static void fwdflat_eval_chan(ngram_search_t *ngs, int frame_idx) { int32 i, w, nw, bestscore; int32 *awl; root_chan_t *rhmm; chan_t *hmm; nw = ngs->n_active_word[frame_idx & 0x1]; awl = ngs->active_word_list[frame_idx & 0x1]; bestscore = WORST_SCORE; ngs->st.n_fwdflat_words += nw; /* Scan all active words. */ for (i = 0; i < nw; i++) { w = *(awl++); rhmm = (root_chan_t *) ngs->word_chan[w]; if (hmm_frame(&rhmm->hmm) == frame_idx) { int32 score = chan_v_eval(rhmm); if ((score BETTER_THAN bestscore) && (w != ps_search_finish_wid(ngs))) bestscore = score; ngs->st.n_fwdflat_chan++; } for (hmm = rhmm->next; hmm; hmm = hmm->next) { if (hmm_frame(&hmm->hmm) == frame_idx) { int32 score = chan_v_eval(hmm); if (score BETTER_THAN bestscore) bestscore = score; ngs->st.n_fwdflat_chan++; } } } ngs->best_score = bestscore; }
static void fwdflat_word_transition(ngram_search_t *ngs, int frame_idx) { int32 cf, nf, b, thresh, pip, i, w, newscore; int32 best_silrc_score = 0, best_silrc_bp = 0; /* FIXME: good defaults? */ bptbl_t *bp; int32 *rcss; root_chan_t *rhmm; int32 *awl; float32 lwf; dict_t *dict = ps_search_dict(ngs); dict2pid_t *d2p = ps_search_dict2pid(ngs); cf = frame_idx; nf = cf + 1; thresh = ngs->best_score + ngs->fwdflatbeam; pip = ngs->pip; best_silrc_score = WORST_SCORE; lwf = ngs->fwdflat_fwdtree_lw_ratio; /* Search for all words starting within a window of this frame. * These are the successors for words exiting now. */ get_expand_wordlist(ngs, cf, ngs->max_sf_win); /* Scan words exited in current frame */ for (b = ngs->bp_table_idx[cf]; b < ngs->bpidx; b++) { xwdssid_t *rssid; int32 silscore; bp = ngs->bp_table + b; ngs->word_lat_idx[bp->wid] = NO_BP; if (bp->wid == ps_search_finish_wid(ngs)) continue; /* DICT2PID location */ /* Get the mapping from right context phone ID to index in the * right context table and the bscore_stack. */ rcss = ngs->bscore_stack + bp->s_idx; if (bp->last2_phone == -1) rssid = NULL; else rssid = dict2pid_rssid(d2p, bp->last_phone, bp->last2_phone); /* Transition to all successor words. */ for (i = 0; ngs->expand_word_list[i] >= 0; i++) { int32 n_used; w = ngs->expand_word_list[i]; /* Get the exit score we recorded in save_bwd_ptr(), or * something approximating it. */ if (rssid) newscore = rcss[rssid->cimap[dict_first_phone(dict, w)]]; else newscore = bp->score; if (newscore == WORST_SCORE) continue; /* FIXME: Floating point... */ newscore += lwf * (ngram_tg_score(ngs->lmset, dict_basewid(dict, w), bp->real_wid, bp->prev_real_wid, &n_used) >> SENSCR_SHIFT); newscore += pip; /* Enter the next word */ if (newscore BETTER_THAN thresh) { rhmm = (root_chan_t *) ngs->word_chan[w]; if ((hmm_frame(&rhmm->hmm) < cf) || (newscore BETTER_THAN hmm_in_score(&rhmm->hmm))) { hmm_enter(&rhmm->hmm, newscore, b, nf); /* DICT2PID: This is where mpx ssids get introduced. */ /* Look up the ssid to use when entering this mpx triphone. */ hmm_mpx_ssid(&rhmm->hmm, 0) = dict2pid_ldiph_lc(d2p, rhmm->ciphone, rhmm->ci2phone, dict_last_phone(dict, bp->wid)); assert(IS_S3SSID(hmm_mpx_ssid(&rhmm->hmm, 0))); E_DEBUG(6,("ssid %d(%d,%d) = %d\n", rhmm->ciphone, dict_last_phone(dict, bp->wid), rhmm->ci2phone, hmm_mpx_ssid(&rhmm->hmm, 0))); bitvec_set(ngs->word_active, w); } } } /* Get the best exit into silence. */ if (rssid) silscore = rcss[rssid->cimap[ps_search_acmod(ngs)->mdef->sil]]; else silscore = bp->score; if (silscore BETTER_THAN best_silrc_score) { best_silrc_score = silscore; best_silrc_bp = b; } } /* Transition to <sil> */ newscore = best_silrc_score + ngs->silpen + pip; if ((newscore BETTER_THAN thresh) && (newscore BETTER_THAN WORST_SCORE)) { w = ps_search_silence_wid(ngs); rhmm = (root_chan_t *) ngs->word_chan[w]; if ((hmm_frame(&rhmm->hmm) < cf) || (newscore BETTER_THAN hmm_in_score(&rhmm->hmm))) { hmm_enter(&rhmm->hmm, newscore, best_silrc_bp, nf); bitvec_set(ngs->word_active, w); } } /* Transition to noise words */ newscore = best_silrc_score + ngs->fillpen + pip; if ((newscore BETTER_THAN thresh) && (newscore BETTER_THAN WORST_SCORE)) { for (w = ps_search_silence_wid(ngs) + 1; w < ps_search_n_words(ngs); w++) { rhmm = (root_chan_t *) ngs->word_chan[w]; /* Noise words that aren't a single phone will have NULL here. */ if (rhmm == NULL) continue; if ((hmm_frame(&rhmm->hmm) < cf) || (newscore BETTER_THAN hmm_in_score(&rhmm->hmm))) { hmm_enter(&rhmm->hmm, newscore, best_silrc_bp, nf); bitvec_set(ngs->word_active, w); } } } /* Reset initial channels of words that have become inactive even after word trans. */ i = ngs->n_active_word[cf & 0x1]; awl = ngs->active_word_list[cf & 0x1]; for (w = *(awl++); i > 0; --i, w = *(awl++)) { rhmm = (root_chan_t *) ngs->word_chan[w]; if (hmm_frame(&rhmm->hmm) == cf) { hmm_clear_scores(&rhmm->hmm); } } }
/** * Find all active words in backpointer table and sort by frame. */ static void build_fwdflat_wordlist(ngram_search_t *ngs) { int32 i, f, sf, ef, wid, nwd; bptbl_t *bp; ps_latnode_t *node, *prevnode, *nextnode; /* No tree-search, use statically allocated wordlist. */ if (!ngs->fwdtree) return; memset(ngs->frm_wordlist, 0, ngs->n_frame_alloc * sizeof(*ngs->frm_wordlist)); /* Scan the backpointer table for all active words and record * their exit frames. */ for (i = 0, bp = ngs->bp_table; i < ngs->bpidx; i++, bp++) { sf = (bp->bp < 0) ? 0 : ngs->bp_table[bp->bp].frame + 1; ef = bp->frame; wid = bp->wid; /* Anything that can be transitioned to in the LM can go in * the word list. */ if (!ngram_model_set_known_wid(ngs->lmset, dict_basewid(ps_search_dict(ngs), wid))) continue; /* Look for it in the wordlist. */ for (node = ngs->frm_wordlist[sf]; node && (node->wid != wid); node = node->next); /* Update last end frame. */ if (node) node->lef = ef; else { /* New node; link to head of list */ node = listelem_malloc(ngs->latnode_alloc); node->wid = wid; node->fef = node->lef = ef; node->next = ngs->frm_wordlist[sf]; ngs->frm_wordlist[sf] = node; } } /* Eliminate "unlikely" words, for which there are too few end points */ for (f = 0; f < ngs->n_frame; f++) { prevnode = NULL; for (node = ngs->frm_wordlist[f]; node; node = nextnode) { nextnode = node->next; /* Word has too few endpoints */ if ((node->lef - node->fef < ngs->min_ef_width) || /* Word is </s> and doesn't actually end in last frame */ ((node->wid == ps_search_finish_wid(ngs)) && (node->lef < ngs->n_frame - 1))) { if (!prevnode) ngs->frm_wordlist[f] = nextnode; else prevnode->next = nextnode; listelem_free(ngs->latnode_alloc, node); } else prevnode = node; } } /* Form overall wordlist for 2nd pass */ nwd = 0; bitvec_clear_all(ngs->word_active, ps_search_n_words(ngs)); for (f = 0; f < ngs->n_frame; f++) { for (node = ngs->frm_wordlist[f]; node; node = node->next) { if (!bitvec_is_set(ngs->word_active, node->wid)) { bitvec_set(ngs->word_active, node->wid); ngs->fwdflat_wordlist[nwd++] = node->wid; } } } ngs->fwdflat_wordlist[nwd] = -1; E_INFO("Utterance vocabulary contains %d words\n", nwd); }