real time to obtain call quality optimization data;

[0011] Step S6: Perform multi-criteria decision fusion analysis on the personalized identification solution data and call quality optimization data to obtain the final anti-fraud identification solution data.

[0012] The present invention can more accurately capture the speaker's unique voiceprint and reduce the possibility of misjudgment and missed judgment by comprehensively analyzing these features, including timbre, frequency characteristics, language rhythm, emotional characteristics and pause patterns. By analyzing multi-dimensional voice feature data, it is predicted whether the sound has been processed by voice change, and voiceprint feature prediction data is obtained; this step helps to detect potential voice change behavior early. Using voiceprint feature prediction data to train a deep learning model can enable the model to learn to recognize various complex voice features and improve the accuracy and robustness of recognition. By mapping the voiceprint state under different call conditions, voiceprint feature graph data is obtained to help the system better understand and analyze the performance of voice features in different environments. The voice change recognition behavior under different call states is simulated and analyzed to obtain voice change recognition simulation data. This helps the system accurately identify voice change behavior under different environments and conditions. According to the environmental parameters corresponding to different call states and the voice change recognition simulation data, a dynamic recognition threshold model is established, and the recognition threshold can be flexibly adjusted to improve the sensitivity and accuracy of recognition. According to the preset target recognition accuracy and dynamic recognition threshold model, the control parameters are adjusted in real time to ensure that the system is always in the best recognition state. Construct a multi-level feedback control system, adjust the data according to the control parameters to continuously optimize the system, and improve the system's adaptability and recognition accuracy. Obtain the user's historical call data, conduct personalized analysis and feature extraction, and establish user portrait data to better understand the user's voice characteristics and behavior patterns. According to the user portrait data and the adaptive recognition model, abnormal behavior prediction is carried out, and the optimal recognition threshold is obtained through machine learning technology to obtain personalized recognition solution data. Obtain the call environment noise data and network quality data, build a call quality model, which can accurately reflect the impact of call quality on recognition. Optimize the call quality model across scenarios and adjust the communication parameters in real time to ensure the recognition performance of the system in different environments. Perform multi-criteria decision fusion analysis on the personalized recognition solution data and call quality optimization data, comprehensively evaluate various factors, and improve the accuracy and reliability of recognition; comprehensively consider multiple factors, make comprehensive recognition decisions, and improve the overall performance of the system; through multi-criteria decision-making, reduce false positives and false negatives, and improve the practicality and user experience of the system.

[0013] Preferably, step S1 comprises the following steps:

[0014] Step S11: Acquire the user's historical voice data and perform short-time Fourier transform on the historical voice data to obtain voice time-frequency data;

[0015] Step S12: extracting Mel-frequency cepstral coefficients from the speech time-frequency data to obtain timbre feature data; extracting frequency features based on fundamental frequency contour from the speech time-frequency data to obtain frequency feature data;

[0016] Step S13: using a voice activity detection algorithm to perform voice and non-voice segment recognition on the voice time-frequency data, thereby obtaining language rhythm feature data;

[0017] Step S14: analyzing the duration of speech segments and the pause patterns of non-speech segments on the speech time-frequency data according to the language rhythm feature data, thereby obtaining pause pattern data;

[0018] Step S15: performing pitch change, energy distribution and speech speed analysis on the speech time-frequency data, and extracting emotional features, thereby obtaining emotional feature data; performing harmonic-to-noise ratio analysis on the speech time-frequency data, and evaluating the harmonic structure of the speech, thereby obtaining harmonic structure data;

[0019] Step S16: merging the timbre feature data, the frequency feature data, the language rhythm feature data, the pause pattern data and the emotion feature data into multi-dimensional speech feature data according to the harmonic structure data;

[0020] Step S17: using a support vector machine to classify the multi-dimensional speech feature data and perform a voice change possibility score, thereby obtaining voiceprint feature prediction data;

[0021] Step S18: Perform deep learning model training based on the voiceprint feature prediction data, and map the voiceprint states and features under different call conditions to obtain voiceprint feature graph data.

[0022] The present invention uses STFT, MFCC, fundamental frequency contour, VAD and other technologies to extract voice features from multiple dimensions and provide high-resolution and high-precision voice analysis. Combined with multi-dimensional features such as timbre, frequency, language rhythm, pause pattern, emotional characteristics and harmonic structure, comprehensive voice feature data is provided to improve the accuracy and robustness of recognition. By dynamically adjusting the recognition threshold and multi-level feedback control system, the adaptability of the system under different call conditions is enhanced to ensure the stability and reliability of recognition. According to the user's historical voice data and personalized analysis, a personalized recognition scheme is provided to improve the system's recognition ability for individual users and reduce false positives and false negatives. A deep learning model is used for training and mapping to capture complex voice features and improve the system's detection ability and robustness for voice change behavior. Various factors are comprehensively evaluated through a multi-criteria decision model to make comprehensive recognition decisions, improve the overall performance of the system, and reduce false positives and false negatives. Overall, this anti-fraud communication method based on voice change recognition significantly improves the security and recognition accuracy of voice communication through multi-dimensional and multi-level comprehensive analysis and optimization, can effectively cope with advanced voice change technology and complex communication environment, and provide more reliable anti-fraud protection.

[0023] Preferably, step S18 comprises the following steps:

[0024] Step S181: adding noise and time stretching to the voiceprint feature prediction data, and performing normalization processing to obtain a voiceprint feature training data set;

[0025] Step S182: training a preset deep learning model according to the voiceprint feature training data set, and adjusting the model parameters by using a cross-validation method, thereby obtaining a voiceprint feature model, wherein the deep learning model includes a multi-layer convolutional neural network structure for extracting local features, a long short-term memory network layer for capturing temporal dependencies, an attention mechanism layer for highlighting important features, and a fully connected layer;

[0026] Step S183: simulating different call conditions to obtain conditional simulation data; applying the voiceprint feature model to the conditional simulation data to obtain conditional response data;

[0027] Step S184: Map the voiceprint states and features under different call conditions according to the conditional response data and the conditional simulation data to obtain voiceprint feature graph data.

[0028] The present invention enhances the diversity of training data and the generalization ability of the model by adding noise and time expansion, so that it can better adapt to various actual call environments. Normalization processing reduces data deviation, improves the stability and efficiency of model training, and ensures that features are compared on the same scale. Extract local features of speech to improve the model's recognition ability of subtle speech features; capture the temporal dependency in speech signals to improve the model's understanding of dynamic changes in speech; highlight important features, enhance the model's attention to key speech features, and improve recognition accuracy. By simulating different call conditions, the performance of the model in various actual environments is comprehensively tested to improve the adaptability of the model; obtain model response data under different conditions to help evaluate and optimize the performance of the model and ensure its accuracy and robustness in various environments. Map the voiceprint states and features under different call conditions to generate voiceprint feature map data, provide intuitive visualization information, help understand and analyze speech features; integrate feature data under various conditions to improve the model's recognition ability of complex speech environments and ensure stability and reliability in practical applications. Optimize model parameters to improve the robustness and generalization ability of the model and ensure stable performance under different call conditions. By adding noise and time stretching, the diversity of training data is enhanced, and the model's adaptability to various actual call environments is improved. Multi-layer convolutional neural networks and long short-term memory network layers are used to extract local features and capture temporal dependencies, improving the model's ability to recognize voice features. The attention mechanism layer highlights important features, enhances the model's attention to key voice features, and improves recognition accuracy. Through cross-validation, the model parameters are optimized to improve the robustness and generalization of the model and ensure stable performance under different call conditions. Different call conditions are simulated to comprehensively test and optimize the model's performance to improve its adaptability and accuracy in various environments. Voiceprint feature map data is generated to provide intuitive visual information to help understand and analyze voice features and improve the model's ability to recognize complex voice environments. Overall, these steps improve the accuracy and robustness of voice change recognition through systematic data processing and model optimization, ensure effectiveness and reliability in practical applications, and provide a solid technical guarantee for anti-fraud communications.

[0029] Preferably, step S2 comprises the following steps:

[0030] Step S21: Performing statistical analysis on the characteristic distribution, clustering and abnormal points of the voiceprint feature graph data, so as to obtain voiceprint feature statistical data;

[0031] Step S22: simulating and analyzing the voice change recognition behavior of the voice print feature graph data according to the voice print feature statistical data, thereby obtaining voice change recognition simulation data;

[0032] Step S23: evaluating the recognition performance of the voice change recognition simulation data, and calculating the recognition accuracy, false alarm rate and missed alarm rate under different conditions, so as to obtain recognition performance evaluation data;

[0033] Step S24: performing a performance impact analysis on different environmental parameters according to the recognition performance evaluation data, and establishing a relationship model between environmental parameters and recognition performance, thereby obtaining an environmental impact model;

[0034] Step S25: Set initial recognition thresholds for different call states according to the environmental impact model and recognition performance evaluation data, adjust the thresholds according to real-time environmental parameters, and perform machine learning model training to obtain a dynamic recognition threshold model.

