Lexical and acoustic modeling of non-native speech in LVSCR

As non-native speakers become more frequent users of speech recognition applications, increasing the tolerance of the system with respect to non-native pronunciation and language use is important and is currently the focus of research in a variety of contexts. Dictionary modiication, acoustic model adaptation, and acoustic model manipulation are a few of the techniques that have been reported successful in improving recognition of non-native speech. In this paper, we address the speciic case of Japanese-accented English, describing the lexical and acoustic mod-eling techniques that give the best recognizer performance. We nd that automatically generated pronunciation variants perform as well as hand-coded \golden" variants in reducing recognizer error, and that a signiicant improvement in system performance can be achieved with acoustic models retrained on a small amount of accented data.