A smart voice recognition and analysis system based on a general-purpose smartphone chip

Abstract

This application discloses an intelligent speech recognition and analysis system based on a general-purpose smartphone chip. The system includes: a data acquisition module, which synchronously captures speech signals and motion sensor data through a built-in microphone array and a motion sensor, generating a raw multimodal data packet with a time stamp; and a noise reduction processing module, which receives the raw multimodal data packet, performs environmental noise spectrum analysis, generates an inverted sound wave, enhances the signal, and outputs a clean speech stream. This invention resolves the contradiction between speech recognition accuracy and privacy security in complex environments through multimodal noise separation, nonlinear signal enhancement, and layered privacy protection. Simultaneously, by leveraging dynamic resource scheduling and a lightweight model, it maximizes the utilization of the limited computing power of the mobile phone chip in speech recognition.

Description

Technical Field

[0001] This application relates to the field of speech recognition, and more particularly to an intelligent speech recognition and analysis system based on a general-purpose smartphone chip. Background Technology

[0002] In recent years, with the popularization of smart terminal devices, voice interaction has gradually become one of the core methods of human-computer interaction, and is widely used in smart homes, in-vehicle systems, mobile healthcare, and other fields. However, existing voice recognition systems face two major technical bottlenecks in complex real-world scenarios:

[0003] 1. Traditional noise reduction algorithms struggle to cope with dynamically changing noise environments (such as public places and vehicles), failing to effectively separate human voices from background noise, resulting in a significant decrease in speech recognition accuracy;

[0004] 2. Voice data contains biometric features (voiceprints) and sensitive semantics (medical and financial information). Existing systems lack end-to-end privacy protection mechanisms in the data collection, transmission, and processing stages, posing a risk of data leakage.

[0005] These two major technical bottlenecks present a technical contradiction in their solutions: improving noise reduction performance requires the collection of multimodal data (such as motion sensor information), but multi-source data fusion may expand the attack surface and increase the risk of privacy leakage. At the same time, the limited computing power of mobile phone chips makes it difficult to simultaneously support high-precision semantic parsing, multimodal emotion modeling, and privacy computing. Summary of the Invention

[0006] To address the aforementioned problems, embodiments of the present invention provide an intelligent speech recognition and analysis system based on a general-purpose smartphone chip. The system includes a data acquisition module, a noise reduction processing module, a semantic parsing module, an emotion modeling module, a computation scheduling module, and a privacy control module.

[0007] The data acquisition module synchronously captures voice signals and motion sensor data through a built-in microphone array and motion sensor, generating a raw multimodal data packet with a time stamp.

[0008] The noise reduction processing module receives the original multimodal data packet and performs environmental noise spectrum analysis, generates an anti-phase sound wave and enhances the signal, and outputs a clean speech stream.

[0009] The semantic parsing module performs real-time speech-to-text conversion and intent extraction on the clean speech stream, generates structured semantic vectors, and outputs parsing results with confidence scores.

[0010] An emotion modeling module receives parsing results and motion sensor data, and constructs a multi-dimensional emotion state matrix that includes acoustic features and body language.

[0011] The computing scheduling module dynamically allocates chip computing resources based on the current speech processing stage and system operating status parameters, and establishes a low-latency processing pipeline from the data acquisition module to the output of the emotion modeling module.

[0012] A privacy control module that performs encryption and desensitization processing during data flow.

[0013] Furthermore, the environmental noise spectrum analysis and anti-phase sound wave generation method includes:

[0014] A physical acoustic path transfer function is constructed using mobile phone vibration sensor data to distinguish between the device's own vibration noise and environmental noise components.

[0015] Nonnegative matrix decomposition is used to separate steady-state noise components from transient noise components in the time-frequency domain;

[0016] For steady-state noise components, the dominant frequency components are extracted by Fourier transform to generate a set of equal-amplitude sine waves with phase reversal; for transient noise components, an adaptive finite impulse response filter is used to predict the noise waveform and generate an anti-phase cancellation signal in advance.