[0035] The present invention provides a comprehensive distribution of feature data to help understand the performance of voice features under different call conditions and improve the basic data quality of the model; identifies and groups similar voiceprint features to help discover potential voice change patterns and improve the accuracy of voice change recognition; detects and analyzes abnormal feature points to enhance the system's ability to recognize abnormal voice behavior and reduce false positives and false negatives. Voice change recognition simulation is performed based on detailed statistical data to provide a comprehensive understanding of voice change behavior and enhance the accuracy and reliability of voice change recognition by the system. Systematically evaluate the accuracy, false positive rate and false negative rate of voice change recognition, provide a detailed understanding of model performance, and guide model optimization and adjustment; optimize the model based on performance evaluation data to improve the overall performance and stability of recognition. Reveal the impact of different environmental parameters on recognition performance, provide an in-depth understanding of environmental factors, and ensure the adaptability and robustness of the model in various environments; establish a relationship model between environmental parameters and recognition performance to provide a basis for dynamically adjusting recognition parameters and ensure continuous optimization of recognition performance. Set reasonable initial recognition thresholds for different call states to improve the initial accuracy of the recognition system, dynamically adjust the recognition threshold according to real-time environmental parameters, and ensure recognition accuracy and robustness under various call conditions; through continuous machine learning model training, continuously optimize the recognition threshold model to improve the system's adaptability and recognition performance. Overall, these steps significantly improve the accuracy and robustness of voice change recognition through systematic data processing, analysis, and model optimization, ensure effectiveness and reliability in practical applications, and provide solid technical support for anti-fraud communications.

[0036] Preferably, step S3 comprises the following steps:

[0037] Step S31: extracting key control parameters and their influence ranges in the dynamic recognition threshold model, thereby obtaining a control parameter influence matrix;

[0038] Step S32: setting initial control parameter values ​​according to a preset target recognition accuracy rate and a control parameter influence matrix, thereby obtaining an initial control parameter set;

[0039] Step S33: performing a parameter sensitivity test on the initial control parameter set based on the degree of influence of each control parameter on the recognition accuracy, thereby obtaining parameter sensitivity data; using a gradient descent method to perform parameter adaptive adjustment according to the parameter sensitivity data, thereby obtaining control parameter adjustment data;

[0040] Step S34: designing a multi-level feedback control loop according to the control parameter adjustment data and real-time monitoring of recognition accuracy, processing delay and resource occupancy indicators, thereby obtaining a multi-level feedback control system, wherein the multi-level feedback control loop includes two levels: fast response and long-term optimization;

[0041] Step S35: running the multi-level feedback control system on the simulation test platform, collecting system response data, and performing control effect analysis to obtain control system performance data;

[0042] Step S36: Encapsulating the multi-level feedback control system into a deployable adaptive recognition model according to the control system performance data, thereby obtaining an adaptive recognition model.

[0043] The present invention extracts key control parameters and their influence ranges, clarifies key variables in the model, and improves the comprehensibility and controllability of the model; the control parameter influence matrix provides the specific influence relationship of each parameter on the recognition performance, and provides a scientific basis for subsequent parameter optimization and adjustment. According to the target recognition accuracy and the control parameter influence matrix, the initial control parameter value is reasonably set to improve the accuracy and performance of the initial operation of the system. The optimized initial control parameter set reduces the model debugging and adjustment time and improves the development efficiency. Through sensitivity testing, the specific influence of each control parameter on the recognition accuracy is clarified, the optimization direction is provided, and the gradient descent method is used to adjust the parameters to improve the accuracy and efficiency of the control parameter adjustment, and ensure the optimization performance of the recognition system. By designing a multi-level feedback control loop with fast response and long-term optimization, it is ensured that the system can respond quickly to short-term fluctuations, and at the same time achieve long-term optimization, real-time monitoring of recognition accuracy, processing delay and resource occupancy, and improve the overall performance and stability of the system. Through the simulation test platform, the multi-level feedback control system is operated, potential problems are discovered and solved in advance, and the successful deployment of the system is ensured; the system response data is collected, the control effect analysis is performed, and comprehensive performance data is provided to provide a scientific basis for system optimization. The optimized multi-level feedback control system is encapsulated as a deployable adaptive recognition model to facilitate practical application and improve the practicality and deployment efficiency of the model; the adaptive recognition model optimized based on comprehensive performance data has higher stability and reliability, ensuring excellent performance in practical applications. Overall, these steps significantly improve the accuracy, robustness and adaptability of the voice change recognition system through systematic parameter extraction, optimization and multi-level feedback control design, providing an efficient and reliable technical solution for anti-fraud communications.

[0044] Preferably, step S4 comprises the following steps:

[0045] Step S41: acquiring the user's historical call data, and performing desensitization processing on the user's historical call data, thereby obtaining desensitized historical call data;

[0046] Step S42: performing time series analysis on the desensitized historical call data and extracting user call pattern features, thereby obtaining user call pattern feature data;

[0047] Step S43: constructing a social network graph of users based on call objects and frequencies according to the desensitized historical call data, thereby obtaining social network feature data;

[0048] Step S44: performing user classification based on a clustering algorithm according to the user's call pattern feature data and social network feature data, thereby obtaining user portrait data;

[0049] Step S45: performing abnormal behavior prediction based on the user portrait data and the adaptive recognition model, thereby obtaining abnormal behavior prediction data;

[0050] Step S46: Optimize the recognition threshold according to the abnormal behavior prediction data through machine learning technology, so as to obtain personalized recognition solution data.

[0051] The present invention effectively protects the user's personal privacy information through desensitization processing, avoids the risk of sensitive data leakage and infringement of user privacy; ensures the security of historical call data during processing and storage, complies with data protection regulations and standards, and enhances the credibility and transparency of data management. Through time series analysis, extract the user's call pattern characteristics, including call frequency, time period preference, etc., and deeply understand the user's behavior habits and communication behavior rules; based on the call pattern feature data, personalized customization services, such as optimizing communication plans, providing customized recommendations, etc., enhance user experience and satisfaction. Build a user's social network graph, analyze the call objects and their frequencies, and reveal the relationship density and activity of the user in the social circle; through social network feature data, identify and understand the user's influence and key roles in the social network, and provide data support for social interaction. Combined with user portrait data and adaptive recognition models, effectively predict abnormal patterns and abnormal events in user communication behavior, and timely identify and respond to potential risks. Improve anti-fraud capabilities, reduce the potential damage caused by abnormal behavior to user and system security, and protect communication security and user interests. Using machine learning technology to predict abnormal behavior data, dynamically adjust the recognition threshold to achieve a personalized anti-fraud recognition solution; by optimizing the recognition threshold, effectively improve the recognition accuracy and real-time response capability of the anti-fraud system, and reduce the false alarm rate and missed alarm rate. In summary, these steps combine data processing, analysis and prediction technologies, provide multi-dimensional data support and analysis capabilities for anti-fraud communication methods, and effectively improve the security, user experience and operational efficiency of the system.

[0052] Preferably, step S5 comprises the following steps:

[0053] Step S51: Acquire call environment noise data and network quality data;

[0054] Step S52: performing call quality evaluation according to the call environment noise data and the network quality data, thereby obtaining a call quality model, wherein the call quality evaluation includes audio clarity evaluation, continuity evaluation, and echo cancellation effect evaluation;

[0055] Step S53: Scenario call data are collected in a variety of real scenarios, and transfer learning technology is used to optimize the model generalization capability of the call quality model according to the scenario call data, thereby obtaining call quality optimization data.

[0056] By acquiring call environment noise data and network quality data, the system can perceive the noise level and network conditions of the real-time communication environment, which helps to adjust the communication parameters to optimize the call quality; it provides basic data support and provides necessary input information for subsequent call quality evaluation and optimization. Based on the acquired environmental data, the system can accurately evaluate the call quality, including key indicators such as audio clarity, call continuity, and echo cancellation effect; the evaluation results can be used to adjust the communication parameters in real time, optimize the call experience, and ensure the good call quality of users under various environmental conditions. By collecting call data in a variety of real scenarios and applying transfer learning technology, the generalization ability of the call quality model is optimized so that it can adapt to different communication scenarios and environmental conditions; the effectiveness and stability of the call quality optimization strategy in various real scenarios are ensured, and the reliability and practicality of the system in practical applications are improved; by continuously collecting and analyzing scene call data, the call quality model is continuously optimized, the system's adaptability to changes in the communication environment is maintained, and stable call service quality is provided. In summary, these steps integrate data collection, analysis, and model optimization technologies, provide key call quality evaluation and optimization capabilities for anti-fraud communication methods, and thus significantly improve the user experience and operation efficiency of the system.

[0057] Preferably, step S6 comprises the following steps:

[0058] Step S61: performing data cleaning and standardization processing on the personalized identification solution data and the call quality optimization data, so as to obtain personalized identification standard data and call quality standard data;

[0059] Step S62: extracting timbre features, emotional features and environmental noise features from the personalized recognition standard data and the call quality standard data, and using the hierarchical analysis method to assign feature weights, thereby obtaining feature weight assignment data;

[0060] Step S63: inputting the feature weight distribution data, the personalized identification standard data and the call quality standard data into a preset multi-criteria decision model, thereby obtaining multi-criteria decision data;

[0061] Step S64: Identify potential abnormal behaviors and fraudulent behaviors on the multi-criteria decision data, and perform risk assessment and threshold adjustment to obtain final anti-fraud identification solution data.

