A Bayesian triphone model with parameter tying

This paper introduces a new statistical framework for constructing triphonic models from models of less context-dependency. The new framework is derived from Bayesian statistics, and represents an alternative to other triphone-by-composition techniques, particularly to the model-interpolation and quasi-triphone approaches. The potential power of this new framework is explored by an implementation based on the hidden Markov modeling technique. It is shown that the new model structure includes the quasi-triphone model as a special case, and leads to more efficient parameter estimation than the model-interpolation method. Two strategies of state-level tying have been investigated within the new model structure. Phone recognition experiments on the TIMIT database show an increase in the accuracy over that obtained by other systems.