Self-supervised learning (SSL) refers to an algorithm that employs information extracted from the input data itself as the label for learning latent representation useful for various downstream tasks. Already shown success in CV and NLP domains, SSL is currently actively growing in the speech domain. In this seminar we will have a brief overview on the recent developments on speech SSL. Specifically, we will explore three main approaches on pretext tasks in speech representation learning and what they have achieved.
Ещё видео!