Hidden Markov models for speech and signal recognition

Hidden Markov methods have become the most widely accepted techniques for speech recognition and modeling. They are based on parametric statistical models which have two components. The first is a Markov chain which produces a sequence of states. This sequence of states characterizes the evolution of a non-stationary process like speech through a set of ''short-time'' stationary events. The sequence of states is ''hidden'' from the observer by the second component of the model, a set of output distributions, which governs the manner in which the sequence of states is converted into a sequence of speech observations. The purpose of this paper is first, to describe hidden Markov models (HMMs) as a general signal modeling procedure, second, to describe the application of HMMs to speech recognition and modeling, and, third, to describe other problems to which HMMs have been successfully applied. It is hoped that this discussion will inspire other applications, especially in the area of modeling biological processes represented as continuous waveforms, such as the electroencephalogram (EEG).