Speech-based depression detection using higher-order spectral features and a multi-level transformer

A precise and prompt automated depression detection system utilizing auditory signals is essential, considering the scarcity of psychiatrists and the exorbitant expense of clinical diagnosis, resulting in approximately 60% of psychiatric patients globally lacking access to mental health care. This study proposes a Depression Detection model utilizing voice, which integrates hierarchical higher-order spectral distribution with deep learning models to overcome these issues. The utilized dataset is the Distress Analysis Interview Corpus, Wizard of Oz (DAIC-WOZ), comprising audio recordings of clinical interviews designed to facilitate the detection of psychological distress disorders such as depression, anxiety, and post-traumatic stress disorder. Two categories of features are extracted: statistical, handmade and bispectral features, as well as deep features utilizing the Multi-level Convolutional Transformer with attention learning (ML-CoT-AL) model. The ML-CoT-AL integrates the Convolutional Transformer (CoT), Channel and Element-wise Attention Module (CEAM), and Negotiator Modules (NM). The study presents the RIME optimization technique utilizing ML-CoT-AL for hyperparameter tuning, leading to improved accuracy in multi-level depression identification.