Improved performance for speaker independent speech recognitions systems requires better modelling of different dialects of the target language. Work on this topic [1, 2], as well as our own results, suggest that separate modelling of dialects is needed to capture accurately the many pronunciation differences that occur. No matter how much dialect data is included in training, however, some speakers will not be covered by the resulting model - for instance, non-native speakers of the language and speakers whose speech patterns have been affected by surgery. We describe a speech recognition system which incorporates both dialect modelling to provide coverage for a high percentage of the target population, and optional speaker adaptation for speakers who are not adequately represented by the model. This comprehensive approach does not require detailed phonetic knowledge of dialectical variations, dramatically increased memory usage, or significantly increased complexity during recognition.
Bibliographic reference. Beattie, V. / Edmondson, S. / Miller, D. / Patel, Y. / Talvola, G. (1995): "An integrated multi-dialect speech recognition system with optional speaker adaptation", In EUROSPEECH-1995, 1123-1126.