[0062] The present invention ensures the consistency and accuracy of personalized identification scheme data and call quality optimization data through data cleaning and standardization, and improves the reliability of subsequent analysis and decision-making; cleaning and standardization can remove noise and errors in the data, making subsequent data analysis and decision-making more accurate and reliable; processing the data according to a unified standard is conducive to the comparison and integration between different data sources, and improving the operability and application value of the data. By extracting timbre features, emotional features and environmental noise features, the key features of call quality and personalized identification schemes are deeply analyzed to provide strong support for subsequent decision-making; using the hierarchical analysis method to allocate feature weights, it is possible to objectively evaluate the contribution of each feature to the final decision, and improve the scientificity and accuracy of the decision; comprehensively considering and weighing different types of data is conducive to establishing a comprehensive and integrated decision-making model, and improving the decision-making efficiency and accuracy of the system. Through the multi-criteria decision-making model, multiple factors such as feature weights, personalized identification standards and call quality standards are comprehensively considered, and the anti-fraud identification scheme can be more comprehensively evaluated and optimized. Based on the scientific decision-making model, subjective bias can be effectively reduced, and the objectivity and credibility of the decision can be improved; through multi-criteria decision-making, the system can more accurately identify potential abnormalities and fraudulent behaviors, take preventive and countermeasures in time, and ensure communication security and user rights. Identifying abnormal and fraudulent behaviors based on multi-criteria decision data can effectively improve the system's ability to perceive and respond to potential risks, conduct in-depth evaluation and analysis of identified abnormal behaviors, quantify risk levels, and provide a basis for developing targeted preventive measures; dynamically adjust the threshold of the anti-fraud identification scheme based on real-time risk assessment results to ensure the stability and efficient operation of the system in different scenarios. In summary, these steps effectively integrate data analysis, decision models, and risk management technologies, providing comprehensive data support and scientific decision-making basis for anti-fraud communication methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0064] Figure 1 The figure is a flowchart of the steps of the anti-fraud communication method based on voice change recognition of the present invention;

[0065] Figure 2 for Figure 1 Detailed step flow diagram of step S1;

[0066] Figure 3 for Figure 1 Detailed step flow chart of step S2 in FIG. DETAILED DESCRIPTION

[0067] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0068] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0069] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0070] To achieve this, please refer to Figures 1 to 3 The present invention provides an anti-fraud communication method based on voice change recognition, the method comprising the following steps:

[0071] Step S1: obtaining multi-dimensional voice feature data of the user, including timbre, frequency features, language rhythm, emotional features, and pause patterns; predicting the possibility of voice change based on the multi-dimensional voice feature data to obtain voiceprint feature prediction data; training a deep learning model based on the voiceprint feature prediction data, and mapping the voiceprint states and features under different call conditions to obtain voiceprint feature graph data;

[0072] Step S2: simulating and analyzing the voice change recognition behavior under different call states according to the voiceprint feature graph data, thereby obtaining voice change recognition simulation data; establishing a dynamic recognition threshold model according to the environmental parameters corresponding to the different call states and the voice change recognition simulation data;

[0073] Step S3: adjusting the control parameters in real time according to the preset target recognition accuracy and the dynamic recognition threshold model, thereby obtaining control parameter adjustment data; constructing a multi-level feedback control system according to the control parameter adjustment data, thereby obtaining an adaptive recognition model;

[0074] Step S4: Obtain the user's historical call data, perform personalized analysis on the user's historical call data, and perform feature extraction to obtain user portrait data; perform abnormal behavior prediction based on the user portrait data and the adaptive recognition model, and obtain the optimal recognition threshold through machine learning technology to obtain personalized recognition solution data;

[0075] Step S5: Acquire call environment noise data and network quality data, and build a call quality model based on the call environment noise data and network quality data; optimize the call quality model across scenarios, and adjust communication parameters in real time to obtain call quality optimization data;

[0076] Step S6: Perform multi-criteria decision fusion analysis on the personalized identification solution data and call quality optimization data to obtain the final anti-fraud identification solution data.

[0077] In the embodiment of the present invention, reference Figure 1 The above is a schematic flow chart of the steps of an anti-fraud communication method based on voice change recognition of the present invention. In this example, the anti-fraud communication method based on voice change recognition includes the following steps:

[0078] Step S1: obtaining multi-dimensional voice feature data of the user, including timbre, frequency features, language rhythm, emotional features, and pause patterns; predicting the possibility of voice change based on the multi-dimensional voice feature data to obtain voiceprint feature prediction data; training a deep learning model based on the voiceprint feature prediction data, and mapping the voiceprint states and features under different call conditions to obtain voiceprint feature graph data;

[0079] The embodiment of the present invention uses an audio processing library such as Librosa or PyAudio to load audio and perform spectrum analysis, extract vocal tract length, formant, sound waveform, etc.; apply Fourier transform or short-time Fourier transform (STFT) to extract spectrum information; use voice activity detection algorithms such as VAD (Voice Activity Detection) to identify speech and non-speech segments; extract emotional information by analyzing speech speed, pitch changes, etc.; use the duration of speech segments and pause pattern analysis of non-speech segments to detect natural pauses in speech. Use machine learning algorithms such as support vector machines (SVM) to classify multi-dimensional voice feature data and predict the voice change possibility score of each voice sample. Based on the predicted voiceprint feature data, design a deep learning model such as convolutional neural network (CNN), long short-term memory network (LSTM), etc., train a voiceprint recognition model, map the voiceprint status and features under different call conditions, and generate voiceprint feature map data for subsequent analysis and simulation.

[0080] Step S2: simulating and analyzing the voice change recognition behavior under different call states according to the voiceprint feature graph data, thereby obtaining voice change recognition simulation data; establishing a dynamic recognition threshold model according to the environmental parameters corresponding to the different call states and the voice change recognition simulation data;

[0081] The embodiment of the present invention uses voiceprint feature map data to simulate and analyze the voice change recognition behavior under different call states, and evaluates the recognition accuracy, false alarm rate and missed alarm rate; combines call environment parameters and voice change recognition simulation data, and uses statistical methods or machine learning algorithms to establish a dynamic recognition threshold model; determines the initial recognition threshold under different call states, and dynamically adjusts it as the real-time environmental parameters change.

[0082] Step S3: adjusting the control parameters in real time according to the preset target recognition accuracy and the dynamic recognition threshold model, thereby obtaining control parameter adjustment data; constructing a multi-level feedback control system according to the control parameter adjustment data, thereby obtaining an adaptive recognition model;

[0083] The embodiment of the present invention adjusts the control parameters in real time according to the preset target recognition accuracy and dynamic recognition threshold model, such as adjusting the learning rate and decision boundary of the model. A multi-level feedback control system is designed according to the control parameter adjustment data, including two levels: fast response and long-term optimization. The fast response level makes rapid adjustments to the real-time call quality and voice change recognition performance, while the long-term optimization level performs more in-depth systematic optimization based on historical data.

[0084] Step S4: Obtain the user's historical call data, perform personalized analysis on the user's historical call data, and perform feature extraction to obtain user portrait data; perform abnormal behavior prediction based on the user portrait data and the adaptive recognition model, and obtain the optimal recognition threshold through machine learning technology to obtain personalized recognition solution data;

[0085] The embodiment of the present invention performs desensitization processing on the user's historical call data to ensure data privacy and security, and uses time series analysis, clustering algorithms and other technologies to extract user call pattern characteristics, such as call frequency, time period preference, etc. Based on user portrait data and adaptive recognition models, the user's abnormal call behavior is predicted, and machine learning technology is used to optimize the personalized recognition threshold to ensure timely recognition and response to abnormal behavior.

[0086] Step S5: Acquire call environment noise data and network quality data, and build a call quality model based on the call environment noise data and network quality data; optimize the call quality model across scenarios, and adjust communication parameters in real time to obtain call quality optimization data;

[0087] The embodiment of the present invention obtains call environment noise data and network quality data, such as audio signal-to-noise ratio, network delay, etc., and uses the collected call environment data to build a call quality evaluation model to evaluate audio clarity, continuity, echo and other effects; and uses transfer learning technology to optimize the model generalization ability to ensure call quality optimization effects in a variety of real scenarios.

[0088] Step S6: Perform multi-criteria decision fusion analysis on the personalized identification solution data and call quality optimization data to obtain the final anti-fraud identification solution data.

[0089] The embodiment of the present invention performs data cleaning and standardization processing on the personalized identification scheme data and call quality optimization data to ensure data quality and consistency; inputs the cleaned data into a preset multi-criteria decision model, such as a hierarchical analysis method, a decision tree, etc., to perform a comprehensive decision analysis; considers factors such as feature weights, user portraits, and call quality to formulate the optimal anti-fraud identification scheme.

[0090] The present invention can more accurately capture the speaker's unique voiceprint and reduce the possibility of misjudgment and missed judgment by comprehensively analyzing these features, including timbre, frequency characteristics, language rhythm, emotional characteristics and pause patterns. By analyzing multi-dimensional voice feature data, it is predicted whether the sound has been processed by voice change, and voiceprint feature prediction data is obtained; this step helps to detect potential voice change behavior early. Using voiceprint feature prediction data to train a deep learning model can enable the model to learn to recognize various complex voice features and improve the accuracy and robustness of recognition. By mapping the voiceprint state under different call conditions, voiceprint feature graph data is obtained to help the system better understand and analyze the performance of voice features in different environments. The voice change recognition behavior under different call states is simulated and analyzed to obtain voice change recognition simulation data. This helps the system accurately identify voice change behavior under different environments and conditions. According to the environmental parameters corresponding to different call states and the voice change recognition simulation data, a dynamic recognition threshold model is established, and the recognition threshold can be flexibly adjusted to improve the sensitivity and accuracy of recognition. According to the preset target recognition accuracy and dynamic recognition threshold model, the control parameters are adjusted in real time to ensure that the system is always in the best recognition state. Construct a multi-level feedback control system, adjust the data according to the control parameters to continuously optimize the system, and improve the system's adaptability and recognition accuracy. Obtain the user's historical call data, conduct personalized analysis and feature extraction, and establish user portrait data to better understand the user's voice characteristics and behavior patterns. According to the user portrait data and the adaptive recognition model, abnormal behavior prediction is carried out, and the optimal recognition threshold is obtained through machine learning technology to obtain personalized recognition solution data. Obtain the call environment noise data and network quality data, build a call quality model, which can accurately reflect the impact of call quality on recognition. Optimize the call quality model across scenarios and adjust the communication parameters in real time to ensure the recognition performance of the system in different environments. Perform multi-criteria decision fusion analysis on the personalized recognition solution data and call quality optimization data, comprehensively evaluate various factors, and improve the accuracy and reliability of recognition; comprehensively consider multiple factors, make comprehensive recognition decisions, and improve the overall performance of the system; through multi-criteria decision-making, reduce false positives and false negatives, and improve the practicality and user experience of the system.

