An EM Algorithm for HMMs with Emission Distributions Represented by HMMs
Samy Bengio, Hervé A. Bourlard, Katrin Weber · Infoscience (Ecole Polytechnique Fédérale de Lausanne) · 2000
. A novel approach to represent emission distributions of Hidden Markov Models is presented in this paper. Whereas they are usually estimated with Gaussian mixtures or neural networks, we propose to estimate them with another HMM, but in feature space. This representation, referred here as HMM 2 , could enable the model to more accurately represent feature correlations with fewer parameters than standard HMMs. A full derivation of an EM algorithm is given in order to globally train all the HMM 2 parameters. Preliminary experiments on speech data show promising results. 2 IDIAP{RR 00-11 1 Introduction Hidden Markov Models (HMMs) are statistical models for sequential data that have been used successfully in many applications in articial intelligence, pattern recognition, speech processing and biological sequence modeling [3]. Emission probabilities of HMMs are typically represented using mixtures of Gaussians or neural networks. In this paper, we propose an alternative approach w...