Voice global feature extraction and numeralization method and system reflecting health state

Abstract

The invention provides a voice global feature extraction and numeralization method and system reflecting a health state. The method comprises the following steps: acquiring a voice signal of a speaker, synchronously acquiring vital sign data and health characteristic data of the speaker, and establishing a database according to the acquired voice signal and the corresponding vital sign data and health characteristic data; defining voice global features related to the vital signs and health features of the speaker, preprocessing the voice signals in the database, and removing interference signals; extracting multi-dimensional voice global features from the preprocessed voice signal to form a feature vector, and performing dimension reduction processing on the feature vector to obtain a dimension-reduced feature vector; and constructing a fusion model by using the feature vector after dimension reduction and the corresponding vital sign data and health feature data, and training the fusion model, so that the fusion model can discriminate the vital sign and health feature of the speaker according to the input voice signal feature vector.

Description

Technical Field

[0001] This application relates to the fields of speech signal processing and health status assessment, and in particular to a method and system for extracting and quantifying global speech features that reflect health status. Background Technology

[0002] In the field of speech signal processing and health assessment, most existing technologies are limited to single-dimensional speech feature analysis (such as extracting only fundamental frequency or speech rate features) or only targeting single objectives such as speaker identification and emotion analysis. They lack multi-dimensional global correlation analysis of the life characteristics (gender, age, region) and health characteristics (vitality, mental state, physical condition, etc.) implicit in speech signals. Traditional methods usually use fixed feature sets or simple statistical models, which cannot effectively capture the complex physiological and health correlation information in speech signals, and have not built an intelligent system for the entire process from data acquisition and feature extraction to health assessment. In addition, the combination of speech processing and health assessment in existing technologies is mostly still in the experimental stage, lacking an engineering-featured numerical analysis framework and adaptive optimization mechanism, making it difficult to meet the needs of dynamic health assessment in real-world scenarios.

[0003] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the Invention

[0004] This application provides a method and system for extracting and quantifying global speech features that reflect health status. It aims to address the problem that the combination of speech processing and health assessment in the prior art is mostly still in the experimental stage, lacks an engineering-featured numerical analysis framework and adaptive optimization mechanism, and is difficult to meet the needs of dynamic health status assessment in real-world scenarios.

[0005] In a first aspect, embodiments of this application provide a method for extracting and quantifying global speech features reflecting health status, the method comprising: Collect the speaker's voice signal and simultaneously collect the speaker's vital signs and health characteristics data. Build a database based on the collected voice signal and the corresponding vital signs and health characteristics data. Define global speech features related to the speaker's vital signs and health characteristics, preprocess the speech signals in the database to remove interference signals, extract multi-dimensional global speech features from the preprocessed speech signals to form feature vectors, and then perform dimensionality reduction processing on the feature vectors to obtain dimensionality-reduced feature vectors. By using the dimensionality-reduced feature vectors and the corresponding vital signs and health features data, a fusion model is constructed and trained to enable the fusion model to identify the speaker's vital signs and health features based on the feature vectors of the input speech signal.

[0006] In some embodiments, the acquisition of the speaker's voice signal and the simultaneous acquisition of the speaker's vital signs data and health characteristics data include: acquiring the speaker's voice signal in a natural vocal state through a microphone array or voice acquisition device; simultaneously acquiring the speaker's gender, age, and region's vital signs data through user registration input, sensor detection, or historical data integration; and acquiring the speaker's vitality, mental state, physical condition, and health status's health characteristics data through questionnaires, physiological indicator detection, or health record extraction; and synchronizing the voice signal with the corresponding vital signs data and health characteristics data using timestamps.

[0007] In some embodiments, the step of building a database based on the collected voice signals and corresponding vital signs data and health features data includes: storing the labeled voice signals and their associated vital signs data and health features data into the database according to a preset data format; performing deduplication, integrity verification and outlier filtering on the data in the database; dividing the data into training datasets, validation datasets and test datasets according to a preset ratio; and establishing a data indexing mechanism for quick retrieval.

[0008] In some embodiments, the definition of global speech features related to the speaker's vital signs and health characteristics includes: defining a multi-dimensional global feature system encompassing speech prosody features, phonological features, temporal statistical features, and frequency domain distribution features, wherein prosodic features include average speech rate, intonation fluctuation, and pause frequency; phonological features include fundamental frequency mean, formant frequency, and harmonic-to-noise ratio; temporal features include short-time energy mean and zero-crossing rate variance; and frequency domain features include spectral entropy and Mel frequency cepstral coefficient mean.

[0009] In some embodiments, the preprocessing of the speech signals in the database to remove interference signals includes: performing endpoint detection on the speech signals to locate valid speech segments, using spectral subtraction or wavelet denoising algorithms to remove environmental noise and equipment interference, performing volume normalization processing on the denoised speech signals, adjusting the signal amplitude to a preset dynamic range, and eliminating volume differences between different acquisition devices.

[0010] In some embodiments, the step of extracting multi-dimensional global speech features from the preprocessed speech signal to form a feature vector includes: performing frame-by-frame windowing processing on the effective speech segments, calculating time-domain feature parameters, frequency-domain feature parameters, and prosodic feature parameters in each frame of the signal, concatenating the time-domain feature parameters, frequency-domain feature parameters, and prosodic feature parameters in a preset order to form an original feature vector, and normalizing the original feature vector to eliminate the influence of dimensions.

[0011] In some embodiments, the dimensionality reduction of the feature vector to obtain the dimensionality-reduced feature vector includes: performing a linear transformation on the feature vector using principal component analysis, calculating the covariance matrix of the feature vector and extracting principal components, filtering principal components according to the cumulative variance contribution rate threshold, and projecting the feature vector onto the principal component space to generate the dimensionality-reduced feature vector; or, using linear discriminant analysis, with vital signs and health feature labels as category supervision information, maximizing inter-class divergence and minimizing intra-class divergence to achieve dimensionality reduction of the feature vector.

