Attention guided learnable time-domain filterbanks for speech depression detection.

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning

Background:

• Effective depression screening remains a challenge globally.
• Current speech depression detection (SDD) models often rely on fixed spectral features, limiting fine-grained analysis.
• There is a need for advanced methods to facilitate large-scale depression screening through speech analysis.

Purpose of the Study:

• To develop and evaluate a novel joint learning framework (DALF) for speech depression detection (SDD).
• To enable large-scale screening of depression by improving the accuracy and interpretability of SDD.
• To identify novel acoustic biomarkers for depression from speech signals.

Main Methods:

• Proposed a joint learning framework (DALF) incorporating depression filterbanks features learning (DFBL) and multi-scale spectral attention learning (MSSA).
• DFBL utilizes learnable time-domain filters to extract biologically meaningful acoustic features.
• MSSA guides filters to focus on relevant frequency sub-bands, enhancing feature representation.
• Introduced a new dataset, Neutral Reading-based Audio Corpus (NRAC), for depression analysis.

Main Results:

• The DALF model achieved state-of-the-art performance, with an F1 score of 78.4% on the DAIC-woz dataset.
• On the NRAC dataset, DALF achieved high F1 scores of 87.3% and 81.7% on its two parts.
• Analysis revealed a key frequency range (600-700Hz) as a potential acoustic biomarker for SDD.

Conclusions:

• The DALF model offers a promising and interpretable approach for speech depression detection.
• Learnable time-domain filters and multi-scale spectral attention can effectively capture depression-related speech characteristics.
• The identified frequency range provides valuable insights for developing more effective depression screening tools.