[0091] Preferably, step S1 comprises the following steps:

[0092] Step S11: Acquire the user's historical voice data and perform short-time Fourier transform on the historical voice data to obtain voice time-frequency data;

[0093] Step S12: extracting Mel-frequency cepstral coefficients from the speech time-frequency data to obtain timbre feature data; extracting frequency features based on fundamental frequency contour from the speech time-frequency data to obtain frequency feature data;

[0094] Step S13: using a voice activity detection algorithm to perform voice and non-voice segment recognition on the voice time-frequency data, thereby obtaining language rhythm feature data;

[0095] Step S14: analyzing the duration of speech segments and the pause patterns of non-speech segments on the speech time-frequency data according to the language rhythm feature data, thereby obtaining pause pattern data;

[0096] Step S15: performing pitch change, energy distribution and speech speed analysis on the speech time-frequency data, and extracting emotional features, thereby obtaining emotional feature data; performing harmonic-to-noise ratio analysis on the speech time-frequency data, and evaluating the harmonic structure of the speech, thereby obtaining harmonic structure data;

[0097] Step S16: merging the timbre feature data, the frequency feature data, the language rhythm feature data, the pause pattern data and the emotion feature data into multi-dimensional speech feature data according to the harmonic structure data;

[0098] Step S17: using a support vector machine to classify the multi-dimensional speech feature data and perform a voice change possibility score, thereby obtaining voiceprint feature prediction data;

[0099] Step S18: Perform deep learning model training based on the voiceprint feature prediction data, and map the voiceprint states and features under different call conditions to obtain voiceprint feature graph data.

[0100] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed step flow diagram of step S1 in the embodiment of the present invention, step S1 includes the following steps:

[0101] Step S11: Acquire the user's historical voice data and perform short-time Fourier transform on the historical voice data to obtain voice time-frequency data;

[0102] The embodiment of the present invention obtains historical voice data from a voice database or user communication records to ensure data security and privacy protection. The voice signal is processed in frames, each frame is usually about 20-40 milliseconds, and the overlapping part can be set to half the frame length. A window function (such as a Hanning window) is applied to each frame of voice to reduce spectrum leakage. The voice signal after each frame of windowing is subjected to FFT (Fast Fourier Transform) to obtain spectrum information. The spectrum information of each frame is represented as time-frequency domain data, that is, the spectrum distribution on the time axis.

[0103] Step S12: extracting Mel-frequency cepstral coefficients from the speech time-frequency data to obtain timbre feature data; extracting frequency features based on fundamental frequency contour from the speech time-frequency data to obtain frequency feature data;

[0104] The embodiment of the present invention applies a Mel filter bank to the time-frequency data of each frame, converts the linear spectrum into a Mel frequency spectrum, takes the logarithm to obtain the power spectrum, and then performs a discrete cosine transform (DCT) to obtain the MFCC coefficients, and extracts the MFCC coefficients as timbre feature data. The fundamental frequency information of the speech signal is analyzed to extract the fundamental frequency contour; frequency features such as fundamental frequency change pattern, frequency distribution, etc. are extracted according to the fundamental frequency contour.

[0105] Step S13: using a voice activity detection algorithm to perform voice and non-voice segment recognition on the voice time-frequency data, thereby obtaining language rhythm feature data;

[0106] The embodiment of the present invention applies a voice activity detection algorithm to the time-frequency data of each frame, identifies the voice segment and the non-voice segment, and analyzes the rhythmic features such as the duration of the voice segment and the pause pattern of the non-voice segment.

[0107] Step S14: analyzing the duration of speech segments and the pause patterns of non-speech segments on the speech time-frequency data according to the language rhythm feature data, thereby obtaining pause pattern data;

[0108] The embodiment of the present invention analyzes patterns such as the duration distribution of speech segments and the pause length of non-speech segments based on language rhythm feature data.

[0109] Step S15: performing pitch change, energy distribution and speech speed analysis on the speech time-frequency data, and extracting emotional features, thereby obtaining emotional feature data; performing harmonic-to-noise ratio analysis on the speech time-frequency data, and evaluating the harmonic structure of the speech, thereby obtaining harmonic structure data;

[0110] The embodiment of the present invention analyzes pitch changes, energy distribution, speech speed, etc. in speech time-frequency data to extract emotional features; calculates the ratio of harmonic components and noise components of the speech signal to evaluate the harmonic structure of the speech.

[0111] Step S16: merging the timbre feature data, the frequency feature data, the language rhythm feature data, the pause pattern data and the emotion feature data into multi-dimensional speech feature data according to the harmonic structure data;

[0112] The embodiment of the present invention integrates timbre features, frequency features, language rhythm features, pause pattern data and emotion feature data into a multi-dimensional speech feature data set.

[0113] Step S17: using a support vector machine to classify the multi-dimensional speech feature data and perform a voice change possibility score, thereby obtaining voiceprint feature prediction data;

[0114] The embodiment of the present invention inputs multi-dimensional speech feature data into an SVM classifier to predict and classify voiceprint features.

[0115] Step S18: Perform deep learning model training based on the voiceprint feature prediction data, and map the voiceprint states and features under different call conditions to obtain voiceprint feature graph data.

[0116] The embodiment of the present invention designs a deep learning model based on voiceprint feature prediction data, such as a convolutional neural network (CNN), a long short-term memory network (LSTM), etc.; under different call conditions, the voiceprint state and features are mapped to generate voiceprint feature map data.

[0117] The present invention uses STFT, MFCC, fundamental frequency contour, VAD and other technologies to extract voice features from multiple dimensions and provide high-resolution and high-precision voice analysis. Combined with multi-dimensional features such as timbre, frequency, language rhythm, pause pattern, emotional characteristics and harmonic structure, comprehensive voice feature data is provided to improve the accuracy and robustness of recognition. By dynamically adjusting the recognition threshold and multi-level feedback control system, the adaptability of the system under different call conditions is enhanced to ensure the stability and reliability of recognition. According to the user's historical voice data and personalized analysis, a personalized recognition scheme is provided to improve the system's recognition ability for individual users and reduce false positives and false negatives. A deep learning model is used for training and mapping to capture complex voice features and improve the system's detection ability and robustness for voice change behavior. Various factors are comprehensively evaluated through a multi-criteria decision model to make comprehensive recognition decisions, improve the overall performance of the system, and reduce false positives and false negatives. Overall, this anti-fraud communication method based on voice change recognition significantly improves the security and recognition accuracy of voice communication through multi-dimensional and multi-level comprehensive analysis and optimization, can effectively cope with advanced voice change technology and complex communication environment, and provide more reliable anti-fraud protection.

[0118] Preferably, step S18 comprises the following steps:

[0119] Step S181: adding noise and time stretching to the voiceprint feature prediction data, and performing normalization processing to obtain a voiceprint feature training data set;

[0120] The embodiment of the present invention selects appropriate noise type and intensity from the original voiceprint feature prediction data and adds it to the speech signal; performs time expansion on the speech signal and adjusts the playback speed of the speech to generate speech segments of different time lengths; ensures that the addition of noise and time expansion does not significantly change the recognizability of the speech signal, that is, the speech features can still maintain recognition. After adding noise and time expansion, normalization is applied to the processed voiceprint feature data; ensure that the normalized data has similar numerical ranges in different feature dimensions to improve the stability and convergence speed of model training.

[0121] Step S182: training a preset deep learning model according to the voiceprint feature training data set, and adjusting the model parameters by using a cross-validation method, thereby obtaining a voiceprint feature model, wherein the deep learning model includes a multi-layer convolutional neural network structure for extracting local features, a long short-term memory network layer for capturing temporal dependencies, an attention mechanism layer for highlighting important features, and a fully connected layer;

[0122] The embodiment of the present invention constructs the network structure of the deep learning model, determines the number of neurons, activation function and connection mode of each layer; defines appropriate loss function (such as cross entropy loss) and optimizer (such as Adam optimizer), sets the correct target for the model training process; uses cross-validation method to adjust the hyperparameters of the model (such as learning rate, batch size, etc.) to obtain the best model performance.

[0123] Step S183: simulating different call conditions to obtain conditional simulation data; applying the voiceprint feature model to the conditional simulation data to obtain conditional response data;

[0124] The embodiment of the present invention creates multiple simulation scenarios based on the voiceprint feature training data set generated in the previous steps; adjusts parameters in the simulation scenarios, such as noise level, network quality, and voice quality, to generate voiceprint feature data under different call conditions; and ensures that the simulation data can reflect the diversity and complexity that may occur in real calls.

[0125] Step S184: Map the voiceprint states and features under different call conditions according to the conditional response data and the conditional simulation data to obtain voiceprint feature graph data.

[0126] The embodiment of the present invention inputs conditional simulation data into a trained deep learning voiceprint feature model; analyzes and maps the voiceprint states and features under different call conditions based on the model output; and generates voiceprint feature graph data, which reflects the changes and responses of voiceprint features under different call conditions.

