Multi-person shared health detection device capable of accurately recognizing user identity based on sound

Abstract

The invention discloses a multi-person shared health detection device capable of accurately recognizing user identity based on sound. The device comprises a microphone array, a physiological parameter sensor, a processor and a memory. The method comprises the following steps of: performing front-end processing on a collected voice signal to obtain an effective voice segment; extracting a voiceprint embedding vector by using a deep neural network; comparing the vector with a local voiceprint template library to obtain a highest score and candidate identities; carrying out living body and playback attack authenticity discrimination by analyzing multi-dimensional acoustic features and utilizing a classification model; and binding and associating the identity label with the health data collected by the sensor when the identity label is judged to be a real living body and the similarity score meets a judgment condition. The method has the advantages of high noise resistance, attack resistance and end-side low delay, supports offline work and model increment self-adaption, is suitable for families, communities, medical institutions and other scenes, and effectively improves the use convenience and data security of shared health equipment.

Description

A multi-user shared health monitoring device that accurately identifies users based on their voice. Technical Field

[0001] This invention belongs to the interdisciplinary field of intelligent health monitoring technology and biometric recognition technology, specifically involving a multi-user shared health detection device based on accurate voice identification of user identity. Background Technology

[0002] In home and public health monitoring devices (such as scales, blood pressure monitors, and pulse oximeters), data attribution currently relies heavily on manual user selection, QR code scanning, Bluetooth binding, or account login. This process is cumbersome and prone to data mismatch. Shared devices in medical institutions often use card swiping, QR codes, or facial recognition for identity verification, but these methods are susceptible to environmental interference such as light obstruction and mask wearing, affecting recognition stability. Meanwhile, although mobile voice assistants possess basic speaker recognition and wake-up capabilities, their reliability significantly decreases in complex scenarios involving multiple users, environmental noise, or playback attacks. Traditional voiceprint recognition solutions face similar challenges, exhibiting high false recognition rates in far-field conditions, when multiple users speak simultaneously, or in the presence of background noise from appliances or televisions, failing to meet the low latency and high accuracy requirements of shared health monitoring devices. Overall, existing devices generally lack a binding mechanism that tightly couples user measurement behavior with their identity verification, resulting in significant deficiencies in the accuracy and robustness of the data attribution process.

[0003] In the process of developing this invention, the inventors discovered that existing technologies have at least the following shortcomings: First, the user experience is noticeably fragmented, requiring additional user selection or login before measurement, a cumbersome process that is easily overlooked and affects ease of use. Second, in home or public multi-user scenarios, identity recognition is unstable, with a high false recognition rate, easily leading to others' data being incorrectly attributed to a particular account, severely impacting the accuracy of subsequent health analysis. Regarding security, existing solutions lack effective liveness detection and adversarial sample protection mechanisms, making them vulnerable to attacks such as audio playback. Furthermore, the system has poor adaptability, high new user registration costs, and the acoustic model struggles to adaptively track the natural drift of user voices over time, affecting long-term recognition performance. Finally, existing solutions have limitations in terms of versatility and scalability, lacking a unified framework to support the effective fusion and binding of multi-sensor data and identity information, hindering the migration and large-scale application of the technology across different devices. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a multi-user shared health detection device based on accurate voice identification of user identity, in order to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, in a first aspect, a multi-user shared health detection device based on accurate voice identification is provided. The device includes: a microphone array for acquiring user voice signals or raw audio signals; one or more physiological parameter sensors for acquiring user health data; one or more processors; and a memory. When a computer program stored in the memory is executed by the processor, the processor performs the following operations: preprocesses the raw audio signals acquired by the microphone array through a voice front-end processing module to obtain effective voice segments; and extracts and embeds voiceprint features using a deep neural network model from the acquired audio signals. The process involves extracting a voiceprint embedding vector from the valid speech segment; comparing the voiceprint embedding vector with a locally stored authorized user voiceprint template library using an identity matching and decision fusion module, and obtaining the highest similarity score and its corresponding candidate identity; extracting multi-dimensional acoustic features using a liveness and replay attack detection module and using a classification model to determine the authenticity of the valid speech segment; and binding the candidate identity with health data collected by the physiological parameter sensor using a multi-person session management and binding module when the liveness and replay attack detection module determines that the candidate is a real live person and the highest similarity score meets predefined judgment conditions.

[0006] Secondly, a method for multi-user shared health detection based on accurate voice identification is provided. The method is executed by a processor and includes the following steps: receiving user voice signals collected by a microphone array; performing front-end preprocessing on the voice signals to obtain valid voice segments; extracting voiceprint embedding vectors from the valid voice segments using a deep neural network model; comparing the voiceprint embedding vectors with a locally stored authorized user voiceprint template library to obtain the highest similarity score and its corresponding candidate identity; extracting multi-dimensional acoustic features and using a classification model to determine the authenticity of the valid voice segments; when the authenticity determination result indicates a real living person and the highest similarity score meets predefined judgment conditions, binding and associating the candidate identity with health data collected by a physiological parameter sensor.

[0007] Thirdly, an electronic device is provided, comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement a multi-user shared health detection method based on accurate voice identification of user identity as described in the second aspect.

[0008] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described in the second aspect.

[0009] Fifthly, a computer program product is provided, comprising a computer-readable storage medium storing a computer program, wherein when the computer program is run, it executes the multi-user shared health detection method based on accurate voice identification described in the second aspect.

[0010] The above technical solution has the following beneficial technical effects: This technical solution accurately extracts voiceprint features through a deep neural network model and combines multi-dimensional acoustic feature analysis for liveness detection, effectively realizing seamless user identification and automatic and accurate attribution of health data in multi-user shared scenarios; its edge processing architecture ensures low-latency response and user privacy security, while the dynamic decision mechanism and anti-attack capability greatly improve the robustness of identification and system security in complex acoustic environments, ultimately providing an efficient, reliable and secure "one person, one test, one binding" intelligent solution for shared health detection scenarios such as families and communities. Attached Figure Description

[0011] Figure 1 is a schematic diagram of the logical structure of a multi-user shared health detection device based on accurate voice identification according to an embodiment of the present invention; Figure 2 is a flowchart of the voice front-end processing module according to an embodiment of the present invention; Figure 3 is a flowchart of the voiceprint feature extraction and embedding module according to an embodiment of the present invention; Figure 4 is a flowchart of the liveness detection and replay attack detection module according to an embodiment of the present invention; Figure 5 is a flowchart of the identity matching and decision fusion module according to an embodiment of the present invention (first flowchart); Figure 6 is a flowchart of the identity matching and decision fusion module according to an embodiment of the present invention (second flowchart); Figure 7 is a flowchart of the data storage and model update module according to an embodiment of the present invention; Figure 8 is a schematic diagram of the structure of the computer system according to an embodiment of the present invention. Detailed Implementation

[0012] The exemplary embodiments of the present invention are described below with reference to the accompanying drawings. Those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] This invention discloses a health monitoring device for multi-user shared scenarios. It employs a microphone array to collect voice data, performs front-end noise reduction and liveness detection, extracts and embeds voiceprints, and matches them against an encrypted database on the device side to output a high-confidence identity result. Through session management, identity is strongly bound to monitoring tasks such as weight, blood pressure, and blood oxygen saturation, enabling automatic data attribution. The device utilizes a dynamic threshold and historical inertia fusion to improve recognition accuracy in complex noisy environments and features replay attack detection and low-latency edge-side capabilities, supporting offline operation and incremental adaptive operation. This solution enhances the usability and security of shared devices and is suitable for scenarios such as homes, communities, and medical institutions.

