A Hybrid Deep Learning Framework for Marathi Speech-Based Stress Detection Using GRU and Handcrafted Audio Features

Stress is one of the major factors affecting mental and physical health and requires the development of new non-invasive, early, and accessible detection methods. Prosodic and acoustic features indicative of emotional states can be employed for voice-based stress detection, which provides a non-invasive method. However, research on regional languages, such as Marathi, is scarce. This paper presented a hybrid deep learning-based framework for stress finding in Marathi speech that combines hand-tuned audio features with deep temporal representation using a Gated Recurrent Unit (GRU)-based network. The system applies pre-processing steps such as removing silent portions and reducing background noise in order to improve overall robustness. System extracted 19 handcrafted features, including MFCCs, Chroma, spectral contrast, zero-crossing rate, and spectral rolloff from each audio clip. Simultaneously, sequences of MFCCs are input to a GRU to model the temporal information. The outputs of the two feature branches are then concatenated and input into the fully connected layers to perform classification. The developed method demonstrated a 92% accuracy with F1-score of roughly 0.92 when examined in the regional Marathi language. This system has also been tested on CNN and RNN with accuracies of 84% and 78%, and the results show that the integration of statistical and temporal features with improved pre-processing leads to an improvement in stress detection performance, using which a scalable solution can be embedded for monitoring mental health in the context of a own-resource language.