·Post-lextree
LM scoring (as above); however:
·Limited,
static lextree replication
·Limits
memory requirements
·No
dynamic lextree management overhead
·Transitions
into lextrees staggered across time:
·At
any time, only one lextree entered
·“-epl” (entries per lextree)
parameter: block of frames one lextree entered, before switching to next
·More
word segmentations (start times) survive
·Full
LM histories; if reached a leaf node for word W:
·Find
all possible LM histories of W
(from BP table)
·Include
LM scores for W
w.r.t. each LM history
·Create
a separate BP table entry for each resulting history