[0014] The purpose of this invention is to provide a health monitoring device and method for multi-user shared scenarios. Through high-precision voiceprint recognition, it automatically achieves an integrated process of "identity binding and data attribution," improving the accuracy of data attribution and ease of use. This solution possesses strong noise resistance and anti-attack capabilities, maintaining stable recognition in complex acoustic environments such as homes or medical institutions. It also integrates liveness detection and playback attack identification mechanisms to enhance system security. The system has a fast response speed, supports localized processing, with the main recognition process completed on the device side. It can return identity recognition results within 300 to 800 ms and has offline working capabilities, ensuring user privacy and real-time usability. Furthermore, the system supports incremental learning and hot model updates, adaptively tracking changes in user voiceprint over time. It also has good scalability, providing a unified framework to integrate data from multiple sensors such as weight, heart rate, blood pressure, and blood oxygen, enabling collaborative work between multimodal identity binding and health monitoring.

[0015] Figure 1 is a schematic diagram of the logical structure of a multi-user shared health detection device based on accurate voice identification according to an embodiment of the present invention. As shown in Figure 1, the device includes: a microphone array for acquiring the user's raw speech signal or raw audio signal; one or more physiological parameter sensors for acquiring the user's health data; one or more processors; a memory; and a computer program stored in the memory. When the computer program is executed by the one or more processors, the processors perform the following operations: preprocess the raw audio signal acquired by the microphone array through a voice front-end processing module to obtain a valid speech segment; extract a voiceprint embedding vector from the valid speech segment using a deep neural network model through a voiceprint feature extraction and embedding module; compare the voiceprint embedding vector with a locally stored authorized user voiceprint template library through an identity matching and decision fusion module to obtain the highest similarity score and its corresponding candidate identity; extract multi-dimensional acoustic features through a liveness and replay attack detection module and use a classification model to determine the authenticity of the valid speech segment; and bind and associate the candidate identity with the health data acquired by the physiological parameter sensors through a multi-person session management and binding module when the liveness and replay attack detection module determines that the user is a real liveness and the highest similarity score meets predefined judgment conditions.

[0016] This embodiment discloses a multi-user shared health monitoring device based on accurate voice identification of users. It is suitable for multi-user shared health monitoring scenarios such as homes, community health service centers, and medical institution physical examination departments. It can achieve automatic identification of multiple users and accurate attribution of health data, balancing ease of use, identification accuracy, and data security. The following example illustrates this: The microphone array adopts a 4-unit linear array structure. The microphone units use MEMS high-fidelity pickups with a sampling rate of 16kHz, a bit depth of 16bit, and an array spacing of 5cm. It receives user voice signals through spatial diversity, covering an effective acquisition range of 1-3 meters, adapting to the user's natural speaking distance, and providing a spatial domain signal foundation for subsequent noise reduction processing.

[0017] The physiological parameter sensors include a weight sensor (range 0-200kg, accuracy ±0.1kg), an electronic blood pressure sensor (measurement range systolic blood pressure 60-200mmHg, diastolic blood pressure 40-130mmHg, accuracy ±3mmHg), a blood oxygen sensor (measurement range 70%-100%SpO2, accuracy ±2%), and a heart rate sensor (measurement range 30-240 beats / minute, accuracy ±1 beat / minute). Each sensor is connected via I... 2 The C interface communicates with the processor, supporting synchronous or asynchronous acquisition of health data to meet the parameter acquisition needs of different detection scenarios.

[0018] The device is equipped with a quad-core processor (1.5GHz) and 2GB LPDDR4 memory. The memory is divided into 128MB dedicated storage area for storing authorized user voiceprint template library, trained neural network model and encrypted health data. The remaining storage space is used for running system programs and temporary data caching to ensure low latency and data security of edge processing.

[0019] When a user triggers the health check process, the computer program is executed by the processor. First, the raw audio signal collected by the microphone array is preprocessed by the voice front-end processing module: a pre-emphasis filter (coefficient 0.97) is used to compensate for high-frequency attenuation of the voice signal, and then frame processing is performed with a frame length of 20ms and a frame shift of 10ms. Hanning windows are superimposed between frames to reduce spectral leakage. Then, a dual-threshold endpoint detection algorithm based on energy and zero-crossing rate is used to remove invalid signal segments such as environmental noise and silence. Finally, adaptive beamforming combined with spectral subtraction is used for noise reduction to suppress background noise and echo interference, and finally outputs an effective voice segment with a signal-to-noise ratio ≥15dB, providing a high-quality signal foundation for subsequent voiceprint recognition and anti-spoofing detection.

[0020] When extracting voiceprint embedding vectors through the voiceprint feature extraction and embedding module, a pre-trained ECAPA-TDNN deep neural network model is used. The input of this model is a pre-processed effective speech segment. After processing through a 10-layer time-delay neural network and a self-attention mechanism layer, a 256-dimensional high-discrimination voiceprint embedding vector is extracted. Optionally, a PLDA model can be introduced to post-process the voiceprint embedding vector to further improve the inter-class discrimination and intra-class aggregation of voiceprint features of different users, and adapt to voiceprint variation scenarios in complex environments.

[0021] The specific execution process of the identity matching and decision fusion module is as follows: First, the cosine similarity score between the embedded vector of the voiceprint to be identified and each template in the local authorized user voiceprint template library is calculated, and the highest similarity score and its corresponding candidate identity are selected. Based on the preset basic similarity threshold (0.7), combined with the consistency ratio (number of successful matches / total number of recognitions) in the last 5 historical recognition records of the candidate identity, the dynamic decision threshold is calculated by the formula "dynamic decision threshold = basic similarity threshold × (1 + 0.1 × consistency ratio)". The higher the consistency ratio, the more lenient the dynamic decision threshold, which is adapted to the voiceprint stability of long-term users. The highest similarity score is compared with the dynamic decision threshold. If the highest similarity score exceeds the dynamic decision threshold, the identity recognition is directly determined to be successful. If the highest similarity score is lower than the dynamic decision threshold, a secondary verification process is triggered. Specifically, the device prompts the user to repeat a specific command (such as confirming a health check) and re-enter the voice, or a candidate identity list pops up on the device touch screen for the user to manually select, reducing the risk of misjudgment.

[0022] The liveness detection and playback attack detection module achieves authenticity discrimination through the collaboration of an anti-spoofing feature extraction unit and a classification model. The anti-spoofing feature extraction unit extracts multi-dimensional anti-spoofing features from valid speech segments, including 40-dimensional constant Q-cepstral coefficients (CQCC), residual signal statistical features (mean, variance, skewness, kurtosis) based on a 12th-order linear predictive coding (LPC) model, frame-level group delay characteristics of the speech signal (mean and standard deviation of group delay per frame), and nonlinear energy distribution features based on the Teager energy operator (instantaneous energy peak, valley, and rate of change). The classification model uses a lightweight MobileNet convolutional neural network, which is trained on the publicly available anti-spoofing dataset ASVspoof2019. The input is a concatenated vector of the above multi-dimensional anti-spoofing features (128 dimensions), and the output is a binary discrimination result of real liveness or spoofing attacks, where spoofing attacks include recording playback, synthesized speech, and other attack types.

[0023] The multi-user session management and binding module is responsible for the precise association between identity and health data. When the liveness detection and replay attack detection module outputs a "real liveness" result, and the identity matching and decision fusion module determines that the identity recognition is successful (including confirmation after secondary verification), this module establishes a temporary session for the current user and assigns a unique session ID. It receives health data collected by physiological parameter sensors in real time, associates and binds the health data with candidate identity, session ID, and collection timestamp, and generates structured data records (data format follows JSON specifications). The bound structured data is encrypted and stored in an encrypted archive in local storage (using the AES-256 encryption algorithm), while the user's recent historical recognition records are updated to provide data support for subsequent dynamic decision threshold calculation. If there is data collected by multiple physiological parameter sensors in the same session, the data is aligned and bound according to the collection timestamp to ensure that all health data of a single user in a single detection is unique and to avoid cross-user data mismatch.

