Estimation of habit-related information from male voice data using machine learning-based methods

Abstract According to a survey on the cause of death among Japanese people, lifestyle-related diseases (such as malignant neoplasms, cardiovascular diseases, and pneumonia) account for 55.8% all deaths. Three habits, namely, drinking, smoking, sleeping, are considered most important factors associated with but it is difficult measure these habits autonomously regularly. Here, we propose machine learning-based approach detecting lifestyle using voice data. We used classifiers probabilistic linear discriminant analysis based acoustic features, such mel-frequency cepstrum coefficients (MFCCs) jitter, extracted from speech dataset developed, an X-vector pre-trained ECAPA-TDNN model. For training models, several implemented in MATLAB 2021b, support vector machines, K-nearest neighbors (KNN), ensemble methods some feature-projection options. Our results show that cubic KNN method features performs well sleep habit classification, while X-vector-based models perform smoking drinking classifications. These suggest X-vectors may help estimate directly affecting vocal cords tracts users (e.g., due drinking), classify chronotypes, which might be informative respect individuals’ cord tract ultrastructure.