No Slide Title

19 Nov 1999

Sphinx Speech Group, CMU-SCS (rkm@cs.cmu.edu)

Pruning in Sphinx-3.2

·Pure beam-pruning has long-tailed distribution of active HMMs/frame

·Absolute pruning to control worst-case performance:

·Max. active HMMs per frame

·Implemented approximately, using “histogram pruning” (avoids expensive sorting step)

·Max. unique words exiting per frame

·Max. LM histories saved in BP table per frame

·Word error rate unaffected

·Additional beam for lextree-internal, cross-HMM transitions

·Unigram “lookahead” scores used in lextree for yet more pruning