Text this: Modelling segmental variability for automatic speech recognition /