Neural networks for statistical inference: Generalizations with applications to speech recognition

The basic principles of the hybrid HMM/MLP (hidden Markov model/multilayer perceptron) approach are reviewed and extended to triphone models. It is also shown how the statistical interpretation of the MLP output values can act upon the development of other algorithms and help them understand their behavior, which is the case with the a priori probabilities and the radial basis function networks. The advantages of a speech recognition system incorporating both MLPs and HMMs are the best discrimination and the ability to incorporate multiple sources of evidence (features, temporal context) without restrictive assumptions of distributions or statistical independence