Modified Baum Welch Algorithm for Hidden Markov Models with Known Structure

Abstract

Hidden Markov Models (HMMs) are widely used in speech and handwriting recognition, behavior prediction in traffic, time series analysis, biostatistics, image and signal processing, and many other fields. For some applications in those real world problems, a-priori knowledge about the structure of the HMM is available. For example the shape of the state transition matrix and/or the observation matrix might be given. We might know that some entries in these matrices are equal and others are zero. For training such a model, we have two options: use the common Baum Welch Algorithm (BWA) and enforce the given structure after training or modify the BWA to enforce it during training. This paper shows several approaches for modifying the BWA and compares the results of all training methods.

Publication
Springer, Cham

Related