Stress is a physiological and psychological response to perceived threats, challenges, or demands and is a prevalent issue in modern society, impacting both mental and physical health. Over the years, methods that utilize biomarkers to determine and measure stress have been extensively explored. Artificial intelligence that uses voice to perceive and detect stress has gained attention, with advancements in voice-based stress detection models offering alternative methods for detecting stress levels. These models use the characteristics of an individual’s voice, which can change under different stress conditions. In this research, we explore the efficiency of voice-based stress detection by experimenting with different combinations of the features that can capture the variations in voice during periods of stress, leading us to extract key acoustic features such as Mel-frequency cepstral coefficients (MFCC), along with zero-crossing rate (ZCR) and root mean square energy (RMSE) from the audio recordings. Our study utilizes two English language audio datasets, focusing on classifying stress into three distinct levels: high, medium, and low stress. We employed a Convolutional Neural Network (CNN) model to train on these datasets with the features extracted, aiming to capture the nuanced variations in the acoustic features associated with different stress levels. Our model achieved an accuracy of 94%, demonstrating the potential of CNN-based approaches in accurate stress detection through voice analysis.