/*ARCHAN, to allow backward compatibility -lm, -lmctlfn coexists. This makes the current implmentation more complicated than necessary. */
void kb_init (kb_t *kb)
{
kbcore_t *kbcore;
mdef_t *mdef;
dict_t *dict;
dict2pid_t *d2p;
lm_t *lm;
lmset_t *lmset;
s3cipid_t sil, ci;
s3wid_t w;
int32 i, n, n_lc;
wordprob_t *wp;
s3cipid_t *lc;
bitvec_t lc_active;
char *str;
int32 cisencnt;
int32 j;
/* Initialize the kb structure to zero, just in case */
memset(kb, 0, sizeof(*kb));
kb->kbcore = NULL;
kb->kbcore = kbcore_init (cmd_ln_float32 ("-logbase"),
cmd_ln_str("-feat"),
cmd_ln_str("-cmn"),
cmd_ln_str("-varnorm"),
cmd_ln_str("-agc"),
cmd_ln_str("-mdef"),
cmd_ln_str("-dict"),
cmd_ln_str("-fdict"),
"", /* Hack!! Hardwired constant
for -compsep argument */
cmd_ln_str("-lm"),
cmd_ln_str("-lmctlfn"),
cmd_ln_str("-lmdumpdir"),
cmd_ln_str("-fillpen"),
cmd_ln_str("-senmgau"),
cmd_ln_float32("-silprob"),
cmd_ln_float32("-fillprob"),
cmd_ln_float32("-lw"),
cmd_ln_float32("-wip"),
cmd_ln_float32("-uw"),
cmd_ln_str("-mean"),
cmd_ln_str("-var"),
cmd_ln_float32("-varfloor"),
cmd_ln_str("-mixw"),
cmd_ln_float32("-mixwfloor"),
cmd_ln_str("-subvq"),
cmd_ln_str("-gs"),
cmd_ln_str("-tmat"),
cmd_ln_float32("-tmatfloor"));
if(kb->kbcore==NULL){
E_FATAL("Initialization of kb failed\n");
}
kbcore = kb->kbcore;
mdef = kbcore_mdef(kbcore);
dict = kbcore_dict(kbcore);
lm = kbcore_lm(kbcore);
lmset=kbcore_lmset(kbcore);
d2p = kbcore_dict2pid(kbcore);
if (NOT_S3WID(dict_startwid(dict)) || NOT_S3WID(dict_finishwid(dict)))
E_FATAL("%s or %s not in dictionary\n", S3_START_WORD, S3_FINISH_WORD);
if(lmset){
for(i=0;i<kbcore_nlm(kbcore);i++){
if (NOT_S3LMWID(lm_startwid(lmset[i].lm)) || NOT_S3LMWID(lm_finishwid(lmset[i].lm)))
E_FATAL("%s or %s not in LM %s\n", S3_START_WORD, S3_FINISH_WORD,lmset[i].name);
}
}else if(lm){
if (NOT_S3LMWID(lm_startwid(lm)) || NOT_S3LMWID(lm_finishwid(lm)))
E_FATAL("%s or %s not in LM\n", S3_START_WORD, S3_FINISH_WORD);
}
/* Check that HMM topology restrictions are not violated */
if (tmat_chk_1skip (kbcore->tmat) < 0)
E_FATAL("Tmat contains arcs skipping more than 1 state\n");
/*
* Unlink <s> and </s> between dictionary and LM, to prevent their
* recognition. They are merely dummy words (anchors) at the beginning
* and end of each utterance.
