int32 align_init ( void ) { int32 k; s3wid_t w; float64 *f64arg; mdef = mdef_getmdef (); tmat = tmat_gettmat (); dict = dict_getdict (); assert (mdef && tmat && dict); startwid = dict_wordid (START_WORD); finishwid = dict_wordid (FINISH_WORD); silwid = dict_wordid (SILENCE_WORD); if ((NOT_WID(startwid)) || (NOT_WID(finishwid))) E_FATAL("%s or %s not in dictionary\n", START_WORD, FINISH_WORD); if (NOT_WID(silwid)) E_ERROR("%s not in dictionary; no optional silence inserted between words\n", SILENCE_WORD); /* Create list of optional filler words to be inserted between transcript words */ fillwid = (s3wid_t *) ckd_calloc ((dict->filler_end - dict->filler_start + 3), sizeof(s3wid_t)); k = 0; if (IS_WID(silwid)) fillwid[k++] = silwid; for (w = dict->filler_start; w <= dict->filler_end; w++) { if ((dict_basewid (w) == w) && (w != silwid) && (w != startwid) && (w != finishwid)) fillwid[k++] = w; } fillwid[k] = BAD_WID; f64arg = (float64 *) cmd_ln_access ("-beam"); beam = logs3 (*f64arg); E_INFO ("logs3(beam)= %d\n", beam); score_scale = (int32 *) ckd_calloc (S3_MAX_FRAMES, sizeof(int32)); hist_head = NULL; align_stseg = NULL; align_phseg = NULL; align_wdseg = NULL; ctr_nstate = counter_new ("NS"); return 0; }
void lmcontext_load (corpus_t *corp, char *uttid, s3wid_t *pred, s3wid_t *succ) { char *str, wd[4096], *strp; s3wid_t w[3]; int32 i, n; dict_t *dict; s3lmwid_t lwid; if ((str = corpus_lookup (corp, uttid)) == NULL) E_FATAL("Couldn't find LM context for %s\n", uttid); dict = dict_getdict (); strp = str; for (i = 0; i < 4; i++) { if (sscanf (strp, "%s%n", wd, &n) != 1) { if (i < 3) E_FATAL("Bad LM context spec for %s: %s\n", uttid, str); else break; } strp += n; if (strcmp (wd, "-") == 0) w[i] = BAD_WID; else { w[i] = dict_wordid (wd); if (NOT_WID(w[i])) E_FATAL("LM context word (%s) for %s not in dictionary\n", wd, uttid); w[i] = dict_basewid(w[i]); switch (i) { case 0: if ((n = dict->word[w[0]].n_comp) > 0) w[0] = dict->word[w[0]].comp[n-1].wid; break; case 1: if ((n = dict->word[w[1]].n_comp) > 0) { w[0] = dict->word[w[1]].comp[n-2].wid; w[1] = dict->word[w[1]].comp[n-1].wid; } break; case 2: if (w[2] != dict_wordid(FINISH_WORD)) E_FATAL("Illegal successor LM context for %s: %s\n", uttid, str); break; default: assert (0); /* Should never get here */ break; } } } if (IS_WID(w[0]) && NOT_WID(w[1])) E_FATAL("Bad LM context spec for %s: %s\n", uttid, str); for (i = 0; i < 3; i++) { if (IS_WID(w[i])) { lwid = lm_lmwid (w[i]); if (NOT_LMWID(lwid)) E_FATAL("LM context word (%s) for %s not in LM\n", wd, uttid); } } pred[0] = w[0]; pred[1] = w[1]; *succ = w[2]; }