ps_nbest_t *
ps_nbest(ps_decoder_t *ps, int sf, int ef,
char const *ctx1, char const *ctx2)
{
ps_lattice_t *dag;
ngram_model_t *lmset;
ps_astar_t *nbest;
float32 lwf;
int32 w1, w2;
if (ps->search == NULL)
return NULL;
if ((dag = ps_get_lattice(ps)) == NULL)
return NULL;
/* FIXME: This is all quite specific to N-Gram search. Either we
* should make N-best a method for each search module or it needs
* to be abstracted to work for N-Gram and FSG. */
if (0 != strcmp(ps_search_name(ps->search), PS_SEARCH_NGRAM)) {
lmset = NULL;
lwf = 1.0f;
} else {
lmset = ((ngram_search_t *)ps->search)->lmset;
lwf = ((ngram_search_t *)ps->search)->bestpath_fwdtree_lw_ratio;
}
w1 = ctx1 ? dict_wordid(ps_search_dict(ps->search), ctx1) : -1;
w2 = ctx2 ? dict_wordid(ps_search_dict(ps->search), ctx2) : -1;
nbest = ps_astar_start(dag, lmset, lwf, sf, ef, w1, w2);
return (ps_nbest_t *)nbest;
}
static void
ngram_fwdflat_expand_all(ngram_search_t *ngs)
{
int n_words, i;
/* For all "real words" (not fillers or <s>/</s>) in the dictionary,
*
* 1) Add the ones which are in the LM to the fwdflat wordlist
* 2) And to the expansion list (since we are expanding all)
*/
ngs->n_expand_words = 0;
n_words = ps_search_n_words(ngs);
bitvec_clear_all(ngs->expand_word_flag, ps_search_n_words(ngs));
for (i = 0; i < n_words; ++i) {
if (!ngram_model_set_known_wid(ngs->lmset,
dict_basewid(ps_search_dict(ngs),i)))
continue;
ngs->fwdflat_wordlist[ngs->n_expand_words] = i;
ngs->expand_word_list[ngs->n_expand_words] = i;
bitvec_set(ngs->expand_word_flag, i);
ngs->n_expand_words++;
}
E_INFO("Utterance vocabulary contains %d words\n", ngs->n_expand_words);
ngs->expand_word_list[ngs->n_expand_words] = -1;
ngs->fwdflat_wordlist[ngs->n_expand_words] = -1;
}
/**
* Free HMM network for one utterance of fwdflat search.
*/
static void
destroy_fwdflat_chan(ngram_search_t *ngs)
{
int32 i, wid;
for (i = 0; ngs->fwdflat_wordlist[i] >= 0; i++) {
root_chan_t *rhmm;
chan_t *thmm;
wid = ngs->fwdflat_wordlist[i];
if (dict_is_single_phone(ps_search_dict(ngs),wid))
continue;
assert(ngs->word_chan[wid] != NULL);
/* The first HMM in ngs->word_chan[wid] was allocated with
* ngs->root_chan_alloc, but this will attempt to free it
* using ngs->chan_alloc, which will not work. Therefore we
* free it manually and move the list forward before handing
* it off. */
rhmm = (root_chan_t *)ngs->word_chan[wid];
thmm = rhmm->next;
listelem_free(ngs->root_chan_alloc, rhmm);
ngs->word_chan[wid] = thmm;
ngram_search_free_all_rc(ngs, wid);
}
}
int
ngram_fwdflat_reinit(ngram_search_t *ngs)
{
/* Reallocate things that depend on the number of words. */
int n_words;
ckd_free(ngs->fwdflat_wordlist);
ckd_free(ngs->expand_word_list);
bitvec_free(ngs->expand_word_flag);
n_words = ps_search_n_words(ngs);
ngs->fwdflat_wordlist = ckd_calloc(n_words + 1, sizeof(*ngs->fwdflat_wordlist));
ngs->expand_word_flag = bitvec_alloc(n_words);
ngs->expand_word_list = ckd_calloc(n_words + 1, sizeof(*ngs->expand_word_list));
/* No tree-search; take care of the expansion list and single phone words. */
if (!ngs->fwdtree) {
/* Free single-phone words. */
ngram_fwdflat_free_1ph(ngs);
/* Reallocate word_chan. */
ckd_free(ngs->word_chan);
ngs->word_chan = ckd_calloc(dict_size(ps_search_dict(ngs)),
sizeof(*ngs->word_chan));
/* Rebuild full expansion list from LM words. */
ngram_fwdflat_expand_all(ngs);
/* Allocate single phone words. */
ngram_fwdflat_allocate_1ph(ngs);
}
/* Otherwise there is nothing to do since the wordlist is
* generated anew every utterance. */
return 0;
}
ps_search_t *
phone_loop_search_init(cmd_ln_t *config,
acmod_t *acmod,