*/
if(lmset){
for(i=0;i<kbcore_nlm(kbcore);i++){
lm_lmwid2dictwid(lmset[i].lm, lm_startwid(lmset[i].lm)) = BAD_S3WID;
lm_lmwid2dictwid(lmset[i].lm, lm_finishwid(lmset[i].lm)) = BAD_S3WID;
for (w = dict_startwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
lmset[i].lm->dict2lmwid[w] = BAD_S3LMWID;
for (w = dict_finishwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
lmset[i].lm->dict2lmwid[w] = BAD_S3LMWID;
}
}else if(lm){ /* No LM is set at this point*/
lm_lmwid2dictwid(lm, lm_startwid(lm)) = BAD_S3WID;
lm_lmwid2dictwid(lm, lm_finishwid(lm)) = BAD_S3WID;
for (w = dict_startwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
kbcore->dict2lmwid[w] = BAD_S3LMWID;
for (w = dict_finishwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
kbcore->dict2lmwid[w] = BAD_S3LMWID;
}
sil = mdef_silphone (kbcore_mdef (kbcore));
if (NOT_S3CIPID(sil))
E_FATAL("Silence phone '%s' not in mdef\n", S3_SILENCE_CIPHONE);
kb->sen_active = (int32 *) ckd_calloc (mdef_n_sen(mdef), sizeof(int32));
kb->rec_sen_active = (int32 *) ckd_calloc (mdef_n_sen(mdef), sizeof(int32));
kb->ssid_active = (int32 *) ckd_calloc (mdef_n_sseq(mdef), sizeof(int32));
kb->comssid_active = (int32 *) ckd_calloc (dict2pid_n_comsseq(d2p), sizeof(int32));
/* Build set of all possible left contexts */
lc = (s3cipid_t *) ckd_calloc (mdef_n_ciphone(mdef) + 1, sizeof(s3cipid_t));
lc_active = bitvec_alloc (mdef_n_ciphone (mdef));
for (w = 0; w < dict_size (dict); w++) {
ci = dict_pron (dict, w, dict_pronlen(dict, w) - 1);
if (! mdef_is_fillerphone (mdef, (int)ci))
bitvec_set (lc_active, ci);
}
ci = mdef_silphone(mdef);
bitvec_set (lc_active, ci);
for (ci = 0, n_lc = 0; ci < mdef_n_ciphone(mdef); ci++) {
if (bitvec_is_set (lc_active, ci))
lc[n_lc++] = ci;
}
lc[n_lc] = BAD_S3CIPID;
E_INFO("Building lextrees\n");
/* Get the number of lexical tree*/
kb->n_lextree = cmd_ln_int32 ("-Nlextree");
if (kb->n_lextree < 1) {
E_ERROR("No. of ugtrees specified: %d; will instantiate 1 ugtree\n",
kb->n_lextree);
kb->n_lextree = 1;
}
/* ARCHAN: This code was rearranged in s3.4 implementation of dynamic LM */
/* Build active word list */
wp = (wordprob_t *) ckd_calloc (dict_size(dict), sizeof(wordprob_t));
if(lmset){
kb->ugtreeMulti = (lextree_t **) ckd_calloc (kbcore_nlm(kbcore)*kb->n_lextree, sizeof(lextree_t *));
/* Just allocate pointers*/
kb->ugtree = (lextree_t **) ckd_calloc (kb->n_lextree, sizeof(lextree_t *));
for(i=0;i<kbcore_nlm(kbcore);i++){
E_INFO("Creating Unigram Table for lm %d name %s\n",i,lmset[i].name);
n=0;
for(j=0;j<dict_size(dict);j++){ /*try to be very careful again */
wp[j].wid=-1;
wp[j].prob=-1;
}
n = lm_ug_wordprob (lmset[i].lm, dict,MAX_NEG_INT32, wp);
E_INFO("Size of word table after unigram + words in class: %d.\n",n);
if (n < 1)
E_FATAL("%d active words in %s\n", n,lmset[i].name);
n = wid_wordprob2alt(dict,wp,n);
E_INFO("Size of word table after adding alternative prons: %d.\n",n);
if (cmd_ln_int32("-treeugprob") == 0) {
for (i = 0; i < n; i++)
wp[i].prob = -1; /* Flatten all initial probabilities */
}
for (j = 0; j < kb->n_lextree; j++) {
kb->ugtreeMulti[i*kb->n_lextree+j] = lextree_build (kbcore, wp, n, lc);
lextree_type (kb->ugtreeMulti[i*kb->n_lextree+j]) = 0;
E_INFO("Lextrees (%d) for lm %d name %s, %d nodes(ug)\n",
kb->n_lextree, i, lmset[i].name,lextree_n_node(kb->ugtreeMulti[i*kb->n_lextree+j]));
}
}
}else if (lm){
E_INFO("Creating Unigram Table\n");
n=0;
n = lm_ug_wordprob (lm, dict,MAX_NEG_INT32, wp);
E_INFO("Size of word table after unigram + words in class: %d\n",n);
if (n < 1)
E_FATAL("%d active words\n", n);