[0017] By integrating device orientation data captured by the mobile phone gyroscope, the beamforming parameters of the multi-microphone array are dynamically adjusted to optimize the spatial directivity of the anti-phase acoustic waves;

[0018] Establish a noise scene classifier and identify typical environments based on a convolutional neural network;

[0019] When a new noise pattern is detected, the federated learning mechanism is triggered to exchange noise reduction parameters with the cloud.

[0020] Personalized noise reduction strategies are stored locally in a secure enclave, generating customized anti-phase acoustic wave templates based on the user's usual environment.

[0021] Furthermore, the structured semantic vector generation method includes:

[0022] S101: Extracting basic semantic features from speech streams using an improved LSTM network;

[0023] S102: Construct a dynamic context association matrix to map the current semantics to the historical dialogue scenario in vector space;

[0024] S103: Inject domain knowledge weighting factors to adjust the direction of semantic parsing;

[0025] S104: Output a structured semantic vector with confidence assessment values.

[0026] Furthermore, methods for adjusting domain knowledge weighting factors include constructing a lightweight industry knowledge graph, setting up a cross-language alignment layer, and deploying a privacy filtering gateway;

[0027] Building a lightweight industry knowledge graph includes using graph neural networks to dynamically update the degree of term association;

[0028] Setting up a cross-language alignment layer includes vector space mapping of Chinese and English technical terms;

[0029] Deploying a privacy filtering gateway includes real-time blocking of sensitive semantic patterns based on reinforcement learning strategies.

[0030] Furthermore, methods for signal enhancement include using nonlinear regression algorithms to eliminate multipath reflection interference. The implementation process of using nonlinear regression algorithms to eliminate multipath reflection interference includes:

[0031] Assume the multipath channel response is a nonlinear function. Assuming the received signal is the original signal, then ;

[0032] ;

[0033] In the formula, The original signal, For the first The attenuation coefficient of the path, For the first The delay of the path, Total number of paths;

[0034] The received signal is subjected to a short-time Fourier transform to extract the time-frequency domain feature matrix. If a microphone array is used, a spatial feature vector is constructed by combining the gyroscope data.

[0035] Neural networks are used to capture the nonlinear relationships of signals in the time and frequency domains. Kernel methods are employed to map multipath components to a high-dimensional space using kernel functions. Volterra series are used to explicitly model the memory effect of the nonlinear system. An objective function for the nonlinear regression model is established, and the objective function is... ,in, It is a nonlinear regression model. For model parameters, The signal actually observed at the receiving end. This is the length of the time window, which is a preset value.

[0036] Predicting multipath interference using a nonlinear regression model The enhanced signal is .

[0037] Furthermore, the method for dynamically allocating chip computing resources includes:

[0038] A task priority queue is established based on the speech processing stage and system operating status parameters. When the battery level is detected to be lower than the priority threshold, the semantic parsing precision is reduced from 32-bit floating point to 8-bit integer calculation, the body language analysis branch in the emotion modeling flow is closed, and the differential privacy mechanism is enabled to perform lightweight updates to the local speech model.

[0039] Furthermore, the encryption and desensitization processing method includes:

[0040] Implement voiceprint feature fragmentation encryption at edge computing nodes, and disperse and store voiceprint biometric templates in secure chip enclaves and trusted execution environments;

[0041] The desensitized semantic vector is transmitted to the cloud through a homomorphic encryption channel, and the voiceprint obfuscation device is automatically activated in public scenarios to inject background noise orthogonal to the user's voice fundamental frequency.

[0042] Furthermore, the privacy filtering gateway triggers sensitive commands when it detects sensitive semantics related to medical and financial matters. These sensitive commands include:

[0043] The local cache clearing command blocks related data from entering the computation scheduling flow;

[0044] The dialogue process continues by generating alternative generalized statements using an adversarial network.

[0045] Record event fingerprints in the security log for use in compliance audits.

