Historically, each of these three individual components was trained separately, then assembled afterwards to do speech recognition. The conformer models that we’re announcing today are based on ...