Cross-task portability of a broadcast news speech recognition system

This paper reports on experiments of porting the ITC-irst Italian broadcast news recognition system to two spontaneous dialogue domains. Porting was investigated by applying state-of-the-art adaptation methods on acoustic and language models, and by evaluating the trade-off between performance and required amount of task specific annotated data. The use of different levels of supervision for acoustic model adaptation was also studied. By employing 2 It of manually annotated speech, word error rates of 26.0% and 28.4% were achieved by the adapted systems. These results are to be compared with the performance of two domain specific baseline systems, 22.6% and 21.2%, respectively, which were developed on much more training data. Finally, a robust method is presented that allows to tune the insertion of spontaneous speech phenomena by the speech decoder. (C) 2001 Elsevier Science B.V. All rights reserved.