prevent bogus transitions into <s> in fwdtree search - makes lattices more sensible and also improves accuracy a wee bit

git-svn-id: svn+ssh://svn.code.sf.net/p/cmusphinx/code/trunk/pocketsphinx@10081 94700074-3cef-4d97-a70e-9c8c206c02f5

src/libpocketsphinx/ngram_search_fwdtree.c

@@ -1341,6 +1341,10 @@ word_transition(ngram_search_t *ngs, int frame_idx)
     /* Now transition to in-LM single phone words */
     for (i = 0; i < ngs->n_1ph_LMwords; i++) {
         w = ngs->single_phone_wid[i];
+        /* Never transition into the start word (for one thing, it is
+           a non-event in the language model.) */
+        if (w == dict_startwid(ps_search_dict(ngs)))
+            continue;
         rhmm = (root_chan_t *) ngs->word_chan[w];
         newscore = ngs->last_ltrans[w].dscr + ngs->pip
             + phone_loop_search_score(pls, rhmm->ciphone);
@@ -1376,6 +1380,10 @@ word_transition(ngram_search_t *ngs, int frame_idx)
     for (w = dict_filler_start(dict); w <= dict_filler_end(dict); w++) {
         if (w == ps_search_silence_wid(ngs))
             continue;
+        /* Never transition into the start word (for one thing, it is
+           a non-event in the language model.) */
+        if (w == dict_startwid(ps_search_dict(ngs)))
+            continue;
         rhmm = (root_chan_t *) ngs->word_chan[w];
         /* If this was not actually a single-phone word, rhmm will be NULL. */
         if (rhmm == NULL)

test/data/wsj/test-hub4-n800-fwdtree.match

@@ -1,4 +1,4 @@
-at any the need for international managers will keep rising (n800_440c0201 -50651981)
+at n. e. c. is needed for international managers will keep rising (n800_440c0201 -50733351)
 the company has five hundred japanese managers in overseas most of them in key positions in expects the number to rise sixteen %percent to the next five years (n800_440c0202 -93289612)
 about half of these managers are in the u. s. (n800_440c0203 -34695378)
 r. l. i. corporation a jury illinois based insurance holding company will begin trading friday on the big board under the symbol r. l. i. (n800_440c0204 -99824669)