Accurate speech-to-text and natural text-to-speech