UniVoice-Net: A Unified Dual-Domain Denoising and Ensemble Feature-Fusion Framework for Voice-Based Depression Screening

Abstract

Automated depression screening from speech has advanced rapidly, yet individual pipelines still specialize in either denoising quality, feature richness, or classifier robustness, rarely all three at once. This working paper proposes UniVoice-Net, a unified voice-analysis framework that combines complementary strengths distilled from two independently validated pipelines into a single four-stage architecture: (i) a dual-domain denoising block that cascades wavelet-based gain-modulated noise suppression with spectral-subtraction and Kalman refinement to handle both stationary and non-stationary background interference; (ii) a fused feature-extraction block that jointly computes cepstral-autocorrelation descriptors and transfer-learned convolutional-recurrent embeddings to capture spectral envelope, pitch dynamics, and contextual temporal structure; (iii) a two-stage feature-selection block that applies a hybrid Grey-Wolf/Krill-Herd metaheuristic ahead of a PCA-t-SNE projection to jointly minimize redundancy and preserve both global and local structure; and (iv) a stacked ensemble classification block that combines a dense neural classifier with a gradient-boosted tree classifier. We describe the architecture, motivate each design choice using the documented, independently reported performance of its constituent components on the DAIC-WOZ clinical interview corpus, and outline the experimental protocol required to validate the unified pipeline empirically.