Friday, 1st March - 11am Herman Kamper : Seminar

Title: Speech systems that emulate human language acquisition  

Abstract:

Most speech systems today rely on supervised learning, where a model is trained on speech paired with transcriptions of the words in the audio. In our group we take a different approach: we try to build speech systems without supervision, taking inspiration from how humans acquire language. In the first part of the talk I will give an example of how this research agenda can be useful from a scientific viewpoint. I will specifically show that a neural model trained on paired speech and images exhibits some of the same biases that young children use to learn words. In the second part of the talk I will show how our research agenda can be useful from an engineering viewpoint. I will talk about using self-supervised models for voice conversion, and how voice conversion can be used to understand what these large spoken language models are learning.

Bio:

Herman is a professor in Electrical and Electronic Engineering at Stellenbosch University, South Africa. Before this, he did a postdoc at TTI-Chicago with Karen Livescu. He obtained his PhD in 2017 from the University of Edinburgh under the supervision of Sharon Goldwater. His group at Stellenbosch works on machine learning methods that would allow machines to acquire language autonomously, using as little supervision as possible. Through this, the hope is to gain new insights into machine and human learning.