Seminario Conversational AI in the Era of LLMs

Abstract: Large Language Models (LLMs) are now part of our daily lives and play a central role in modern conversational AI. Despite their impressive abilities, they still have important limitations. In this talk, I will discuss the key challenges our lab is addressing, with a particular focus on multimodality and model interpretability. First, I will highlight the need for LLMs that integrate speech and audio more effectively, presenting our recent progress on low-bit-rate audio tokenization methods such as FocalCodec. I will then discuss the growing importance of interpretability and share our recent ideas for making deep learning models more transparent. Finally, I will present our plans for SpeechBrain 2.0, our open-source toolkit for speech processing, and show how we are updating it for the LLM era.

Bio: Mirco Ravanelli is Assistant Professor at Concordia University (Gina Cody School of Engineering and Computer Science) and Adjunct Professor at Université de Montréal. He is an Associate Member at Mila - Quebec AI Institute and creator of SpeechBrain, an open-source toolkit for speech processing. His research focuses on deep learning for conversational AI, with over 80 published papers in the field. He received his Ph.D. with distinction from the University of Trento in 2017 and was honored with the 2022 Amazon Research Award. https://scholar.google.com/citations?user=-6Pj3IYAAAAJ&hl=en