Depression detection from speech data using deep learning-based optimized temporal-frequency-channel attention with...

Detecting depression from voice recordings is challenging because acoustic indicators are often highly subtle and vary broadly among individuals. A key drawback of current models is their poor cross-lingual generalization; they often degrade sharply when applied to languages or real-world datasets other than the one used for training. To address these limitations, we propose a lightweight, interpretable deep learning framework that infers depressive states directly from raw speech audio, eliminating the need for transcriptions or visual cues. The framework segments each recording into overlapping short clips, converts them into spectrograms via time-frequency analysis, and feeds them to a streamlined ResNet-18 enhanced with a Temporal-Frequency-Channel Attention (TFCA) unit that selectively highlights diagnostically relevant temporal, frequency, and channel patterns. Another major advancement is the integration of a curvature-sensitive Parameter Optimization with Conscious Allocation using Iterative Intelligence (POCAII) optimization strategy that adaptively tunes key hyperparameters from performance feedback. This yields faster convergence, stronger robustness, and better adaptability to varied acoustic conditions. This approach was evaluated on two established depression speech datasets. On the DAIC-WOZ dataset, we achieved an accuracy of 89.38 % per segment and 93.94 % per subject, with a segment-level Area Under the Receiver Operating Characteristic Curve (AUC) of 95.3 %. On the Androids Corpus, the system reached 89.96 % and 93.23 %, respectively, with a segment-level Area Under the Receiver Operating Characteristic Curve of 95.7 %. Attention visualizations show the model focuses on acoustic cues such as reduced mid-frequency energy and prolonged pauses. These findings demonstrate interpretability and diagnostic utility, offering a solution for depression screening.