I run the full SRI system on SVB 500 task subsets. This system uses swbd and Fisher data for both acoustic and language modeling. It is essentially the best system around here, and includes MPE, HLDA, speaker adaptation, and cross-system adaptation, tandem features, and so on. So the WERs below are kind of a lower bound, with the caveat that decoding was done with a some 50k vocab, which might have introduced some errors:

     1st   final
A    40.0  24.2
B    44.3  28.2
C    43.0  N/A (crashed)
D    41.6  26.2
E    42.4  26.8

1st is the WER from first-pass decoding with nothing else such as adaptation, with mfcc+mlp features. final WER is the WER after adaptation and everything.

-- KarenLivescu - 25 Jul 2006