n = wid_wordprob2alt (dict, wp, n); /* Add alternative pronunciations */
/* Retain or remove unigram probs from lextree, depending on option */
if (cmd_ln_int32("-treeugprob") == 0) {
for (i = 0; i < n; i++)
wp[i].prob = -1; /* Flatten all initial probabilities */
}
/* Create the desired no. of unigram lextrees */
kb->ugtree = (lextree_t **) ckd_calloc (kb->n_lextree, sizeof(lextree_t *));
for (i = 0; i < kb->n_lextree; i++) {
kb->ugtree[i] = lextree_build (kbcore, wp, n, lc);
lextree_type (kb->ugtree[i]) = 0;
}
E_INFO("Lextrees(%d), %d nodes(ug)\n",
kb->n_lextree, lextree_n_node(kb->ugtree[0]));
}
/* Create filler lextrees */
/* ARCHAN : only one filler tree is supposed to be build even for dynamic LMs */
n = 0;
for (i = dict_filler_start(dict); i <= dict_filler_end(dict); i++) {
if (dict_filler_word(dict, i)) {
wp[n].wid = i;
wp[n].prob = fillpen (kbcore->fillpen, i);
n++;
}
}
kb->fillertree = (lextree_t **)ckd_calloc(kb->n_lextree,sizeof(lextree_t*));
for (i = 0; i < kb->n_lextree; i++) {
kb->fillertree[i] = lextree_build (kbcore, wp, n, NULL);
lextree_type (kb->fillertree[i]) = -1;
}
ckd_free ((void *) wp);
ckd_free ((void *) lc);
bitvec_free (lc_active);
E_INFO("Lextrees(%d), %d nodes(filler)\n",
kb->n_lextree,
lextree_n_node(kb->fillertree[0]));
if (cmd_ln_int32("-lextreedump")) {
if(lmset){
E_FATAL("Currently, doesn't support -lextreedump for multiple-LMs\n");
}
for (i = 0; i < kb->n_lextree; i++) {
fprintf (stderr, "UGTREE %d\n", i);
lextree_dump (kb->ugtree[i], dict, stderr);
}
for (i = 0; i < kb->n_lextree; i++) {
fprintf (stderr, "FILLERTREE %d\n", i);
lextree_dump (kb->fillertree[i], dict, stderr);
}
fflush (stderr);
}
kb->ascr = ascr_init (mgau_n_mgau(kbcore_mgau(kbcore)),
kbcore->dict2pid->n_comstate);
kb->beam = beam_init (cmd_ln_float64("-subvqbeam"),
cmd_ln_float64("-beam"),
cmd_ln_float64("-pbeam"),
cmd_ln_float64("-wbeam"));
E_INFO("Beam= %d, PBeam= %d, WBeam= %d, SVQBeam= %d\n",
kb->beam->hmm, kb->beam->ptrans, kb->beam->word, kb->beam->subvq);
/*Sections of optimization related parameters*/
kb->ds_ratio=cmd_ln_int32("-ds");
E_INFO("Down Sampling Ratio = %d\n",kb->ds_ratio);
kb->rec_bstcid=-1;
kb->skip_count=0;
kb->cond_ds=cmd_ln_int32("-cond_ds");
E_INFO("Conditional Down Sampling Parameter = %d\n",kb->cond_ds);
if(kb->cond_ds>0&&kb->kbcore->gs==NULL) E_FATAL("Conditional Down Sampling require the use of Gaussian Selection map\n");
kb->gs4gs=cmd_ln_int32("-gs4gs");
E_INFO("GS map would be used for Gaussian Selection? = %d\n",kb->gs4gs);
kb->svq4svq=cmd_ln_int32("-svq4svq");
E_INFO("SVQ would be used as Gaussian Score ?= %d\n",kb->svq4svq);
kb->ci_pbeam=-1*logs3(cmd_ln_float32("-ci_pbeam"));
E_INFO("CI phone beam to prune the number of parent CI phones in CI-base GMM Selection = %d\n",kb->ci_pbeam);
if(kb->ci_pbeam>10000000){
E_INFO("Virtually no CI phone beam is applied now. (ci_pbeam>1000000)\n");
}
kb->wend_beam=-1*logs3(cmd_ln_float32("-wend_beam"));
E_INFO("Word-end pruning beam: %d\n",kb->wend_beam);
kb->pl_window=cmd_ln_int32("-pl_window");
E_INFO("Phoneme look-ahead window size = %d\n",kb->pl_window);
kb->pl_window_start=0;
kb->pl_beam=logs3(cmd_ln_float32("-pl_beam"));
E_INFO("Phoneme look-ahead beam = %d\n",kb->pl_beam);
for(cisencnt=0;cisencnt==mdef->cd2cisen[cisencnt];cisencnt++)
;
kb->cache_ci_senscr=(int32**)ckd_calloc_2d(kb->pl_window,cisencnt,sizeof(int32));
kb->cache_best_list=(int32*)ckd_calloc(kb->pl_window,sizeof(int32));
kb->phn_heur_list=(int32*)ckd_calloc(mdef_n_ciphone (mdef),sizeof(int32));
if ((kb->feat = feat_array_alloc(kbcore_fcb(kbcore),S3_MAX_FRAMES)) == NULL)
