Context dependent tree based transforms for phonetic speech recognition

Abstract

This paper presents a novel method for modeling phonetic context using linear context transforms. Initial investigations have shown the feasibility of synthesising context dependent models from context independent models through weighted interpolation of the peripheral states of a given hidden markov model with its adjacent model. This idea can be further extended, to maximum likelihood estimation of not only single weights, but a matrix of weights or a transform. This paper outlines the application of Maximum Likelihood Linear Regression (MLLR) as a means of modeling context dependency in continuous density Hidden Markov Models (HMM).