Posted September 29, 2025
Pizzas and salads will be delivered at approx. 3:30p; the talk will begin at 4p.
WHAT: TWed Talk: "Toward Fluid AI Conversation with Natural Turn-taking: Full-duplex Modeling with Audio Codec LMs"
WHEN: 4p, Weds, 01 Oct
WHERE: Winslow 1140
VIDEO: TBD
EVENT PAGE: https://bit.ly/472kXTG
Please join us this Wednesday as Abraham Sanders presents what promises to be a hugely interesting TWed Talk, "Toward Fluid AI Conversation with Natural Turn-taking: Full-duplex Modeling with Audio Codec LMs." Pizza arrives approx. 3:30p, the talk begins at 4p.
DESCRIPTION: Humans naturally converse in a “Full-duplex” manner, simultaneously listening, thinking, and speaking at will. Humans continuously perform ultra-low-latency turn-taking decisions that result in a conversation that is mostly coordinated but contains instances of both accidental and intentional overlap. Such overlap includes phenomena such as simultaneous laughter, sentence-completion, backchannel acknowledgements (mhm, yeah!) and interruptions. This talk will lay out the foundation for constructing naturally full-duplex AI conversational agents that replicate these phenomena, delivering a conversational experience that feels more human than AI. I will focus on language modeling techniques for interleaving the audio and text modalities, discuss recent advances in this area, and showcase the system that I am building for my dissertation research.
BIO: Abraham is a 5th year PhD candidate in Cognitive Science working with Dr. Tomek Strzalkowski. Abraham’s research interests include many topics in conversational AI with specific focus on making spoken dialogue systems more human-like. Abraham has previously worked on dialogue systems for DARPA, AFRL & IBM, and developed NLP methods for large-scale social media analytics. Prior to RPI, Abraham worked as a software engineer in the medical industry for nine years.
Pizzas and salads will be delivered at approx. 3:30p; the talk will begin at 4p.