Maxine Eskenazi

Maxine Eskenazi

Carnegie Mellon University, retiring on December 31, 2022, now at max-esk.com

Research Goals

To create intelligent agents (using spoken dialogue architectures, automatic speech recognition and synthesis) using knowledge of the speech signal and of human cognition. To confront research with real human users and, in turn, provide a real benefit to those users. This endeavor implies studying groups of users, input conditions and speaking styles, the manner in which humans and systems can entrain to one another, and how we can assess the systems we build, often profiting from the wisdom of the crowd.

Invited Speaker:

Invited Speaker: Workshop on conversational AI at NIPS 2018

Best Paper Award:

Zhao, Tiancheng and Eskenazi, Maxine, Zero-Shot Dialog Generation with Cross-Domain Latent Actions, Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dialogue

Published Book:

Eskenazi, Maxine and Devillers, Laurence and Mariani, Joseph, Advanced Social Interaction with Agents , Springer, 2018

Keynote Speaker:

Keynote speech -"Les systemes de dialogue oral: avancees et limites", Journees d'Etudes sur la Parole, Avignon France, June 5, 2018.

Nominated for Best Paper:

Tiancheng Zhao and Maxine Eskenazi, 2016, Towards End-to-End Learning for Dialog State Tracking and Management using Deep Reinforcement Learning, in Proceedings of SIGDIAL 2016 Conference.

Challenge Winner:

Lee, S., Eskenazi, M., 2013, Recipe for building robust spoken dialog state trackers: Dialog state tracking challenge system description, Proceedings of SIGDIAL 2013, Metz. WINNER OF THE 2013 DIALOG STATE TRACKING CHALLENGE

Long Term Real World Applications:

The Letâ€™s Go system answered the phone daily for 11 years (2005-2014) for the Port Authority of Allegheny County for bus scheduling inquiries.

End-to-end Neural Dialog Systems:

The DialPort team was FIRST to apply deep reinforcement learning to end-to-end task-oriented dialog systems. Achieved best results in learning dialog policy and dialog state tracking jointly

Multi-domain Portal:

The DialPort team was FIRST dialog platform to aggregate heterogeneous dialog systems. Links in real time to systems from Cambridge U, USC, U Santa Cruz and soon more.

Real Users:

the DialPort Portal is the ONLY publicly available source of real user dialog data

Entrainment:

Letâ€™s Go was the only system to entrain to the way the user speaks, using their vocabulary, for example. Also was able to get the user to entrain to the way it wants them to speak so that it can better understand them.