Voice voiceprint analysis and numeralization method and system for reflecting yang deficiency state of human body

Abstract

The invention provides a voice voiceprint analysis and numeralization method and system for reflecting the yang deficiency state of a human body. The method comprises the steps that a preset number of human body voice data are acquired through voice acquisition equipment, the voice data comprise voice data of people in a yang-deficiency state and voice data of people in a non-yang-deficiency state, and yang-deficiency state labeling information of human body states corresponding to the voice data is acquired; for each piece of voice data, voice global features are extracted from the voice data by using a feature extraction algorithm, and the voice global features comprise a fundamental frequency feature, an energy feature and a Mel-frequency cepstrum coefficient feature; screening the extracted voice global features by using a feature selection algorithm to obtain a feature subset related to the yang deficiency state; constructing a yang deficiency state feature voiceprint group model based on the feature subset; and inputting to-be-analyzed voice data into the yang deficiency state feature voiceprint group model, and calculating and outputting a numerical result reflecting the yang deficiency state of the human body.

Description

Technical Field

[0001] This application relates to the technical field, and in particular to a method and system for voiceprint analysis and quantification that reflects the state of Yang deficiency in the human body. Background Technology

[0002] In Traditional Chinese Medicine (TCM) theory, the diagnosis of Yang deficiency mainly relies on the patient's subjective description of symptoms (such as aversion to cold, cold limbs, fatigue, etc.) and the TCM practitioner's subjective judgment of tongue and pulse characteristics, which presents the following technical bottlenecks: 1. Lack of objective quantitative standards: Traditional diagnostic methods rely on human experience, and different physicians have different judgment standards, making it difficult to achieve standardization and traceability of diagnostic results; 2. Data collection limitations: Existing technologies have not established a correlation between Yang deficiency and voice characteristics, making it impossible to remotely collect physiological data through non-invasive means, which limits the feasibility of large-scale population screening. 3. Lack of technological application: Currently, voiceprint analysis technology is mainly used in fields such as identity recognition and sentiment analysis. However, there is no technical solution to model the correlation between global speech features such as fundamental frequency features, energy features, and Mel frequency cepstral coefficients and the Yang deficiency state. Furthermore, there is no technical path to construct a voiceprint group model of the Yang deficiency state features through machine learning algorithms and output numerical evaluation results.

[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 voiceprint analysis and quantification method and system that reflects the state of Yang deficiency in the human body. It aims to solve the problems in traditional Chinese medicine theory that the diagnosis of Yang deficiency mainly relies on the patient's subjective description of symptoms (such as aversion to cold, cold limbs, fatigue, etc.) and the subjective judgment of TCM doctors on tongue and pulse, which lacks objective quantitative standards, has limitations in data collection, and has gaps in technology application.

[0005] In a first aspect, embodiments of this application provide a method for voiceprint analysis and quantification reflecting a state of Yang deficiency in the human body, the method comprising: A preset number of human voice data are acquired through a voice acquisition device, wherein the voice data includes voice data of people in a state of Yang deficiency and voice data of people in a state of non-Yang deficiency, and the Yang deficiency status labeling information of the human body corresponding to the voice data is acquired. For each piece of speech data, a feature extraction algorithm is used to extract global speech features from the speech data. The global speech features include fundamental frequency features, energy features, and Mel frequency cepstral coefficient features. A feature selection algorithm is used to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state. Based on the feature subset, a machine learning algorithm is used to train the speech data of people with the Yang deficiency state and the speech data of people without the Yang deficiency state to construct a Yang deficiency state feature voiceprint group model. The speech data to be analyzed is input into the Yang deficiency state characteristic voiceprint group model. The Yang deficiency state characteristic voiceprint group model calculates and outputs a numerical result reflecting the Yang deficiency state of the human body. The numerical result is used to characterize the probability or degree of the human body being in a Yang deficiency state.

[0006] In some embodiments, acquiring a preset number of human voice data through a voice acquisition device includes: stratifying the target population according to age, gender, and region using an intelligent stratified sampling algorithm to control the sample ratio between people in a Yang deficiency state and those in a non-Yang deficiency state; synchronously acquiring environmental noise data corresponding to the voice data using the voice acquisition device, and performing noise reduction preprocessing on the original voice data using an adaptive filtering algorithm; and analyzing symptom descriptions in electronic health records using a natural language processing algorithm to automatically associate the voice data with Yang deficiency state annotation information, wherein the annotation information includes Yang deficiency, non-Yang deficiency, and Yang deficiency degree grading labels.

[0007] In some embodiments, the step of extracting global speech features from each speech data using a feature extraction algorithm includes: after performing frame-by-frame windowing processing on the speech data, aligning speech segments at different speech rates using a dynamic time warping algorithm; in addition to fundamental frequency features, energy features, and Mel frequency cepstral coefficient features, extracting first-order and second-order difference features of fundamental frequency features, energy features, and Mel frequency cepstral coefficient features to characterize dynamic feature changes; and extracting features from the framed speech spectrogram using a pre-trained convolutional neural network model to generate a deep feature vector containing frequency domain context information, and concatenating and fusing it with traditional speech features.

[0008] In some embodiments, the step of using a feature selection algorithm to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state includes: calculating the importance score of each feature using a random forest algorithm and filtering features with an importance score higher than a preset threshold; using a recursive feature elimination algorithm combined with a logistic regression model to iteratively optimize the feature subset, removing the features that contribute the least to the model performance each time; and using an automatic feature selection neural network with the classification accuracy of the feature subset as the objective function, using a reinforcement learning algorithm to search for the optimal feature combination.

[0009] In some embodiments, the step of training the speech data of people in a state of Yang deficiency and those in a state of non-Yang deficiency using machine learning algorithms based on the feature subset to construct a characteristic voiceprint group model for the Yang deficiency state includes: dividing the feature subset into a training set, a validation set, and a test set; optimizing the model hyperparameters using a k-fold cross-validation strategy; constructing a temporal feature modeling module based on a long short-term memory neural network; automatically learning the feature weights of different speech segments by combining an attention mechanism; and using an ensemble learning framework to fuse the prediction results of support vector machines, gradient boosting decision trees, and deep belief networks, and generating the final characteristic voiceprint group model for the Yang deficiency state through a voting mechanism.

[0010] In some embodiments, the step of inputting the speech data to be analyzed into the Yang deficiency state feature voiceprint group model and calculating and outputting a numerical result reflecting the Yang deficiency state of the human body through the Yang deficiency state feature voiceprint group model includes: performing the speech data to be analyzed through the same preprocessing and feature extraction process as the training data to generate an input feature vector; calculating a probability value through the Yang deficiency state feature voiceprint group model, the probability value representing the probability that the human body is in a Yang deficiency state; constructing a posterior probability correction module based on a Bayesian network, calibrating the model output probability by combining prior information of user age and gender, and outputting a numerical result containing a confidence interval.

