Abstract

Hidden Markov Models are commonly trained using algorithms derived from gradient-based methods such as the Baum-Welch procedure. We describe a representation of discrete observation HMMs that permits them to be trained using Population-Based Incremental Learning (PBIL), a variant of genetic learning. In this paper we examine the training speed and recognition performance of PBIL on a spoken digit recognition task. We demonstrate maximum likelihood training of HMMs can be achieved using PBIL, although the Baum-Welch procedure is significantly faster. However, unlike the Baum-Welch procedure, PBIL is easily extended to perform discriminative training. Furthermore, PBIL training extends to applications like speech intelligibility assessment that lack a differentiable or analytical optimization function.

Citations