[0012] In some embodiments, the construction of a fusion model using the dimensionality-reduced feature vectors and corresponding vital sign data and health sign data includes: selecting a neural network model, a random forest model, or a support vector machine model as the base model; using the dimensionality-reduced feature vectors as the model input layer; designing the output layer structure according to the number of categories or numerical range of vital sign and health sign data; and establishing a fusion model architecture that includes a feature fusion layer and a classification regression layer. The feature fusion layer is used to integrate multi-dimensional feature information, and the classification regression layer is used to output the prediction results of vital sign categories and the prediction results of health sign values.

[0013] In some embodiments, training the fusion model to enable it to determine the speaker's vital signs and health characteristics based on the input speech signal feature vector includes: inputting a training dataset into the fusion model; using a cross-entropy loss function as the classification task loss function and a mean squared error loss function as the regression task loss function; updating the model parameters of the fusion model using a stochastic gradient descent algorithm or an adaptive optimization algorithm; periodically evaluating the model's classification accuracy and regression error using a validation dataset during training; adjusting the model's hyperparameters based on the evaluation results until the fusion model's performance on the validation set reaches a preset convergence condition; and verifying the generalization ability of the fusion model using a test dataset.

[0014] Secondly, this application provides a system for extracting and quantifying global speech features reflecting health status, the system comprising: The signal acquisition unit is used to acquire the speaker's voice signal and simultaneously acquire the speaker's vital signs data and health characteristics data, and build a database based on the acquired voice signal and the corresponding vital signs data and health characteristics data; The feature definition unit is used to define global speech features related to the speaker's vital signs and health characteristics. It preprocesses the speech signals in the database to remove interference signals; extracts multi-dimensional global speech features from the preprocessed speech signals to form feature vectors; and then performs dimensionality reduction processing on the feature vectors to obtain dimensionality-reduced feature vectors. The feature discrimination unit is used to construct a fusion model using the dimensionality-reduced feature vectors and the corresponding vital signs and health features data, and to train the fusion model so that it can discriminate the speaker's vital signs and health features based on the feature vectors of the input speech signal.

[0015] This application overcomes the limitations of single-feature analysis by defining multi-dimensional global speech features encompassing prosody, timbre, temporal domain, and frequency domain, thus providing a more comprehensive reflection of the speaker's physiological state and health information. By constructing a fusion model combining dimensionality reduction and multi-task learning, it integrates vital sign classification with numerical prediction of health characteristics, improving assessment accuracy and interpretability. Through software system integration of data acquisition, feature processing, model discrimination, and dynamic optimization functions, it forms an engineering-ready health assessment solution, supporting automated analysis and continuous optimization in real-world scenarios. It is compatible with speech data from different age groups and regions, and can be extended to various scenarios such as medical health monitoring, mental health screening, and occupational health assessment.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart illustrating the steps of a method for extracting and quantifying global speech features reflecting health status, provided in an embodiment of this application. Figure 2 This is a schematic block diagram of a speech global feature extraction and quantification system reflecting health status provided in an embodiment of this application; Figure 3 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0022] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0023] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0025] In the field of speech signal processing and health assessment, most existing technologies are limited to single-dimensional speech feature analysis (such as extracting only fundamental frequency or speech rate features) or only targeting single objectives such as speaker identification and emotion analysis. They lack multi-dimensional global correlation analysis of the life characteristics (gender, age, region) and health characteristics (vitality, mental state, physical condition, etc.) implicit in speech signals. Traditional methods usually use fixed feature sets or simple statistical models, which cannot effectively capture the complex physiological and health correlation information in speech signals, and have not built an intelligent system for the entire process from data acquisition and feature extraction to health assessment. In addition, the combination of speech processing and health assessment in existing technologies is mostly still in the experimental stage, lacking an engineering-featured numerical analysis framework and adaptive optimization mechanism, making it difficult to meet the needs of dynamic health assessment in real-world scenarios.

[0026] Therefore, a method is urgently needed to solve at least one of the above problems.

[0027] To resolve the above issues, please refer to... Figure 1 This application provides a method for extracting and quantifying global voice features reflecting health status, applied to computer devices. The computer devices can be deployed on a single server or server cluster, or on handheld terminals, laptops, wearable devices, or robots. It should be noted that all information involved in the method provided in this application is extracted with the authorization of the relevant user and in accordance with relevant regulations, and will not infringe on user privacy.

[0028] The provided method for global speech feature extraction and quantification reflecting health status includes steps S101 to S103. Details are as follows: Step S101. Collect the speaker's voice signal and simultaneously collect the speaker's vital signs data and health characteristics data. Build a database based on the collected voice signal and the corresponding vital signs data and health characteristics data.

[0029] Specifically, by collecting and synchronously labeling multi-source data, a database containing voice signals and their associated vital and health characteristics is constructed, solving the problems of single data dimensions and insufficient correlation in existing technologies.

[0030] Voice signal acquisition utilizes microphone arrays, mobile phone recording modules, or wearable devices (such as smart bracelets) to collect the speaker's natural voice signals. It supports single-channel or multi-channel acquisition, covering various vocal scenarios such as dialogue, reading aloud, and free expression. Environmental parameters (such as noise decibels and acquisition device model) are recorded simultaneously during the acquisition process for reference in subsequent data preprocessing.

[0031] Vital characteristics can be obtained by entering gender, age, and region (such as place of origin or long-term residence) through the user registration interface, or automatically through ID card OCR recognition; regional characteristics can be further refined into dialect type, accent characteristics, etc.

[0032] Health characteristics include: Subjective data: Collecting subjective indicators such as vitality (e.g., fatigue) and mental state (e.g., anxiety level) through standardized questionnaires (e.g., health self-assessment scales), supporting questionnaire completion on mobile or PC. Objective data: Obtaining physiological indicators (heart rate, blood oxygen, body frequency) by connecting to smart medical devices (e.g., heart rate belts, body fat scales), or extracting historical diagnostic records and medication information from electronic health records. Data synchronization and labeling: Adding timestamps to voice signals and establishing unique association identifiers (e.g., UUIDs) with vital signs and health characteristics data to ensure accurate alignment of multi-source data in the time dimension. Manual or automatic annotation of voice signals, marking key event points (e.g., voice interruptions, emotional fluctuations) to assist in subsequent feature extraction.

[0033] The database is constructed using relational databases (such as MySQL) or non-relational databases (such as MongoDB) to store data. Voice signals are stored in binary format, and text data is stored in JSON or CSV format.

