19 Nov
1999
Sphinx
Speech Group, CMU-SCS (rkm@cs.cmu.edu)
Conclusions
·10x
CHMM system available (thanks to P-!!!)
·With
lextree implementation, only about 1-2% of total HMMs active per
frame
·Order
of magnitude fewer compared to flat lexicon search
·Lextree
replication improves WER noticeably
·Absolute
pruning parameters improve worst-case behavior significantly, without penalizing
accuracy
·When active search space grows beyond some
threshold, no hope of correct recognition anyway