imho native audio models aren't ready for prime time except in few narrow use cases where tools and accurate transcription aren't required