·Most
 active HMMs are word-initial models, decaying rapidly subsequently
 ·On
 60K-word Hub-4 task, 55% of active HMMs are word-initial
 ·(Same
 reason for handling left/right contexts differently.)
 ·But,
 no. of distinct word-initial model types much fewer:
 ·
 ·
 ·
 ·
 ·Use
 a “prefix-tree” structure to maximize sharing among words