Third International Conference on Spoken Language Processing (ICSLP 94)
This paper describes a new technique for performing Linear Discriminant Analysis (LDA) on Hidden Markov Models(HMMs) incorporating state specific mixture densities. Previous work with LDA in speech recognition has used models comprising a unimodal multivariate Gaussian density per state or semi-continuous models using tied densities across states. As the number of models in the mixture densities in a HMM are increased and speech frames are mapped into the new feature space the LDA does not produce the expected unit variance distributions. By treating each state as a single class the assumption that a state can be described as a single multivariate Gaussian is violated. A new technique has been developed which maintains the mapping into the LDA feature space and constructs new HMMs directly from the transformed speech frames, considerably reducing the computation required. An 11% reduction in error rate has been achieved over the standard LDA technique for monophone HMMs and 25% of the system parameters can be discarded without loss in performance.
Bibliographic reference. Parris, Eluned S. / Carey, Michael J. (1994): "Estimating linear discriminant parameters for continuous density hidden Markov models", In ICSLP-1994, 215-218.