[0011] In some embodiments, the method further includes: establishing a real-time speech data acquisition queue; triggering an online learning mechanism when the number of newly added labeled speech data reaches a preset amount; updating and training the Yangxu state feature voiceprint group model through an incremental learning algorithm; using gradient accumulation technology to avoid overfitting of the Yangxu state feature voiceprint group model on a small amount of new data; and using a model performance monitoring algorithm to evaluate the classification accuracy and loss value of the updated Yangxu state feature voiceprint group model on the validation set in real time, and automatically rolling back to the previous version of the Yangxu state feature voiceprint group model when the performance indicators do not meet the standards.

[0012] In some embodiments, the method further includes: constructing a cross-domain voiceprint pre-training model and performing unsupervised pre-training using a large-scale non-medical voiceprint dataset; during the training process of the Yang deficiency state feature voiceprint group model, transferring the high-level feature extraction layer parameters of the cross-domain voiceprint pre-training model to the current Yang deficiency state feature voiceprint group model, and fine-tuning it for medical speech data; reducing the distribution difference between medical speech data and non-medical speech data through a domain adversarial neural network, thereby improving the generalization ability of the Yang deficiency state feature voiceprint group model on small sample Yang deficiency data.

[0013] In some embodiments, obtaining the labeling information of the Yang deficiency state corresponding to the human body state of the voice data includes: parsing the diagnostic text in the electronic medical record using a natural language processing algorithm, extracting keywords related to Yang deficiency symptoms, and generating a labeling confidence score based on the frequency of keyword occurrence and semantic relevance; the keywords related to Yang deficiency symptoms include at least aversion to cold, cold limbs, and fatigue; using an expert system to perform secondary verification of the automatic labeling results, establishing a labeling conflict resolution mechanism, and manually reviewing samples with confidence scores below a threshold; and using blockchain technology to store the labeling process and results to ensure the immutability and traceability of the labeling information.

[0014] Secondly, this application provides a voiceprint analysis and quantification system reflecting the state of Yang deficiency in the human body, the system comprising: The data acquisition unit is used to acquire a preset number of human voice data through a voice acquisition device, wherein the voice data includes voice data of people in a state of Yang deficiency and voice data of people in a state of non-Yang deficiency, and to acquire the Yang deficiency status label information of the human body corresponding to the voice data. The model building unit is used to extract global speech features from each speech data using a feature extraction algorithm. The global speech features include fundamental frequency features, energy features, and Mel-frequency cepstral coefficient features. A feature selection algorithm is used to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state. Based on the feature subset, a machine learning algorithm is used to train the speech data of people with the Yang deficiency state and people without the Yang deficiency state to construct a Yang deficiency state feature voiceprint group model. The result output unit is used to input the speech data to be analyzed into the Yang deficiency state characteristic voiceprint group model, and calculate and output a numerical result reflecting the Yang deficiency state of the human body through the Yang deficiency state characteristic voiceprint group model. The numerical result is used to characterize the probability or degree of the human body being in a Yang deficiency state.

[0015] This application transforms the Yang deficiency state into quantifiable numerical results (such as probability values ​​or severity grading) through feature extraction and machine learning modeling of speech data, reducing subjective errors in traditional diagnosis and improving the scientific rigor and consistency of diagnostic results. Data acquisition can be completed using speech acquisition devices without relying on specialized medical equipment or invasive examinations, making it suitable for scenarios such as telemedicine and sub-health screening, thus lowering the diagnostic threshold. By extracting multi-dimensional speech features such as fundamental frequency, energy, and Mel-frequency cepstral coefficients, and combining them with feature selection algorithms to screen key indicators, it can capture subtle dynamic changes in the Yang deficiency state in speech, improving the diagnostic accuracy of the model. It constructs a complete process from data acquisition and feature processing to model training, providing a standardized technical solution for large-scale population studies and clinical auxiliary diagnosis of the Yang deficiency state.

[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 voiceprint analysis and quantification method reflecting the state of Yang deficiency in the human body, provided in an embodiment of this application. Figure 2 This is a schematic block diagram of a voiceprint analysis and quantification system that reflects the state of Yang deficiency in the human body, 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 Traditional Chinese Medicine (TCM) theory, the diagnosis of Yang deficiency mainly relies on the patient's subjective description of symptoms (such as aversion to cold, cold limbs, fatigue, etc.) and the TCM practitioner's subjective judgment of tongue and pulse characteristics, which presents the following technical bottlenecks: 1. Lack of objective quantitative standards: Traditional diagnostic methods rely on human experience, and different physicians have different judgment standards, making it difficult to achieve standardization and traceability of diagnostic results; 2. Data collection limitations: Existing technologies have not established a correlation between Yang deficiency and voice characteristics, making it impossible to remotely collect physiological data through non-invasive means, which limits the feasibility of large-scale population screening. 3. Lack of technological application: Currently, voiceprint analysis technology is mainly used in fields such as identity recognition and sentiment analysis. However, there is no technical solution to model the correlation between global speech features such as fundamental frequency features, energy features, and Mel frequency cepstral coefficients and the Yang deficiency state. Furthermore, there is no technical path to construct a voiceprint group model of the Yang deficiency state features through machine learning algorithms and output numerical evaluation results.

[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 voiceprint analysis and quantification reflecting a state of Yang deficiency in the human body, applied to computer equipment. The computer equipment can be deployed on a single server or server cluster. It can also be deployed on handheld terminals, laptops, wearable devices, or robots, etc. It should be noted that every piece of 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 voiceprint analysis and quantification reflecting the state of Yang deficiency in the human body includes steps S101 to S103. Details are as follows: Step S101. Acquire a preset number of human voice data through a voice acquisition device, wherein the voice data includes voice data of people in a state of Yang deficiency and voice data of people in a state of non-Yang deficiency, and acquire the Yang deficiency status label information of the human body corresponding to the voice data.

[0029] Specifically, voice data of people with Yang deficiency and those without Yang deficiency are acquired through voice acquisition devices, and corresponding Yang deficiency status annotation information is associated with them, thus solving the limitations of data acquisition and the subjectivity of annotation in traditional diagnosis.

[0030] Voice acquisition devices such as microphone arrays, smartphones, and wearable devices (e.g., smartwatches) are used to collect audio data, supporting real-time recording or uploading of pre-stored audio files. The system collects user speech while reading specified text (e.g., standard speech passages) or engaging in natural conversation in a quiet environment, ensuring the speech content contains sufficient rhythm and intonation features. Stratified sampling by age (e.g., children, adults, elderly), gender, and region is employed to broaden sample coverage and avoid bias from single-group interactions.

