This material covers the internal Project Briefing talk by John Bridle and Li Deng on Wed15Jul98

• Background    The basic idea and reasons why some of us are interested in it
• Block Diagram of the basic model: A speech pattern generator based on hidden segmental dynamics with built-in co-articulation
• Research Issues    Some of the choices to be made, and alternatives to explore
• Evaluation and training.
• We start with two different methods, which have different strengths.  Although they were developed independently (by Deng, Schuster and Ma on the one hand and Richards and Bridle on the other) we now have a way of putting them both on the same page. R&B and DSM
• The R&B generative model is really very simple.  Here are some illustrations of the kind of internal dynamics we are using to start with
• The R&B model can be trained using an extension of a multi-layer perceptron (MLP) training program
• The DSM model, which learns the dynamics, and has extra stochastic elements, uses an Extended Kalman Filter algorithm
• The basic R&B model has been trained on synthetic test patterns and on one "real" utterance.
• Status of the DSM work