[0034] Remove duplicate samples, filter outliers (such as obviously erroneous age data), and fill in missing values ​​(such as by interpolating from neighboring samples); divide the training set, validation set, and test set into a 7:2:1 ratio to ensure that each subset is evenly distributed across dimensions such as age and gender.

[0035] Step S102. Define global speech features related to the speaker's vital signs and health characteristics, preprocess the speech signals in the database to remove interference signals, extract multi-dimensional global speech features from the preprocessed speech signals to form feature vectors, and then perform dimensionality reduction processing on the feature vectors to obtain dimensionality-reduced feature vectors.

[0036] Specifically, by constructing a multi-dimensional global speech feature system, enhancing signal quality through preprocessing, and eliminating feature redundancy using dimensionality reduction algorithms, the problems of single features and high computational complexity in existing technologies can be solved.

[0037] By establishing a four-dimensional feature system: prosodic features: average speech rate (words / second), intonation fluctuation (fundamental frequency standard deviation), percentage of pause duration between sentences, and stress frequency; phonological features: mean fundamental frequency (F0), first formant frequency (F1), harmonic noise ratio (HNR), jitter, and shimmer; temporal features: mean short-time energy, variance of zero-crossing rate, and percentage of speech activity detection (VAD) duration; and frequency domain features: mean and variance of Mel frequency cepstral coefficients (MFCC), spectral entropy, and percentage of high-frequency energy (>5000Hz).

[0038] Speech signal preprocessing includes: Endpoint detection: using a dual-threshold energy detection algorithm or speech activity detection based on a Hidden Markov Model (HMM) to locate effective speech segments and remove silent and environmental noise segments. Noise reduction: estimating the noise spectrum and subtracting it from the speech signal using spectral subtraction, or using wavelet transform to remove high-frequency interference; for specific scenarios (such as vehicle noise), using adaptive filters (such as the LMS algorithm) to cancel environmental noise. Normalization: standardizing the volume of the speech signal (e.g., adjusting the peak amplitude to -3dBFS) and performing Z-score normalization on the feature parameters to eliminate amplitude differences between different acquisition devices.

[0039] Feature extraction and feature vector construction involve segmenting the preprocessed speech signal into frames (25ms frame length, 10ms frame shift), with a Hamming window added to each frame to reduce spectral leakage. Within each frame, time-domain, frequency-domain, prosodic, and phonological feature parameters are calculated to form a single-frame feature vector. A global feature vector is generated using a sliding window, calculating statistical measures such as the mean, variance, and extrema of the entire speech segment. The multi-dimensional feature parameters are then concatenated into the original feature vector in a preset order (e.g., time-domain first, then frequency-domain), typically with 50-200 dimensions.

[0040] Feature dimensionality reduction includes: Unsupervised dimensionality reduction: Principal component analysis (PCA) is used to calculate the covariance matrix of the eigenvectors, retaining principal components with a cumulative variance contribution rate ≥95%, and projecting the original eigenvectors to a lower-dimensional space (e.g., reducing to 20-50 dimensions). Supervised dimensionality reduction: For data labeled with vital or health characteristics, linear discriminant analysis (LDA) is used to maximize inter-class distance and minimize intra-class distance to achieve discriminative feature extraction.

[0041] Step S103. Using the dimensionality-reduced feature vectors and the corresponding vital signs data and health features data, construct a fusion model and train the fusion model so that it can identify the speaker's vital signs and health features based on the input speech signal feature vectors.

[0042] Specifically, a fusion model based on multi-dimensional features is designed. Through joint training of classification and regression tasks, it can realize the identification of vital characteristics and the numerical prediction of health characteristics, thus solving the problems of weak model generalization ability and single evaluation dimension in existing technologies.

[0043] Model architecture design includes: basic model selection: for small datasets, support vector machines (SVM) or random forest models are used to avoid overfitting; for large datasets, deep learning models such as multilayer perceptrons (MLP), convolutional neural networks (CNN, suitable for spectral features) or recurrent neural networks (RNN, suitable for temporal features) are used.

[0044] The multi-task output layer includes: Vital feature classification: gender (binary classification), age (multi-class classification after binning), and region (dialect-based multi-class classification). The output layer uses the Softmax activation function, and the loss function is cross-entropy. Health feature regression includes: Vital energy index, mental state score (such as anxiety scale score), and physical fitness index (such as fatigue value). The output layer uses the linear activation function, and the loss function is mean squared error (MSE).

[0045] The model training process includes: Input layer: The reduced feature vector (e.g., 30-dimensional) is used as input, and feature transformation is performed through a fully connected layer or embedding layer. Feature fusion layer: Attention mechanism or feature concatenation method is used to integrate the correlation information of features of different dimensions, such as capturing the temporal correlation between prosodic features and health features through gated recurrent units (GRU).

[0046] Joint training optimizes both classification and regression losses simultaneously, with the total loss function being a weighted sum of the two (the weights can be tuned using the validation set). Batch normalization is used to mitigate gradient vanishing, and dropout layers are used to prevent overfitting.

[0047] The optimization algorithm uses the Adam optimizer with an initial learning rate of 0.001, which is dynamically adjusted based on the validation set loss (e.g., using a learning rate decay strategy). If the validation set loss does not decrease for 10 consecutive epochs, training is stopped to avoid overfitting.

[0048] Performance evaluation includes: Vital trait classification: calculating accuracy, precision, recall, and F1 score; Health trait regression: calculating root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The effects of factors such as age and region on health characteristic assessment are analyzed by visualizing the prediction results using confusion matrices and scatter plots.

[0049] Model deployment and updates are achieved by encapsulating the trained model into an API interface, supporting calls from multiple languages ​​such as Python and Java, and deploying it on servers or edge computing devices (such as smart speakers). An online model update mechanism is established: when new data collected in practical applications reaches a preset threshold (e.g., 1000 data points), incremental training is automatically triggered, allowing for rapid adaptation to new scenario data through transfer learning.