[0031] Combining TCM diagnostic results from electronic health records (EHRs) (such as confirmed cases of Yang deficiency syndrome), TCM physicians manually annotate user symptoms (such as aversion to cold and cold limbs) according to the standards of "Traditional Chinese Medicine Diagnostics," establishing a binary label of "Yang deficiency" or "non-Yang deficiency," or a graded classification of Yang deficiency severity (such as mild, moderate, and severe). Voice data is anonymized (e.g., using hash values ​​to replace identity identifiers) to ensure data collection complies with relevant regulations and is used only within the scope authorized by the user.

[0032] Step S102. For each piece of speech data, a feature extraction algorithm is used to extract global speech features from the speech data. The global speech features include fundamental frequency features, energy features, and Mel frequency cepstral coefficient features. A feature selection algorithm is used to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state. Based on the feature subset, a machine learning algorithm is used to train the speech data of people with the Yang deficiency state and the speech data of people without the Yang deficiency state to construct a Yang deficiency state feature voiceprint group model.

[0033] Specifically, multi-dimensional global features are extracted from voice data, key indicators are selected through feature selection, and machine learning algorithms are used to build a classification model for Yang deficiency state, thus solving the problems of lack of objective quantitative standards and gaps in technology application.

[0034] Fundamental frequency characteristics are derived by calculating the mean, variance, maximum, and minimum values ​​of the fundamental frequency of a speech signal, reflecting the pitch and stability (a state of Yang deficiency may be accompanied by a low and reduced pitch and fluctuation in speech). Energy features are extracted from speech frame energy and short-time average energy to characterize vocal intensity (people with Yang deficiency may exhibit low speech energy and weak breath). Mel frequency cepstral coefficients (MFCCs) are used to simulate the characteristics of human hearing through a Mel filter bank, extracting 12-40 dimensional MFCC features to capture the dynamic changes in the spectral envelope. First-order and second-order differences are calculated for these features to characterize the rate of change over time, enhancing the ability to model dynamic features.

[0035] The filtering method uses statistical methods such as Pearson correlation coefficient and mutual information to calculate the correlation between each feature and the positive / negative label, and removes features with low correlation. The packaging method uses recursive feature elimination (RFE) combined with a logistic regression model to iteratively select the optimal feature subset. The embedding method uses algorithms such as feature importance scoring of random forest and L1 regularization to automatically filter redundant features and reduce model complexity.

[0036] The labeled speech data was divided into training, validation, and test sets in a 7:2:1 ratio, and k-fold cross-validation (e.g., 10-fold) was used to evaluate the model's generalization ability. Traditional machine learning models include Support Vector Machines (SVM) and Gradient Boosting Decision Trees (GBDT), suitable for small sample scenarios; deep learning models include Recurrent Neural Networks (RNN) and Long Short-Term Memory Networks (LSTM), which capture the temporal dependencies of speech sequences; ensemble learning combines the prediction results of multiple single models (e.g., voting method, stacking method) to improve classification accuracy.

[0037] Using the Yang deficiency state label as a supervision signal, the model's classification accuracy, AUC-ROC and other indicators are optimized, and the output is a "feature voiceprint group model" that can distinguish between Yang deficiency and non-Yang deficiency states.

[0038] Step S103. Input the speech data to be analyzed into the Yang deficiency state characteristic voiceprint group model, and calculate and output a numerical result reflecting the Yang deficiency state of the human body through the Yang deficiency state characteristic voiceprint group model. The numerical result is used to characterize the probability or degree of the human body being in a Yang deficiency state.

[0039] Specifically, the trained model of the voice input to be analyzed is used to output a numerical result reflecting the Yang deficiency state through feature mapping and algorithm calculation, thereby achieving objective quantification and standardization of the diagnostic process.

[0040] Input preprocessing follows the same preprocessing procedure as the training data for the speech to be analyzed: noise reduction (such as Wiener filtering), frame windowing (frame length 25ms, frame shift 10ms), and endpoint detection (removal of silent segments) to ensure that the input feature format is consistent.

[0041] Extract global speech features of the same type (fundamental frequency, energy, MFCC, etc.) and normalize them using the mean / variance of the training set statistics.

[0042] The model calculates the probability value (e.g., 0.1-0.9) of input features belonging to the "Yang deficiency state" through forward propagation; a higher value indicates a higher probability of Yang deficiency. Based on preset thresholds, the probability values ​​are mapped to specific degrees (e.g., probability ≤0.3 for "non-Yang deficiency," 0.3-0.7 for "suspected Yang deficiency," and ≥0.7 for "Yang deficiency"). The numerical results are displayed on the computer interface, along with a ranking of feature importance (e.g., which speech features contribute most to the Yang deficiency state), to help physicians understand the model's decision-making logic. The model's performance (e.g., accuracy, recall) is periodically evaluated using newly collected labeled data. When these metrics decline, model retraining is triggered to ensure long-term effectiveness.

[0043] In some embodiments, acquiring a preset number of human voice data through a voice acquisition device includes: stratifying the target population according to age, gender, and region using an intelligent stratified sampling algorithm to control the sample ratio between people in a Yang deficiency state and those in a non-Yang deficiency state; synchronously acquiring environmental noise data corresponding to the voice data using the voice acquisition device, and performing noise reduction preprocessing on the original voice data using an adaptive filtering algorithm; and analyzing symptom descriptions in electronic health records using a natural language processing algorithm to automatically associate the voice data with Yang deficiency state annotation information, wherein the annotation information includes Yang deficiency, non-Yang deficiency, and Yang deficiency degree grading labels.

[0044] By employing intelligent stratified sampling, noise preprocessing, and automated labeling, we can address the issues of data collection bias and low labeling efficiency, thereby improving sample representativeness and labeling accuracy.

[0045] Intelligent stratified sampling divides the target population into three dimensions: age (<18 years, 18-60 years, >60 years), gender (male / female), and region (geographical divisions such as North China / East China / South China). Proportional stratified sampling is used to control the ratio of Yang deficiency to non-Yang deficiency samples (e.g., 1:3) to avoid a single group dominating model training. Example: Voice samples were collected from 500 Yang deficiency patients (evenly distributed across age groups) and 1500 non-Yang deficiency individuals in North China to ensure a balance of regional characteristics.

[0046] Simultaneously collect environmental noise (such as background human voices and equipment noise), and use Wiener filtering or deep learning-based noise reduction models (such as WaveNet) to denoise the original speech while preserving effective voiceprint features (such as fundamental frequency fluctuations). When collecting speech via mobile phone in a home environment, it automatically identifies and filters interference such as television noise and air conditioner fan noise.

