Conversation about Turn-taking: My Research on Learning Turn-taking for Conversational Systems
Time: Tue 2021-11-02 15.15
Location: Fantum and zoom
Participating: Erik Ekstedt
In this seminar I will share my research on modeling conversational behaviours, namely turn-taking, for conversational systems. Turn-taking is how humans organize who is speaking and who is listening throughout a conversation. While I am not completely certain how humans manage this, I try to reason about how AI/Deep-/Machine-learning/linear-algebra-machines could learn this behaviour from data. The lack of high quality, annotated data for (spontaneous) spoken dialog requires thinking about models/modeling/learning that can utilize representations/patterns learned from more accessible, unstructured, data. An endeavour referred to as Self-Supervised Learning, SSL.
My research is about applying Self-supervised learning algorithms (and thinking) on conversational data (text and speech) to model turn-taking behaviour for conversational systems. I will present our two papers about training and using text-based self-supervised models for turn-taking, my current work in the acoustic domain, and future work combining the two.
Have a nice day!