"Eigenlips" for robust speech recognition | IEEE Conference Publication | IEEE Xplore