[0050] In some embodiments, the acquisition of the speaker's voice signal and the simultaneous acquisition of the speaker's vital signs data and health characteristics data include: acquiring the speaker's voice signal in a natural vocal state through a microphone array or voice acquisition device; simultaneously acquiring the speaker's gender, age, and region's vital signs data through user registration input, sensor detection, or historical data integration; and acquiring the speaker's vitality, mental state, physical condition, and health status's health characteristics data through questionnaires, physiological indicator detection, or health record extraction; and synchronizing the voice signal with the corresponding vital signs data and health characteristics data using timestamps.

[0051] By defining a multi-source data synchronous acquisition mechanism, spatiotemporal alignment of voice signals with vital / health characteristics is achieved.

[0052] The voice acquisition module acquires 16kHz / 16bit PCM format voice signals through a microphone array (such as an 8-channel linear array) or a device with an integrated voice chip (such as a smart speaker), supporting real-time acquisition or file import. Acquisition triggering methods include button activation and voice wake-up (such as "start detection"), ensuring that the speaker is in a natural speaking state (such as daily conversation or reading a specified text).

[0053] Vital characteristics acquisition includes: Gender / Age: manually entered by the user through the registration interface, or automatically recognized by ID card OCR (calling the Optical Character Recognition API); Regional characteristics: automatically labeled based on IP address location or manual selection of administrative division, combined with dialect characteristics of speech recognition (such as Mandarin / Cantonese).

[0054] Health characteristics acquisition includes: Subjective data: Before / after voice collection, standardized scales (such as the PHQ-9 depression scale) are pushed through mobile H5 questionnaires or APPs. After the user selects the scale, a mental status score is automatically generated; Objective data: Real-time heart rate and body movement frequency are obtained through Bluetooth connection to a smartwatch, or blood lipid and blood sugar indicators are retrieved from the hospital's physical examination report over the past year through the API interface.

[0055] A high-precision clock (such as a GPS-synchronized clock) is used to add a UTC timestamp to the voice file. Life / health data is indexed and associated with the voice file through UUID and stored in a distributed file system (such as HDFS).

[0056] In some embodiments, the step of building a database based on the collected voice signals and corresponding vital signs data and health features data includes: storing the labeled voice signals and their associated vital signs data and health features data into the database according to a preset data format; performing deduplication, integrity verification and outlier filtering on the data in the database; dividing the data into training datasets, validation datasets and test datasets according to a preset ratio; and establishing a data indexing mechanism for quick retrieval.

[0057] By building a standardized database, data cleaning, database sharding, and rapid retrieval can be achieved.

[0058] The data storage architecture includes: voice signals stored in WAV format in an object storage service (such as Amazon S3), metadata (timestamps, UUIDs, device IDs) stored in a relational database (MySQL), and life / health characteristics stored in JSON format in a document database (MongoDB), with cross-database associations achieved through ETL tools.

[0059] The data cleaning process includes: Deduplication: Based on a combination query of voice hash value (such as MD5) and user ID, duplicate samples are deleted; Integrity verification: Scan for field missing rates, and mark samples with a missing rate >30% as "to be supplemented" and remind users to complete the information via email; Outlier filtering: Use the Z-score method to detect outliers in numerical features such as age (e.g., >120 years old) and heart rate (<30 or >200 beats / min), automatically mark them as "suspicious data" and manually review them.

[0060] The dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio. Stratified sampling is used to ensure that the gender and age distribution of each subset is consistent with the population (e.g., the training set contains 5000 cases, the validation set contains 1500 cases, and the test set contains 500 cases).

[0061] The indexing mechanism creates a composite index for "user ID + collection time" in MySQL and a text index for "health characteristic labels" in MongoDB, supporting second-level retrieval of data for specific populations (such as samples with "age > 60 years old and anxiety score > 7 points").

[0062] In some embodiments, the definition of global speech features related to the speaker's vital signs and health characteristics includes: defining a multi-dimensional global feature system encompassing speech prosody features, phonological features, temporal statistical features, and frequency domain distribution features, wherein prosodic features include average speech rate, intonation fluctuation, and pause frequency; phonological features include fundamental frequency mean, formant frequency, and harmonic-to-noise ratio; temporal features include short-time energy mean and zero-crossing rate variance; and frequency domain features include spectral entropy and Mel frequency cepstral coefficient mean.

[0063] A multi-dimensional speech feature system is established, covering global statistics in prosody, tone quality, and time-frequency domain.

[0064] Prosodic feature calculation includes: average speech rate: obtaining text transcription results through speech recognition APIs (such as Google Speech-to-Text) and calculating the ratio of total word count to speech duration (words / second); intonation fluctuation: extracting the fundamental frequency curve (using the Praat algorithm) and calculating the absolute mean of the fundamental frequency difference between adjacent syllables; pause frequency: based on endpoint detection results, counting the number of pauses between sentences per minute (pause duration > 200ms).

[0065] The calculation of sound quality features includes: mean fundamental frequency: the average fundamental frequency value of the effective speech segment is taken, which is usually 85-155Hz for males and 165-255Hz for females; formant frequencies: the F1, F2, and F3 frequencies are estimated using linear predictive coding (LPC), reflecting the correlation between vocal tract shape and age and gender; harmonic noise ratio (HNR): the harmonic components and noise components in the speech are separated by harmonic subtraction, and the energy ratio (dB) is calculated.

[0066] The time-domain / frequency-domain feature calculation includes: short-time energy mean: calculate the energy value of each frame of the framed signal (frame length 25ms) and take the global mean; Mel frequency cepstral coefficient (MFCC) mean: convert the linear spectrum to Mel spectrum through the Mel filter bank, and take the mean of the first 13 coefficients after discrete cosine transform.

[0067] In some embodiments, the preprocessing of the speech signals in the database to remove interference signals includes: performing endpoint detection on the speech signals to locate valid speech segments, using spectral subtraction or wavelet denoising algorithms to remove environmental noise and equipment interference, performing volume normalization processing on the denoised speech signals, adjusting the signal amplitude to a preset dynamic range, and eliminating volume differences between different acquisition devices.

[0068] The reliability of feature extraction is improved by designing a speech signal preprocessing pipeline.

