7th International Conference on Spoken Language Processing
September 16-20, 2002
We investigate a highly effective and extremely simple noise-robust front end based on novel post-processing of standard MFCC features on the Aurora databases. It performs remarkably well on both the Aurora 2.0 and Aurora 3.0 databases without requiring any increase in model complexity. Our experiments on Aurora 2.0 have been reported in . In this paper, we evaluate this technique on the Aurora 3.0 corpus, and present updated results on Aurora 2.0. Results in the past have shown that endpointing (i.e., pre-segmentation) on Aurora 3.0 can yield significant improvements. Our experiments reported herein show that our approach integrates well with this endpointing, namely we obtain additional significant improvements. Overall, on Aurora 3.0 we obtain a 47.17% improvement over the segmented baseline. Also, our most recent Aurora 2.0 results show an overall improvement of 41.09% over the baseline for the matched training conditions, and 65.07% for the mis-matched conditions.
Bibliographic reference. Chen, Chia-Ping / Filali, Karim / Bilmes, Jeff A. (2002): "Frontend post-processing and backend model enhancement on the Aurora 2.0/3.0 databases", In ICSLP-2002, 241-244.