·Tree
 lexicon still not as accurate as flat lexicon baseline
 ·Residual
 word segmentation problems?
 ·Try
 lextree replication?
 ·Use
 of composite SSID model at leaf nodes?
 ·Parameters
 not close to optimal?
 ·HMM
 state acoustic score computation now dominant
 ·Back
 to efficient Gaussian selection/computation