[0069] Endpoint detection (VAD) employs a dual-threshold algorithm: setting a high-energy threshold (e.g., -30dB) and a low-energy threshold (-50dB), and combining zero-crossing rate features (>100 times / second) to locate speech segments and remove silence and non-speech noise (e.g., coughing, keyboard sounds).

[0070] Noise reduction processing includes: Spectral subtraction: estimating the noise spectrum in the silent segment of speech and subtracting the noise estimate from the noisy speech spectrum, which is suitable for stationary noise (such as fan noise); Wavelet denoising: performing wavelet decomposition on the speech signal and setting a soft threshold in the high-frequency subband to remove impulse noise (such as knocking sound), which improves the signal-to-noise ratio of the reconstructed signal by 3-5dB.

[0071] Volume normalization uses peak normalization to scale the signal amplitude to the range of [-1, 1] (corresponding to -32768~32767 for 16-bit quantization), or uses root mean square (RMS) normalization to keep the average energy consistent (e.g., -23 LUFS).

[0072] In some embodiments, the step of extracting multi-dimensional global speech features from the preprocessed speech signal to form a feature vector includes: performing frame-by-frame windowing processing on the effective speech segments, calculating time-domain feature parameters, frequency-domain feature parameters, and prosodic feature parameters in each frame of the signal, concatenating the time-domain feature parameters, frequency-domain feature parameters, and prosodic feature parameters in a preset order to form an original feature vector, and normalizing the original feature vector to eliminate the influence of dimensions.

[0073] By implementing temporal modeling and vector construction of multi-dimensional features.

[0074] Framing and windowing processing divides the effective speech segment into 25ms frames, with a frame shift of 10ms (i.e., 15ms overlap). Hamming windows are used to reduce spectral leakage, generating an N-frame signal sequence (N = speech duration / 10ms). Frame-by-frame feature calculation: For each frame, the following are calculated: Time domain: short-time energy, zero-crossing rate; Frequency domain: MFCC coefficients, spectral entropy (reflecting frequency distribution uniformity); Prosody: current frame fundamental frequency, speech rate change rate (difference from the previous frame's speech rate).

[0075] The global feature vector is constructed by taking statistics (mean, standard deviation, maximum, minimum, median) of each dimension of the N frames to form 4×5=20-dimensional features (4 dimensions × 5 statistics); the original feature vector (e.g., 20-dimensional) is concatenated in the order of "time domain-frequency domain-prosody-tone quality" and the values ​​are scaled to the [0, 1] interval by Min-Max normalization.

[0076] In some embodiments, the dimensionality reduction of the feature vector to obtain the dimensionality-reduced feature vector includes: performing a linear transformation on the feature vector using principal component analysis, calculating the covariance matrix of the feature vector and extracting principal components, filtering principal components according to the cumulative variance contribution rate threshold, and projecting the feature vector onto the principal component space to generate the dimensionality-reduced feature vector; or, using linear discriminant analysis, with vital signs and health feature labels as category supervision information, maximizing inter-class divergence and minimizing intra-class divergence to achieve dimensionality reduction of the feature vector.

[0077] Feature dimensionality reduction methods based on statistical and supervised learning. Principal Component Analysis (PCA) process calculates the covariance matrix of the original feature vectors, solves for eigenvalues ​​and eigenvectors, and sorts the eigenvalues ​​from largest to smallest; selects principal components with a cumulative variance contribution rate ≥ 90% (such as the first 10 principal components), and generates a dimensionality reduction matrix W∈R^(D×d) (D = original dimension, d = target dimension); the dimensionality reduction vector Z = X·W, where X is the normalized original feature matrix (N samples × D dimensions).

[0078] The Linear Discriminant Analysis (LDA) process uses vital characteristic labels (such as gender and age binning) as categories y, calculates the within-class scatter matrix S_w and the between-class scatter matrix S_b, solves the generalized eigenvalue problem S_b w = λS_w w, and selects the eigenvectors corresponding to the top k largest eigenvalues ​​(k ≤ C-1, where C is the number of categories). For health characteristic regression tasks, the LDA-reduced features and regression labels are used to construct a linear model to achieve discriminative feature extraction.

[0079] In some embodiments, the construction of a fusion model using the dimensionality-reduced feature vectors and corresponding vital sign data and health sign data includes: selecting a neural network model, a random forest model, or a support vector machine model as the base model; using the dimensionality-reduced feature vectors as the model input layer; designing the output layer structure according to the number of categories or numerical range of vital sign and health sign data; and establishing a fusion model architecture that includes a feature fusion layer and a classification regression layer. The feature fusion layer is used to integrate multi-dimensional feature information, and the classification regression layer is used to output the prediction results of vital sign categories and the prediction results of health sign values.

[0080] A fusion model architecture supporting multi-task learning was constructed. The basic model selection included: Lightweight scenarios: a random forest model with 100 trees and a maximum depth of 8, suitable for edge devices (such as smart bracelets); High-dimensional feature scenarios: a three-layer fully connected neural network (input layer D-dimensional → hidden layer 128-dimensional → hidden layer 64-dimensional → output layer), with ReLU activation function. Output layer design: Vital feature classification: Gender (2 categories) - 2 neurons in the output layer, Softmax activation; Age (5 levels) - 5 neurons in the output layer, Softmax activation; Health feature regression: Vitality index (0-100 points) - 1 neuron in the output layer, linear activation; Mental state score (continuous value) - 1 neuron in the output layer, linear activation.

[0081] The feature fusion layer uses a concatenate fusion method to concatenate PCA-reduced features with LDA-reduced features (e.g., 20+10=30 dimensions), or calculates the weights of features in each dimension through an attention mechanism (Attention Layer) to highlight features that are strongly correlated with health (e.g., fundamental frequency mean, MFCC variance).

[0082] In some embodiments, training the fusion model to enable it to determine the speaker's vital signs and health characteristics based on the input speech signal feature vector includes: inputting a training dataset into the fusion model; using a cross-entropy loss function as the classification task loss function and a mean squared error loss function as the regression task loss function; updating the model parameters of the fusion model using a stochastic gradient descent algorithm or an adaptive optimization algorithm; periodically evaluating the model's classification accuracy and regression error using a validation dataset during training; adjusting the model's hyperparameters based on the evaluation results until the fusion model's performance on the validation set reaches a preset convergence condition; and verifying the generalization ability of the fusion model using a test dataset.

