A Machine Learning-Based Approach For Stress Detection In Sports Students Using Vocal Analysis

Abstract

Sports students frequently encounter considerable stress due to the dual demands of intense physical training and academic responsibilities. Conventional methods for detecting stress, such as self-reports and clinical evaluations, tend to be subjective and unsuitable for real-time applications. This study investigates the use of machine learning to objectively detect stress among different students from sports background examining their vocal traits. Speech samples were gathered from 670 undergraduates at the North Eastern Regional Centre (NERC) of the Lakshmibai National Institute of Physical Education (LNIPE), which is located in Sonapur, near Guwahati, and stress levels were confirmed using the Beck Depression Inventory (BDI). Essential acoustic features, such as Mel-Frequency Cepstral Coefficients (MFCC), pitch, Zero-Crossing Rate (ZCR), and Formant frequencies, were derived from the audio recordings. Three machine learning algorithms - Support Vector Machine (SVM), K-Nearest Neighbours (KNN), and Long Short-Term Memory (LSTM) RNN - were utilized to categorize stress levels. The findings showed that KNN surpassed the other models, achieving the highest accuracy (94.72%) and F1-score (94.67%), followed by LSTM at 90.90% accuracy. The SVM exhibited the lowest performance (62.01% accuracy), underscoring its challenges in managing intricate vocal stress features. These results indicate that machine learning-based vocal analysis offers a promising method for real-time stress detection in sports students, potentially facilitating early intervention and better stress management strategies.