[0024] In this embodiment, the device supports offline operation mode. Authorized user voiceprint template library, neural network model, and encrypted health data are all stored locally, enabling identity recognition and data binding without relying on the network. Through model quantization and edge computing optimization, the response time of the entire identity recognition process (including preprocessing, feature extraction, matching, and anti-spoofing detection) is controlled within 300-800ms, meeting real-time usage requirements. It also supports incremental learning and hot model updates. When a user's voiceprint features change naturally over time, 5-10 valid voice samples can be actively submitted to trigger model fine-tuning, adaptively updating the user's voiceprint template to ensure long-term recognition stability. Furthermore, the device reserves I... 2 The C and SPI expansion interfaces are compatible with newly added physiological parameter acquisition modules such as blood glucose sensors and body temperature sensors. Through unified data format adaptation, multimodal health monitoring and identity binding can be achieved in a collaborative manner.

[0025] This multi-user shared health monitoring device, which accurately identifies users based on their voice, collects voice signals through a microphone array and preprocesses and purifies effective voice segments. It then uses a deep neural network model to extract highly recognizable voiceprint embedding vectors, which are accurately compared with an authorized user voiceprint template library to achieve identity matching. Combined with a multi-dimensional acoustic feature classification-based liveness and playback attack detection mechanism, it mitigates security risks. Finally, when a real liveness is confirmed and the identity matching meets the standards, the candidate identity is bound and associated with the health data collected by physiological parameter sensors. This achieves accurate identification and secure verification of user identity in multi-user shared scenarios, while ensuring the unique correspondence between health data and user identity, effectively avoiding data mismatch and impersonation. It balances ease of use, accuracy of identity recognition, and data security, significantly improving the practical value and reliability of multi-user shared health monitoring scenarios.

[0026] As shown in Figure 2, in Embodiment 2, the voice front-end processing module is used to perform sequence processing on the raw audio signal acquired by the microphone array. The sequence processing includes acoustic echo cancellation, adaptive noise reduction, automatic gain control, beamforming, and voice endpoint detection to extract effective voice segments suitable for voiceprint recognition and liveness detection from the raw audio signal. The output of the voice front-end processing module is simultaneously connected to the input of both the voiceprint feature extraction and embedding module and the liveness and playback attack detection module, providing the extracted effective voice segments in parallel to both modules to shorten the overall processing time of identity recognition through parallel processing. Furthermore, during the user registration phase, the voice front-end processing module performs the sequence processing on multiple acquired registration voice segments to obtain multiple registration voice segments.

[0027] As shown in Figure 3, in some embodiments, the voiceprint feature extraction and embedding module may include: an acoustic feature conversion unit, used to convert the effective speech segments output by the speech front-end processing module into acoustic feature maps, wherein the acoustic feature maps are log-Mel spectrograms or Mel frequency cepstral coefficients; a deep feature extraction unit, used to input the acoustic feature maps into a pre-trained deep neural network model to map the effective speech segments of different lengths into fixed-dimensional voiceprint embedding vectors; wherein the deep neural network model is a time-delay neural network ECAPA-TDNN model or an x-vector model that emphasizes channel attention, propagation, and aggregation; the deep feature extraction unit is also used to extract multiple voiceprint embedding vectors from multiple registered speech segments of the same user during the user registration stage; and a voiceprint template generation unit, used to perform cluster analysis on the multiple voiceprint embedding vectors to obtain the set of voiceprint embedding vectors corresponding to the largest cluster, and generate an initial user voiceprint template by calculating the average value or centroid of all voiceprint embedding vectors in the set of voiceprint embedding vectors corresponding to the largest cluster.

[0028] As shown in Figure 4, in some embodiments, the liveness and playback attack detection module and the voiceprint feature extraction and embedding module are set in parallel to process the same valid speech segment output by the speech front-end processing module simultaneously. The liveness and playback attack detection module includes: an anti-spoofing feature extraction unit, used to extract anti-spoofing features from the valid speech segment, the anti-spoofing features including any multiple of the following: constant Q cepstral coefficients, residual signal statistical features based on a linear predictive coding model, group delay characteristics of the speech signal, and nonlinear energy distribution features based on the Teager energy operator; a spectrum discrimination information extraction unit, used to obtain spectrum discrimination information by analyzing the short-time spectrum of the valid speech segment, the spectrum discrimination information including at least high-frequency energy attenuation characteristics obtained through frequency band analysis, envelope distortion characteristics obtained through spectrum distortion detection, and background noise statistical characteristics obtained through signal-to-noise ratio analysis; and a classification and discrimination unit, connected to the anti-spoofing feature extraction unit and the spectrum discrimination information extraction unit, used to input the anti-spoofing features and the spectrum discrimination information into a classification model, and output a discrimination result or confidence score indicating the authenticity of the speech sample.

[0029] As shown in Figure 5, in some embodiments, the identity matching and decision fusion module is configured to perform the following operations: calculate the similarity score between the voiceprint embedding vector to be identified and the pre-stored user voiceprint template, and determine the highest similarity score and its corresponding candidate identity; compare the similarity score with a preset high threshold and a preset low threshold, wherein the low threshold is lower than the high threshold, and a fuzzy decision interval is defined between the high threshold and the low threshold; when the highest similarity score is higher than the high threshold, directly adopt the corresponding candidate identity as the identification result; when the highest similarity score falls into the fuzzy decision interval, start the historical decision inertial filtering process; wherein, the historical decision inertial filtering process includes: calculating the consistency ratio of the candidate identity in the recent preset number of successful identification records, and comparing the consistency ratio with a preset inertial threshold; if the consistency ratio exceeds the inertial threshold, adopt the candidate identity; otherwise, reject the current identification result and trigger a secondary confirmation process, wherein the secondary confirmation process includes prompting the user to re-enter the voice or start an alternative identity verification method.

[0030] In other embodiments, the preset high threshold is replaced by a dynamic decision threshold, which is calculated as follows: based on the preset basic similarity threshold and combined with the consistency ratio of the current candidate identity in recent historical identification records, the dynamic decision threshold is calculated.

[0031] As shown in Figure 6, in some embodiments, the identity matching and decision fusion module is configured to perform the following operations: calculate the similarity score between the voiceprint embedding vector to be identified and the pre-stored user voiceprint template, and determine the highest similarity score and its corresponding candidate identity; calculate a dynamic decision threshold based on a preset basic similarity threshold and combined with the consistency ratio of the current candidate identity in recent historical identification records; compare the highest similarity score with the dynamic decision threshold; if the highest similarity score exceeds the dynamic decision threshold, the identity recognition is determined to be successful; if the highest similarity score is lower than or equal to the dynamic decision threshold, a secondary verification process is triggered, the secondary verification process including prompting the user to re-enter the voice command or providing a manual identity selection interface.

[0032] In some embodiments, the dynamic decision threshold T dynamic Calculated using the following formula: T dynamic =T base ×1+α (R 0.5); where T base R is the basic similarity threshold, R is the consistency ratio of the current candidate identity in the most recent N successful identification records, and α is a preset adjustment coefficient used to control the intensity of historical influence.

[0033] The adjustment coefficient α is determined in one of the following ways: In method one, the adjustment coefficient α is a preset fixed value, ranging from 0.05 to 0.3; In method two, the adjustment coefficient α is configured according to the application scenario of the device, wherein a first α value is configured when used for medical diagnosis or financial payment scenarios, and a second α value is configured when used for home health monitoring scenarios, and the first α value is greater than the second α value; In method three, the adjustment coefficient α is dynamically adjusted based on the number of authorized users, and the more authorized users there are, the smaller the value of α becomes.

[0034] In some embodiments, the calculation of the similarity score between the voiceprint embedding vector to be identified and the pre-stored user voiceprint template employs a weighted similarity algorithm based on multi-scale feature fusion, specifically including: performing multi-scale feature decomposition on the voiceprint embedding vector and the user voiceprint template to obtain corresponding global features and local detail features respectively; calculating a first similarity between the global features of the voiceprint embedding vector to be identified and the global features of the user voiceprint template as the global similarity; and calculating a second similarity between the local detail features of the voiceprint embedding vector to be identified and the local detail features of the user voiceprint template as the local similarity; dynamically adjusting the weight ratio of the global features and the local detail features according to the current environmental noise level, wherein the higher the current environmental noise level, the greater the weight ratio of the local detail features; and linearly fusing the weighted global similarity and the local similarity to obtain the final similarity score.

