University of Washington
Department of Electrical Engineering
Speech and Signal Processing Seminar

Spontaneous speech poses many challenges to the speech recognition and understanding problem. In this talk we will address the language modeling issues for speech recognition and understanding for large vocabulary spoken dialog systems. We will show how finite state representation and stochastic modeling provide rich tools to model different language models: from n-grams, to word phrases, phrase grammars and head-dependecy structures. Phrase-based models are powerful in that they enhance the traditional n-gram model and allow for a tight integration with the understanding features ("Recognition for Understanding"). We will also present our latest results on automatically learned head-dependency grammars and speech disfluency-based language models. We show that people's responses to computer prompts vary over time and (dialog) state and we propose a framework to track time-varying statistical parameters of a spoken dialog system. The set of algorithms presented will be evaluated over the 100K human-machine dialogs for different ( in time and topic ) "How May I Help You?" speech databases.

Modelling sequential data is important in many areas of science and engineering. Hidden Markov models (HMMs) are popular for this because they are simple and flexible. Dynamic Bayesian Networks (DBNs) generalize HMMs by allowing the state space to be represented in factored form. This gives us the ability to represent much richer models. In this talk, I will discuss the following topics, which illustrate the versatility and difficulties involved with using DBNs. - We show how to represent hierarchical HMMs as DBNs, and thereby speedup inference from O(T^3) to O(T) time, where T is the length of the sequence. We then apply HHMMs to the problem of learning hierarchical models of behavior from some people-tracking data. - Time-space tradeoffs for exact inference in DBNs. We show how to reduce the space requirements from O(T) to O(log T), if we increase running time by a O(log T) factor. This allows us to learn DBNs from very long sequences of data. - Rao-Blackwellised particle filtering (RBPF) for approximate inference in DBNs. RBPF combines exact inference with sequential importance sampling. We illustrate RBPF by applying it to the problem of simultaneous localization and mapping (SLAM) for a simulated mobile robot.

• Winter Quarter, 2002
• Fall Quarter, 2001
• Spring Quarter, 2001
• Fall Quarter, 2000
• Spring Quarter, 2000