dict_t *dict)
{
phone_loop_search_t *pls;
/* Allocate and initialize. */
pls = ckd_calloc(1, sizeof(*pls));
ps_search_init(ps_search_base(pls), &phone_loop_search_funcs,
config, acmod, dict, NULL);
phone_loop_search_reinit(ps_search_base(pls), ps_search_dict(pls),
ps_search_dict2pid(pls));
return ps_search_base(pls);
}
static void
ngram_fwdflat_free_1ph(ngram_search_t *ngs)
{
int i, w;
int n_words = ps_search_n_words(ngs);
for (i = w = 0; w < n_words; ++w) {
if (!dict_is_single_phone(ps_search_dict(ngs), w))
continue;
hmm_deinit(&ngs->rhmm_1ph[i].hmm);
++i;
}
ckd_free(ngs->rhmm_1ph);
ngs->rhmm_1ph = NULL;
ckd_free(ngs->single_phone_wid);
}
ps_search_t *
phone_loop_search_init(cmd_ln_t *config,
acmod_t *acmod,
dict_t *dict)
{
phone_loop_search_t *pls;
/* Allocate and initialize. */
pls = (phone_loop_search_t *)ckd_calloc(1, sizeof(*pls));
ps_search_init(ps_search_base(pls), &phone_loop_search_funcs,
PS_SEARCH_TYPE_PHONE_LOOP, PS_DEFAULT_PL_SEARCH,
config, acmod, dict, NULL);
phone_loop_search_reinit(ps_search_base(pls), ps_search_dict(pls),
ps_search_dict2pid(pls));
return ps_search_base(pls);
}
static void
ngram_fwdflat_allocate_1ph(ngram_search_t *ngs)
{
dict_t *dict = ps_search_dict(ngs);
int n_words = ps_search_n_words(ngs);
int i, w;
/* Allocate single-phone words, since they won't have
* been allocated for us by fwdtree initialization. */
ngs->n_1ph_words = 0;
for (w = 0; w < n_words; w++) {
if (dict_is_single_phone(dict, w))
++ngs->n_1ph_words;
}
ngs->single_phone_wid = ckd_calloc(ngs->n_1ph_words,
sizeof(*ngs->single_phone_wid));
ngs->rhmm_1ph = ckd_calloc(ngs->n_1ph_words, sizeof(*ngs->rhmm_1ph));
i = 0;
for (w = 0; w < n_words; w++) {
if (!dict_is_single_phone(dict, w))
continue;
/* DICT2PID location */
ngs->rhmm_1ph[i].ciphone = dict_first_phone(dict, w);
ngs->rhmm_1ph[i].ci2phone = bin_mdef_silphone(ps_search_acmod(ngs)->mdef);
hmm_init(ngs->hmmctx, &ngs->rhmm_1ph[i].hmm, TRUE,
/* ssid */ bin_mdef_pid2ssid(ps_search_acmod(ngs)->mdef,
ngs->rhmm_1ph[i].ciphone),
/* tmatid */ bin_mdef_pid2tmatid(ps_search_acmod(ngs)->mdef,
ngs->rhmm_1ph[i].ciphone));
ngs->rhmm_1ph[i].next = NULL;
ngs->word_chan[w] = (chan_t *) &(ngs->rhmm_1ph[i]);
ngs->single_phone_wid[i] = w;
i++;
}
}
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);
}
}
}
static void
fwdflat_prune_chan(ngram_search_t *ngs, int frame_idx)
{
int32 i, cf, nf, w, pip, newscore, thresh, wordthresh;
int32 *awl;
root_chan_t *rhmm;
chan_t *hmm, *nexthmm;
cf = frame_idx;
nf = cf + 1;
i = ngs->n_active_word[cf & 0x1];
awl = ngs->active_word_list[cf & 0x1];
bitvec_clear_all(ngs->word_active, ps_search_n_words(ngs));
thresh = ngs->best_score + ngs->fwdflatbeam;
wordthresh = ngs->best_score + ngs->fwdflatwbeam;
pip = ngs->pip;
E_DEBUG(3,("frame %d thresh %d wordthresh %d\n", frame_idx, thresh, wordthresh));
/* Scan all active words. */
for (w = *(awl++); i > 0; --i, w = *(awl++)) {
rhmm = (root_chan_t *) ngs->word_chan[w];
/* Propagate active root channels */
if (hmm_frame(&rhmm->hmm) == cf
&& hmm_bestscore(&rhmm->hmm) BETTER_THAN thresh) {
hmm_frame(&rhmm->hmm) = nf;
bitvec_set(ngs->word_active, w);
/* Transitions out of root channel */
newscore = hmm_out_score(&rhmm->hmm);
if (rhmm->next) {
assert(!dict_is_single_phone(ps_search_dict(ngs), w));
newscore += pip;
if (newscore BETTER_THAN thresh) {
hmm = rhmm->next;
/* Enter all right context phones */
if (hmm->info.rc_id >= 0) {
for (; hmm; hmm = hmm->next) {
if ((hmm_frame(&hmm->hmm) < cf)
|| (newscore BETTER_THAN hmm_in_score(&hmm->hmm))) {
hmm_enter(&hmm->hmm, newscore,
hmm_out_history(&rhmm->hmm), nf);
}