[0047] Automated annotation technology utilizes Named Entity Recognition (NER) technology from Natural Language Processing (NLP) to parse Electronic Health Records (EHRs), extracting keywords related to Yang deficiency symptoms such as "cold intolerance," "cold limbs," and "fatigue." These keywords are then combined with a rule engine to generate annotation labels (e.g., if ≥2 of these keywords appear and a Traditional Chinese Medicine practitioner diagnoses "Yang deficiency"). Based on keyword frequency and semantic relevance (e.g., the co-occurrence probability of "cold intolerance" and "Yang deficiency"), a confidence score of 0-1 is generated. Scores ≥0.8 automatically pass, while scores <0.8 proceed to manual review.

[0048] In some embodiments, the step of extracting global speech features from each speech data using a feature extraction algorithm includes: after performing frame-by-frame windowing processing on the speech data, aligning speech segments at different speech rates using a dynamic time warping algorithm; in addition to fundamental frequency features, energy features, and Mel frequency cepstral coefficient features, extracting first-order and second-order difference features of fundamental frequency features, energy features, and Mel frequency cepstral coefficient features to characterize dynamic feature changes; and extracting features from the framed speech spectrogram using a pre-trained convolutional neural network model to generate a deep feature vector containing frequency domain context information, and concatenating and fusing it with traditional speech features.

[0049] By using speech rate alignment, dynamic feature expansion, and deep feature fusion, the problems of speech temporal differences and single feature dimensions are solved, thereby improving the model's ability to capture subtle features of the Yang deficiency state.

[0050] Speech rate normalization processing involves framing the speech data (25ms) and then applying a window (Hamming window). The Dynamic Time Warping (DTW) algorithm is then used to align speech segments at different speech rates, ensuring that the acoustic features of the same semantic content are aligned on the time axis (e.g., the fundamental frequency curves of the word "fatigue" are aligned when read quickly / slowly).

[0051] Dynamic feature extension calculates the first-order difference (Δ) and second-order difference (Δ) based on the fundamental frequency, energy, and MFCC. 2 This represents the rate of change of a characteristic over time. Example: Individuals with Yang deficiency may exhibit a low fundamental frequency mean and a first-order difference approaching 0 (a flat tone of voice), as indicated by Δ... 2 The feature can capture anomalies in its fundamental frequency stability.

[0052] Deep feature fusion converts speech into a Mel spectrogram as input and uses a pre-trained CNN model (such as VGGish) to extract high-level frequency domain features (such as spectral texture patterns). These features are then concatenated with traditional features (such as MFCC) to form a multimodal feature vector (e.g., 40-dimensional traditional features + 128-dimensional deep features). CNNs can automatically identify patterns of low-frequency energy concentration in weak or effeminate speech (e.g., enhancement of low-frequency components in the spectrum due to reduced vocal cord vibration efficiency).

[0053] In some embodiments, the step of using a feature selection algorithm to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state includes: calculating the importance score of each feature using a random forest algorithm and filtering features with an importance score higher than a preset threshold; using a recursive feature elimination algorithm combined with a logistic regression model to iteratively optimize the feature subset, removing the features that contribute the least to the model performance each time; and using an automatic feature selection neural network with the classification accuracy of the feature subset as the objective function, using a reinforcement learning algorithm to search for the optimal feature combination.

[0054] By combining multiple feature selection algorithms, the problems of feature redundancy and omission of key indicators are solved, thereby improving model training efficiency and classification accuracy.

[0055] Random forest feature ranking uses a random forest model to calculate the Gini importance score of each feature, selects the top 50% of features (such as the fundamental frequency mean, the MFCC third coefficient, etc.), and removes indicators with low correlation to the Yang deficiency state (such as the short-term zero crossing rate).

[0056] Recursive Feature Elimination (RFE) uses logistic regression as the base model and iteratively removes the features that contribute the least to AUC-ROC: initially all features are retained, and each time 10% of the least important features are removed until the performance of the validation set no longer improves, resulting in the optimal feature subset (e.g., retaining 20 key features).

[0057] The reinforcement learning-based automatic selection neural network (FS-Net) uses the classification accuracy of a subset of features as the reward function and searches for the optimal feature combination (e.g., selecting 15 dimensions from 30 features) through the Q-learning algorithm, supporting end-to-end feature selection and model training.

[0058] In some embodiments, the step of training the speech data of people in a state of Yang deficiency and those in a state of non-Yang deficiency using machine learning algorithms based on the feature subset to construct a characteristic voiceprint group model for the Yang deficiency state includes: dividing the feature subset into a training set, a validation set, and a test set; optimizing the model hyperparameters using a k-fold cross-validation strategy; constructing a temporal feature modeling module based on a long short-term memory neural network; automatically learning the feature weights of different speech segments by combining an attention mechanism; and using an ensemble learning framework to fuse the prediction results of support vector machines, gradient boosting decision trees, and deep belief networks, and generating the final characteristic voiceprint group model for the Yang deficiency state through a voting mechanism.

[0059] By employing temporal modeling, attention mechanisms, and ensemble learning, we address the issues of insufficient modeling of speech temporal dependencies and weak generalization ability of single models, thereby improving model robustness.

[0060] The LSTM+attention mechanism constructs a bidirectional LSTM network to capture the contextual dependencies of speech sequences (such as the intonation variations within a sentence) and automatically assigns weights to different speech segments through an attention layer (e.g., focusing on the energy decay features at the end of a sentence). Example: When a person with Yang deficiency states "feeling weak and listless," the energy at the end of the sentence drops sharply; the attention mechanism can enhance the feature weight of this segment.

[0061] Multi-model ensemble learning combines the predictions of SVM (which handles linearly separable features), GBDT (which captures non-linear relationships), and Deep Belief Network (DBN, which extracts hierarchical features). A soft voting mechanism (probability-weighted averaging) is used to generate the final output, reducing the risk of overfitting from a single model. The SVM output probability accounts for 30% of the weight, GBDT for 40%, and DBN for 30%, with the weight parameters fine-tuned using a validation set.

[0062] In some embodiments, the step of inputting the speech data to be analyzed into the Yang deficiency state feature voiceprint group model and calculating and outputting a numerical result reflecting the Yang deficiency state of the human body through the Yang deficiency state feature voiceprint group model includes: performing the speech data to be analyzed through the same preprocessing and feature extraction process as the training data to generate an input feature vector; calculating a probability value through the Yang deficiency state feature voiceprint group model, the probability value representing the probability that the human body is in a Yang deficiency state; constructing a posterior probability correction module based on a Bayesian network, calibrating the model output probability by combining prior information of user age and gender, and outputting a numerical result containing a confidence interval.

