19th March 2021 - 11am - Hao Tang: Seminar

TITLE:  Self-Supervised Speech Representation Learning with Autoregressive Predictive Coding

ABSTRACT: 

Self-supervised learning has been shown to be a scalable and effective approach to learn representations that are widely applicable to a variety of tasks. One form of self-supervised learning is to predict what comes next in the future, a simple yet powerful idea where all the great advances in learning text representations originated. In this talk, I will discuss a slight variant of the simple idea, termed autoregressive predictive coding (APC), for learning representations from unlabeled speech. I will then discuss a vector-quantized extension, or VQ-APC, for learning discrete representations from speech. In addition to demonstrating the usefulness of these representations, I will show analyses to understand what is learned in the representation, why certain information is learned, and how the information is represented. Finally, I will discuss what is impossible to learn with this approach.

BIOGRAPHY:

Hao Tang is a Lecturer in the School of Informatics at the University of Edinburgh. He was a postdoctoral associate working with Jim Glass at Massachusetts Institute of Technology, and he completed his PhD under the advisement of Karen Livescu at Toyota Technological Institute at Chicago. His research focuses on speech processing and related aspects in machine learning. He has received best paper awards in Interspeech and ICASSP for his work on self-supervised learning, low-resource automatic speech recognition, and automatic sign language recognition.

Add to your calendar

 vCal  iCal