Without some degree of acuracy, learners could end up producing unrecognisable or incomprehensible language. I don't think we need to try to turn people into sounding like native speakers, but we do need to get them to be capable of being understood by a variety of listeners, which involves a degree of approximation to the core sounds.