·Exhaustive
 search over large vocabulary too expensive, and unnecessary
 ·Use a “beam” to
 “prune” the set of active HMMs:
 ·At
 start of each frame, find best available path-score S
 ·Use
 a scale-factor f  (< 1.0)
 to set a pruning threshold T
 = S*f
 ·Deactivate
 an HMM if no state in it has path score >= T
 ·Effect:
 No. of active HMMs larger if no clear frontrunner
 ·Two kinds of
 beams:
 ·To
 control active set of HMMs
 ·No.
 of active HMMs per frame typically 10-20% of total space
 ·To control word exits
 taken (and recorded in BP table)
 ·No.
 of words exited typically 10-20 per frame
 ·Recognition accuracy
 essentially unaffected