[0127] The present invention enhances the diversity of training data and the generalization ability of the model by adding noise and time expansion, so that it can better adapt to various actual call environments. Normalization processing reduces data deviation, improves the stability and efficiency of model training, and ensures that features are compared on the same scale. Extract local features of speech to improve the model's recognition ability of subtle speech features; capture the temporal dependency in speech signals to improve the model's understanding of dynamic changes in speech; highlight important features, enhance the model's attention to key speech features, and improve recognition accuracy. By simulating different call conditions, the performance of the model in various actual environments is comprehensively tested to improve the adaptability of the model; obtain model response data under different conditions to help evaluate and optimize the performance of the model and ensure its accuracy and robustness in various environments. Map the voiceprint states and features under different call conditions to generate voiceprint feature map data, provide intuitive visualization information, help understand and analyze speech features; integrate feature data under various conditions to improve the model's recognition ability of complex speech environments and ensure stability and reliability in practical applications. Optimize model parameters to improve the robustness and generalization ability of the model and ensure stable performance under different call conditions. By adding noise and time stretching, the diversity of training data is enhanced, and the model's adaptability to various actual call environments is improved. Multi-layer convolutional neural networks and long short-term memory network layers are used to extract local features and capture temporal dependencies, improving the model's ability to recognize voice features. The attention mechanism layer highlights important features, enhances the model's attention to key voice features, and improves recognition accuracy. Through cross-validation, the model parameters are optimized to improve the robustness and generalization of the model and ensure stable performance under different call conditions. Different call conditions are simulated to comprehensively test and optimize the model's performance to improve its adaptability and accuracy in various environments. Voiceprint feature map data is generated to provide intuitive visual information to help understand and analyze voice features and improve the model's ability to recognize complex voice environments. Overall, these steps improve the accuracy and robustness of voice change recognition through systematic data processing and model optimization, ensure effectiveness and reliability in practical applications, and provide a solid technical guarantee for anti-fraud communications.

[0128] Preferably, step S2 comprises the following steps:

[0129] Step S21: Performing statistical analysis on the characteristic distribution, clustering and abnormal points of the voiceprint feature graph data, so as to obtain voiceprint feature statistical data;

[0130] Step S22: simulating and analyzing the voice change recognition behavior of the voice print feature graph data according to the voice print feature statistical data, thereby obtaining voice change recognition simulation data;

[0131] Step S23: evaluating the recognition performance of the voice change recognition simulation data, and calculating the recognition accuracy, false alarm rate and missed alarm rate under different conditions, so as to obtain recognition performance evaluation data;

[0132] Step S24: performing a performance impact analysis on different environmental parameters according to the recognition performance evaluation data, and establishing a relationship model between environmental parameters and recognition performance, thereby obtaining an environmental impact model;

[0133] Step S25: Set initial recognition thresholds for different call states according to the environmental impact model and recognition performance evaluation data, adjust the thresholds according to real-time environmental parameters, and perform machine learning model training to obtain a dynamic recognition threshold model.

[0134] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 1 Detailed step flow diagram of step S2 in the embodiment of the present invention, step S2 includes the following steps:

[0135] Step S21: Performing statistical analysis on the characteristic distribution, clustering and abnormal points of the voiceprint feature graph data, so as to obtain voiceprint feature statistical data;

[0136] The embodiments of the present invention can understand the distribution of voiceprint features by calculating various statistical indicators of voiceprint feature data (such as mean, variance, skewness, kurtosis, etc.). Statistical methods and functions, such as descriptive statistics, histograms, kernel density estimation, etc., can be used to analyze the distribution characteristics of voiceprint features. Cluster analysis is performed on voiceprint features using clustering algorithms (such as K-means clustering, hierarchical clustering, etc.) to discover clusters and patterns in the data. Samples can be clustered into different groups based on the similarity of voiceprint features, and the feature differences and similarities of each group can be analyzed. Outliers in voiceprint feature data are identified through outlier detection algorithms (such as box plots, Z scores, isolation forests, etc.). Outliers may indicate abnormal conditions in the data collection or processing process, which require special attention and processing.

[0137] Step S22: simulating and analyzing the voice change recognition behavior of the voice print feature graph data according to the voice print feature statistical data, thereby obtaining voice change recognition simulation data;

[0138] The embodiment of the present invention uses a voiceprint feature extraction tool to extract voiceprint features from sound data, such as feature extraction based on MFCC (Mel frequency cepstral coefficient) or other voiceprint feature extraction methods. These features may include spectral features, time domain features, energy features, etc. of the sound. The voiceprint features are simulated by using a simulation analysis tool to simulate voice changes, and different sound changes are simulated by changing parameters such as the frequency, amplitude, and resonance of the sound. Digital signal processing technology and sound synthesis algorithms can be used to achieve voice change simulation. The voiceprint features after voice change simulation are compared and analyzed with the original voiceprint features to obtain simulated data for voice change recognition. Machine learning algorithms and classifiers can be used to classify and identify voiceprint features and evaluate the performance and accuracy of voice change recognition.

[0139] Step S23: evaluating the recognition performance of the voice change recognition simulation data, and calculating the recognition accuracy, false alarm rate and missed alarm rate under different conditions, so as to obtain recognition performance evaluation data;

[0140] The embodiment of the present invention selects appropriate evaluation indicators to measure the performance of voice change recognition, such as accuracy, false positive rate, false negative rate, precision, recall, etc. These indicators can help evaluate the overall performance and recognition ability of the system. The voice change recognition simulation data is divided into a training set and a test set, the model is trained using the training set, and then the performance is evaluated using the test set. Techniques such as cross-validation can be used to reduce the deviation of the evaluation results and improve the reliability of the results. Evaluation indicators such as accuracy, false positive rate and false negative rate are calculated under different conditions based on the test results. These indicators can be calculated using a confusion matrix based on the number of true positives, true negatives, false positives and false negatives.

[0141] Step S24: performing a performance impact analysis on different environmental parameters according to the recognition performance evaluation data, and establishing a relationship model between environmental parameters and recognition performance, thereby obtaining an environmental impact model;

[0142] The embodiment of the present invention collects different environmental parameters (such as noise level, echo condition, voice quality, etc.) and corresponding recognition performance evaluation data. Organize the data for analysis and modeling. Analyze the impact of different environmental parameters on recognition performance through statistical analysis and data visualization methods. The relationship between environmental parameters and performance indicators can be determined by correlation analysis, variance analysis and other methods. Based on the analysis results, a relationship model between environmental parameters and recognition performance is established. Machine learning algorithms such as linear regression, logistic regression, and support vector machines can be used to establish a model, and model training and verification can be performed. Evaluate the performance and accuracy of the established environmental impact model, and optimize and adjust it as needed. Cross-validation and model evaluation indicators (such as R square value, mean square error, etc.) can be used to evaluate the quality of the model.

[0143] Step S25: Set initial recognition thresholds for different call states according to the environmental impact model and recognition performance evaluation data, adjust the thresholds according to real-time environmental parameters, and perform machine learning model training to obtain a dynamic recognition threshold model.

[0144] The embodiment of the present invention organizes the voiceprint features and the voice change recognition simulation data into the format required by the machine learning model, usually taking the feature data as the input variable and the voice change recognition result as the target variable. Ensure that the data set contains enough samples and labels for effective model training. Further process and transform the voiceprint features as needed. This may include operations such as feature scaling, dimensionality reduction, and feature selection to improve the performance and generalization ability of the model. Select a suitable machine learning model based on the requirements of the task and the characteristics of the data. Common voiceprint recognition models include support vector machines (SVM), random forests, deep neural networks, etc. Use the prepared data set to train the selected model. This involves dividing the data set into a training set and a validation set, and iteratively optimizing the parameters and hyperparameters of the model so that the model can best adapt to the training data. Use the test set to evaluate the performance of the trained model. Common evaluation indicators include accuracy, precision, recall, F1 score, etc. The model can be adjusted and optimized based on the evaluation results to improve the performance of the model. Once the requirements are met, the trained model can be saved for subsequent voiceprint recognition tasks.

[0145] The present invention provides a comprehensive distribution of feature data to help understand the performance of voice features under different call conditions and improve the basic data quality of the model; identifies and groups similar voiceprint features to help discover potential voice change patterns and improve the accuracy of voice change recognition; detects and analyzes abnormal feature points to enhance the system's ability to recognize abnormal voice behavior and reduce false positives and false negatives. Voice change recognition simulation is performed based on detailed statistical data to provide a comprehensive understanding of voice change behavior and enhance the accuracy and reliability of voice change recognition by the system. Systematically evaluate the accuracy, false positive rate and false negative rate of voice change recognition, provide a detailed understanding of model performance, and guide model optimization and adjustment; optimize the model based on performance evaluation data to improve the overall performance and stability of recognition. Reveal the impact of different environmental parameters on recognition performance, provide an in-depth understanding of environmental factors, and ensure the adaptability and robustness of the model in various environments; establish a relationship model between environmental parameters and recognition performance to provide a basis for dynamically adjusting recognition parameters and ensure continuous optimization of recognition performance. Set reasonable initial recognition thresholds for different call states to improve the initial accuracy of the recognition system, dynamically adjust the recognition threshold according to real-time environmental parameters, and ensure recognition accuracy and robustness under various call conditions; through continuous machine learning model training, continuously optimize the recognition threshold model to improve the system's adaptability and recognition performance. Overall, these steps significantly improve the accuracy and robustness of voice change recognition through systematic data processing, analysis, and model optimization, ensure effectiveness and reliability in practical applications, and provide solid technical support for anti-fraud communications.