[0083] By designing a multi-task joint training process, the classification and regression performance is optimized.

[0084] The loss function is defined as follows: Classification loss: L_cls = -Σ(y_true·log(y_pred)), where y_true is the one-hot encoded label and y_pred is the Softmax output probability; Regression loss: L_reg = Σ(y_true - y_pred)^2, where y_true is the true value of the health feature and y_pred is the model's predicted value; Total loss: L_total = L_cls + λ·L_reg, with the hyperparameter λ tuned using the validation set (initial value set to 1.0). The optimization algorithm uses the AdamW optimizer (with weight decay), with an initial learning rate of 1e-3, decaying by a factor of 0.9 every 5 epochs, and a weight decay coefficient of 1e-5; the batch size is set to 32, and gradient clipping is used to prevent gradient explosion (threshold set to 5.0).

[0085] Model evaluation and tuning include: After each training epoch, calculating the following metrics on the validation set: classification metrics: accuracy, F1-score (macro-average); regression metrics: RMSE, MAE, R²; if the loss on the validation set does not decrease after three consecutive epochs, automatically adjusting hyperparameters (e.g., increasing hidden layer neurons, adjusting λ value), or triggering an early stopping mechanism to save the optimal model; finally, validating generalization ability on the test set, requiring gender classification accuracy ≥95%, age grouping accuracy ≥80%, and health feature regression R². 2 ≥0.75.

[0086] Please see Figure 2 As shown, Figure 2 This is a schematic diagram of the structure of a speech global feature extraction and quantification system 200 reflecting health status provided in an embodiment of this application. This speech global feature extraction and quantification system 200 reflects health status and is used to execute the steps of the speech global feature extraction and quantification methods reflecting health status shown in the above embodiments. The speech global feature extraction and quantification system 200 reflecting health status can be a single server or a server cluster, or it can be a terminal, such as a handheld terminal, a laptop computer, a wearable device, or a robot.

[0087] like Figure 2 As shown, the speech global feature extraction and quantification system 200 reflecting health status includes: The signal acquisition unit 201 is used to acquire the speaker's voice signal and simultaneously acquire the speaker's vital signs data and health characteristics data, and build a database based on the acquired voice signal and the corresponding vital signs data and health characteristics data; The feature definition unit 202 is used to define global speech features related to the speaker's vital signs and health characteristics, preprocess the speech signals in the database to remove interference signals, extract multi-dimensional global speech features from the preprocessed speech signals to form feature vectors, and then perform dimensionality reduction processing on the feature vectors to obtain dimensionality-reduced feature vectors. The feature discrimination unit 203 is used to construct a fusion model using the dimensionality-reduced feature vector and the corresponding vital feature data and health feature data, and to train the fusion model so that the fusion model can discriminate the speaker's vital features and health features based on the feature vector of the input speech signal.

[0088] In some embodiments, the acquisition of the speaker's voice signal and the simultaneous acquisition of the speaker's vital signs data and health characteristics data include: acquiring the speaker's voice signal in a natural vocal state through a microphone array or voice acquisition device; simultaneously acquiring the speaker's gender, age, and region's vital signs data through user registration input, sensor detection, or historical data integration; and acquiring the speaker's vitality, mental state, physical condition, and health status's health characteristics data through questionnaires, physiological indicator detection, or health record extraction; and synchronizing the voice signal with the corresponding vital signs data and health characteristics data using timestamps.

[0089] In some embodiments, the step of building a database based on the collected voice signals and corresponding vital signs data and health features data includes: storing the labeled voice signals and their associated vital signs data and health features data into the database according to a preset data format; performing deduplication, integrity verification and outlier filtering on the data in the database; dividing the data into training datasets, validation datasets and test datasets according to a preset ratio; and establishing a data indexing mechanism for quick retrieval.

[0090] In some embodiments, the definition of global speech features related to the speaker's vital signs and health characteristics includes: defining a multi-dimensional global feature system encompassing speech prosody features, phonological features, temporal statistical features, and frequency domain distribution features, wherein prosodic features include average speech rate, intonation fluctuation, and pause frequency; phonological features include fundamental frequency mean, formant frequency, and harmonic-to-noise ratio; temporal features include short-time energy mean and zero-crossing rate variance; and frequency domain features include spectral entropy and Mel frequency cepstral coefficient mean.

[0091] In some embodiments, the preprocessing of the speech signals in the database to remove interference signals includes: performing endpoint detection on the speech signals to locate valid speech segments, using spectral subtraction or wavelet denoising algorithms to remove environmental noise and equipment interference, performing volume normalization processing on the denoised speech signals, adjusting the signal amplitude to a preset dynamic range, and eliminating volume differences between different acquisition devices.

[0092] In some embodiments, the step of extracting multi-dimensional global speech features from the preprocessed speech signal to form a feature vector includes: performing frame-by-frame windowing processing on the effective speech segments, calculating time-domain feature parameters, frequency-domain feature parameters, and prosodic feature parameters in each frame of the signal, concatenating the time-domain feature parameters, frequency-domain feature parameters, and prosodic feature parameters in a preset order to form an original feature vector, and normalizing the original feature vector to eliminate the influence of dimensions.

[0093] In some embodiments, the dimensionality reduction of the feature vector to obtain the dimensionality-reduced feature vector includes: performing a linear transformation on the feature vector using principal component analysis, calculating the covariance matrix of the feature vector and extracting principal components, filtering principal components according to the cumulative variance contribution rate threshold, and projecting the feature vector onto the principal component space to generate the dimensionality-reduced feature vector; or, using linear discriminant analysis, with vital signs and health feature labels as category supervision information, maximizing inter-class divergence and minimizing intra-class divergence to achieve dimensionality reduction of the feature vector.

[0094] In some embodiments, the construction of a fusion model using the dimensionality-reduced feature vectors and corresponding vital sign data and health sign data includes: selecting a neural network model, a random forest model, or a support vector machine model as the base model; using the dimensionality-reduced feature vectors as the model input layer; designing the output layer structure according to the number of categories or numerical range of vital sign and health sign data; and establishing a fusion model architecture that includes a feature fusion layer and a classification regression layer. The feature fusion layer is used to integrate multi-dimensional feature information, and the classification regression layer is used to output the prediction results of vital sign categories and the prediction results of health sign values.

