·Search
 all words in parallel
 ·Initialize
 start state of every word with path-score 1.0
 ·For
 each frame of input speech:
 ·Update
 path scores within each HMM
 ·Propagate
 exit state score from one HMM to initial state of its successor (using
 Viterbi criterion)
 ·Select word with best
 exit state path-score