Is non-native pronunciation modelling necessary ?
Silke Goronzy, Marina Sahakyan, Wolfgang Wokurek · 2001
It is difficult to recognize accented or non-native speech with speech recognition systems that are trained using native speech. While standard acoustic speaker adaptation techniques are of-ten applied in these cases, they can only reduce the recognition errors that are due to mispronunciations on the phoneme level. They are not able to handle severe deviations from the expected pronunciation. Also, there has been a lot of interest in native pronunciation modelling recently. However results often were not as good as expected. This paper concentrates on non-native speakers and examines, if a special treatment of these speakers is necessary. The effect of adding special non-native pronun-ciation variants to the pronunciation lexicon is investigated. In contrast to native pronunciation modelling the results show that for the non-native case the enhanced dictionary is really neces-sary to obtain acceptable recognition rates. Recognition rates on the Interactive Spoken Language Education corpus (ISLE) were improved by up to 10 % for German and even up to 28% for Italian learners of English. When combining this with Max-imum Likelihood Linear Regression (MLLR) adaptation, these results can be further improved. 1.