E_FATAL("feat_array_alloc() failed\n");
kb->vithist = vithist_init(kbcore, kb->beam->word, cmd_ln_int32("-bghist"));
ptmr_init (&(kb->tm_sen));
ptmr_init (&(kb->tm_srch));
ptmr_init (&(kb->tm_ovrhd));
kb->tot_fr = 0;
kb->tot_sen_eval = 0.0;
kb->tot_gau_eval = 0.0;
kb->tot_hmm_eval = 0.0;
kb->tot_wd_exit = 0.0;
kb->hmm_hist_binsize = cmd_ln_int32("-hmmhistbinsize");
if(lmset)
n = ((kb->ugtreeMulti[0]->n_node) + (kb->fillertree[0]->n_node)) * kb->n_lextree;
else
n = ((kb->ugtree[0]->n_node) + (kb->fillertree[0]->n_node)) * kb->n_lextree;
n /= kb->hmm_hist_binsize;
kb->hmm_hist_bins = n+1;
kb->hmm_hist = (int32 *) ckd_calloc (n+1, sizeof(int32)); /* Really no need for +1 */
/* Open hypseg file if specified */
str = cmd_ln_str("-hypseg");
kb->matchsegfp = NULL;
if (str) {
#ifdef WIN32
if ((kb->matchsegfp = fopen(str, "wt")) == NULL)
#else
if ((kb->matchsegfp = fopen(str, "w")) == NULL)
#endif
E_ERROR("fopen(%s,w) failed; use FWDXCT: from std logfile\n", str);
}
str = cmd_ln_str("-hyp");
kb->matchfp = NULL;
if (str) {
#ifdef WIN32
if ((kb->matchfp = fopen(str, "wt")) == NULL)
#else
if ((kb->matchfp = fopen(str, "w")) == NULL)
#endif
E_ERROR("fopen(%s,w) failed; use FWDXCT: from std logfile\n", str);
}
}
dtree_t *
mk_tree_comp(float32 ****mixw,
/* ADDITION FOR CONTINUOUS_TREES, 18 May 98 */
float32 ****means,
float32 ****vars,
uint32 *veclen,
/* END ADDITIONS FOR CONTINUOUS_TREES */
uint32 n_model,
uint32 n_state,
uint32 n_stream,
uint32 n_density,
float32 *stwt,
uint32 *id,
uint32 n_id,
quest_t *all_q,
uint32 n_all_q,
pset_t *pset,
uint32 n_base_phone,
uint32 **dfeat,
uint32 n_dfeat,
uint32 split_min,
uint32 split_max,
float32 split_thr,
uint32 split_min_comp,
uint32 split_max_comp,
float32 split_thr_comp,
float32 mwfloor)
{
dtree_t *comp_tree;
dtree_node_t *root, *b_n;
uint32 i;
comp_tree = ckd_calloc(1, sizeof(dtree_t));
comp_tree->node = ckd_calloc(2*split_max_comp+1, sizeof(dtree_node_t));
comp_tree->n_node = 0;
comp_tree->node[0].node_id = 0;
comp_tree->n_node = 1;
root = &comp_tree->node[0];
/* MODIFICATION FOR CONTINUOUS_TREES 18 May 98, pass means and var along with mixw */
mk_node(root, 0,
id, n_id,
mixw, means, vars, veclen,
n_model, n_state, n_stream, n_density, stwt, mwfloor);
/* END MODIFICATION FOR CONTINUOUS_TREES */
/* MODIFICATION FOR CONTINUOUS_TREES 18 May 98, pass means and var along with mixw */
root->q = (void *)mk_comp_quest(&root->wt_ent_dec,
mixw, means, vars, veclen,
n_model, n_state, n_stream, n_density, stwt,
id, n_id,
all_q, n_all_q, pset, n_base_phone,
dfeat, n_dfeat,
split_min, split_max, split_thr,
mwfloor);
/* END MODIFICATION FOR CONTINUOUS_TREES */
for (i = 0; i < split_max_comp; i++) {
b_n = best_leaf_node(root);
E_INFO("Comp split %u\n", i);
if (b_n == NULL) {
E_INFO("stop. leaf nodes are specific\n");
break;
}
if (b_n->wt_ent_dec <= 0) {
E_INFO("stop. b_n->wt_ent_dec (%.3e) <= 0\n",
b_n->wt_ent_dec);
break;
}
if ((i > split_min_comp) &&
(b_n->wt_ent_dec < split_thr_comp * b_n->wt_ent)) {
E_INFO("stop. b_n->wt_ent_dec <= split_thr_comp * b_n->wt_ent. %.3e <= %.3e\n",
b_n->wt_ent_dec, split_thr_comp * b_n->wt_ent);
break;
}
/* MODIFICATION FOR CONTINUOUS_TREES 18 May 98, pass means and var along with mixw */
split_node_comp(comp_tree, b_n->node_id,
mixw, means, vars, veclen,
n_model, n_state, n_stream, n_density, stwt,
all_q, n_all_q, pset, n_base_phone,
dfeat, n_dfeat,
split_min, split_max, split_thr, mwfloor);
/* END MODIFICATION FOR CONTINUOUS_TREES */
#if 0
printf("Comp Split %u:\n", i);
print_tree_comp(stderr, "*", root, pset, 0);
fprintf(stderr, "\n");
#endif
}
#if 0
