Unsupervised Instance Discriminative Learning for Depression Detection from Speech Signals.

Area of Science:

• Computational Linguistics
• Machine Learning
• Psychiatry

Background:

• Major Depressive Disorder (MDD) affects millions globally, necessitating early and accurate diagnosis.
• Voice analysis offers a non-invasive method for depression detection, crucial due to limited labeled data.
• Unsupervised pre-training methods are vital for extracting meaningful features from scarce datasets.

Purpose of the Study:

• To develop an unsupervised pre-training technique for extracting augment-invariant and instance-spread-out embeddings from speech signals.
• To enhance the accuracy of Major Depressive Disorder detection using voice biomarkers.
• To address the challenge of limited labeled data in depression detection research.

Main Methods:

• A modified Instance Discriminative Learning (IDL) method was employed for unsupervised pre-training.
• Speech augmentation techniques were investigated, with time-masking identified as optimal for learning augment-invariant embeddings.
• Distinct speaker-based sampling and a novel Pseudo Instance-based Sampling (PIS) strategy were explored for learning instance-spread-out embeddings.

Main Results:

• Time-masking proved effective for learning augment-invariant speech embeddings.
• Distinct speaker-based sampling outperformed random sampling, suggesting the importance of preserved speaker information.
• The proposed PIS strategy, combined with DepAudioNet, demonstrated statistically significant improvements in MDD detection on both English and Mandarin datasets.

Conclusions:

• The modified IDL approach, particularly with PIS, enhances the extraction of discriminative voice embeddings for MDD detection.
• This method offers a promising, non-invasive tool for early depression diagnosis, especially valuable in low-data scenarios.
• The findings highlight the potential of unsupervised learning and advanced sampling strategies in clinical voice analysis.