14th International Congress of Phonetic Sciences (ICPhS-14)
San Francisco, CA, USA
The aim of this study is to improve automatic speech recognition performance when the speech material to be recognised is produced by non-native speakers and therefore prone to a foreign accent. A foreign accent generally causes a drop in recognition performance. In the this study, it is shown that adding modelling units from the native language into the phonetic pronunciation network can largely improve the recognition of non-native speakers without significantly degrading the performance of native speakers. Moreover, when some extra pronunciation variants are also added into the word phonetic transcription network, the recognition improvement becomes even larger. On the different corpora used in this paper, the error rate reduction ranged from 11% to 71%.
Bibliographic reference. Bartkova, Katarina / Jouvet, Denis (1999): "Language based phone model combination for ASR adaptation to foreign accent", In ICPhS-14, 1725-1728.