[0095] In some embodiments, training the fusion model to enable it to determine the speaker's vital signs and health characteristics based on the input speech signal feature vector includes: inputting a training dataset into the fusion model; using a cross-entropy loss function as the classification task loss function and a mean squared error loss function as the regression task loss function; updating the model parameters of the fusion model using a stochastic gradient descent algorithm or an adaptive optimization algorithm; periodically evaluating the model's classification accuracy and regression error using a validation dataset during training; adjusting the model's hyperparameters based on the evaluation results until the fusion model's performance on the validation set reaches a preset convergence condition; and verifying the generalization ability of the fusion model using a test dataset.

[0096] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the speech global feature extraction and quantification system reflecting health status and each module described above can be referred to the corresponding content in the various embodiments of the speech global feature extraction and quantification method reflecting health status, and will not be repeated here.

[0097] The aforementioned method for extracting and quantifying global speech features reflecting health status can be implemented as a computer program, which can be used in various ways, such as... Figure 2 It runs on the device shown.

[0098] Please see Figure 3 , Figure 3 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0099] The storage medium may store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform any method for extracting and quantifying global speech features that reflect the health status.

[0100] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0101] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When executed by a processor, the computer program enables the processor to perform any method for extracting and quantifying global speech features that reflect the health status.

[0102] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0103] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0104] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: Collect the speaker's voice signal and simultaneously collect the speaker's vital signs and health characteristics data. Build a database based on the collected voice signal and the corresponding vital signs and health characteristics data. Define global speech features related to the speaker's vital signs and health characteristics, preprocess the speech signals in the database to remove interference signals, extract multi-dimensional global speech features from the preprocessed speech signals to form feature vectors, and then perform dimensionality reduction processing on the feature vectors to obtain dimensionality-reduced feature vectors. By using the dimensionality-reduced feature vectors and the corresponding vital signs and health features data, a fusion model is constructed and trained to enable the fusion model to identify the speaker's vital signs and health features based on the feature vectors of the input speech signal.

[0105] In some embodiments, the acquisition of the speaker's voice signal and the simultaneous acquisition of the speaker's vital signs data and health characteristics data include: acquiring the speaker's voice signal in a natural vocal state through a microphone array or voice acquisition device; simultaneously acquiring the speaker's gender, age, and region's vital signs data through user registration input, sensor detection, or historical data integration; and acquiring the speaker's vitality, mental state, physical condition, and health status's health characteristics data through questionnaires, physiological indicator detection, or health record extraction; and synchronizing the voice signal with the corresponding vital signs data and health characteristics data using timestamps.

[0106] In some embodiments, the step of building a database based on the collected voice signals and corresponding vital signs data and health features data includes: storing the labeled voice signals and their associated vital signs data and health features data into the database according to a preset data format; performing deduplication, integrity verification and outlier filtering on the data in the database; dividing the data into training datasets, validation datasets and test datasets according to a preset ratio; and establishing a data indexing mechanism for quick retrieval.

[0107] In some embodiments, the definition of global speech features related to the speaker's vital signs and health characteristics includes: defining a multi-dimensional global feature system encompassing speech prosody features, phonological features, temporal statistical features, and frequency domain distribution features, wherein prosodic features include average speech rate, intonation fluctuation, and pause frequency; phonological features include fundamental frequency mean, formant frequency, and harmonic-to-noise ratio; temporal features include short-time energy mean and zero-crossing rate variance; and frequency domain features include spectral entropy and Mel frequency cepstral coefficient mean.

[0108] In some embodiments, the preprocessing of the speech signals in the database to remove interference signals includes: performing endpoint detection on the speech signals to locate valid speech segments, using spectral subtraction or wavelet denoising algorithms to remove environmental noise and equipment interference, performing volume normalization processing on the denoised speech signals, adjusting the signal amplitude to a preset dynamic range, and eliminating volume differences between different acquisition devices.

[0109] In some embodiments, the step of extracting multi-dimensional global speech features from the preprocessed speech signal to form a feature vector includes: performing frame-by-frame windowing processing on the effective speech segments, calculating time-domain feature parameters, frequency-domain feature parameters, and prosodic feature parameters in each frame of the signal, concatenating the time-domain feature parameters, frequency-domain feature parameters, and prosodic feature parameters in a preset order to form an original feature vector, and normalizing the original feature vector to eliminate the influence of dimensions.

[0110] In some embodiments, the dimensionality reduction of the feature vector to obtain the dimensionality-reduced feature vector includes: performing a linear transformation on the feature vector using principal component analysis, calculating the covariance matrix of the feature vector and extracting principal components, filtering principal components according to the cumulative variance contribution rate threshold, and projecting the feature vector onto the principal component space to generate the dimensionality-reduced feature vector; or, using linear discriminant analysis, with vital signs and health feature labels as category supervision information, maximizing inter-class divergence and minimizing intra-class divergence to achieve dimensionality reduction of the feature vector.

[0111] In some embodiments, the construction of a fusion model using the dimensionality-reduced feature vectors and corresponding vital sign data and health sign data includes: selecting a neural network model, a random forest model, or a support vector machine model as the base model; using the dimensionality-reduced feature vectors as the model input layer; designing the output layer structure according to the number of categories or numerical range of vital sign and health sign data; and establishing a fusion model architecture that includes a feature fusion layer and a classification regression layer. The feature fusion layer is used to integrate multi-dimensional feature information, and the classification regression layer is used to output the prediction results of vital sign categories and the prediction results of health sign values.

[0112] In some embodiments, training the fusion model to enable it to determine the speaker's vital signs and health characteristics based on the input speech signal feature vector includes: inputting a training dataset into the fusion model; using a cross-entropy loss function as the classification task loss function and a mean squared error loss function as the regression task loss function; updating the model parameters of the fusion model using a stochastic gradient descent algorithm or an adaptive optimization algorithm; periodically evaluating the model's classification accuracy and regression error using a validation dataset during training; adjusting the model's hyperparameters based on the evaluation results until the fusion model's performance on the validation set reaches a preset convergence condition; and verifying the generalization ability of the fusion model using a test dataset.

