I've done this with Google's video transcription model, which works surprisingly well.