[0063] By ensuring consistency in preprocessing and using Bayesian posterior correction, the problems of input distribution bias and insufficient interpretability of results are solved, thereby improving the credibility of clinical applications.

[0064] The standardized preprocessing workflow ensures that the speech to be analyzed undergoes a processing chain that is completely consistent with the training data: noise reduction → framing → endpoint detection → feature extraction → normalization, thus ensuring that the input feature distribution is consistent (such as the same MFCC parameter configuration and the same normalized mean / variance).

[0065] Construct a Bayesian network, combine the model output probability (e.g., the probability of male deficiency is 0.6) with prior information such as user age and gender (e.g., the probability of male deficiency is higher in the elderly), calculate the posterior probability using the Bayesian formula (e.g., the posterior probability of elderly males is calibrated to 0.7), and output a 95% confidence interval (0.65-0.75).

[0066] Application scenarios include situations where teenagers have a high fundamental frequency in their speech, which the model may misjudge as not having Yang deficiency. Combining age priors can correct this bias.

[0067] In some embodiments, the method further includes: establishing a real-time speech data acquisition queue; triggering an online learning mechanism when the number of newly added labeled speech data reaches a preset amount; updating and training the Yangxu state feature voiceprint group model through an incremental learning algorithm; using gradient accumulation technology to avoid overfitting of the Yangxu state feature voiceprint group model on a small amount of new data; and using a model performance monitoring algorithm to evaluate the classification accuracy and loss value of the updated Yangxu state feature voiceprint group model on the validation set in real time, and automatically rolling back to the previous version of the Yangxu state feature voiceprint group model when the performance indicators do not meet the standards.

[0068] By employing online learning, incremental updates, and model monitoring, we address data distribution drift and model aging issues, ensuring long-term diagnostic performance.

[0069] The online learning mechanism establishes a real-time data queue. When ≥1000 new labeled speech data are added, incremental learning is triggered: the model is iteratively updated using stochastic gradient descent (SGD), and the gradient of the new data is calculated and accumulated only, avoiding the high computational cost of retraining with all data.

[0070] Gradient accumulation is used to avoid updating model parameters immediately when there is a small amount of new data (e.g., <500 data points). Instead, the gradients are accumulated in multiple batches and then updated uniformly, which reduces the interference of noisy data on the model.

[0071] The automated rollback mechanism monitors the performance of the validation set in real time. If the model accuracy drops by more than 5% or the loss value increases by more than 10% after the update, it automatically rolls back to the previous version of the model and triggers manual investigation (such as checking the quality of the new data annotation).

[0072] In some embodiments, the method further includes: constructing a cross-domain voiceprint pre-training model and performing unsupervised pre-training using a large-scale non-medical voiceprint dataset; during the training process of the Yang deficiency state feature voiceprint group model, transferring the high-level feature extraction layer parameters of the cross-domain voiceprint pre-training model to the current Yang deficiency state feature voiceprint group model, and fine-tuning it for medical speech data; reducing the distribution difference between medical speech data and non-medical speech data through a domain adversarial neural network, thereby improving the generalization ability of the Yang deficiency state feature voiceprint group model on small sample Yang deficiency data.

[0073] By using pre-trained models and domain adversarial learning, we address the scarcity of labeled data for Yang deficiency and improve the model's generalization ability in small sample scenarios by utilizing non-medical voiceprint data.

[0074] Cross-domain pre-training uses large-scale non-medical voiceprint datasets such as VoxCeleb to pre-train CNN-LSTM models, learn general voiceprint features (such as spectral patterns in speaker recognition), and freeze the parameters of the first 3 convolutional layers of the pre-trained model as the backbone network for feature extraction.

[0075] Domain-adaptive fine-tuning introduces a Domain Adversarial Neural Network (DANN) when fine-tuning the model on hypothetical medical data. Through adversarial training, the model is forced to ignore the distribution differences between medical and non-medical speech (such as differences in recording equipment) and focus on learning common features related to hypothetical medical data (such as fundamental frequency stability). The training objective includes simultaneously minimizing the hypothetical medical data classification loss and the accuracy of the domain discriminator, achieving "domain-invariant" feature learning.

[0076] In some embodiments, obtaining the labeling information of the Yang deficiency state corresponding to the human body state of the voice data includes: parsing the diagnostic text in the electronic medical record using a natural language processing algorithm, extracting keywords related to Yang deficiency symptoms, and generating a labeling confidence score based on the frequency of keyword occurrence and semantic relevance; the keywords related to Yang deficiency symptoms include at least aversion to cold, cold limbs, and fatigue; using an expert system to perform secondary verification of the automatic labeling results, establishing a labeling conflict resolution mechanism, and manually reviewing samples with confidence scores below a threshold; and using blockchain technology to store the labeling process and results to ensure the immutability and traceability of the labeling information.

[0077] By using semantic parsing, expert verification, and blockchain notarization, the problems of subjective labeling and untraceable data are solved, ensuring that the labeling process is trustworthy and controllable.

[0078] Semantic augmentation annotation uses pre-trained language models such as BERT to parse electronic medical record text, extract semantic vectors of Yang deficiency symptoms (such as semantic similarity calculation between "fear of cold and cold limbs" and "insufficient Yang Qi"), and combines them with a rule engine to generate annotation labels, avoiding the limitations of simple keyword matching.

[0079] Human-machine collaborative verification involves constructing an expert system to assign samples with a confidence level <0.7 in the automatic annotation results to TCM doctors for manual review. The final label is determined through an annotation conflict resolution mechanism (such as majority expert voting), reducing the subjective bias of individual doctors.

[0080] Blockchain-based evidence storage technology stores the annotation process (such as annotation time, annotator, and modification records) and results on the blockchain. It uses a hash chain to ensure that the data is immutable and supports tracing the source of annotation during subsequent audits, thus meeting the compliance requirements for medical data.

[0081] Please see Figure 2 As shown, Figure 2This is a schematic diagram of the structure of a voiceprint analysis and quantification system 200 reflecting a state of Yang deficiency in the human body, provided in an embodiment of this application. This voiceprint analysis and quantification system 200 is used to execute the steps of the voiceprint analysis and quantification methods for reflecting a state of Yang deficiency in the human body shown in the above embodiments. The voiceprint analysis and quantification system 200 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.

