Deep Learning based Classification of Depression and Suicidal Risk Among Normal Speakers using Delta-Spectral Cepstral Coefficients

This study utilizes deep learning techniques to develop a precise model for identifying audio recordings, of female speakers for their mental health conditions. The female speakers were categorized into three volunteer groups, which are remitted (RMT), depressed ( DPR), and high-risk suicidal (HRK) by psychiatrist. The various deep learning techniques such as CNN1D, CNN2D, SVM and LSTM were thoroughly trained and validated with our feature samples which are the Delta MFCC (∆MFCC), representing the vocal-tract frequency response associated with different categorized emotional illnesses of three studied volunteer groups. The experimental results show that the CNN2D model achieved the highest accuracies of 0.95 and 0.99 among studied models in classifying ∆MFCC samples between RMT and DPR groups, and between RMT and HRK groups, respectively. Moreover, precision and recall scores were also robust for the CNN models. In contrast, the SVM model achieved with accuracy of 0.83 as compare to CNN models while its precision and recall found to be adequately high. The findings can significantly contribute to understanding of affected vocal characteristics of spoken sound samples associated with mental illness conditions.