Speaker-trained recognition using allophonic enrollment models | IEEE Conference Publication | IEEE Xplore