[0035] This technical solution effectively improves the accuracy of voiceprint recognition through multi-scale feature decomposition and fusion; enhances the system's anti-interference ability and robustness by dynamically adjusting feature weights based on environmental noise; and finally, it fuses global and local similarities in a linear weighted manner, making the identity recognition results more reliable, and is particularly suitable for health detection equipment in complex noise environments.

[0036] Furthermore, the multi-scale feature decomposition of the voiceprint embedding vector and the user voiceprint template to obtain corresponding global features and local detail features specifically includes: taking the voiceprint embedding vector and the user voiceprint template as inputs, extracting their first K principal components using a principal component analysis algorithm to form global features representing the overall acoustic characteristics; K is a positive integer; dividing the voiceprint embedding vector and the user voiceprint template into several sub-vector segments, calculating the statistical features of each sub-vector segment to form local detail features representing acoustic detail changes; wherein, the global features and local detail features together constitute a multi-scale feature representation.

[0037] This technical solution extracts global features through principal component analysis to capture overall acoustic characteristics, and combines them with local detail features obtained by sub-vector segmentation statistics to form a complementary multi-scale voiceprint representation. This feature decomposition method can effectively improve the robustness of the voiceprint recognition system in noisy environments, while enhancing the sensitivity to changes in sound details, thereby achieving a recognition accuracy and adaptability that surpasses single feature methods in overall performance.

[0038] Further, the consistency ratio of the current candidate identity in recent historical identification records is determined through the following steps: obtaining the occurrence records of the current candidate identity in the most recent preset number of successful identification events from the historical records; assigning a time-decay-based weight factor to each occurrence record, wherein the weight factor value corresponding to the occurrence record closer to the current time is larger; calculating the sum of all weight factors corresponding to the candidate identity as the actual weight sum; calculating the theoretical maximum weight sum, which is the sum of the maximum weight factors that can be obtained assuming that the candidate identity appears in all the most recent preset number of successful identification events; dividing the actual weight sum by the theoretical maximum weight sum, and using the quotient as the consistency ratio.

[0039] This technical solution calculates the consistency ratio by introducing a time decay weight factor, enabling the system to more accurately reflect the user's recent real usage habits and effectively improve the timeliness and accuracy of identity verification decisions. Compared with traditional simple statistical methods, its dynamic weighting mechanism can effectively reduce the risk of misjudgment caused by accidental matching or interference from historical data, thereby achieving more reliable voiceprint identity authentication in complex application scenarios.

[0040] In some embodiments, the multi-person session management and binding module includes: a speaker log unit, used to process the collected audio data using a speaker separation algorithm to separate the speech segments corresponding to each speaker from the mixed speech; a measurement behavior perception unit, used to detect valid measurement behaviors triggered by user operations; and a binding logic unit, connected to the speaker log unit and the measurement behavior perception unit, configured to: in response to detecting a valid measurement behavior, acquire audio data output by the voice front-end processing module within a predetermined time period before and after the occurrence of the valid measurement behavior; control the speaker log unit to process the audio data to obtain at least one separated speech segment; sequentially provide each separated speech segment to the voiceprint feature extraction and embedding module and the identity matching and decision fusion module for processing to obtain corresponding user identity information; and bind and associate the successfully identified user identity information with the physiological data collected by the physiological parameter sensor in the current valid measurement behavior.

[0041] Specifically, the physiological parameter sensors may include at least one of the following: a photoacoustic blood flow spectrum sensor, used to simultaneously measure hemoglobin concentration and blood flow velocity non-invasively by detecting the photoacoustic effect signal generated by laser irradiation of skin tissue in a specified band; a millimeter-wave radar vital sign monitoring sensor, used to obtain heart rate and respiratory rate non-contactly by resolving the micro-movements on the body surface caused by heartbeat and respiration from the micro-Doppler effect by transmitting and receiving frequency-modulated continuous waves; a bioelectrical impedance spectrum analysis sensor, used to analyze body composition parameters, including body fat percentage, muscle mass, and total body water, by applying multi-frequency weak currents to the human body and measuring the impedance response; and a skin conductance and skin temperature coupling sensor, used to assess the excitability of the autonomic nervous system or the vasomotor state by measuring changes in skin conductivity and surface temperature fluctuations.

[0042] As shown in Figure 7, in some embodiments, the device further includes a data storage and model update module, which includes: a secure storage unit for securely storing the user's voiceprint template, voiceprint embedding vector, and associated personal configuration locally using encryption; and a template update unit for incrementally updating the user's voiceprint template using an exponential moving average algorithm after each successful authentication, wherein the updated voiceprint template vector v updated Calculate according to the following formula corresponding to the exponential moving average algorithm: v updated =β v old +(1 β) v new ; where v old v is the currently stored voiceprint template vector. new The voiceprint embedding vector successfully verified in this study is β, which is a preset smoothing factor. The model update unit is used to fine-tune the deep neural network model used by the voiceprint feature extraction and embedding module using the locally accumulated anonymized voiceprint embedding vector after obtaining user authorization. The cloud synchronization unit is used to securely synchronize the anonymized voiceprint features or the updated deep neural network model parameters to the cloud after obtaining user authorization, so as to realize cross-device roaming.

[0043] The associated personal configurations include at least one of the following: user's health data preference settings, physiological parameter alarm thresholds, device display and voice prompt preferences, and commonly used measurement task types.

[0044] The smoothing factor β is a preset fixed value, ranging from 0.9 to 0.99; or, the smoothing factor β is an adaptive value dynamically determined based on the user's voiceprint stability index and / or the ambient noise level index of the device's usage environment; the value of the smoothing factor β is a monotonically increasing function of the voiceprint stability index, and the value of the smoothing factor β is negatively correlated with the ambient noise level index. Specifically, when the voiceprint stability is high or the ambient noise level is low, a relatively large β value is assigned; when the voiceprint stability is low or the ambient noise level is high, a relatively small β value is assigned.

[0045] The user voiceprint stability index is obtained by calculating the average cosine distance or average Euclidean distance between multiple voiceprint embedding vectors generated when the current user successfully verifies the system the most recent preset number of times; wherein, the smaller the value of the average cosine distance or average Euclidean distance, the higher the voiceprint stability; the device uses the environmental noise level index to obtain the signal-to-noise ratio estimated when the voice front-end processing module performs adaptive noise reduction processing.

[0046] Furthermore, the model update unit is configured to fine-tune in the following manner: using the pre-trained deep neural network model adopted by the voiceprint feature extraction and embedding module as the base model; triggering the fine-tuning process when the device is idle and connected to an external power source; the fine-tuning process includes: updating the parameters of the last fully connected layer and / or the last few hidden layers of the base model using a dataset composed of locally stored anonymized voiceprint embedding vectors that have been decoupled from any user's personal identity information; the update uses a mini-batch gradient descent algorithm and applies regularization constraints to prevent forgetting of the generalized features learned by the base model on general corpora; after the fine-tuning is completed, the updated model parameters are hot-reloaded to the voiceprint feature extraction and embedding module so that they take effect in subsequent recognition.

[0047] This fine-tuning scheme effectively balances the conflict between model optimization and terminal device resource constraints by limiting updates to when the device is idle and connected to power, and by selectively fine-tuning only high-level network parameters. It employs anonymized data and mini-batch gradient descent to achieve incremental model optimization while protecting user privacy. Combined with regularization constraints and a hot reload mechanism, it significantly improves the model's adaptability to the acoustic characteristics of specific user groups, while avoiding catastrophic forgetting and ensuring continuous and stable performance improvement.

[0048] Specifically, the parameters of the last fully connected layer and / or the last few hidden layers of the base model are updated, and the updated parameters include the weight matrix and bias vector.

