What are the best practices for doing the reverse: taking audio and producing text? I don't mind the translation to be rough, the error rate can be quite high for my purposes, but I want the process not to get stuck and recover so it processes a full length talk.