[0046] The technical effects and advantages of the intelligent speech recognition and analysis system provided by this invention are as follows:

[0047] This invention resolves the conflict between speech recognition accuracy and privacy security in complex environments by using multimodal noise separation, nonlinear signal enhancement, and layered privacy protection; at the same time, it achieves high-efficiency real-time processing on mobile devices by leveraging dynamic resource scheduling and lightweight models. This invention integrates device vibration data with microphone arrays to distinguish between device vibration noise and environmental noise. It generates anti-phase sound waves to cancel out different noise types (steady-state and transient), thus accurately capturing user speech even in noisy environments (such as subways and cafes), improving speech recognition success rates and enhancing user experience. By combining speech content, historical dialogue scenarios, and industry knowledge graphs, it dynamically adjusts the semantic parsing direction to support continuous dialogue in complex scenarios (such as medical consultations and device control), improving the naturalness of interaction. Through encrypted storage of voiceprint features, real-time filtering of sensitive semantics, and encrypted transmission after data anonymization, it ensures user privacy and security while achieving intelligent interaction, meeting compliance requirements (such as GDPR). It dynamically adjusts calculation precision based on task priority (e.g., high-precision mode → power-saving mode), disabling non-core functions to prioritize basic voice functions in low-power scenarios, avoiding interaction interruptions. When detecting new noise patterns, it updates the noise reduction strategy through federated learning and stores personalized configurations locally. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the connection of an intelligent voice recognition and analysis system based on a general-purpose smartphone chip;

[0049] Figure 2 A flowchart illustrating the method for environmental noise spectrum analysis and antiphase acoustic wave generation;

[0050] Figure 3 Flowchart of the method for generating structured semantic vectors. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Please see Figure 1 As shown in this embodiment, an intelligent voice recognition and analysis system based on a general-purpose smartphone chip is described. The system includes:

[0053] The data acquisition module synchronously captures voice signals and motion sensor data through a built-in microphone array and motion sensor, generating raw multimodal data packets with time stamps;

[0054] The noise reduction processing module receives the original multimodal data packets and performs environmental noise spectrum analysis, generates anti-phase sound waves and enhances the signal, and outputs a clean speech stream;

[0055] The semantic parsing module performs real-time speech-to-text conversion and intent extraction on the clean speech stream, generates structured semantic vectors, and outputs parsing results with confidence scores.

[0056] The emotion modeling module receives the analysis results and motion sensor data, and constructs a multi-dimensional emotion state matrix that includes acoustic features and body language.

[0057] The computing scheduling module dynamically allocates chip computing resources based on the current speech processing stage and system operating status parameters, and establishes a low-latency processing pipeline from the data acquisition module to the output of the emotion modeling module.

[0058] The privacy control module implements encryption and desensitization processing during data flow.

[0059] like Figure 2 As shown, the environmental noise spectrum analysis and anti-phase sound wave generation methods include:

[0060] By using mobile phone vibration sensor data to construct a physical acoustic path transfer function, the device's own vibration noise component can be distinguished from the environmental noise component.

[0061] Mobile phone vibration sensors are typically triaxial MEMS accelerometers that continuously monitor the vibration acceleration of the device body in the X / Y / Z directions. When the device is in operation, vibration noise sources include, but are not limited to, the operation of the haptic feedback motor, the vibration of the processor cooling fan, and the micro-movements of the user's hand when holding the device. The physical acoustic path transfer function characterizes the attenuation characteristics of vibration energy transmitted from the inside of the device to the microphone, such as the complete path modeling of motor vibration → device shell → air conduction → microphone diaphragm. When the mobile phone vibration sensor data is collected, the vibration signal can be converted into corresponding data using Fourier transform. At the same time, the device's own vibration noise has a significant cross-correlation peak (e.g., >0.7) with the accelerometer signal in the time domain. The vibration noise energy is concentrated in the effective frequency band of the transfer function (10-500Hz). Therefore, the distinction is based on the time-domain synchronization of vibration noise with the acceleration signal, while environmental noise exhibits spatial diffusion characteristics.