[0049] Furthermore, the encryption method employed by the secure storage unit includes a key derivation and data encryption binding process based on the user's voiceprint biometric features. Specifically, it includes: a key deriver, used to extract consistency features from the user's voiceprint embedding vector, and process the consistency features and the user's preset auxiliary password through a key derivation function to generate a unique user-specific encryption key; an encryption engine, connected to the key derivation module, used locally to use the user-specific encryption key to encrypt the user's voiceprint template, voiceprint embedding vector, and associated personal configuration using a symmetric encryption algorithm conforming to national cryptographic standards, generating ciphertext data; and a secure isolation storage area, used to isolate and store the ciphertext data and the necessary parameters required for decryption in a protected area of ​​the hardware security element, and the user-specific encryption key is only used within the hardware security element and never leaves the secure isolation storage area in an exposed form.

[0050] Specifically, extracting consistency features from user voiceprint embedding vectors means: during user registration, calculating statistical features from multiple voiceprint embedding vectors generated from multiple registration voice segments of the same user as the consistency features; the statistical features are the average vector of these multiple voiceprint embedding vectors or the first principal component vector obtained through principal component analysis; processing through a key derivation function means: concatenating the consistency features and the user's preset auxiliary password, and using the concatenated data as input, generating a fixed-length user-specific encryption key through iterative hash function calculation.

[0051] Furthermore, the cross-device roaming implemented by the cloud synchronization unit specifically includes: after obtaining user authorization, encrypting the anonymized voiceprint features or updated deep neural network model parameters to form a synchronization data packet, and securely uploading it to the cloud server for storage; when the user logs into another new device equipped with the same system, after identity authentication, the new device securely downloads the synchronization data packet from the cloud server; the new device decrypts the downloaded synchronization data packet locally, and loads the decrypted voiceprint features or model parameters into the local data storage and model update module, enabling the user to achieve seamless identity recognition and health data management on the new device without re-registering.

[0052] In some embodiments, the device further includes a workflow orchestration and system management module, configured to: define and control the workflow of a complete health measurement, the workflow being configured to: prevent physiological parameter sensors from collecting data before the identity matching and decision fusion module determines successful identity recognition based on its dynamic decision threshold or historical decision inertial filtering process; and immediately trigger data collection and bind it to the recognition result after successful identity recognition; manage the computing resources of the device to ensure the execution priority of the voiceprint feature extraction and embedding module and the identity matching and decision fusion module, thereby controlling the end-to-end latency of the identity recognition process within a predetermined duration threshold, ensuring low-latency recognition within 300 to 800 milliseconds on the end side.

[0053] Example 3 provides a more detailed description of the structure and operation of the device in this example. The device specifically includes: a voice front-end processing module, a voiceprint feature extraction and embedding module, a liveness and replay attack detection module, an identity matching and decision fusion module, a multi-person conversation management and binding module, a data storage and model update module, and a service flow orchestration and system management module.

[0054] The voice front-end processing module preprocesses the raw audio signals acquired by the device's microphone array to improve audio signal quality and provide favorable conditions for subsequent voiceprint recognition. Specifically, it includes the following sub-functions: employing acoustic echo cancellation algorithms to suppress interference from the device's own playback sound; utilizing adaptive noise reduction algorithms to filter out steady-state and non-steady-state noise in the environment (such as appliance operation sounds, television sounds, etc.); maintaining the stability of the input audio signal amplitude through automatic gain control algorithms; when configuring the microphone array, combining beamforming algorithms to enhance the voice signal from the direction of the target speaker while suppressing interference from other directions; and integrating voice endpoint detection functionality to accurately determine the start and end positions of voice segments by analyzing the short-time energy and zero-crossing rate of the audio signal, thereby avoiding processing invalid silence segments and improving system efficiency.

[0055] The voiceprint feature extraction and embedding module is used to extract highly discriminative voiceprint feature representations from preprocessed clean speech. This process first calculates the Log-Mel spectrogram or MFCC (Mel-Frequency Cepstral Coefficients) features of the speech signal as the front-end acoustic representation. Subsequently, this front-end acoustic representation is input into a deep neural network model (e.g., ECAPA-TDNN (Emphasized Channel Attention, Propagation and Aggregation Time Delay Neural Network) or x-vector) pre-trained on a large-scale corpus. This network model acts as a feature extractor, mapping clean, effective speech segments of varying lengths obtained from the speech front-end processing module into a fixed-dimensional (192- to 256-dimensional) voiceprint embedding vector. This voiceprint embedding vector can fully represent the speaker's identity information and has a certain degree of robustness to factors such as speech content and channel changes, providing a basis for subsequent identity matching.

[0056] The liveness detection and playback attack detection module is used to distinguish whether the input speech comes from a real, living person or is a deception attack such as a recording playback. This module works in parallel with the voiceprint feature extraction process. It directly extracts anti-spoofing features from the preprocessed audio signal that can effectively capture traces of recording playback and analyzes the spectral characteristics of the speech to obtain identification information. Subsequently, the extracted anti-spoofing features and spectral identification information are used as input data to a lightweight classifier (such as a convolutional neural network or a gated recurrent unit network) for authenticity judgment. The classifier ultimately outputs a binary judgment result or a corresponding confidence score indicating whether the input speech is from a "real, living person" or a "spoofing attack." The output of this module serves as an important reference for the credibility of the identity recognition result and is directly used in the subsequent identity binding decision process, effectively reducing the risk of the system being subjected to spoofing attacks. Among these methods, spectral characteristic analysis uses acoustic feature analysis algorithms to detect short-time spectral data of speech to obtain discriminative information. This mainly includes: checking for abnormal energy attenuation in high-frequency components due to frequency response limitations of the recording and playback equipment through frequency band analysis; analyzing the smoothness of the spectral envelope and the presence of specific distortion patterns introduced by the playback system through spectral distortion detection; and detecting whether background noise exhibits statistical characteristics not characteristic of a natural acquisition environment through signal-to-noise ratio analysis. These analytical methods work together to reveal the acoustic differences between recorded playback and real live speech from different dimensions.

[0057] Specifically, the anti-spoofing features include constant Q-cepstral coefficients, residual signal statistical features based on a linear predictive coding model, group delay characteristics of the speech signal, and nonlinear energy distribution features based on the Teager energy operator.

[0058] The residual signal statistical characteristics based on the Linear Predictive Coding (LPC) model enable liveness detection at the level of the excitation source mechanism of speech generation. The LPC model estimates a digital filter based on a speech signal to simulate the resonance characteristics of the human vocal tract. When the original speech signal passes through this inverse filter, a signal called the "residual" or "excitation" is obtained, which ideally represents the original pulse sequence generated by vocal cord vibration. In real human speech, the LPC residual signal exhibits distinct non-Gaussian statistical characteristics, such as high kurtosis, and a clear, non-periodic pulse shape. However, when speech is recorded and then played back by a speaker, it undergoes a secondary electroacoustic and electroacoustic conversion process. During this process, the frequency response and linearity of the devices (microphone, speaker) distort and smooth out these subtle pulse characteristics in the residual signal, making its statistical distribution (e.g., kurtosis) flatter and weakening its pulse characteristics. Therefore, by analyzing higher-order statistics (such as kurtosis and skewness) or pulse shape indices of the residual signal, the essential differences between real human speech and recorded playback at the sound source excitation level can be effectively captured. This is a deep detection dimension that effectively complements spectral features (such as CQCC).

