This talk briefly reviews the probabilistic framework of modern HMM systems and then presents an overview of the refinements needed to achieve state of the art performance. These include the use of quinphones, soft-tied decision trees, full variance normalisation, speaker normalisation, speaker adaptation, model merging, interpolated language models, and word posterior probability decoding. The talk concludes by highlighting some specific areas where current architectures are particularly constraining and suggests possible future areas of development.

Mathematical Foundations of Speech Processing and Recognition

2000-2001 Program: Mathematics in Multimedia