[0082] like Figure 2 As shown, the voiceprint analysis and numerical system 200 reflecting the state of Yang deficiency in the human body includes: The data acquisition unit 201 is used to acquire a preset number of human voice data through a voice acquisition device, wherein the voice data includes voice data of people in a state of Yang deficiency and voice data of people in a state of non-Yang deficiency, and to acquire the Yang deficiency status label information of the human body corresponding to the voice data. The model building unit 202 is used to extract global speech features from each speech data using a feature extraction algorithm. The global speech features include fundamental frequency features, energy features, and Mel frequency cepstral coefficient features. The extracted global speech features are then filtered using a feature selection algorithm to obtain a feature subset related to the Yang deficiency state. Based on the feature subset, a machine learning algorithm is used to train the speech data of people with the Yang deficiency state and the speech data of people without the Yang deficiency state to construct a Yang deficiency state feature voiceprint group model. The result output unit 203 is used to input the speech data to be analyzed into the Yang deficiency state characteristic voiceprint group model, and calculate and output a numerical result reflecting the Yang deficiency state of the human body through the Yang deficiency state characteristic voiceprint group model. The numerical result is used to characterize the probability or degree of the human body being in a Yang deficiency state.

[0083] In some embodiments, acquiring a preset number of human voice data through a voice acquisition device includes: stratifying the target population according to age, gender, and region using an intelligent stratified sampling algorithm to control the sample ratio between people in a Yang deficiency state and those in a non-Yang deficiency state; synchronously acquiring environmental noise data corresponding to the voice data using the voice acquisition device, and performing noise reduction preprocessing on the original voice data using an adaptive filtering algorithm; and analyzing symptom descriptions in electronic health records using a natural language processing algorithm to automatically associate the voice data with Yang deficiency state annotation information, wherein the annotation information includes Yang deficiency, non-Yang deficiency, and Yang deficiency degree grading labels.

[0084] In some embodiments, the step of extracting global speech features from each speech data using a feature extraction algorithm includes: after performing frame-by-frame windowing processing on the speech data, aligning speech segments at different speech rates using a dynamic time warping algorithm; in addition to fundamental frequency features, energy features, and Mel frequency cepstral coefficient features, extracting first-order and second-order difference features of fundamental frequency features, energy features, and Mel frequency cepstral coefficient features to characterize dynamic feature changes; and extracting features from the framed speech spectrogram using a pre-trained convolutional neural network model to generate a deep feature vector containing frequency domain context information, and concatenating and fusing it with traditional speech features.

[0085] In some embodiments, the step of using a feature selection algorithm to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state includes: calculating the importance score of each feature using a random forest algorithm and filtering features with an importance score higher than a preset threshold; using a recursive feature elimination algorithm combined with a logistic regression model to iteratively optimize the feature subset, removing the features that contribute the least to the model performance each time; and using an automatic feature selection neural network with the classification accuracy of the feature subset as the objective function, using a reinforcement learning algorithm to search for the optimal feature combination.

[0086] In some embodiments, the step of training the speech data of people in a state of Yang deficiency and those in a state of non-Yang deficiency using machine learning algorithms based on the feature subset to construct a characteristic voiceprint group model for the Yang deficiency state includes: dividing the feature subset into a training set, a validation set, and a test set; optimizing the model hyperparameters using a k-fold cross-validation strategy; constructing a temporal feature modeling module based on a long short-term memory neural network; automatically learning the feature weights of different speech segments by combining an attention mechanism; and using an ensemble learning framework to fuse the prediction results of support vector machines, gradient boosting decision trees, and deep belief networks, and generating the final characteristic voiceprint group model for the Yang deficiency state through a voting mechanism.

[0087] In some embodiments, the step of inputting the speech data to be analyzed into the Yang deficiency state feature voiceprint group model and calculating and outputting a numerical result reflecting the Yang deficiency state of the human body through the Yang deficiency state feature voiceprint group model includes: performing the speech data to be analyzed through the same preprocessing and feature extraction process as the training data to generate an input feature vector; calculating a probability value through the Yang deficiency state feature voiceprint group model, the probability value representing the probability that the human body is in a Yang deficiency state; constructing a posterior probability correction module based on a Bayesian network, calibrating the model output probability by combining prior information of user age and gender, and outputting a numerical result containing a confidence interval.

[0088] In some embodiments, the method further includes: establishing a real-time speech data acquisition queue; triggering an online learning mechanism when the number of newly added labeled speech data reaches a preset amount; updating and training the Yangxu state feature voiceprint group model through an incremental learning algorithm; using gradient accumulation technology to avoid overfitting of the Yangxu state feature voiceprint group model on a small amount of new data; and using a model performance monitoring algorithm to evaluate the classification accuracy and loss value of the updated Yangxu state feature voiceprint group model on the validation set in real time, and automatically rolling back to the previous version of the Yangxu state feature voiceprint group model when the performance indicators do not meet the standards.

[0089] In some embodiments, the method further includes: constructing a cross-domain voiceprint pre-training model and performing unsupervised pre-training using a large-scale non-medical voiceprint dataset; during the training process of the Yang deficiency state feature voiceprint group model, transferring the high-level feature extraction layer parameters of the cross-domain voiceprint pre-training model to the current Yang deficiency state feature voiceprint group model, and fine-tuning it for medical speech data; reducing the distribution difference between medical speech data and non-medical speech data through a domain adversarial neural network, thereby improving the generalization ability of the Yang deficiency state feature voiceprint group model on small sample Yang deficiency data.

[0090] In some embodiments, obtaining the labeling information of the Yang deficiency state corresponding to the human body state of the voice data includes: parsing the diagnostic text in the electronic medical record using a natural language processing algorithm, extracting keywords related to Yang deficiency symptoms, and generating a labeling confidence score based on the frequency of keyword occurrence and semantic relevance; the keywords related to Yang deficiency symptoms include at least aversion to cold, cold limbs, and fatigue; using an expert system to perform secondary verification of the automatic labeling results, establishing a labeling conflict resolution mechanism, and manually reviewing samples with confidence scores below a threshold; and using blockchain technology to store the labeling process and results to ensure the immutability and traceability of the labeling information.

[0091] 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 voiceprint analysis and quantification system and its modules that reflect the state of Yang deficiency in the human body described above can be referred to the corresponding content in the various embodiments of the voiceprint analysis and quantification method that reflects the state of Yang deficiency in the human body, and will not be repeated here.

[0092] The aforementioned voiceprint analysis and numerical method reflecting the state of Yang deficiency in the human body can be implemented as a computer program, which can be used in, for example... Figure 2 It runs on the device shown.

[0093] Please see Figure 3 , Figure 3This 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.

[0094] The storage medium can store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform any method of voiceprint analysis and quantification reflecting a state of Yang deficiency in the human body.

[0095] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0096] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to perform any voiceprint analysis and quantification method that reflects the state of Yang deficiency in the human body.

[0097] 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.

[0098] 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.