[0059] The group delay characteristic of speech signals provides discriminative information from the unique dimension of the phase domain. Traditional liveness detection methods mostly focus only on the amplitude spectrum (e.g., CQCC) of speech signals, neglecting the equally information-rich phase spectrum. Physically defined as the negative derivative of the phase spectrum, group delay reflects the time difference in arrival of sound waves at different frequencies and is closely related to the shape of the signal's spectral envelope. The phase changes in speech produced by a real human are natural and continuous. However, the recording and playback process introduces additional electroacoustic and electro-acoustic conversion stages. The physical characteristics and nonlinear responses of the devices in these stages (e.g., speaker diaphragms, microphone diaphragms) introduce specific, nonlinear phase distortions. This distortion disrupts the continuity and natural pattern of the original speech phase / group delay. By analyzing the group delay spectrum of the speech signal, this distortion introduced by the playback system can be detected, such as abnormal peaks or uneven regions appearing on the group delay spectrum. This feature provides a novel detection perspective independent of the amplitude spectrum, starting from the signal's phase characteristics.

[0060] Based on the nonlinear energy distribution characteristics of the Teager energy operator, this feature is used to identify the instantaneous nonlinear dynamic changes in speech energy. Traditional short-time energy analysis can only reflect the sum of squares of signal amplitudes and cannot accurately express the energy dynamics of the complex nonlinear physical process of speech production. The Teager energy operator is a nonlinear operator that can more accurately track and estimate the instantaneous changes in mechanical or acoustic energy involved in speech production (such as vocal cord vibration and air turbulence), and is highly sensitive to the amplitude-frequency change rate of the signal. The onset and decay processes of real human speech contain specific nonlinear energy dynamic patterns. However, the dynamic range, response speed, and nonlinear characteristics of the energy changes in recording and playback systems (especially loudspeakers) when responding to electrical signals inherently differ from those of real human vocal organs. This difference is reflected in the Teager energy trajectory; for example, the energy changes in the reproduced speech are smoother, lack certain instantaneous spikes, or have different transition characteristics. Therefore, by analyzing the time-frequency energy distribution based on the Teager energy operator, subtle differences in the dynamic processes of energy generation and release between real speech and recorded playback can be captured. This is a more refined and effective feature than simple short-time energy analysis.

[0061] The identity matching and decision fusion module compares the current speech's embedded voiceprint vector with a pre-stored authorized user voiceprint template library on the local device. The comparison algorithm can employ methods such as cosine similarity calculation or probabilistic linear discriminant analysis. To avoid random errors in single-shot recognition, this identity matching and decision fusion module introduces a confidence gating mechanism and a historical decision inertial filtering strategy: when the confidence or similarity of the current comparison exceeds a set threshold, it is directly adopted; if the confidence is in a fuzzy range, the results of several recent recognitions are considered for inertial decision-making, thereby improving the stability and reliability of identity binding in complex scenarios. Specifically, the fuzzy range refers to a preset confidence score range, with its lower limit higher than the level of random guessing but lower than the threshold required for reliable confirmation, and its upper limit lower than the high threshold for direct adoption. For example, if the threshold for direct adoption is 0.85, the fuzzy range can be defined as [0.60, 0.85). When the highest confidence of the current comparison falls into this range, the system does not immediately make a final decision but instead initiates the historical decision inertial filtering process. The specific process of historical decision-making inertia filtering is as follows: The system retrieves the occurrences of the current candidate identity (i.e., the identity with the highest similarity in this instance) in the most recent N (e.g., N=5) successful recognition events and calculates its historical consistency ratio R (R = the number of times the identity has appeared in recent history / N). If the consistency ratio R exceeds a preset inertia threshold (e.g., 0.6), the system tends to maintain the historical trend and adopt the candidate identity identified this time; if the consistency ratio R is lower than the inertia threshold, the system tends to consider the recognition result unreliable, rejects the match, and requires the user to re-enter voice or initiate other verification methods. This strategy effectively improves the stability and reliability of identity binding in complex scenarios by introducing statistical regularities in the time dimension.

[0062] The multi-person session management and binding module is used to handle situations where multiple people may take measurements continuously or nearly simultaneously in shared scenarios. It has basic speaker logging capabilities, enabling the separation and segmentation of different speakers within an audio clip. This module works in conjunction with the voice endpoint detection and identity matching module to intelligently bind each detected valid measurement action (e.g., a user stepping onto a scale or pressing a blood pressure monitor button) to the voice identities captured before and after that action, ensuring "whoever takes the measurement, gets the data," thus achieving a seamless and accurate association between measurement actions and user identities.

[0063] The data storage and model update module manages all identity-related data. User voiceprint templates, feature embedding vectors, and associated personal configurations on the local device are securely stored using encryption. With user authorization, the system supports securely synchronizing anonymized voiceprint features or model parameters to the cloud for cross-device roaming. Simultaneously, this data storage and model update module supports a hot-update mechanism for the model, allowing for incremental learning-based fine-tuning of the voiceprint model on the device or in the cloud based on subsequent user usage data, enabling it to adaptively track the long-term, slow changes in the user's voiceprint.

[0064] The Business Flow Orchestration and System Management module serves as the system's control center, coordinating the execution flow of various software modules and system resources. It defines and executes a standard procedure for a complete health measurement, including voice activity monitoring, identity verification, sensor data acquisition, result binding, and storage. This module also handles anomalies (such as recognition failure or measurement interruption) and triggers retry mechanisms. Simultaneously, it comprehensively manages device power status and network connectivity, prioritizing low-latency, high-reliability operation of the core identity verification process on the device side.

[0065] The device's operation involves the user initiating a measurement command via voice in front of the health monitoring device (e.g., saying "Start measuring blood pressure" or using the device's default prompt). The built-in microphone array simultaneously collects the voice signal. The raw audio is processed by the voice front-end module for noise reduction and enhancement, outputting a valid voice segment. Simultaneously, a liveness detection and playback attack detection module works in parallel to assess the bioactivity and authenticity of the voice sample, ensuring it is not a recording attack. Subsequently, a voiceprint feature extraction and embedding module extracts the voiceprint embedding vector from the valid voice, and an identity matching and decision fusion module performs a rapid similarity comparison with a locally encrypted user voiceprint archive, outputting the user identity with the highest matching degree and its confidence level. If the confidence level is lower than a preset threshold, the system automatically triggers a secondary confirmation mechanism (e.g., prompting the user to repeat the voice command) to retry, ensuring recognition reliability. After identity verification, a multi-user session management module binds the identity identifier to the currently initiated measurement task (e.g., blood pressure measurement) in real time. Subsequently, the device's sensors collect physiological data (e.g., blood pressure values), and the data management module ultimately associates and stores the measurement data with the corresponding user identity information. Throughout the process, the various algorithm modules communicate asynchronously through a message bus or event bus, and are organized into a pipelined parallel processing architecture based on a unified API interface, thereby achieving low-latency identity recognition and data binding within 300 to 800 milliseconds on the edge.

[0066] In a specific embodiment, as an example, the complete working process of the device is as follows: Step 1. User Registration and Voiceprint Profile Establishment Process. When a new user registers, the system guides the user to read aloud 3 to 5 fixed text phrases specified by the system in a relatively quiet environment, with the total recording time controlled between 15 and 30 seconds. The collected multiple speech segments are first enhanced and effective speech segments are filtered by the speech front-end processing module, and the liveness detection module ensures that they are the actual speech of the real user. Subsequently, the voiceprint feature extraction and embedding module extracts fixed-dimensional voiceprint embedding vectors from each effective speech segment. The system uses density-based (e.g., DBSCAN) or distance-based (e.g., K-means) clustering algorithms to perform clustering analysis on these multiple voiceprint embedding vectors from the same user to effectively eliminate abnormal speech segments caused by misreading or environmental interference, and finally generates a stable and representative user voiceprint template by calculating its average value or centroid. The user's voiceprint template is stored in a secure storage area on the device using a high-strength encryption algorithm such as AES-256. To facilitate synchronous use across multiple devices, an optional function of securely backing up the encrypted data to the cloud is available, provided that privacy compliance regulations are strictly followed (e.g., with explicit user authorization).

