Variational Bayesian learning of speech GMMS for feature enhancement based on Algonquin