SonoMind: deep learning-based voice analysis for mental health monitoring

Depressive disorder is a major global health concern and is often underdiagnosed due to stigma, unreliable self-reporting, and limited access to proper mental health screening tools. Despite recent advances, most existing automated approaches depend on questionnaires or multi-modal data, while efficient and reliable voice-only clinical detection frameworks remain limited, creating a clear research gap. Motivated by the need for a non-invasive, objective, and privacy-preserving diagnostic alternative, this study proposes SonoMind, an adaptive deep learning framework for early depression detection using voice signals. The methodology incorporates Adaptable Spectral Pairing for effective noise reduction, SynchroSonic Learning for synchronized feature extraction, and Adaptive Krill-Wolf Optimization for optimal feature selection, followed by a neural classification stage. The framework was evaluated using the publicly available DAIC-WOZ clinical interview dataset. Experimental results show that SonoMind achieves 97.22% accuracy, 100% precision, 95.92% recall, 97.92% F1-score, MAE of 0.027, and RMSE of 0.1666. These results confirm the robustness and reliability of the system, demonstrating its potential as a scalable and supportive tool for mental health professionals in voice-based depression screening.