[0067] The voiceprint embedding vector (VIN) is a fixed-dimensional numerical vector (e.g., 192-dimensional or 256-dimensional) extracted from a speech signal using a deep neural network model (e.g., ECAPA-TDNN). This VIN uniquely and stably represents the speaker's individual identity in a high-dimensional space. Its generation process involves first converting the speech signal into acoustic features (e.g., Log-Mel spectrum), then performing nonlinear transformations and mappings by the neural network, ultimately outputting a dense floating-point vector. A key characteristic of this vector is that the VIN corresponding to different statements by the same speaker has relatively close cosine or Euclidean distances in high-dimensional space, while the VIN of different speakers is much farther apart. This allows for accurate identity recognition and verification by calculating the similarity between the vectors. This embedding representation has strong speaker discrimination capabilities and is robust to changes in speech content and short-term noise.

[0068] Step 2. Measurement and Identity Recognition Process When a user uses the device (e.g., stepping on the scale or pressing the start button), the measurement and identity recognition process begins: the device automatically plays a voice prompt and controls the microphone array to start collecting voice signals or raw audio data. The system processes the collected raw audio data. First, the voice front-end processing module performs noise reduction and enhancement operations to obtain a clean and valid voice segment. Subsequently, the liveness and playback attack detection module performs biometric and anti-spoofing verification on the valid voice segment. If the voice sample fails the security test corresponding to the verification, the system will guide the user to try speaking again through voice prompts or indicator lights, or provide an alternative option of manually selecting an identity.

[0069] After a voice sample passes security checks, the voiceprint feature extraction and embedding module extracts the corresponding voiceprint embedding vector from the voice sample. The system compares this voiceprint embedding vector with a locally encrypted voiceprint template archive, calculating the similarity score and corresponding candidate identity information for each user's voiceprint template in the archive. The system presets a basic similarity threshold (e.g., 0.75 to 0.85, which can be adjusted according to the specific device model and application scenario) and simultaneously incorporates the "historical inertia" formed by the user's recent historical recognition results to generate a dynamic decision threshold. If the identity confidence corresponding to the highest similarity score in this comparison exceeds the dynamic decision threshold, the identity recognition is deemed successful.

[0070] The similarity score is primarily obtained by calculating the cosine of the angle between the speaker's embedded vector and the pre-stored user speaker template vector in a high-dimensional space; this is known as the cosine similarity algorithm. The algorithm calculates the cosine of the angle between the two vectors, with the result ranging from -1 to 1. A value closer to 1 indicates greater similarity. In practical applications, to improve robustness, models such as Probabilistic Linear Discriminant Analysis (PLDA) can be used to further incorporate probabilistic statistical information about the distribution of the same speaker's vectors, thereby generating a more discriminative likelihood ratio score as the final similarity rating.

[0071] The generation process of the dynamic decision threshold is as follows: The system first sets a basic similarity threshold T. base (e.g., 0.75). To improve recognition continuity and reduce false positives, a historical inertia factor H based on the user's recent historical recognition results is introduced for dynamic adjustment. This factor H is calculated by statistically analyzing the proportion R (consistency ratio) of the number of times the current user's identity matches the current candidate in the most recent N (e.g., N=5) successful recognition records, i.e., R = number of matches / N. The inertia factor H can be defined as H = 1 + α (R 0.5), where α is an adjustment coefficient (e.g., α = 0.1), used to control the strength of historical influence. Finally, the dynamic decision threshold T... dynamic From formula T dynamic =T base The threshold is calculated using ×H. When historical consistency is high (R>0.5), H>1, and the threshold is appropriately relaxed to facilitate continuous identification; when consistency is low (R<0.5), H<1, and the threshold is tightened to improve security requirements. This method achieves adaptive adjustment of the threshold according to user habits.

[0072] If the system detects multiple people speaking simultaneously during audio analysis, it will activate a basic speaker separation algorithm to separate the mixed speech and independently match the identities of each individual speaker's voice segment. If the current measurement task logically only supports single-user operation (e.g., weight measurement), the system will proactively prompt the user to use the functions sequentially. Within the entire measurement task's time window, from start to finish, the session management module is responsible for strongly associating the various physiological data finally collected by the sensors (e.g., weight, blood pressure, heart rate, blood oxygen saturation, etc.) with the successfully identified user's identity, establishing an immutable one-to-one correspondence, and storing the associated data in a local or cloud database. The speaker separation algorithm, also known as speech separation, is a computational method for handling scenarios where multiple people speak simultaneously. Its goal is to separate the speech stream of each individual speaker from the mixed speech signal collected by one or more microphones. The process first utilizes spatial information (e.g., time difference of arrival) and spectral features of the speech obtained from the microphone array to classify speech frames belonging to different speakers using clustering methods (e.g., depth-based clustering). Then, a mask is generated based on the clustering results and applied to the spectrum of the mixed speech to estimate the clean speech signal of each speaker. End-to-end deep learning models (e.g., Conv-TasNet) can directly model the mapping relationship from the mixed speech to the individual speaker's speech, improving the separation effect. Finally, the speaker separation algorithm outputs several separated audio trajectories, each containing, as far as possible, the voice of only one target speaker, so that the subsequent voiceprint recognition module can perform independent identity matching for each trajectory.

[0073] Step 3. Model Adaptive Update Mechanism To improve the system's adaptability to long-term changes in user voiceprints, this solution designs a model adaptive update mechanism. This mechanism includes the following two aspects: On the device side, an incremental learning-based template update strategy is adopted. After each successful recognition, the system merges the successfully verified voiceprint embedding vector with a small weight (e.g., using an exponential moving average algorithm) into the corresponding user voiceprint template, thereby achieving gradual optimization of the template and enabling it to track the natural and slow changes in user voiceprints. Specifically, after each successful authentication, the system triggers the template update process. Let the currently stored user voiceprint template be vector v. old The extracted voiceprint embedding vector was successfully verified in this study. new The system employs an exponential moving average algorithm, using a smoothing factor β close to 1 (e.g., β = 0.95), to calculate the updated template vector v according to the following formula. updated :v updated =β v old +(1 β) v new This algorithm assigns higher weights to historical templates while incorporating new voiceprint features with lower weights. This makes the evolution of user voiceprint templates smooth and stable, gradually absorbing slow changes in user voiceprints (such as changes caused by age or health conditions) while effectively suppressing noise caused by occasional fluctuations or momentary interference in a single voice acquisition. Thus, while maintaining the stability of the main features of the user voiceprint template, it achieves progressive optimization and adaptive tracking of its features.

[0074] At the system level, the system supports parameter updates for voiceprint recognition models (such as the ECAPA-TDNN model). The system supports periodically pushing lightweight model update files (e.g., files containing pre-trained update parameters) to devices via the cloud. Anonymous data refers to data processed on the device that cannot be traced back to a specific user's identity. With user authorization, the system can locally decouple the embedding vectors corresponding to successfully verified voiceprint features from any personal identification information, retaining only their essence as voiceprint feature vectors. Subsequently, when the device is idle, this locally accumulated anonymized data can be used to fine-tune the voiceprint recognition model. Fine-tuning here refers to: based on the pre-trained voiceprint recognition model, using local anonymized data to make small adjustments and optimizations to all or some layer parameters of the voiceprint recognition model, thereby enabling the voiceprint recognition model to better adapt to the common acoustic characteristics of specific user groups or new speech patterns, ensuring that the voiceprint recognition algorithm (i.e., the algorithm used to determine the speaker's identity) continues to maintain optimal performance.

[0075] Step 4. Security and Privacy Protection Strategy This solution strictly adheres to security and privacy protection principles in its design. All user voiceprint embedding vectors and user voiceprint templates are encrypted and stored using keys provided by the device's built-in Hardware Security Module (a physical chip dedicated to key management and secure encryption operations, providing protected storage and resistance to physical attacks). All communication data with external servers (e.g., the cloud) is encrypted and transmitted using security protocols such as Transport Layer Security (TLS). The system follows the principle of data minimization, collecting only the minimum data necessary to implement its functions.