[0099] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: A preset number of human voice data are acquired through a voice acquisition device, wherein the voice data includes voice data of people in a state of Yang deficiency and voice data of people in a state of non-Yang deficiency, and the Yang deficiency status labeling information of the human body corresponding to the voice data is acquired. For each piece of speech data, a feature extraction algorithm is used to extract global speech features from the speech data. The global speech features include fundamental frequency features, energy features, and Mel frequency cepstral coefficient features. A feature selection algorithm is used to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state. Based on the feature subset, a machine learning algorithm is used to train the speech data of people with the Yang deficiency state and the speech data of people without the Yang deficiency state to construct a Yang deficiency state feature voiceprint group model. The speech data to be analyzed is input into the Yang deficiency state characteristic voiceprint group model. The Yang deficiency state characteristic voiceprint group model calculates and outputs a numerical result reflecting the Yang deficiency state of the human body. The numerical result is used to characterize the probability or degree of the human body being in a Yang deficiency state.

[0100] In some embodiments, acquiring a preset number of human voice data through a voice acquisition device includes: stratifying the target population according to age, gender, and region using an intelligent stratified sampling algorithm to control the sample ratio between people in a Yang deficiency state and those in a non-Yang deficiency state; synchronously acquiring environmental noise data corresponding to the voice data using the voice acquisition device, and performing noise reduction preprocessing on the original voice data using an adaptive filtering algorithm; and analyzing symptom descriptions in electronic health records using a natural language processing algorithm to automatically associate the voice data with Yang deficiency state annotation information, wherein the annotation information includes Yang deficiency, non-Yang deficiency, and Yang deficiency degree grading labels.

[0101] In some embodiments, the step of extracting global speech features from each speech data using a feature extraction algorithm includes: after performing frame-by-frame windowing processing on the speech data, aligning speech segments at different speech rates using a dynamic time warping algorithm; in addition to fundamental frequency features, energy features, and Mel frequency cepstral coefficient features, extracting first-order and second-order difference features of fundamental frequency features, energy features, and Mel frequency cepstral coefficient features to characterize dynamic feature changes; and extracting features from the framed speech spectrogram using a pre-trained convolutional neural network model to generate a deep feature vector containing frequency domain context information, and concatenating and fusing it with traditional speech features.

[0102] In some embodiments, the step of using a feature selection algorithm to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state includes: calculating the importance score of each feature using a random forest algorithm and filtering features with an importance score higher than a preset threshold; using a recursive feature elimination algorithm combined with a logistic regression model to iteratively optimize the feature subset, removing the features that contribute the least to the model performance each time; and using an automatic feature selection neural network with the classification accuracy of the feature subset as the objective function, using a reinforcement learning algorithm to search for the optimal feature combination.

[0103] In some embodiments, the step of training the speech data of people in a state of Yang deficiency and those in a state of non-Yang deficiency using machine learning algorithms based on the feature subset to construct a characteristic voiceprint group model for the Yang deficiency state includes: dividing the feature subset into a training set, a validation set, and a test set; optimizing the model hyperparameters using a k-fold cross-validation strategy; constructing a temporal feature modeling module based on a long short-term memory neural network; automatically learning the feature weights of different speech segments by combining an attention mechanism; and using an ensemble learning framework to fuse the prediction results of support vector machines, gradient boosting decision trees, and deep belief networks, and generating the final characteristic voiceprint group model for the Yang deficiency state through a voting mechanism.

[0104] In some embodiments, the step of inputting the speech data to be analyzed into the Yang deficiency state feature voiceprint group model and calculating and outputting a numerical result reflecting the Yang deficiency state of the human body through the Yang deficiency state feature voiceprint group model includes: performing the speech data to be analyzed through the same preprocessing and feature extraction process as the training data to generate an input feature vector; calculating a probability value through the Yang deficiency state feature voiceprint group model, the probability value representing the probability that the human body is in a Yang deficiency state; constructing a posterior probability correction module based on a Bayesian network, calibrating the model output probability by combining prior information of user age and gender, and outputting a numerical result containing a confidence interval.

[0105] In some embodiments, the method further includes: establishing a real-time speech data acquisition queue; triggering an online learning mechanism when the number of newly added labeled speech data reaches a preset amount; updating and training the Yangxu state feature voiceprint group model through an incremental learning algorithm; using gradient accumulation technology to avoid overfitting of the Yangxu state feature voiceprint group model on a small amount of new data; and using a model performance monitoring algorithm to evaluate the classification accuracy and loss value of the updated Yangxu state feature voiceprint group model on the validation set in real time, and automatically rolling back to the previous version of the Yangxu state feature voiceprint group model when the performance indicators do not meet the standards.

[0106] In some embodiments, the method further includes: constructing a cross-domain voiceprint pre-training model and performing unsupervised pre-training using a large-scale non-medical voiceprint dataset; during the training process of the Yang deficiency state feature voiceprint group model, transferring the high-level feature extraction layer parameters of the cross-domain voiceprint pre-training model to the current Yang deficiency state feature voiceprint group model, and fine-tuning it for medical speech data; reducing the distribution difference between medical speech data and non-medical speech data through a domain adversarial neural network, thereby improving the generalization ability of the Yang deficiency state feature voiceprint group model on small sample Yang deficiency data.

[0107] In some embodiments, obtaining the labeling information of the Yang deficiency state corresponding to the human body state of the voice data includes: parsing the diagnostic text in the electronic medical record using a natural language processing algorithm, extracting keywords related to Yang deficiency symptoms, and generating a labeling confidence score based on the frequency of keyword occurrence and semantic relevance; the keywords related to Yang deficiency symptoms include at least aversion to cold, cold limbs, and fatigue; using an expert system to perform secondary verification of the automatic labeling results, establishing a labeling conflict resolution mechanism, and manually reviewing samples with confidence scores below a threshold; and using blockchain technology to store the labeling process and results to ensure the immutability and traceability of the labeling information.

[0108] 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 voiceprint analysis and quantification method reflecting the state of Yang deficiency in the human body as provided in any embodiment of this application.

[0109] 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.

[0110] 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 voiceprint analysis and quantification reflecting the state of Yang deficiency in the human body, characterized in that, include: A preset number of human voice data are acquired through a voice acquisition device, wherein the voice data includes voice data of people in a state of Yang deficiency and voice data of people in a state of non-Yang deficiency, and the Yang deficiency status labeling information of the human body corresponding to the voice data is acquired. For each piece of speech data, a feature extraction algorithm is used to extract global speech features from the speech data. The global speech features include fundamental frequency features, energy features, and Mel frequency cepstral coefficient features. A feature selection algorithm is used to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state. Based on the feature subset, a machine learning algorithm is used to train the speech data of people with the Yang deficiency state and the speech data of people without the Yang deficiency state to construct a Yang deficiency state feature voiceprint group model. The speech data to be analyzed is input into the Yang deficiency state characteristic voiceprint group model. The Yang deficiency state characteristic voiceprint group model calculates and outputs a numerical result reflecting the Yang deficiency state of the human body. The numerical result is used to characterize the probability or degree of the human body being in a Yang deficiency state.

