Gaborformer: A method for depression detection through hybrid acoustic feature extraction and fusion

Depression, a complex and heterogeneous mental disorder, poses significant challenges for timely and effective detection, highlighting the need for advanced methodologies. Current technologies often fail to fully leverage deep learning's potential, potentially overlooking critical speech features. To address this, we propose Gaborformer, a novel framework for depression detection through hybrid acoustic feature extraction and fusion. Our approach introduces classified iterative neighborhood component analysis (CINCA) for feature selection, reducing dimensionality and extracting highly correlated depressive speech features. We design GaborNet, integrating learnable Gabor filters with CNN to fuse features, and incorporate a Conformer model to capture depression-related characteristics from high-dimensional signals. Extensive validation on the DAIC-WOZ and MODMA datasets demonstrates the efficacy of our approach, significantly outperforming previous studies. Our analysis underscores the importance of features such as MFCC and LSP in speech-based depression detection (SDD). Code is available at https://github.com/staygoldboy/gaborformer.