Yoshua Bengio
Dept. Informatique et Recherche Opérationnelle
Université de Montréal, Montreal, Qc H3C-3J7, Canada
bengioy@iro.umontreal.ca
31st, July, 1997
SUBMITTED TO STATISTICAL SCIENCE
Hidden Markov Models (HMMs) are statistical models of sequential data that have been used successfully in several applications, such as in speech recognition, molecular biology, econometrics, and fault-detection. We first summarize the basics of HMMs and of their application to speech recognition, and then review several recent related learning algorithms and extensions of HMMs, including hybrids of HMMs with artificial neural networks, Input-Output HMMs, weighted transducers, variable-length Markov models and Markov switching state-space models. Finally, we discuss some of the challenges of future research in this area. In particular we point out the move away from models that are meant to represent the true probability distribution of the data and towards algorithmic tools for taking decisions with good generalization properties.
keywords: hidden Markov models, learning algorithms, artificial
neural networks, weighted transducers, state space models,
input-output hidden Markov models, Markov switching models.