[0146] Preferably, step S3 comprises the following steps:

[0147] Step S31: extracting key control parameters and their influence ranges in the dynamic recognition threshold model, thereby obtaining a control parameter influence matrix;

[0148] The embodiment of the present invention selects a suitable dynamic recognition threshold model as a research object, such as a model based on probability distribution or a classifier model based on machine learning; determines possible key control parameters, such as thresholds, weights, learning rates, etc. Design a set of experiments, change the value of a control parameter in each experiment, and keep other parameters unchanged. Record the recognition accuracy and the value of the control parameter of each experiment. Analyze the influence of each control parameter on the recognition accuracy based on the experimental data to form a control parameter influence matrix. Statistical methods, sensitivity analysis and other techniques can be used to evaluate the influence range of control parameters.

[0149] Step S32: setting initial control parameter values ​​according to a preset target recognition accuracy rate and a control parameter influence matrix, thereby obtaining an initial control parameter set;

[0150] The embodiment of the present invention determines a preset target recognition accuracy rate, and sets a target accuracy rate as a reference according to specific application requirements and performance requirements. Combined with the control parameter influence matrix and the target recognition accuracy rate, the initial control parameter value is inferred according to the preset target accuracy rate. Considering the adjustable range of the model and the actual application scenario, the initial control parameters are appropriately adjusted and optimized to ensure that the expected performance can be achieved in actual operation.

[0151] Step S33: performing a parameter sensitivity test on the initial control parameter set based on the degree of influence of each control parameter on the recognition accuracy, thereby obtaining parameter sensitivity data; using a gradient descent method to perform parameter adaptive adjustment according to the parameter sensitivity data, thereby obtaining control parameter adjustment data;

[0152] The embodiment of the present invention performs a sensitivity test on each parameter in the initial control parameter set to evaluate the influence of each parameter on the recognition accuracy. The test can be performed by fixing the other parameters one by one, changing the value of one parameter, and then observing the change in the recognition accuracy. According to the parameter sensitivity data, the control parameters are adaptively adjusted using methods such as gradient descent so that the recognition accuracy reaches the expected target; optimization algorithms such as gradient descent can calculate the gradient of the parameters according to the parameter sensitivity data and update the parameter values.

[0153] Step S34: designing a multi-level feedback control loop according to the control parameter adjustment data and real-time monitoring of recognition accuracy, processing delay and resource occupancy indicators, thereby obtaining a multi-level feedback control system, wherein the multi-level feedback control loop includes two levels: fast response and long-term optimization;

[0154] In the fast response level, the embodiment of the present invention sets a real-time monitoring system to collect data on indicators such as recognition accuracy, processing delay and resource occupancy. Based on the real-time monitoring data, a control loop is designed to quickly respond to changes in system performance; the control parameter adjustment data and the real-time monitoring data are used to adjust the control parameters using a feedback control algorithm (such as a proportional-integral-differential controller) to keep the recognition accuracy, processing delay and resource occupancy within an acceptable range. In the long-term optimization level, historical data is collected, including control parameter adjustment data, real-time monitoring data and system performance evaluation indicators; based on the historical data, performance evaluation and analysis are performed to understand the long-term performance trend and existing problems of the system, and optimization algorithms (such as genetic algorithms, particle swarm optimization algorithms, etc.) are used in combination with historical data to optimize the control parameters to improve the performance and stability of the system. Based on the optimization results, the control parameters are updated to gradually approach the optimal value, and the performance of the system is continuously monitored to make necessary adjustments and improvements.

[0155] Step S35: running the multi-level feedback control system on the simulation test platform, collecting system response data, and performing control effect analysis to obtain control system performance data;

[0156] The embodiment of the present invention designs an experiment to verify the performance of the multi-level feedback control system. A set of typical input samples can be selected, and indicators such as recognition accuracy, processing delay, and resource occupancy can be evaluated through the actual operation of the system. Experimental data, including indicators such as recognition accuracy, processing delay, and resource occupancy, are collected, and the experimental data are analyzed and evaluated to verify the performance and effectiveness of the multi-level feedback control system; based on the experimental results, necessary adjustments and improvements are made to the system to further improve performance and optimize control strategies.

[0157] Step S36: Encapsulating the multi-level feedback control system into a deployable adaptive recognition model according to the control system performance data, thereby obtaining an adaptive recognition model.

[0158] The embodiment of the present invention analyzes the advantages and disadvantages of the system based on the experimental verification results and performance evaluation, and makes improvements and optimizations; according to the experimental data and performance evaluation, the control parameters in the multi-level feedback control system are adjusted and optimized to further improve the performance and stability of the system. After optimization and adjustment, the performance of the system is re-evaluated, and necessary iterations and improvements are made according to the experimental verification results. Finally, an optimized dynamic recognition threshold model is obtained, which has good recognition performance and stability, and can adaptively adjust the threshold according to real-time changes to adapt to different environments and application requirements.

[0159] The present invention extracts key control parameters and their influence ranges, clarifies key variables in the model, and improves the comprehensibility and controllability of the model; the control parameter influence matrix provides the specific influence relationship of each parameter on the recognition performance, and provides a scientific basis for subsequent parameter optimization and adjustment. According to the target recognition accuracy and the control parameter influence matrix, the initial control parameter value is reasonably set to improve the accuracy and performance of the initial operation of the system. The optimized initial control parameter set reduces the model debugging and adjustment time and improves the development efficiency. Through sensitivity testing, the specific influence of each control parameter on the recognition accuracy is clarified, the optimization direction is provided, and the gradient descent method is used to adjust the parameters to improve the accuracy and efficiency of the control parameter adjustment, and ensure the optimization performance of the recognition system. By designing a multi-level feedback control loop with fast response and long-term optimization, it is ensured that the system can respond quickly to short-term fluctuations, and at the same time achieve long-term optimization, real-time monitoring of recognition accuracy, processing delay and resource occupancy, and improve the overall performance and stability of the system. Through the simulation test platform, the multi-level feedback control system is operated, potential problems are discovered and solved in advance, and the successful deployment of the system is ensured; the system response data is collected, the control effect analysis is performed, and comprehensive performance data is provided to provide a scientific basis for system optimization. The optimized multi-level feedback control system is encapsulated as a deployable adaptive recognition model to facilitate practical application and improve the practicality and deployment efficiency of the model; the adaptive recognition model optimized based on comprehensive performance data has higher stability and reliability, ensuring excellent performance in practical applications. Overall, these steps significantly improve the accuracy, robustness and adaptability of the voice change recognition system through systematic parameter extraction, optimization and multi-level feedback control design, providing an efficient and reliable technical solution for anti-fraud communications.

[0160] Preferably, step S4 comprises the following steps:

[0161] Step S41: acquiring the user's historical call data, and performing desensitization processing on the user's historical call data, thereby obtaining desensitized historical call data;

[0162] The embodiment of the present invention obtains the user's historical call data, which can be obtained from the communication service provider or system log, to ensure the privacy protection of the user's call data, and adopts desensitization processing technology to remove sensitive information, such as using a hash function or encryption algorithm to desensitize the phone number; after desensitization, the desensitized historical call data is saved for subsequent analysis and processing.

[0163] Step S42: performing time series analysis on the desensitized historical call data and extracting user call pattern features, thereby obtaining user call pattern feature data;

[0164] The embodiment of the present invention performs time series analysis on the desensitized historical call data, including statistics and analysis of call time, call duration, call frequency, etc.; based on the results of the time series analysis, the user's call pattern characteristics are extracted, such as the average call duration, the distribution of the number of calls, call activity, etc., and the extracted user call pattern characteristic data is saved for subsequent social network construction and user classification.

[0165] Step S43: constructing a social network graph of users based on call objects and frequencies according to the desensitized historical call data, thereby obtaining social network feature data;

[0166] The embodiment of the present invention constructs a call relationship graph between users based on desensitized historical call data, with each user as a node and call relationships as edges connecting nodes. The weight of the edge is determined based on the call frequency and call object information to indicate the intimacy or communication frequency between users. Based on the constructed call relationship graph, social network feature data is extracted, such as node degree centrality, betweenness centrality and other indicators, for subsequent user classification and personalized identification solutions.

[0167] Step S44: performing user classification based on a clustering algorithm according to the user's call pattern feature data and social network feature data, thereby obtaining user portrait data;

[0168] The embodiment of the present invention combines the user's call pattern feature data and social network feature data to form a user's feature vector, and uses a clustering algorithm (such as K-means, hierarchical clustering, etc.) to cluster the user's feature vector to classify similar users into the same category; based on the clustering results, a user classification label is assigned to each user for subsequent abnormal behavior prediction and personalized identification solutions.

[0169] Step S45: performing abnormal behavior prediction based on the user portrait data and the adaptive recognition model, thereby obtaining abnormal behavior prediction data;

[0170] The embodiment of the present invention establishes an abnormal behavior prediction model based on user portrait data. A machine learning algorithm (such as a support vector machine, a decision tree, etc.) can be used to train the model to predict the user's normal behavior pattern; an adaptive recognition model is used to combine the user's personalized features and historical call data to predict and judge the user's current call behavior to determine whether there is abnormal behavior. Based on the prediction results, abnormal behavior prediction data is obtained to identify whether the user has abnormal behavior.

[0171] Step S46: Optimize the recognition threshold according to the abnormal behavior prediction data through machine learning technology to obtain personalized recognition solution data.

