We describe the first version of the Baum-Welch training module of our public
domain speech recognition system. This training module is designed to estimate
the parameters of a set of Hidden Markov Models (HMMs) using observation
sequences, which represent the actual speech utterances, and their associated
transcriptions. It is now capable of training both context-independent and
context-dependent models. Other standard features include the capability of
estimating multiple Gaussian mixture components and the use of phone and word
level transcriptions. A preliminary experiment gave us a word error rate (WER)
54% on OGI Alphadigits using monophone models with 8 Gaussian mixture
components per state.