E_INFO("Final Comp Tree %u:\n", i);
print_tree_comp(stderr, "", root, pset, 0);
fprintf(stderr, "\n");
#endif
return comp_tree;
}
fillpen_t *fillpen_init (dict_t *dict, char *file, float64 silprob, float64 fillprob,
float64 lw, float64 wip)
{
s3wid_t w, bw;
float64 prob;
FILE *fp;
char line[1024], wd[1024];
int32 k;
fillpen_t *_fillpen;
_fillpen = (fillpen_t *) ckd_calloc (1, sizeof(fillpen_t));
_fillpen->dict = dict;
_fillpen->lw = lw;
_fillpen->wip = wip;
if (dict->filler_end >= dict->filler_start)
_fillpen->prob = (int32 *) ckd_calloc (dict->filler_end - dict->filler_start + 1,
sizeof(int32));
else
_fillpen->prob = NULL;
/* Initialize all words with filler penalty (HACK!! backward compatibility) */
prob = fillprob;
for (w = dict->filler_start; w <= dict->filler_end; w++)
_fillpen->prob[w - dict->filler_start] = (int32) ((logs3(prob) + logs3(wip)) * lw);
/* Overwrite silence penalty (HACK!! backward compatibility) */
w = dict_wordid (dict, S3_SILENCE_WORD);
if (NOT_S3WID(w) || (w < dict->filler_start) || (w > dict->filler_end))
E_FATAL("%s not a filler word in the given dictionary\n", S3_SILENCE_WORD);
prob = silprob;
_fillpen->prob[w - dict->filler_start] = (int32) ((logs3(prob) + logs3(wip)) * lw);
/* Overwrite with filler prob input file, if specified */
if (! file)
return _fillpen;
E_INFO("Reading filler penalty file: %s\n", file);
if ((fp = fopen (file, "r")) == NULL)
E_FATAL("fopen(%s,r) failed\n", file);
while (fgets (line, sizeof(line), fp) != NULL) {
if (line[0] == '#') /* Skip comment lines */
continue;
k = sscanf (line, "%s %lf", wd, &prob);
if ((k != 0) && (k != 2))
E_FATAL("Bad input line: %s\n", line);
w = dict_wordid(dict, wd);
if (NOT_S3WID(w) || (w < dict->filler_start) || (w > dict->filler_end))
E_FATAL("%s not a filler word in the given dictionary\n", S3_SILENCE_WORD);
_fillpen->prob[w - dict->filler_start] = (int32) ((logs3(prob) + logs3(wip)) * lw);
}
fclose (fp);
/* Replicate fillpen values for alternative pronunciations */
for (w = dict->filler_start; w <= dict->filler_end; w++) {
bw = dict_basewid (dict, w);
if (bw != w)
_fillpen->prob[w-dict->filler_start] = _fillpen->prob[bw-dict->filler_start];
}
return _fillpen;
}
int
ps_start_utt(ps_decoder_t *ps, char const *uttid)
{
FILE *mfcfh = NULL;
FILE *rawfh = NULL;
int rv;
if (ps->search == NULL) {
E_ERROR("No search module is selected, did you forget to "
"specify a language model or grammar?\n");
return -1;
}
ptmr_reset(&ps->perf);
ptmr_start(&ps->perf);
if (uttid) {
ckd_free(ps->uttid);
ps->uttid = ckd_salloc(uttid);
}
else {
char nuttid[16];
ckd_free(ps->uttid);
sprintf(nuttid, "%09u", ps->uttno);
ps->uttid = ckd_salloc(nuttid);
++ps->uttno;
}
/* Remove any residual word lattice and hypothesis. */
ps_lattice_free(ps->search->dag);
ps->search->dag = NULL;
ps->search->last_link = NULL;
ps->search->post = 0;
ckd_free(ps->search->hyp_str);
ps->search->hyp_str = NULL;
if ((rv = acmod_start_utt(ps->acmod)) < 0)
return rv;
/* Start logging features and audio if requested. */
if (ps->mfclogdir) {
char *logfn = string_join(ps->mfclogdir, "/",
ps->uttid, ".mfc", NULL);
E_INFO("Writing MFCC log file: %s\n", logfn);
if ((mfcfh = fopen(logfn, "wb")) == NULL) {
E_ERROR_SYSTEM("Failed to open MFCC log file %s", logfn);
ckd_free(logfn);
return -1;
}
ckd_free(logfn);
acmod_set_mfcfh(ps->acmod, mfcfh);
}
if (ps->rawlogdir) {
char *logfn = string_join(ps->rawlogdir, "/",
ps->uttid, ".raw", NULL);
E_INFO("Writing raw audio log file: %s\n", logfn);
if ((rawfh = fopen(logfn, "wb")) == NULL) {
E_ERROR_SYSTEM("Failed to open raw audio log file %s", logfn);