[0062] Non-negative matrix factorization (NMF) is used to separate steady-state noise components and transient noise components in the time-frequency domain. NMF can decompose the spectrum of mixed noise into a steady-state basis and a transient coding matrix. Steady-state noise exists stably for a long time (such as the low-frequency hum of an air conditioner) and is represented as a continuous energy band in the basis matrix. Transient noise is sudden short-term noise (such as the sound of a door closing) and exhibits pulse-like activation in the coding matrix. Therefore, they can be separated in the time-frequency domain.

[0063] For steady-state noise components, the dominant frequency components are extracted by Fourier transform to generate a set of equal-amplitude sine waves with phase reversal.

[0064] The formulas for generating phase reversal include:

[0065] ;

[0066] In the formula, The input signal vector contains current and historical input samples. This refers to the prediction error (the difference between the actual noise and the predicted value). Forgetting factor (controlling the degree of influence of historical data). It is the inverse correlation matrix. It is the transpose (row vector) of the input signal vector. For the filter at time The coefficient vector represents the filter parameter configuration at the current moment, and is used to perform weighted summation on the input signal to generate the output signal.

[0067] For transient noise components, an adaptive finite impulse response (FIR) filter is used to predict the noise waveform and generate an anti-phase cancellation signal in advance. By analyzing the statistical characteristics of the noise signal (such as the autocorrelation function) in real time, the coefficients of the FIR filter are adjusted to predict the transient noise waveform at future moments (such as impact sounds and keyboard taps), and an anti-phase signal with the same amplitude but opposite phase is generated in advance. This cancels the noise in the sound field superposition, solving the problem of delayed response to sudden noise in traditional noise reduction systems and achieving millisecond-level cancellation.

[0068] By integrating device orientation data captured by the phone's gyroscope, the beamforming parameters of the multi-microphone array are dynamically adjusted to optimize the spatial directivity of the anti-phase sound waves. By utilizing the phone's gyroscope to monitor the device's pitch or yaw angle in real time, the beamforming weight matrix of the microphone array is dynamically corrected, ensuring that the main lobe direction of the noise-canceling anti-phase sound waves is always aligned with the physical location of the noise source (such as changes in noise direction caused by device rotation during a call). This improves the spatial directivity accuracy of the noise cancellation system in mobile scenarios and avoids noise cancellation failure caused by changes in device posture.

[0069] A noise scene classifier is established to identify typical environments (office, street, transportation, etc.) based on a convolutional neural network. The noise signal collected by the microphone is converted into a Mel spectrogram and input into a convolutional neural network (CNN) to extract time-frequency features. The probability distribution of typical scenes (such as identifying "subway" and "coffee shop") is output through Softmax, which provides a basis for subsequent noise reduction parameter selection, thereby realizing environment-adaptive noise reduction mode switching (such as focusing on low-frequency noise reduction in transportation scenes and optimizing human voice preservation in office scenes).

[0070] When a new noise pattern is detected, the federated learning mechanism is triggered to exchange noise reduction parameters with the cloud. When a local device detects an unrecorded noise pattern (such as high-frequency noise from new appliances), only the key parameter gradients of the noise reduction model are encrypted and uploaded to the cloud. The cloud aggregates data from multiple devices to update the global model and then securely distributes the new parameters to each terminal. This allows for the continuous evolution of noise reduction capabilities while protecting user privacy, avoiding the risk of data leakage associated with traditional centralized training.

[0071] Personalized noise reduction strategies are stored locally in a secure enclave, generating customized inverse sound wave templates based on the user's usual environment. A noise feature library of the user's usual environment (such as the background noise spectrum of home address and office location) is built in the mobile phone security chip (such as TEE), generating a customized inverse waveform template library, which is quickly matched and loaded when used, thereby reducing the misjudgment of personalized scenarios by general noise reduction algorithms (such as the preservation of specific frequency bands of pet barking at home), and improving the subjective experience of noise reduction effect.