[0113] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the steps of the speech global feature extraction and quantification method reflecting health status as provided in any embodiment of this application.

[0114] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0115] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for extracting and quantifying global speech features reflecting health status, characterized in that, include: Collect the speaker's voice signal and simultaneously collect the speaker's vital signs and health characteristics data. Build a database based on the collected voice signal and the corresponding vital signs and health characteristics data. Define global speech features related to the speaker's vital signs and health characteristics, preprocess the speech signals in the database to remove interference signals, extract multi-dimensional global speech features from the preprocessed speech signals to form feature vectors, and then perform dimensionality reduction processing on the feature vectors to obtain dimensionality-reduced feature vectors. By using the dimensionality-reduced feature vectors and the corresponding vital signs and health features data, a fusion model is constructed and trained to enable the fusion model to identify the speaker's vital signs and health features based on the feature vectors of the input speech signal.

2. The method according to claim 1, characterized in that, The process of collecting the speaker's voice signal and simultaneously collecting the speaker's vital signs and health data includes: The system collects the speaker's voice signal in a natural speaking state using a microphone array or voice acquisition device. Simultaneously, it obtains the speaker's gender, age, and region through user registration input, sensor detection, or historical data integration. It also obtains the speaker's vitality, mental state, physical condition, and health status through questionnaires, physiological indicator detection, or health record extraction. The voice signal is then time-stamped and synchronized with the corresponding vital sign data and health sign data.

3. The method according to claim 1, characterized in that, The process of building a database based on the collected voice signals and corresponding vital sign data and health sign data includes: The labeled voice signals and their associated vital signs and health characteristics are stored in a database according to a preset data format. The data in the database is deduplicated, verified for integrity, and filtered for outliers. The data is divided into training datasets, validation datasets, and test datasets according to a preset ratio, and a data indexing mechanism is established for quick retrieval.

4. The method according to claim 1, characterized in that, The definition of global speech features related to the speaker's vital signs and health characteristics includes: The definition encompasses a multi-dimensional global feature system that includes prosodic features, phonological features, temporal statistical features, and frequency distribution features. Prosodic features include average speech rate, intonation fluctuation, and pause frequency; phonological features include mean fundamental frequency, formant frequency, and harmonic-to-noise ratio; temporal features include mean short-time energy and zero-crossing rate variance; and frequency features include spectral entropy and mean Mel frequency cepstral coefficients.

5. The method according to claim 1, characterized in that, The preprocessing of the speech signals in the database to remove interference signals includes: Endpoint detection is performed on the speech signal to locate the effective speech segment. Spectral subtraction or wavelet denoising algorithm is used to remove environmental noise and equipment interference. The denoised speech signal is then normalized to adjust the signal amplitude to a preset dynamic range, eliminating volume differences between different acquisition devices.

6. The method according to claim 1, characterized in that, The step of extracting multi-dimensional global speech features from the preprocessed speech signal to form a feature vector includes: The effective speech segments are processed by frame-by-frame windowing. Time-domain feature parameters, frequency-domain feature parameters, and prosodic feature parameters are calculated in each frame of signal. The time-domain feature parameters, frequency-domain feature parameters, and prosodic feature parameters are concatenated into the original feature vector in a preset order. The original feature vector is then normalized to eliminate the influence of dimensions.

7. The method according to claim 1, characterized in that, The dimensionality reduction process of the feature vector to obtain the dimensionality-reduced feature vector includes: Principal component analysis (PCA) is used to linearly transform the eigenvectors, calculate the covariance matrix of the eigenvectors and extract the principal components. The principal components are then selected based on the cumulative variance contribution rate threshold. The eigenvectors are projected onto the principal component space to generate dimensionality-reduced eigenvectors. Alternatively, linear discriminant analysis (LCA) is used, with life characteristic and health characteristic labels as category supervision information, to maximize inter-class divergence and minimize intra-class divergence, thereby achieving dimensionality reduction of the eigenvectors.

8. The method according to claim 1, characterized in that, The process of constructing a fusion model using the dimensionality-reduced feature vectors and corresponding vital sign data and health sign data includes: Neural network model, random forest model or support vector machine model are selected as the base model. The dimensionality-reduced feature vectors are used as the model input layer. The output layer structure is designed according to the number of categories or numerical range of vital features and health features. A fusion model architecture including feature fusion layer and classification regression layer is established. The feature fusion layer is used to integrate multi-dimensional feature information, and the classification regression layer is used to output the prediction results of vital feature categories and the prediction results of health feature values.

9. The method according to claim 1, characterized in that, The training of the fusion model enables it to determine the speaker's vital signs and health characteristics based on the feature vector of the input speech signal, including: The training dataset is input into the fusion model. Cross-entropy loss function is used as the loss function for classification tasks, and mean squared error loss function is used as the loss function for regression tasks. The model parameters of the fusion model are updated using stochastic gradient descent algorithm or adaptive optimization algorithm. During the training process, the model's classification accuracy and regression error are evaluated periodically using the validation dataset. The model hyperparameters are adjusted according to the evaluation results until the performance of the fusion model on the validation set reaches the preset convergence condition. The generalization ability of the fusion model is verified by using the test dataset.

10. A system for extracting and quantifying global speech features reflecting health status, characterized in that, The method applied to any one of claims 1-9 includes: The signal acquisition unit is used to acquire the speaker's voice signal and simultaneously acquire the speaker's vital signs data and health characteristics data, and build a database based on the acquired voice signal and the corresponding vital signs data and health characteristics data; The feature definition unit is used to define global speech features related to the speaker's vital signs and health characteristics. It preprocesses the speech signals in the database to remove interference signals; extracts multi-dimensional global speech features from the preprocessed speech signals to form feature vectors; and then performs dimensionality reduction processing on the feature vectors to obtain dimensionality-reduced feature vectors. The feature discrimination unit is used to construct a fusion model using the dimensionality-reduced feature vectors and the corresponding vital signs and health features data, and to train the fusion model so that it can discriminate the speaker's vital signs and health features based on the feature vectors of the input speech signal.

Images