Task-Aware Depression Detection from Speech: Feature Extraction and Model Integration in Deep Learning

This study focuses on the 1st SpeechWellness Challenge, aiming to detect depression through speech signals using deep learning methodologies. Emotional cues in speech are critical for accurate depression detection, as they reflect the psychological states of individuals with depression. The proposed framework employs wav2vec 2.0 and HuBERT to extract highlevel acoustic and semantic representations, followed by LightSERNet for emotion-specific feature extraction. The aggregated features are then refined by ECAPA-TDNN for classification. In five-fold cross-validation, wav2vec 2.0 achieved an average accuracy of $64 \%$, while HuBERT attained $61.6 \%$. Model integration yielded an optimal accuracy of $67 \%$. These results demonstrate the efficacy of our approach and underscore the significance of emotion-aware feature extraction in speech-based depression detection.