[0172] The embodiments of the present invention use machine learning techniques, such as classification algorithms (such as logistic regression, support vector machine, etc.) or clustering algorithms (such as K-means, DBSCAN, etc.) to process abnormal behavior prediction data; use training data sets and test data sets to train and evaluate the model, and select appropriate evaluation indicators (such as accuracy, recall rate, F1 score, etc.) to evaluate the performance of the model. Adjust the parameters of the classifier or clusterer, and optimize the model through methods such as cross-validation to obtain better performance; select the optimal recognition threshold or clustering threshold based on the evaluation results of the model to divide users into normal behavior and abnormal behavior. Save the personalized recognition scheme data, including the selected optimal threshold and corresponding model parameters, for subsequent abnormal behavior recognition and application.

[0173] The present invention effectively protects the user's personal privacy information through desensitization processing, avoids the risk of sensitive data leakage and infringement of user privacy; ensures the security of historical call data during processing and storage, complies with data protection regulations and standards, and enhances the credibility and transparency of data management. Through time series analysis, extract the user's call pattern characteristics, including call frequency, time period preference, etc., and deeply understand the user's behavior habits and communication behavior rules; based on the call pattern feature data, personalized customization services, such as optimizing communication plans, providing customized recommendations, etc., enhance user experience and satisfaction. Build a user's social network graph, analyze the call objects and their frequencies, and reveal the relationship density and activity of the user in the social circle; through social network feature data, identify and understand the user's influence and key roles in the social network, and provide data support for social interaction. Combined with user portrait data and adaptive recognition models, effectively predict abnormal patterns and abnormal events in user communication behavior, and timely identify and respond to potential risks. Improve anti-fraud capabilities, reduce the potential damage caused by abnormal behavior to user and system security, and protect communication security and user interests. Using machine learning technology to predict abnormal behavior data, dynamically adjust the recognition threshold to achieve a personalized anti-fraud recognition solution; by optimizing the recognition threshold, effectively improve the recognition accuracy and real-time response capability of the anti-fraud system, and reduce the false alarm rate and missed alarm rate. In summary, these steps combine data processing, analysis and prediction technologies, provide multi-dimensional data support and analysis capabilities for anti-fraud communication methods, and effectively improve the security, user experience and operational efficiency of the system.

[0174] Preferably, step S5 comprises the following steps:

[0175] Step S51: Acquire call environment noise data and network quality data;

[0176] The embodiment of the present invention can use professional noise sensors or microphone devices to measure the noise level in the call environment, and the noise data can be expressed as decibel (dB) values ​​or other noise indicators. Network performance monitoring tools or network quality measurement equipment can be used to measure network delay, packet loss rate, bandwidth and other indicators during calls, and these data can be used to evaluate network stability and communication quality.

[0177] Step S52: performing call quality evaluation according to the call environment noise data and the network quality data, thereby obtaining a call quality model, wherein the call quality evaluation includes audio clarity evaluation, continuity evaluation, and echo cancellation effect evaluation;

[0178] The embodiment of the present invention uses the call environment noise data to analyze the signal-to-noise ratio (SNR) of the audio signal. A lower SNR value generally indicates poor audio clarity. Signal processing techniques (such as filtering, noise reduction algorithms, etc.) can be used to enhance and denoise the audio signal to improve audio clarity. The continuity of the call connection is evaluated based on indicators such as delay and packet loss rate in the network quality data. Higher delay or packet loss rate may cause call interruption or discontinuous sound. Network optimization technology (such as congestion control, traffic scheduling, etc.) can be used to improve call continuity. The effect of the echo cancellation algorithm is evaluated by analyzing the echo signal in the call and the sound signal collected by the microphone. Echo cancellation technology can be used to eliminate echo sounds in calls and improve call quality.

[0179] Step S53: Scenario call data are collected in a variety of real scenarios, and transfer learning technology is used to optimize the model generalization capability of the call quality model according to the scenario call data, thereby obtaining call quality optimization data.

[0180] The embodiments of the present invention conduct calls in a variety of real scenarios, covering different environmental noise levels and network quality conditions, and collect call data, including call environmental noise data, network quality data, and call quality evaluation results.

[0181] By using transfer learning technology, the previously trained call quality model is combined with the newly collected scenario call data to optimize the model and improve its generalization ability. You can use methods such as knowledge distillation and domain adaptation in transfer learning to transfer the previous model knowledge to the new scenario and improve the performance of the model in the new scenario. Use the optimized call quality model to predict and evaluate call quality optimization data to obtain more accurate call quality optimization guidance and suggestions.

[0182] By acquiring call environment noise data and network quality data, the system can perceive the noise level and network conditions of the real-time communication environment, which helps to adjust the communication parameters to optimize the call quality; it provides basic data support and provides necessary input information for subsequent call quality evaluation and optimization. Based on the acquired environmental data, the system can accurately evaluate the call quality, including key indicators such as audio clarity, call continuity, and echo cancellation effect; the evaluation results can be used to adjust the communication parameters in real time, optimize the call experience, and ensure the good call quality of users under various environmental conditions. By collecting call data in a variety of real scenarios and applying transfer learning technology, the generalization ability of the call quality model is optimized so that it can adapt to different communication scenarios and environmental conditions; the effectiveness and stability of the call quality optimization strategy in various real scenarios are ensured, and the reliability and practicality of the system in practical applications are improved; by continuously collecting and analyzing scene call data, the call quality model is continuously optimized, the system's adaptability to changes in the communication environment is maintained, and stable call service quality is provided. In summary, these steps integrate data collection, analysis, and model optimization technologies, provide key call quality evaluation and optimization capabilities for anti-fraud communication methods, and thus significantly improve the user experience and operation efficiency of the system.

[0183] Preferably, step S6 comprises the following steps:

[0184] Step S61: performing data cleaning and standardization processing on the personalized identification solution data and the call quality optimization data, so as to obtain personalized identification standard data and call quality standard data;

[0185] The embodiment of the present invention cleans the personalized identification scheme data and call quality optimization data, including processing missing values, abnormal values, and duplicate values, etc., and can use data cleaning tools or data processing libraries (such as pandas) in programming languages ​​(such as Python) to perform data cleaning operations. The cleaned data is standardized and converted into a unified scale and range to eliminate dimensional differences between different data. Commonly used standardization methods include mean-variance standardization and maximum-minimum standardization.

[0186] Step S62: extracting timbre features, emotional features and environmental noise features from the personalized recognition standard data and the call quality standard data, and using the hierarchical analysis method to assign feature weights, thereby obtaining feature weight assignment data;

[0187] The embodiment of the present invention uses audio processing technology to extract timbre features from the personalized recognition standard data and call quality standard data. Common timbre features include pitch, audio spectrum features (such as frequency, energy, spectral centroid, etc.) and time domain and frequency domain features of sound. Emotional features are extracted from the personalized recognition standard data and call quality standard data using emotional analysis technology. Emotional features may include features such as emotional color, speech speed, and intonation of speech. Environmental noise-related features are extracted from the personalized recognition standard data and call quality standard data. Noise analysis algorithms or noise recognition technologies may be used to extract features such as noise level and noise spectrum.

[0188] Step S63: inputting the feature weight distribution data, the personalized identification standard data and the call quality standard data into a preset multi-criteria decision model, thereby obtaining multi-criteria decision data;

[0189] The embodiment of the present invention selects appropriate multi-criteria decision-making methods, such as weighted summation, TOPSIS (Technique for Order Preference by Similarity to Ideal Solution), fuzzy comprehensive evaluation, etc., according to specific needs and problems, and constructs a multi-criteria decision-making model based on the weights of feature weight allocation data, personalized identification standard data, and call quality standard data. The feature weight allocation data, personalized identification standard data, and call quality standard data are input into the multi-criteria decision-making model for calculation and evaluation. According to the calculation rules and weights of the model, the multi-criteria decision data is obtained for subsequent analysis and judgment.

[0190] Step S64: Identify potential abnormal behaviors and fraudulent behaviors on the multi-criteria decision data, and perform risk assessment and threshold adjustment to obtain final anti-fraud identification solution data.

[0191] The embodiment of the present invention uses machine learning technology or a rule engine to identify potential abnormal behaviors and fraudulent behaviors in multi-criteria decision-making data. It can perform supervised learning based on known abnormal behaviors and fraudulent patterns, or use unsupervised learning technology to detect abnormal patterns in the data. Risk assessment is performed based on the identified abnormal behaviors and fraudulent behaviors. The assessment method may include calculating a risk score or probability, and assessing the risk level based on the severity of the abnormal behavior and the possibility of fraud. According to the results of the risk assessment, the threshold used to judge abnormal behaviors and fraudulent behaviors is adjusted. According to the actual situation and needs, the risk of false positives and false negatives is balanced by changing the threshold to achieve the best anti-fraud effect. According to the results after risk assessment and threshold adjustment, the final anti-fraud identification solution data is obtained, including the identification and related information identified as abnormal or fraudulent.