ckd_free(logfn);
return -1;
}
ckd_free(logfn);
acmod_set_rawfh(ps->acmod, rawfh);
}
/* Start auxiliary phone loop search. */
if (ps->phone_loop)
ps_search_start(ps->phone_loop);
return ps_search_start(ps->search);
}
acmod_t *
acmod_init(cmd_ln_t *config, logmath_t *lmath, fe_t *fe, feat_t *fcb)
{
acmod_t *acmod;
char const *featparams;
acmod = ckd_calloc(1, sizeof(*acmod));
acmod->config = cmd_ln_retain(config);
acmod->lmath = lmath;
acmod->state = ACMOD_IDLE;
/* Look for feat.params in acoustic model dir. */
if ((featparams = cmd_ln_str_r(acmod->config, "-featparams"))) {
if (NULL !=
cmd_ln_parse_file_r(acmod->config, feat_defn, featparams, FALSE))
E_INFO("Parsed model-specific feature parameters from %s\n",
featparams);
}
/* Initialize feature computation. */
if (fe) {
if (acmod_fe_mismatch(acmod, fe))
goto error_out;
fe_retain(fe);
acmod->fe = fe;
}
else {
/* Initialize a new front end. */
acmod->fe = fe_init_auto_r(config);
if (acmod->fe == NULL)
goto error_out;
if (acmod_fe_mismatch(acmod, acmod->fe))
goto error_out;
}
if (fcb) {
if (acmod_feat_mismatch(acmod, fcb))
goto error_out;
feat_retain(fcb);
acmod->fcb = fcb;
}
else {
/* Initialize a new fcb. */
if (acmod_init_feat(acmod) < 0)
goto error_out;
}
/* Load acoustic model parameters. */
if (acmod_init_am(acmod) < 0)
goto error_out;
/* The MFCC buffer needs to be at least as large as the dynamic
* feature window. */
acmod->n_mfc_alloc = acmod->fcb->window_size * 2 + 1;
acmod->mfc_buf = (mfcc_t **)
ckd_calloc_2d(acmod->n_mfc_alloc, acmod->fcb->cepsize,
sizeof(**acmod->mfc_buf));
/* Feature buffer has to be at least as large as MFCC buffer. */
acmod->n_feat_alloc = acmod->n_mfc_alloc + cmd_ln_int32_r(config, "-pl_window");
acmod->feat_buf = feat_array_alloc(acmod->fcb, acmod->n_feat_alloc);
acmod->framepos = ckd_calloc(acmod->n_feat_alloc, sizeof(*acmod->framepos));
acmod->utt_start_frame = 0;
/* Senone computation stuff. */
acmod->senone_scores = ckd_calloc(bin_mdef_n_sen(acmod->mdef),
sizeof(*acmod->senone_scores));
acmod->senone_active_vec = bitvec_alloc(bin_mdef_n_sen(acmod->mdef));
acmod->senone_active = ckd_calloc(bin_mdef_n_sen(acmod->mdef),
sizeof(*acmod->senone_active));
acmod->log_zero = logmath_get_zero(acmod->lmath);
acmod->compallsen = cmd_ln_boolean_r(config, "-compallsen");
return acmod;
error_out:
acmod_free(acmod);
return NULL;
}
kbcore_t *kbcore_init (float64 logbase,
char *feattype,
char *cmn,
char *varnorm,
char *agc,
char *mdeffile,
char *dictfile,
char *fdictfile,
char *compsep,
char *lmfile,
char *fillpenfile,
float64 silprob,
float64 fillprob,
float64 langwt,
float64 inspen,
float64 uw,
char *meanfile,
char *varfile,
float64 varfloor,
char *mixwfile,
float64 mixwfloor,
char *subvqfile,
char *tmatfile,
float64 tmatfloor)
{
kbcore_t *kb;
E_INFO("Initializing core models:\n");
kb = (kbcore_t *) ckd_calloc (1, sizeof(kbcore_t));
kb->fcb = NULL;
kb->mdef = NULL;
kb->dict = NULL;
kb->dict2pid = NULL;
kb->lm = NULL;
kb->fillpen = NULL;
kb->dict2lmwid = NULL;
kb->mgau = NULL;
kb->svq = NULL;
kb->tmat = NULL;
logs3_init (logbase);
if (feattype) {
if ((kb->fcb = feat_init (feattype, cmn, varnorm, agc)) == NULL)
E_FATAL("feat_init(%s) failed\n", feattype);
if (feat_n_stream(kb->fcb) != 1)
E_FATAL("#Feature streams(%d) != 1\n", feat_n_stream(kb->fcb));
}
if (mdeffile) {
if ((kb->mdef = mdef_init (mdeffile)) == NULL)
E_FATAL("mdef_init(%s) failed\n", mdeffile);
}
if (dictfile) {
if (! compsep)
compsep = "";
else if ((compsep[0] != '\0') && (compsep[1] != '\0')) {
E_FATAL("Compound word separator(%s) must be empty or single character string\n",
