PACB

This project is targeting Self-supervised learning in Audio and Speech workshop, ICML.

Click here for Project Slides

• Proposed a novel self-supervised training scheme to better leverage the large corpus of unlabeled audio data; Designed a problem-agnostic cluster-based pretext task to pretrain the feature extractor;

• Tested the scheme on source separation and speaker classification datasets; the models converge faster and result in higher accuracy in both tasks.