 |
ITRW on
|
 |
|
ITRW on
|
|
Tasks involving the recognition of native as well as foreign names and being accessed by native as well as non-native speakers are not handled very well by a baseline ASR system with single language acoustical models and one transcription per word in its lexicon. One possible solution is lexical adaptation: alternative transcriptions allowing for a closer match with actual pronunciations from the speakers, are added to the lexicon. In this paper it is shown how transcriptions created by foreign language grapheme-tophoneme converters can be integrated adequately in the recognizer and serve the objected purpose. The experiments reported here show that this approach can effectively deal with foreign names and non-native speakers, as WER reductions of almost 60% relative are attained.
Bibliographic reference. Cremelie, Nick / Bosch, Louis ten (2001): "Improving the recognition of foreign names and non-native speech by combining multiple grapheme-to-phoneme converters", In Adaptation-2001, 151-154.