[0072] like Figure 3 As shown, the methods for generating structured semantic vectors include:

[0073] S101: Extracting basic semantic features from clean speech streams using an improved LSTM network;

[0074] The improved LSTM network is a bidirectional gated enhanced LSTM network. It adds a context gate to the traditional input gate, forget gate and output gate to dynamically filter noisy words (such as interjections and repeated words). At the same time, it embeds phoneme sequences and part-of-speech tagging information to solve homophonic ambiguity problems (such as "apple" as fruit vs. brand). In this way, it obtains a primary semantic vector group containing grammatical structure and temporal dependence, which is the basic semantic feature.

[0075] S102: Construct a dynamic context association matrix to map the current semantics to the historical dialogue scenario in vector space;

[0076] For example:

[0077] The semantic vectors of the most recent 10 rounds of dialogue (such as the mathematical representations of keywords like "coffee", "meeting room", and "appointment") are cached in memory;

[0078] Compare the current dialogue vector with the historical vector using cosine similarity;

[0079] (For example: the current question "What time does it start?" is highly similar to the historical question "The meeting starts at 3 pm").

[0080] Check if metadata such as time, people, and location are continuous;

[0081] (For example: the conversation has been about "meeting room reservations" for three consecutive rounds.)

[0082] Fill the similarity values ​​into the matrix, where rows represent the current semantics and columns represent the historical semantics;

[0083] Example matrix:

[0084] History 1 (Meeting) | History 2 (Coffee) | History 3 (Weather)

[0085] ------|-----------|-----------|----------

[0086] Current semantics | 0.9 | 0.2 | 0.1;

[0087] Retain the strongest correlations (such as values ​​> 0.7), and set the rest to zero to avoid interference;

[0088] Each new dialogue triggers a recalculation of the matrix; old dialogues that have not been associated for more than 5 rounds are automatically removed from the matrix.

[0089] S103: Inject domain knowledge weighting factors to adjust the direction of semantic parsing;

[0090] S104: Output a structured semantic vector with confidence assessment values.

[0091] Methods for adjusting domain knowledge weight factors include constructing a lightweight industry knowledge graph, setting up a cross-language alignment layer, and deploying a privacy filtering gateway;

[0092] Building a lightweight industry knowledge graph involves dynamically updating terminology associations using graph neural networks (GNNs); compressing the scale of the industry knowledge graph using knowledge distillation techniques, retaining only high-frequency terminology nodes on mobile devices (such as the "symptom-drug" association in the medical field); and dynamically updating the association weights between nodes through the message passing mechanism of GNNs. For example, when a user repeatedly mentions "headache" and "ibuprofen," the GNN automatically strengthens the edge weights between these two nodes. Simultaneously, combined with incremental learning algorithms, when a new term is detected (such as a new drug name), a local graph structure expansion is triggered instead of a global reconstruction, ensuring knowledge evolution within chip memory limitations. This lightweight industry knowledge graph enables the semantic parsing module to automatically associate statements like "long-term sleep deprivation leads to migraines" with the medical knowledge node "recommendation to consult a neurologist," improving the accuracy of intent recognition.

[0093] Setting up a cross-language alignment layer involves mapping Chinese and English technical terms into a vector space. The cross-language alignment layer uses a dual-tower architecture design to build a parallel encoder that maps Chinese and English terms to a unified semantic space. It optimizes vector distance through contrastive learning. For example, the cosine similarity between "diabetes (Chinese)" and "diabetes (English)" is >0.9. The corresponding word vector matrix is ​​automatically loaded according to the phone's system language settings. When users use a mix of Chinese and English (such as "This MRI report shows..."), semantic coherence is still maintained. The INT8 quantization instruction of the phone's NPU is used to accelerate vector similarity calculation, which significantly reduces the response time of cross-language retrieval.