[0192] The present invention ensures the consistency and accuracy of personalized identification scheme data and call quality optimization data through data cleaning and standardization, and improves the reliability of subsequent analysis and decision-making; cleaning and standardization can remove noise and errors in the data, making subsequent data analysis and decision-making more accurate and reliable; processing the data according to a unified standard is conducive to the comparison and integration between different data sources, and improving the operability and application value of the data. By extracting timbre features, emotional features and environmental noise features, the key features of call quality and personalized identification schemes are deeply analyzed to provide strong support for subsequent decision-making; using the hierarchical analysis method to allocate feature weights, it is possible to objectively evaluate the contribution of each feature to the final decision, and improve the scientificity and accuracy of the decision; comprehensively considering and weighing different types of data is conducive to establishing a comprehensive and integrated decision-making model, and improving the decision-making efficiency and accuracy of the system. Through the multi-criteria decision-making model, multiple factors such as feature weights, personalized identification standards and call quality standards are comprehensively considered, and the anti-fraud identification scheme can be more comprehensively evaluated and optimized. Based on the scientific decision-making model, subjective bias can be effectively reduced, and the objectivity and credibility of the decision can be improved; through multi-criteria decision-making, the system can more accurately identify potential abnormalities and fraudulent behaviors, take preventive and countermeasures in time, and ensure communication security and user rights. Identifying abnormal and fraudulent behaviors based on multi-criteria decision data can effectively improve the system's ability to perceive and respond to potential risks, conduct in-depth evaluation and analysis of identified abnormal behaviors, quantify risk levels, and provide a basis for developing targeted preventive measures; dynamically adjust the threshold of the anti-fraud identification scheme based on real-time risk assessment results to ensure the stability and efficient operation of the system in different scenarios. In summary, these steps effectively integrate data analysis, decision models, and risk management technologies, providing comprehensive data support and scientific decision-making basis for anti-fraud communication methods.

[0193] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0194] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A fraud prevention communication method based on voice change recognition, characterized in that: The following steps are involved: Step S1: obtaining multi-dimensional voice feature data of the user, including timbre, frequency features, language rhythm, emotional features, and pause patterns; predicting the possibility of voice change based on the multi-dimensional voice feature data to obtain voiceprint feature prediction data; training a deep learning model based on the voiceprint feature prediction data, and mapping the voiceprint states and features under different call conditions to obtain voiceprint feature graph data; Step S2: simulating and analyzing the voice change recognition behavior under different call states according to the voiceprint feature graph data, thereby obtaining voice change recognition simulation data; establishing a dynamic recognition threshold model according to the environmental parameters corresponding to the different call states and the voice change recognition simulation data; Step S3: adjusting the control parameters in real time according to the preset target recognition accuracy and the dynamic recognition threshold model, thereby obtaining control parameter adjustment data; constructing a multi-level feedback control system according to the control parameter adjustment data, thereby obtaining an adaptive recognition model; Step S4: Obtain the user's historical call data, perform personalized analysis on the user's historical call data, and perform feature extraction to obtain user portrait data; Abnormal behavior prediction is performed based on user portrait data and adaptive recognition models, and the optimal recognition threshold is obtained through machine learning technology to obtain personalized recognition solution data; Step S5: Acquire call environment noise data and network quality data, and construct a call quality model according to the call environment noise data and the network quality data; Optimize call quality models across scenarios and adjust communication parameters in real time to obtain call quality optimization data; Step S6: Perform multi-criteria decision fusion analysis on the personalized identification solution data and call quality optimization data to obtain the final anti-fraud identification solution data.

2. The anti-fraud communication method based on voice change recognition according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: Acquire the user's historical voice data and perform short-time Fourier transform on the historical voice data to obtain voice time-frequency data; Step S12: extracting Mel-frequency cepstral coefficients from the speech time-frequency data to obtain timbre feature data; extracting frequency features based on fundamental frequency contour from the speech time-frequency data to obtain frequency feature data; Step S13: using a voice activity detection algorithm to perform voice and non-voice segment recognition on the voice time-frequency data, thereby obtaining language rhythm feature data; Step S14: analyzing the duration of speech segments and the pause patterns of non-speech segments on the speech time-frequency data according to the language rhythm feature data, thereby obtaining pause pattern data; Step S15: performing pitch change, energy distribution and speech speed analysis on the speech time-frequency data, and extracting emotional features, thereby obtaining emotional feature data; performing harmonic-to-noise ratio analysis on the speech time-frequency data, and evaluating the harmonic structure of the speech, thereby obtaining harmonic structure data; Step S16: merging the timbre feature data, the frequency feature data, the language rhythm feature data, the pause pattern data and the emotion feature data into multi-dimensional speech feature data according to the harmonic structure data; Step S17: using a support vector machine to classify the multi-dimensional speech feature data and perform a voice change possibility score, thereby obtaining voiceprint feature prediction data; Step S18: Perform deep learning model training based on the voiceprint feature prediction data, and map the voiceprint states and features under different call conditions to obtain voiceprint feature graph data.

3. The anti-fraud communication method based on voice change recognition according to claim 2, characterized in that: Step S18 includes the following steps: Step S181: adding noise and time stretching to the voiceprint feature prediction data, and performing normalization processing to obtain a voiceprint feature training data set; Step S182: training a preset deep learning model according to the voiceprint feature training data set, and adjusting the model parameters by using a cross-validation method, thereby obtaining a voiceprint feature model, wherein the deep learning model includes a multi-layer convolutional neural network structure for extracting local features, a long short-term memory network layer for capturing temporal dependencies, an attention mechanism layer for highlighting important features, and a fully connected layer; Step S183: simulating different call conditions to obtain conditional simulation data; applying the voiceprint feature model to the conditional simulation data to obtain conditional response data; Step S184: Map the voiceprint states and features under different call conditions according to the conditional response data and the conditional simulation data to obtain voiceprint feature graph data.

4. The anti-fraud communication method based on voice change recognition according to claim 3 is characterized in that: Step S2 includes the following steps: Step S21: Performing statistical analysis on the characteristic distribution, clustering and abnormal points of the voiceprint feature graph data, so as to obtain voiceprint feature statistical data; Step S22: simulating and analyzing the voice change recognition behavior of the voice print feature graph data according to the voice print feature statistical data, thereby obtaining voice change recognition simulation data; Step S23: evaluating the recognition performance of the voice change recognition simulation data, and calculating the recognition accuracy, false alarm rate and missed alarm rate under different conditions, so as to obtain recognition performance evaluation data; Step S24: performing a performance impact analysis on different environmental parameters according to the recognition performance evaluation data, and establishing a relationship model between environmental parameters and recognition performance, thereby obtaining an environmental impact model; Step S25: Set initial recognition thresholds for different call states according to the environmental impact model and recognition performance evaluation data, adjust the thresholds according to real-time environmental parameters, and perform machine learning model training to obtain a dynamic recognition threshold model.

5. The anti-fraud communication method based on voice change recognition according to claim 4, characterized in that: Step S3 includes the following steps: Step S31: extracting key control parameters and their influence ranges in the dynamic recognition threshold model, thereby obtaining a control parameter influence matrix; Step S32: setting initial control parameter values ​​according to a preset target recognition accuracy rate and a control parameter influence matrix, thereby obtaining an initial control parameter set; Step S33: performing a parameter sensitivity test on the initial control parameter set based on the degree of influence of each control parameter on the recognition accuracy, thereby obtaining parameter sensitivity data; using a gradient descent method to perform parameter adaptive adjustment according to the parameter sensitivity data, thereby obtaining control parameter adjustment data; Step S34: designing a multi-level feedback control loop according to the control parameter adjustment data and real-time monitoring of recognition accuracy, processing delay and resource occupancy indicators, thereby obtaining a multi-level feedback control system, wherein the multi-level feedback control loop includes two levels: fast response and long-term optimization; Step S35: running the multi-level feedback control system on the simulation test platform, collecting system response data, and performing control effect analysis to obtain control system performance data; Step S36: Encapsulating the multi-level feedback control system into a deployable adaptive recognition model according to the control system performance data, thereby obtaining an adaptive recognition model.

6. The anti-fraud communication method based on voice change recognition according to claim 5, characterized in that: Step S4 includes the following steps: Step S41: acquiring the user's historical call data, and performing desensitization processing on the user's historical call data, thereby obtaining desensitized historical call data; Step S42: performing time series analysis on the desensitized historical call data and extracting user call pattern features, thereby obtaining user call pattern feature data; Step S43: constructing a social network graph of users based on call objects and frequencies according to the desensitized historical call data, thereby obtaining social network feature data; Step S44: performing user classification based on a clustering algorithm according to the user's call pattern feature data and social network feature data, thereby obtaining user portrait data; Step S45: performing abnormal behavior prediction based on the user portrait data and the adaptive recognition model, thereby obtaining abnormal behavior prediction data; Step S46: Optimize the recognition threshold according to the abnormal behavior prediction data through machine learning technology to obtain personalized recognition solution data.

7. The anti-fraud communication method based on voice change recognition according to claim 6, characterized in that: Step S5 includes the following steps: Step S51: Acquire call environment noise data and network quality data; Step S52: performing call quality evaluation according to the call environment noise data and the network quality data, thereby obtaining a call quality model, wherein the call quality evaluation includes audio clarity evaluation, continuity evaluation, and echo cancellation effect evaluation; Step S53: Scenario call data are collected in a variety of real scenarios, and transfer learning technology is used to optimize the model generalization capability of the call quality model according to the scenario call data, thereby obtaining call quality optimization data.

8. The anti-fraud communication method based on voice change recognition according to claim 7, characterized in that: Step S6 includes the following steps: Step S61: performing data cleaning and standardization processing on the personalized identification solution data and the call quality optimization data, so as to obtain personalized identification standard data and call quality standard data; Step S62: extracting timbre features, emotional features and environmental noise features from the personalized recognition standard data and the call quality standard data, and using the hierarchical analysis method to assign feature weights, thereby obtaining feature weight assignment data; Step S63: inputting the feature weight distribution data, the personalized identification standard data and the call quality standard data into a preset multi-criteria decision model, thereby obtaining multi-criteria decision data; Step S64: Identify potential abnormal behaviors and fraudulent behaviors on the multi-criteria decision data, and perform risk assessment and threshold adjustment to obtain final anti-fraud identification solution data.

Images