The extended Baum-Welch (EBW) is the most popular algorithm for discriminative training of speech recognition acoustic models. The EBW algorithm is usually controlled with heuristic rules, which are used to determine the smoothing parameters of the algorithm. In this paper we propose a control method for EBW which is based on optimization of an error measure over a small control set. The large vocabulary speech recognition experiments show this to have clear benefits over the heuristic methods.

Index Terms: speech recognition, discriminative training, extended Baum-Welch, optimization