3-stream LTG model with asynchrony variables, one state per phone per stream; i.e. Karen's old model, with tweaks (e.g. Chris's rearranged wordTransition structure).

-- KarenLivescu - 03 Aug 2006

256+ mixtures not yet tried. 
run on UIUC.
cbeam used for training andtesting


This model does worse than the Monofeat No asycrony model, and shouldn't
possible reasons:
lm optimum is off the edge of the tested grid
narrower beam pruning during testing drops more test utterances, and accounts for 3.2 accuracy hit
karen thinks that low occupancy of some async states can also contribute error.

WER     comp    lm_sc   lm_pen
83.4    128     4       -4
85      64      4       -8

-- ArtKantor - 03 Aug 2006