Enhancing accuracy and privacy in speech-based depression detection through speaker disentanglement

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Psychiatry

Background:

• Speech signals are increasingly recognized as valuable biomarkers for mental health assessment, particularly for the automatic detection of Major Depressive Disorder (MDD).
• Current methods often rely on speaker identity features, raising concerns about patient privacy and potential biases that can impair diagnostic accuracy.
• There is a need for advanced techniques that can extract depression-related information from speech without compromising sensitive speaker-specific data.

Purpose of the Study:

• To develop and evaluate novel methods for disentangling speaker-identity information from depression-related information in speech signals.
• To improve the accuracy of automatic Major Depressive Disorder (MDD) detection while simultaneously enhancing voice privacy.
• To overcome the limitations of existing approaches that over-rely on speaker identity features.

Main Methods:

• Proposed four distinct disentanglement methods: adversarial speaker identification (SID)-loss maximization (ADV), SID-loss equalization with variance (LEV), SID-loss equalization using Cross-Entropy (LECE), and SID-loss equalization using KL divergence (LEKLD).
• Conducted experiments using diverse input features (e.g., ComparE16, raw-audio) and model architectures on the DAIC-WOZ (English) and EATD (Mandarin) datasets.
• Quantified improvements in MDD detection performance (F1-Scores) and voice-privacy attributes using Gain in Voice Distinctiveness ($$) and De-Identification Scores (DeID).

Main Results:

• The LECE method with ComparE16 features achieved state-of-the-art (SOTA) audio-only F1-Scores of 80% for MDD detection on the DAIC-WOZ dataset, with a $$ of -1.1 dB and DeID of 85%.
• The ADV method with raw-audio signals attained an F1-Score of 72.38% on the EATD dataset, surpassing multi-modal SOTA, alongside a $$ of -0.89 dB and DeID of 51.21%.
• All proposed disentanglement methods demonstrated improvements in both depression detection accuracy and voice-privacy metrics compared to baseline approaches.

Conclusions:

• Disentangling speaker-identity information from depression-related cues in speech is a viable strategy for privacy-preserving mental health assessment.
• The proposed methods offer a promising direction for developing robust and ethical speech-based depression detection systems.
• Reducing reliance on speaker identity features enhances system performance and safeguards sensitive patient information.