N-best search algorithm is very similar to the time synchronous
Viterbi search []. Since the purpose of the N-best method is
to find the optimum speech unit sequence instead of the optimum state
sequence, a summing operation should be done instead of the maximum finding
operation. However if we completely drop the maximum finding operation
it will become the Forward algorithm, and we go back again to the
start. Therefore a pruning is performed at every state, (in addition to
the pruning of the beam) keeping only the first *N* paths with the
highest scores. Therefore even this algorithm does not give the
theoretically optimum sentence. At the end, the algorithm gives *N*
most likely sentences, and for a simple task without post processors
,*N*=1 is enough [].

Fri May 10 20:35:10 MET DST 1996