Abstract
We set out a methodology for the automated generation of hidden Markov models (HMMs) of observed feature-space transitions in a noisy experimental environment that is maximally generalising under the assumed experimental constraints. Specifically, we provide an ICA-based feature-selection technique for determining the number, and the transition sequence of the underlying hidden states, along with the observed-state emission characteristics when the specified noise model assumptions are fulfilled. In retaining correlation information between features, the method is potentially more general than the commonly employed Gaussian mixture model HMM parameterisation methods, to which we demonstrate that our method reduces when an arbitrary separation of features, or an experimentally-limited feature-space is imposed. A practical demonstration of the application of this method to automated sign-language classification is given, for which we demonstrate that a performance improvement of the order of 40% over naive Markovian modelling of the observed transitions is possible.