Are you able to provide more information on the fine tuning? Any improvement in WER and what language it was fine tuned in and the size of the dataset used?