compsep);
}
if ((kb->dict = dict_init (kb->mdef, dictfile, fdictfile, compsep[0])) == NULL)
E_FATAL("dict_init(%s,%s,%s) failed\n", dictfile,
fdictfile ? fdictfile : "", compsep);
}
if (lmfile) {
if ((kb->lm = lm_read (lmfile, langwt, inspen, uw)) == NULL)
E_FATAL("lm_read(%s, %e, %e, %e) failed\n", lmfile, langwt, inspen, uw);
}
if (fillpenfile || (lmfile && kb->dict)) {
if (! kb->dict) /* Sic */
E_FATAL("No dictionary for associating filler penalty file(%s)\n", fillpenfile);
if ((kb->fillpen = fillpen_init (kb->dict, fillpenfile, silprob, fillprob,
langwt, inspen)) == NULL)
E_FATAL("fillpen_init(%s) failed\n", fillpenfile);
}
if (meanfile) {
if ((! varfile) || (! mixwfile))
E_FATAL("Varfile or mixwfile not specified along with meanfile(%s)\n", meanfile);
kb->mgau = mgau_init (meanfile, varfile, varfloor, mixwfile, mixwfloor, TRUE);
if (kb->mgau == NULL)
E_FATAL("gauden_init(%s, %s, %e) failed\n", meanfile, varfile, varfloor);
if (subvqfile) {
if ((kb->svq = subvq_init (subvqfile, varfloor, -1, kb->mgau)) == NULL)
E_FATAL("subvq_init (%s, %e, -1) failed\n", subvqfile, varfloor);
}
}
if (tmatfile) {
if ((kb->tmat = tmat_init (tmatfile, tmatfloor)) == NULL)
E_FATAL("tmat_init (%s, %e) failed\n", tmatfile, tmatfloor);
}
if (kb->dict && kb->lm) { /* Initialize dict2lmwid */
if ((kb->dict2lmwid = wid_dict_lm_map (kb->dict, kb->lm)) == NULL)
E_FATAL("Dict/LM word-id mapping failed\n");
}
if (kb->mdef && kb->dict) { /* Initialize dict2pid */
kb->dict2pid = dict2pid_build (kb->mdef, kb->dict);
}
/* ***************** Verifications ***************** */
E_INFO("Verifying models consistency:\n");
if (kb->fcb && kb->mgau) {
/* Verify feature streams against gauden codebooks */
if (feat_stream_len(kb->fcb, 0) != mgau_veclen(kb->mgau))
E_FATAL("Feature streamlen(%d) != mgau streamlen(%d)\n",
feat_stream_len(kb->fcb, 0), mgau_veclen(kb->mgau));
}
if (kb->mdef && kb->mgau) {
/* Verify senone parameters against model definition parameters */
if (kb->mdef->n_sen != mgau_n_mgau(kb->mgau))
E_FATAL("Mdef #senones(%d) != mgau #senones(%d)\n",
kb->mdef->n_sen, mgau_n_mgau(kb->mgau));
}
if (kb->mdef && kb->tmat) {
/* Verify transition matrices parameters against model definition parameters */
if (kb->mdef->n_tmat != kb->tmat->n_tmat)
E_FATAL("Mdef #tmat(%d) != tmatfile(%d)\n", kb->mdef->n_tmat, kb->tmat->n_tmat);
if (kb->mdef->n_emit_state != kb->tmat->n_state)
E_FATAL("Mdef #states(%d) != tmat #states(%d)\n",
kb->mdef->n_emit_state, kb->tmat->n_state);
}
return kb;
}
int
ngram_fwdflat_search(ngram_search_t *ngs, int frame_idx)
{
int16 const *senscr;
int32 nf, i, j;
int32 *nawl;
/* Activate our HMMs for the current frame if need be. */
if (!ps_search_acmod(ngs)->compallsen)
compute_fwdflat_sen_active(ngs, frame_idx);
/* Compute GMM scores for the current frame. */
senscr = acmod_score(ps_search_acmod(ngs), &frame_idx);
ngs->st.n_senone_active_utt += ps_search_acmod(ngs)->n_senone_active;
/* Mark backpointer table for current frame. */
ngram_search_mark_bptable(ngs, frame_idx);
/* If the best score is equal to or worse than WORST_SCORE,
* recognition has failed, don't bother to keep trying. */
if (ngs->best_score == WORST_SCORE || ngs->best_score WORSE_THAN WORST_SCORE)
return 0;
/* Renormalize if necessary */
if (ngs->best_score + (2 * ngs->beam) WORSE_THAN WORST_SCORE) {
E_INFO("Renormalizing Scores at frame %d, best score %d\n",
frame_idx, ngs->best_score);
fwdflat_renormalize_scores(ngs, frame_idx, ngs->best_score);
}
ngs->best_score = WORST_SCORE;
hmm_context_set_senscore(ngs->hmmctx, senscr);
/* Evaluate HMMs */
