In my experience Viterbi training alone is often enough to get reasonable accuracy, at least in speech recognition.
Instead of doing the more costly Baum-Welch training you can spend your time better elsewhere, e.g. use deep neural networks instead of GMMs or collect more data.
Interesting - I've never tried Viterbi training. Maybe it is worth implementing after all. I plan to do a hybrid DNN-HMM (or whatever it is called now) with pylearn2 in a followup post.
> Large data sets are available on the internets. Something tells me Google may even have access to large quantities of data already.
You don't just need audio data, you also need the correct transcriptions to learn anything from it. This reduces the amount of available data significantly. And producing correct transcriptions is time consuming and expensive.