[0094] Deploying a privacy filtering gateway involves real-time blocking of sensitive semantic patterns based on reinforcement learning strategies. The state-action mechanism of the privacy filtering gateway defines the state space as a semantic vector sensitivity score, taking patient medical care as an example (such as including ID number, medical history, etc.). The action space includes blocking, replacement, and allowing. The decision model is trained in a local enclave using the Q-learning algorithm, and the risk of the scene is judged by combining motion sensor data. For example, in public places (by detecting walking status through a gyroscope), the filtering intensity is automatically increased, voiceprint replacement is performed on voices such as "My medical record number is...", and adversarial examples (such as deliberately blurring the expression of sensitive information) are injected to enhance the robustness of the model and prevent "cerebral infarction history" from being rewritten as "cerebrovascular accident history" and still being identified.

[0095] Through a three-layer collaboration of "knowledge enhancement, cross-language understanding, and privacy protection," accurate and secure semantic parsing is achieved with limited computing power.

[0096] For example:

[0097] When a user says, "My recent HbA1c test result is high":

[0098] The knowledge graph identifies "HbA1c" as a key indicator of diabetes, thereby increasing the parsing weight of related terms such as "blood glucose monitoring" and "insulin".

[0099] The cross-language layer ensures that the system correctly understands "HbA1c" or "glycated hemoglobin", regardless of whether it is described in Chinese or English later.

[0100] The privacy gateway detects that the "HbA1c value" is sensitive health data. Based on the current GPS location (within the hospital), it reduces the filtering intensity and allows the raw data to be uploaded to the authorized medical cloud. In other situations, it replaces the data with a vague description.

[0101] Methods for signal enhancement include nonlinear regression algorithms to eliminate multipath reflection interference. In voice signals or wireless communications, signals travel from the transmitter to the receiver (such as a mobile phone microphone) through different paths (direct, reflected, diffracted), resulting in different arrival times for signals along different paths, causing signal superposition and inter-symbol interference. Differences in attenuation of different frequency components lead to signal distortion, and the material properties of reflective surfaces (such as walls) may introduce nonlinear phase distortion. Traditional linear filtering (such as the LMS algorithm) is difficult to model complex nonlinear multipath effects and cannot enhance the signal, while nonlinear regression algorithms can achieve signal enhancement.

[0102] The implementation process of eliminating multipath reflection interference using a nonlinear regression algorithm includes:

[0103] Assume the multipath channel response is a nonlinear function. Assuming the received signal is the original signal, then ;

[0104] ;

[0105] In the formula, The original signal, For the first The attenuation coefficient of the path, For the first The delay of the path, Total number of paths;

[0106] The received signal is subjected to a short-time Fourier transform (STFT) to extract the time-frequency domain feature matrix. If a microphone array is used, the spatial feature vector is constructed by combining the gyroscope data to capture the signal arrival direction information.

[0107] Neural networks (such as 1D-CNN+LSTM) are used to capture the nonlinear relationships of signals in the time and frequency domains. Kernel methods (such as Support Vector Regression and SVR) are used to map to a high-dimensional space through kernel functions to separate multipath components. Volterra series is used to explicitly model the memory effect of the nonlinear system. An objective function for the nonlinear regression model is established, where the objective function is... ,in It is a nonlinear regression model. For model parameters, The signal actually observed at the receiving end (including multipath interference and noise). The time window length (number of signal frames) is a preset value.

[0108] Predicting multipath interference using a nonlinear regression model The enhanced signal is .

[0109] By using a nonlinear regression algorithm, the system can effectively separate the nonlinear distortion component in multipath reflection signals, significantly improving voice quality and communication reliability. The flexibility of this method makes it suitable for complex dynamic environments, such as smart homes and in-vehicle voice systems.

[0110] The methods of encryption and de-identification include:

[0111] Voiceprint feature fragmentation encryption is implemented at edge computing nodes, distributing voiceprint biometric templates across chip-secure enclaves and trusted execution environments. User voiceprint biometric templates (such as MFCC coefficients and fundamental frequency trajectories) are divided into multiple parts and stored in different secure areas to ensure that a single point of leakage cannot restore the complete voiceprint. The secure enclave is a hardware-secure area isolated within the chip (such as ARM TrustZone) used to store the main fragments and perform encryption operations. The trusted execution environment is a secure computing environment (such as Intel SGX) running outside the enclave, handling some of the computational tasks of the fragmented data.

