Accent specific modifications for continuous speech recognition based on a sub-word lattice approach.