The compression could be quite extensive, it wouldn't even have to be convertible back to speech if all that was parsing it on the other side was ML.