[0112] The desensitized semantic vector is transmitted to the cloud through a homomorphic encryption channel, and the voiceprint obfuscation device is automatically activated in public scenarios to inject background noise orthogonal to the user's voice fundamental frequency.

[0113] By combining layered encryption (fragmented storage + homomorphic transmission) with proactive defense (voiceprint obfuscation), comprehensive protection of user privacy is achieved in complex environments.

[0114] Methods for dynamically allocating chip computing resources include:

[0115] A task priority queue is established based on the speech processing stage and system operating status parameters. When the battery level is detected to be lower than the priority threshold, the semantic parsing precision is reduced from 32-bit floating point to 8-bit integer calculation, the body language analysis branch in the emotion modeling flow is closed, and the differential privacy mechanism is enabled to perform lightweight updates to the local speech model.

[0116] Through a dynamic resource allocation strategy, the system achieves triple optimization of performance, power consumption, and privacy when power is limited, ensuring the continuous availability of core voice interaction functions while meeting privacy protection requirements.

[0117] When the privacy filtering gateway detects sensitive semantics related to healthcare and finance:

[0118] Trigger a local cache clearing command to block related data from entering the computation scheduling flow;

[0119] The dialogue continues by generating alternative generalized statements through a generative adversarial network; based on the Transformer-based Seq2Seq model, the input is the original sensitive statement, and the output is a generalized statement. For example, if the input is "My bank card number is 6214-8600-1234-5678", the output is "Do you need to check your account service?".

[0120] Event fingerprints are recorded in the security log for compliance auditing. The system automatically generates audit reports according to regulatory requirements (such as GDPR and HIPAA). Key fields include the occurrence rate of sensitive events (times / day), data erasure validity verification, and the accuracy of alternative statement generation. At the same time, the PIA process is executed regularly to assess whether there is a risk of indirect identifier leakage in the logs.

[0121] By employing a three-tiered protection system of real-time blocking, intelligent replacement, and audit trail, the risk of data breaches will be reduced when sensitive medical and financial information is detected, while simultaneously meeting stringent compliance requirements.

[0122] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0123] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present application, based on the technical solution and concept of the present application, should be covered within the scope of protection of the present application.

Claims

1. A smart voice recognition and analysis system based on a general-purpose smartphone chip, characterized in that, The system includes: The data acquisition module synchronously captures voice signals and motion sensor data through a built-in microphone array and motion sensor, generating a raw multimodal data packet with a time stamp. The noise reduction processing module receives the original multimodal data packet and performs environmental noise spectrum analysis, generates an anti-phase sound wave and enhances the signal, and outputs a clean speech stream. The semantic parsing module performs real-time speech-to-text conversion and intent extraction on the clean speech stream, generates structured semantic vectors, and outputs parsing results with confidence scores. An emotion modeling module receives parsing results and motion sensor data, and constructs a multi-dimensional emotion state matrix that includes acoustic features and body language. The computing scheduling module dynamically allocates chip computing resources based on the current speech processing stage and system operating status parameters, and establishes a low-latency processing pipeline from the data acquisition module to the output of the emotion modeling module. A privacy control module that performs encryption and desensitization processing during data flow.

2. The intelligent voice recognition and analysis system based on a general-purpose smartphone chip according to claim 1, characterized in that, The environmental noise spectrum analysis and anti-phase sound wave generation method includes: A physical acoustic path transfer function is constructed using mobile phone vibration sensor data to distinguish between the device's own vibration noise and environmental noise components. Nonnegative matrix decomposition is used to separate steady-state noise components from transient noise components in the time-frequency domain; For steady-state noise components, the dominant frequency components are extracted by Fourier transform to generate a set of equal-amplitude sine waves with phase reversal; for transient noise components, an adaptive finite impulse response filter is used to predict the noise waveform and generate an anti-phase cancellation signal in advance. By integrating device orientation data captured by the mobile phone gyroscope, the beamforming parameters of the multi-microphone array are dynamically adjusted to optimize the spatial directivity of the anti-phase acoustic waves; Establish a noise scene classifier and identify typical environments based on a convolutional neural network; When a new noise pattern is detected, the federated learning mechanism is triggered to exchange noise reduction parameters with the cloud. Personalized noise reduction strategies are stored locally in a secure enclave, generating customized anti-phase acoustic wave templates based on the user's usual environment.