2. The method according to claim 1, characterized in that, The acquisition of a preset number of human voice data through a voice acquisition device includes: The target population is stratified according to age, gender and region through intelligent stratified sampling algorithm to control the sample ratio between people with Yang deficiency and those without Yang deficiency. The environmental noise data corresponding to the voice data is collected synchronously using a voice acquisition device, and the original voice data is preprocessed for noise reduction using an adaptive filtering algorithm. By analyzing symptom descriptions in electronic health records using natural language processing algorithms, the system automatically associates voice data with Yang deficiency status annotation information, which includes labels for Yang deficiency, non-Yang deficiency, and Yang deficiency degree grading.

3. The method according to claim 1, characterized in that, For each piece of speech data, a feature extraction algorithm is used to extract global speech features from the speech data, including: After performing frame-by-frame windowing on the speech data, a dynamic time warping algorithm is used to align speech segments at different speaking rates. In addition to the fundamental frequency features, energy features, and Mel frequency cepstral coefficient features, first-order and second-order difference features of the fundamental frequency features, energy features, and Mel frequency cepstral coefficient features are extracted to characterize the dynamic changes of the features; The pre-trained convolutional neural network model extracts features from the framed speech spectrogram, generating a deep feature vector containing frequency domain context information, which is then concatenated and fused with traditional speech features.

4. The method according to claim 1, characterized in that, The feature selection algorithm is used to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state, including: The importance score of each feature is calculated using the random forest algorithm, and features with an importance score higher than a preset threshold are selected. The recursive feature elimination algorithm is combined with a logistic regression model to perform iterative optimization of feature subsets, eliminating the features that contribute the least to the model performance each time. By using an automatic feature selection neural network, with the classification accuracy of the feature subset as the objective function, a reinforcement learning algorithm is employed to search for the optimal feature combination.

5. The method according to claim 1, characterized in that, Based on the aforementioned feature subset, a machine learning algorithm is used to train the speech data of individuals in a state of Yang deficiency and those in a state of non-Yang deficiency, thereby constructing a characteristic voiceprint group model for a state of Yang deficiency, including: The feature subset is divided into training set, validation set and test set, and the model hyperparameters are optimized by k-fold cross-validation strategy; A temporal feature modeling module is constructed based on a long short-term memory neural network, and the feature weights of different speech segments are automatically learned by combining an attention mechanism. An ensemble learning framework is adopted, which integrates the prediction results of support vector machine, gradient boosting decision tree and deep belief network, and generates the final voiceprint group model of Yang deficiency state through a voting mechanism.

6. The method according to claim 1, characterized in that, The process of inputting the speech data to be analyzed into the Yang deficiency state characteristic voiceprint group model, and calculating and outputting a numerical result reflecting the Yang deficiency state of the human body through the Yang deficiency state characteristic voiceprint group model, includes: The speech data to be analyzed undergoes the same preprocessing and feature extraction process as the training data to generate an input feature vector. The probability value is calculated by the voiceprint group model of Yang deficiency state, and the probability value represents the probability that the human body is in a Yang deficiency state. A posterior probability correction module is constructed based on a Bayesian network. It combines prior information about the user's age and gender to calibrate the model's output probability, and outputs a numerical result containing a confidence interval.

7. The method according to claim 1, characterized in that, The method further includes: Establish a real-time voice data collection queue. When the number of newly added labeled voice data reaches a preset amount, trigger an online learning mechanism. The voiceprint group model of Yang deficiency state is updated and trained by incremental learning algorithm, and the gradient accumulation technique is used to avoid overfitting of the voiceprint group model of Yang deficiency state on a small amount of new data. A model performance monitoring algorithm is used to evaluate the classification accuracy and loss value of the updated Yang-deficient state feature voiceprint group model on the validation set in real time. When the performance indicators fail to meet the standards, the model is automatically rolled back to the previous version of the Yang-deficient state feature voiceprint group model.

8. The method according to claim 1, characterized in that, The method further includes: Construct a cross-domain voiceprint pre-training model and perform unsupervised pre-training using a large-scale non-medical voiceprint dataset; During the training of the Yang deficiency state characteristic voiceprint group model, the high-level feature extraction layer parameters of the cross-domain voiceprint pre-training model are transferred to the current Yang deficiency state characteristic voiceprint group model, and fine-tuning is performed for medical speech data. By using domain adversarial neural networks to reduce the distribution differences between medical and non-medical speech data, the generalization ability of the Yang deficiency state feature voiceprint group model on small sample Yang deficiency data is improved.

9. The method according to claim 1, characterized in that, The step of obtaining the Yang deficiency state annotation information corresponding to the human body state of the voice data includes: The diagnostic text in the electronic medical record was parsed using a natural language processing algorithm to extract keywords related to Yang deficiency symptoms. Based on the frequency of keyword occurrence and semantic relevance, a label confidence score was generated. Keywords related to Yang deficiency symptoms include at least aversion to cold, cold limbs, and fatigue. The automatic annotation results are validated a second time using an expert system, an annotation conflict resolution mechanism is established, and samples with confidence levels below the threshold are manually reviewed. Blockchain technology is used to record the labeling process and results, ensuring the immutability and traceability of the labeling information.

10. A voiceprint analysis and quantification system reflecting the state of Yang deficiency in the human body, characterized in that, The method applied to any one of claims 1-9 includes: The data acquisition unit is used to acquire a preset number of human voice data through a voice acquisition device, wherein the voice data includes voice data of people in a state of Yang deficiency and voice data of people in a state of non-Yang deficiency, and to acquire the Yang deficiency status label information of the human body corresponding to the voice data. The model building unit is used to extract global speech features from each speech data using a feature extraction algorithm. The global speech features include fundamental frequency features, energy features, and Mel-frequency cepstral coefficient features. A feature selection algorithm is used to filter the extracted global speech features to obtain a feature subset related to the Yang deficiency state. Based on the feature subset, a machine learning algorithm is used to train the speech data of people with the Yang deficiency state and people without the Yang deficiency state to construct a Yang deficiency state feature voiceprint group model. The result output unit is used to input the speech data to be analyzed into the Yang deficiency state characteristic voiceprint group model, and calculate and output a numerical result reflecting the Yang deficiency state of the human body through the Yang deficiency state characteristic voiceprint group model. The numerical result is used to characterize the probability or degree of the human body being in a Yang deficiency state.

Images