[Univ of Cambridge]alt[Dept of Engineering]

MIL Speech Seminars 2004-2005

The MIL Speech Seminar series schedule for the Long Vacation 2005 was as follows:

July 5th 2005 Heiga Zen (PhD student, Nagoya Institute of Technology, Japan) Reformulating the HMM as a trajectory model by imposing explicit relationship between static and dynamic features A trajectory model, derived from the HMM by imposing explicit relationship between static and dynamic features, is developed and evaluated. The derived model, named "trajectory-HMM", can alleviate some limitations of the standard HMM, which are i) piece-wise constant statistics within a state and ii) conditional independence assumption of state output probabilities, without increasing the number of model parameters. In this talk, a Viterbi-type training algorithm is also derived. This model was evaluated both in speech recognition and synthesis experiments. In speaker-dependent continuous speech recognition experiment, the trajectory-HMM achieved error reduction over the standard HMM. The experimental results of subjective listening tests shows that introduction of the trajectory-HMM can improve the quality of synthetic speech generated from HMM-based speech synthesis system which we have proposed.
Auguest 8th 2005 Pascal Poupart, Assistant Professor, School of Computer > Science, University of Waterloo, Canada A Decision-Theoretic Approach to Task Assistance for Persons with Dementia People suffering from dementia (e.g., Alzheimer's disease) often have difficulty completing even simple activities of daily living (ADL) such as toileting, dressing, eating, taking medication, etc. Cognitive assistive technologies hold the promise to provide people suffering from dementia with an increased level of independence. In this talk, I will first describe a system that guides patients with memory deficiencies through the steps of a simple task: handwashing. The system monitors patient progress in the task of handwashing with video-cameras and when necessary, prompts the next step with a verbal cue. The talk will focus on how to design effective prompting strategies that take into account the uncertainty due to the inherent noise of the video sequence as well as the fact that patients do not always follow the prompts. More specifically, I will explain how to model robust prompting strategies with partially observable Markov decision processes. In a second part of my talk, I will present a new technique for scaling POMDP optimization algorithms in domains with continuous observations spaces such as assistive technologies. The rich observation spaces often produced by sensing devices such as video-cameras, microphones or sonars pose significant problems for standard POMDP algorithms that require explicit enumeration of the observations. This problem is usually approached by imposing an a priori discretisation of the observation space, which can be sub-optimal for the decision making task. However, since only those observations that would change the course of action need to be distinguished, the planning task induces a lossless partitioning of the observation space. In this talk, I will explain how to find this partition, and how the resulting discretisation reveals the relevant observation features of the application domain. This will be demonstrated with the handwashing guidance task presented in the first part of the talk. This is work done in collaboration with Jesse Hoey, Alex Mihailidis, Jennifer Boger, Craig Boutilier and Geoff Fernie at the University of Toronto.