[0076] The liveness detection and playback attack detection functions are enabled by default during each recognition. For voice samples deemed suspicious, the system will force manual identity verification or directly reject the data binding operation, thus constructing a multi-layered security defense. Specifically, the determination of suspicious voice samples is achieved through a dedicated classifier algorithm: the system first extracts features for anti-spoofing detection (such as constant Q-cepstral coefficients, spectral distortion features, etc.) from the input voice, and then inputs them into a pre-trained lightweight classification model (such as a convolutional neural network or a gated recurrent unit network) for authenticity judgment. The classifier finally outputs a confidence score representing whether the voice sample is "live" or "spoofed"; if this confidence score is lower than a preset security threshold, the voice sample is determined to be "suspicious". For voice samples determined to be suspicious, the system will force manual identity verification (such as prompting for password input or user selection) or directly reject the data binding operation, thus constructing a multi-layered security defense.

[0077] The technical advantages of this invention are mainly reflected in the following aspects: First, it achieves a login-free and manual user selection-free identity recognition experience, seamlessly integrating identity binding with the measurement process, shortening the operation process, and greatly improving the usability of multi-user shared health detection devices. Second, by adopting strong representation models such as ECAPA-TDNN and combining a dynamic discrimination threshold with an inertial fusion mechanism of historical recognition results, the system can maintain high recognition accuracy and robustness even in complex scenarios such as home noise and far-field speech. In terms of security, it integrates liveness detection and playback attack recognition functions, effectively defending against attacks such as recording spoofing and enhancing system reliability. Simultaneously, the system has low latency and offline processing capabilities. With the help of edge-side pipelined parallel computing, it can return identity recognition results within 300 to 800 milliseconds and work normally in weak network or offline environments, ensuring user privacy and real-time usage. Furthermore, this solution is easily expandable, providing a flexible identity binding framework to support the integration of multiple types of sensors and business functions, and utilizing an adaptive update strategy to reduce model maintenance costs, exhibiting good long-term applicability and system scalability.

[0078] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the methods described above. This invention also provides an electronic device comprising: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the methods provided by this invention. As shown in FIG8, a computer system 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The RAM 803 also stores various programs and data required for the operation of the computer system 800. The CPU 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804. The following components are connected to I / O interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to I / O interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 810 as needed so that computer programs read from it can be installed into storage section 808 as needed. According to embodiments disclosed in this invention, the processes described in the above main step diagrams can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the main step diagrams. This computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by the central processing unit 801, it performs the functions defined in this invention.

[0079] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A multi-user shared health monitoring device based on accurate voice identification, characterized in that, include: Microphone array, used to capture the user's raw audio signal; Physiological parameter sensors are used to collect users' health data; processor; Memory; The processor executes a computer program in the memory to perform the following operations: preprocessing the original audio signal through a voice front-end processing module to obtain a valid voice segment; extracting a voiceprint embedding vector from the valid voice segment using a deep neural network model through a voiceprint feature extraction and embedding module; comparing the voiceprint embedding vector with the authorized user voiceprint template library through an identity matching and decision fusion module to obtain the highest similarity score and its corresponding candidate identity; extracting multi-dimensional acoustic features through a liveness and replay attack detection module and using a classification model to determine the authenticity of the valid voice segment; and binding the candidate identity with the health data through a multi-person conversation management and binding module when the determination result is a real liveness and the highest similarity score meets the judgment conditions.

2. The apparatus according to claim 1, characterized in that, The voice front-end processing module is used to perform sequence processing on the original audio signal to extract effective voice segments suitable for voiceprint recognition and liveness detection from the original audio signal; wherein, the voice front-end processing module is used to provide the extracted effective voice segments in parallel to the voiceprint feature extraction and embedding module and the liveness and playback attack detection module.

3. The apparatus according to claim 1, characterized in that, The voiceprint feature extraction and embedding module includes: an acoustic feature conversion unit for converting the effective speech segment into an acoustic feature map; a deep feature extraction unit for inputting the acoustic feature map into a deep neural network model to map the effective speech segment into a fixed-dimensional voiceprint embedding vector; the deep feature extraction unit is also used to extract multiple voiceprint embedding vectors from multiple registered speech segments of the same user during the user registration stage; and a voiceprint template generation unit for performing cluster analysis on the multiple voiceprint embedding vectors to obtain the set of voiceprint embedding vectors corresponding to the largest cluster, and calculating the average value or centroid of all voiceprint embedding vectors in the set of voiceprint embedding vectors corresponding to the largest cluster to generate an initial user voiceprint template.

4. The apparatus according to claim 1, characterized in that, The liveness and replay attack detection module includes: an anti-spoofing feature extraction unit, used to extract anti-spoofing features from the effective speech segment; a spectrum identification information extraction unit, used to obtain spectrum identification information by analyzing the short-time spectrum of the effective speech segment; and a classification and discrimination unit, used to input the anti-spoofing features and the spectrum identification information into a classification model and output a discrimination result or confidence score indicating authenticity.

5. The apparatus according to claim 1, characterized in that, The identity matching and decision fusion module is used to: calculate the similarity score between the voiceprint embedding vector to be identified and the pre-stored user voiceprint template, and determine the highest similarity score and its corresponding candidate identity; compare the similarity score with a preset high threshold and a low threshold, and define a fuzzy decision interval between the high threshold and the low threshold; when the highest similarity score is higher than the high threshold, directly adopt the corresponding candidate identity as the identification result; when the highest similarity score falls into the fuzzy decision interval, start the historical decision inertia filtering process.

6. The apparatus according to claim 1, characterized in that, The identity matching and decision fusion module is used to: calculate the similarity score between the voiceprint embedding vector to be identified and the pre-stored user voiceprint template, and determine the highest similarity score and its corresponding candidate identity; calculate the dynamic decision threshold based on the preset basic similarity threshold and combined with the consistency ratio of the current candidate identity in recent historical identification records; and compare the highest similarity score with the dynamic decision threshold. If the highest similarity score exceeds the dynamic decision threshold, the identity recognition is deemed successful; if the highest similarity score is lower than or equal to the dynamic decision threshold, a secondary verification process is triggered.

7. The apparatus according to claim 6, characterized in that, The dynamic decision threshold T dynamic Calculated using the following formula: T dynamic =T base ×1+α (R 0.5); where T base R is the basic similarity threshold, R is the consistency ratio of the current candidate identity in the most recent N successful recognition records, and α is a preset adjustment coefficient.

8. The apparatus according to claim 6, characterized in that, The calculation of the similarity score between the voiceprint embedding vector to be identified and the pre-stored user voiceprint template specifically includes: performing multi-scale feature decomposition on the voiceprint embedding vector and the user voiceprint template to obtain corresponding global features and local detail features respectively; calculating the first similarity between the global features of the voiceprint embedding vector to be identified and the global features of the user voiceprint template as the global similarity; calculating the second similarity between the local detail features of the voiceprint embedding vector to be identified and the local detail features of the user voiceprint template as the local similarity; dynamically adjusting the weight ratio of the global features and the local detail features according to the current environmental noise level; and linearly fusing the weighted global similarity and the local similarity to obtain the final similarity score.

9. The apparatus according to claim 1, characterized in that, The multi-user conversation management and binding module includes: a speaker log unit for processing collected audio data; a measurement behavior perception unit for detecting valid measurement behaviors triggered by user operations; and a binding logic unit for: in response to detecting a valid measurement behavior, acquiring audio data output by the voice front-end processing module within a predetermined time period before and after the valid measurement behavior occurs; controlling the speaker log unit to process the audio data to obtain at least one separated speech segment; sequentially providing each separated speech segment to the voiceprint feature extraction and embedding module and the identity matching and decision fusion module for processing to obtain corresponding user identity information; and binding and associating the successfully identified user identity information with the physiological data collected by the physiological parameter sensor in this valid measurement behavior.

10. The apparatus according to claim 1, characterized in that, The device also includes a data storage and model update module for managing all identity-related data.

Images