3. The intelligent voice recognition and analysis system based on a general-purpose smartphone chip according to claim 1, characterized in that, The structured semantic vector generation method includes: S101: Extracting basic semantic features from speech streams using an improved LSTM network; S102: Construct a dynamic context association matrix to map the current semantics to the historical dialogue scenario in vector space; S103: Inject domain knowledge weighting factors to adjust the direction of semantic parsing; S104: Output a structured semantic vector with confidence assessment values.

4. The intelligent voice recognition and analysis system based on a general-purpose smartphone chip according to claim 3, characterized in that, Methods for adjusting domain knowledge weight factors include constructing a lightweight industry knowledge graph, setting up a cross-language alignment layer, and deploying a privacy filtering gateway; Building a lightweight industry knowledge graph includes using graph neural networks to dynamically update the degree of term association; Setting up a cross-language alignment layer includes vector space mapping of Chinese and English technical terms; Deploying a privacy filtering gateway includes real-time blocking of sensitive semantic patterns based on reinforcement learning strategies.

5. The intelligent voice recognition and analysis system based on a general-purpose smartphone chip according to claim 1, characterized in that, Methods for signal enhancement include using nonlinear regression algorithms to eliminate multipath reflection interference. The implementation process of using nonlinear regression algorithms to eliminate multipath reflection interference includes: Assume the multipath channel response is a nonlinear function. Assuming the received signal is the original signal, then ; ； In the formula, The original signal, For the first The attenuation coefficient of the path, For the first The delay of the path, Total number of paths; The received signal is subjected to a short-time Fourier transform to extract the time-frequency domain feature matrix. If a microphone array is used, a spatial feature vector is constructed by combining the gyroscope data. Neural networks are used to capture the nonlinear relationships of signals in the time and frequency domains. Kernel methods are employed to map multipath components to a high-dimensional space using kernel functions. Volterra series are used to explicitly model the memory effect of the nonlinear system. An objective function for the nonlinear regression model is established, and the objective function is... ,in, It is a nonlinear regression model. For model parameters, The signal actually observed at the receiving end. This is the length of the time window, which is a preset value. Predicting multipath interference using a nonlinear regression model The enhanced signal is .

6. The intelligent voice recognition and analysis system based on a general-purpose smartphone chip according to claim 1, characterized in that, The method for dynamically allocating chip computing resources includes: A task priority queue is established based on the speech processing stage and system operating status parameters. When the battery level is detected to be lower than the priority threshold, the semantic parsing precision is reduced from 32-bit floating point to 8-bit integer calculation, the body language analysis branch in the emotion modeling flow is closed, and the differential privacy mechanism is enabled to perform lightweight updates to the local speech model.

7. The intelligent voice recognition and analysis system based on a general-purpose smartphone chip according to claim 1, characterized in that, The encryption and desensitization processing methods include: Implement voiceprint feature fragmentation encryption at edge computing nodes, and disperse and store voiceprint biometric templates in secure chip enclaves and trusted execution environments; The desensitized semantic vector is transmitted to the cloud through a homomorphic encryption channel, and the voiceprint obfuscation device is automatically activated in public scenarios to inject background noise orthogonal to the user's voice fundamental frequency.

8. The intelligent voice recognition and analysis system based on a general-purpose smartphone chip according to claim 4, characterized in that, The privacy filtering gateway triggers a sensitive command when it detects sensitive semantics related to medical and financial matters. The sensitive command includes: The local cache clearing command blocks related data from entering the computation scheduling flow; The dialogue process continues by generating alternative generalized statements using an adversarial network. Record event fingerprints in the security log for use in compliance audits.

Images