}
}
/* Just a normal word internal phone */
else {
if ((hmm_frame(&hmm->hmm) < cf)
|| (newscore BETTER_THAN hmm_in_score(&hmm->hmm))) {
hmm_enter(&hmm->hmm, newscore,
hmm_out_history(&rhmm->hmm), nf);
}
}
}
}
else {
assert(dict_is_single_phone(ps_search_dict(ngs), w));
/* Word exit for single-phone words (where did their
* whmms come from?) (either from
* ngram_search_fwdtree, or from
* ngram_fwdflat_allocate_1ph(), that's where) */
if (newscore BETTER_THAN wordthresh) {
ngram_search_save_bp(ngs, cf, w, newscore,
hmm_out_history(&rhmm->hmm), 0);
}
}
}
/* Transitions out of non-root channels. */
for (hmm = rhmm->next; hmm; hmm = hmm->next) {
if (hmm_frame(&hmm->hmm) >= cf) {
/* Propagate forward HMMs inside the beam. */
if (hmm_bestscore(&hmm->hmm) BETTER_THAN thresh) {
hmm_frame(&hmm->hmm) = nf;
bitvec_set(ngs->word_active, w);
newscore = hmm_out_score(&hmm->hmm);
/* Word-internal phones */
if (hmm->info.rc_id < 0) {
newscore += pip;
if (newscore BETTER_THAN thresh) {
nexthmm = hmm->next;
/* Enter all right-context phones. */
if (nexthmm->info.rc_id >= 0) {
for (; nexthmm; nexthmm = nexthmm->next) {
if ((hmm_frame(&nexthmm->hmm) < cf)
|| (newscore BETTER_THAN
hmm_in_score(&nexthmm->hmm))) {
hmm_enter(&nexthmm->hmm,
newscore,
hmm_out_history(&hmm->hmm),
nf);
}
}
}
/* Enter single word-internal phone. */
else {
if ((hmm_frame(&nexthmm->hmm) < cf)
|| (newscore BETTER_THAN
hmm_in_score(&nexthmm->hmm))) {
hmm_enter(&nexthmm->hmm, newscore,
hmm_out_history(&hmm->hmm), nf);
}
}
}
}
/* Right-context phones - apply word beam and exit. */
else {
if (newscore BETTER_THAN wordthresh) {
ngram_search_save_bp(ngs, cf, w, newscore,
hmm_out_history(&hmm->hmm),
hmm->info.rc_id);
}
}
}
/* Zero out inactive HMMs. */
else if (hmm_frame(&hmm->hmm) != nf) {
hmm_clear_scores(&hmm->hmm);
}
}
}
}
}
/**
* Build HMM network for one utterance of fwdflat search.
*/
static void
build_fwdflat_chan(ngram_search_t *ngs)
{
int32 i, wid, p;
root_chan_t *rhmm;
chan_t *hmm, *prevhmm;
dict_t *dict;
dict2pid_t *d2p;
dict = ps_search_dict(ngs);
d2p = ps_search_dict2pid(ngs);
/* Build word HMMs for each word in the lattice. */
for (i = 0; ngs->fwdflat_wordlist[i] >= 0; i++) {
wid = ngs->fwdflat_wordlist[i];
/* Single-phone words are permanently allocated */
if (dict_is_single_phone(dict, wid))
continue;
assert(ngs->word_chan[wid] == NULL);
/* Multiplex root HMM for first phone (one root per word, flat
* lexicon). diphone is irrelevant here, for the time being,
* at least. */
rhmm = listelem_malloc(ngs->root_chan_alloc);
rhmm->ci2phone = dict_second_phone(dict, wid);
rhmm->ciphone = dict_first_phone(dict, wid);
rhmm->next = NULL;
hmm_init(ngs->hmmctx, &rhmm->hmm, TRUE,
bin_mdef_pid2ssid(ps_search_acmod(ngs)->mdef, rhmm->ciphone),
bin_mdef_pid2tmatid(ps_search_acmod(ngs)->mdef, rhmm->ciphone));
/* HMMs for word-internal phones */
prevhmm = NULL;
for (p = 1; p < dict_pronlen(dict, wid) - 1; p++) {
hmm = listelem_malloc(ngs->chan_alloc);
hmm->ciphone = dict_pron(dict, wid, p);
hmm->info.rc_id = (p == dict_pronlen(dict, wid) - 1) ? 0 : -1;
hmm->next = NULL;
hmm_init(ngs->hmmctx, &hmm->hmm, FALSE,
dict2pid_internal(d2p,wid,p),
bin_mdef_pid2tmatid(ps_search_acmod(ngs)->mdef, hmm->ciphone));
if (prevhmm)
prevhmm->next = hmm;
else
rhmm->next = hmm;
prevhmm = hmm;
}
/* Right-context phones */
ngram_search_alloc_all_rc(ngs, wid);
/* Link in just allocated right-context phones */
if (prevhmm)
prevhmm->next = ngs->word_chan[wid];
else
rhmm->next = ngs->word_chan[wid];
ngs->word_chan[wid] = (chan_t *) rhmm;
}
}
/**
* 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);
}