Can you go into more detail on this? I’ve been putting together the data pipeline for an ASR fine tune but your approach sounds more interesting.