This paper focuses on the use of linguistic constraints in continuous speech decoding. In particular, it aims at verifying the advantages of the delayed use of linguistic constraints with respect to their use during the forward search. We evaluate the accuracy loss of the best "backward" derived sentence with respect to the one obtained by forward decoding using grammars. A "backward" de- rived sentence is obtained growing a complete path starting from the end of the sentence using a time asynchronous A* search which exploits the word lattice scores computed during a classical forward search without grammar. Results are evaluated on a 718 word, speaker independent recognition task. Applying linguistic constraints during forward decoding gives slightly better results than delaying them to the backward search, but the difference is small, and it has to be traded with the advantage of having N solutions instead of just one.