fwdflat_eval_chan(ngs, frame_idx);
/* Prune HMMs and do phone transitions. */
fwdflat_prune_chan(ngs, frame_idx);
/* Do word transitions. */
fwdflat_word_transition(ngs, frame_idx);
/* Create next active word list */
nf = frame_idx + 1;
nawl = ngs->active_word_list[nf & 0x1];
for (i = 0, j = 0; ngs->fwdflat_wordlist[i] >= 0; i++) {
if (bitvec_is_set(ngs->word_active, ngs->fwdflat_wordlist[i])) {
*(nawl++) = ngs->fwdflat_wordlist[i];
j++;
}
}
for (i = ps_search_start_wid(ngs); i < ps_search_n_words(ngs); i++) {
if (bitvec_is_set(ngs->word_active, i)) {
*(nawl++) = i;
j++;
}
}
if (!ngs->fwdtree)
++ngs->n_frame;
ngs->n_active_word[nf & 0x1] = j;
/* Return the number of frames processed. */
return 1;
}
/* Read a Sphinx3 mean or variance file. */
static int32
s3_read_mgau(s2_semi_mgau_t *s, const char *file_name, float32 ***out_cb)
{
char tmp;
FILE *fp;
int32 i, blk, n;
int32 n_mgau;
int32 n_feat;
int32 n_density;
int32 *veclen;
int32 byteswap, chksum_present;
char **argname, **argval;
uint32 chksum;
E_INFO("Reading S3 mixture gaussian file '%s'\n", file_name);
if ((fp = fopen(file_name, "rb")) == NULL)
E_FATAL("fopen(%s,rb) failed\n", file_name);
/* Read header, including argument-value info and 32-bit byteorder magic */
if (bio_readhdr(fp, &argname, &argval, &byteswap) < 0)
E_FATAL("bio_readhdr(%s) failed\n", file_name);
/* Parse argument-value list */
chksum_present = 0;
for (i = 0; argname[i]; i++) {
if (strcmp(argname[i], "version") == 0) {
if (strcmp(argval[i], MGAU_PARAM_VERSION) != 0)
E_WARN("Version mismatch(%s): %s, expecting %s\n",
file_name, argval[i], MGAU_PARAM_VERSION);
}
else if (strcmp(argname[i], "chksum0") == 0) {
chksum_present = 1; /* Ignore the associated value */
}
}
bio_hdrarg_free(argname, argval);
argname = argval = NULL;
chksum = 0;
/* #Codebooks */
if (bio_fread(&n_mgau, sizeof(int32), 1, fp, byteswap, &chksum) != 1)
E_FATAL("fread(%s) (#codebooks) failed\n", file_name);
if (n_mgau != 1) {
E_ERROR("%s: #codebooks (%d) != 1\n", file_name, n_mgau);
fclose(fp);
return -1;
}
/* #Features/codebook */
if (bio_fread(&n_feat, sizeof(int32), 1, fp, byteswap, &chksum) != 1)
E_FATAL("fread(%s) (#features) failed\n", file_name);
if (s->n_feat == 0)
s->n_feat = n_feat;
else if (n_feat != s->n_feat)
E_FATAL("#Features streams(%d) != %d\n", n_feat, s->n_feat);
/* #Gaussian densities/feature in each codebook */
if (bio_fread(&n_density, sizeof(int32), 1, fp,
byteswap, &chksum) != 1)
E_FATAL("fread(%s) (#density/codebook) failed\n", file_name);
if (s->n_density == 0)
s->n_density = n_density;
else if (n_density != s->n_density)
E_FATAL("%s: Number of densities per feature(%d) != %d\n",
file_name, n_mgau, s->n_density);
/* Vector length of feature stream */
veclen = ckd_calloc(s->n_feat, sizeof(int32));
if (bio_fread(veclen, sizeof(int32), s->n_feat,
fp, byteswap, &chksum) != s->n_feat)
E_FATAL("fread(%s) (feature vector-length) failed\n", file_name);
for (i = 0, blk = 0; i < s->n_feat; ++i) {
if (s->veclen[i] == 0)
s->veclen[i] = veclen[i];
else if (veclen[i] != s->veclen[i])
E_FATAL("feature stream length %d is inconsistent (%d != %d)\n",
i, veclen[i], s->veclen[i]);
blk += veclen[i];
}
/* #Floats to follow; for the ENTIRE SET of CODEBOOKS */
if (bio_fread(&n, sizeof(int32), 1, fp, byteswap, &chksum) != 1)
E_FATAL("fread(%s) (total #floats) failed\n", file_name);
if (n != n_mgau * n_density * blk)
E_FATAL
("%s: #float32s(%d) doesn't match dimensions: %d x %d x %d\n",
file_name, n, n_mgau, n_density, blk);
*out_cb = ckd_calloc(s->n_feat, sizeof(float32 *));
for (i = 0; i < s->n_feat; ++i) {
(*out_cb)[i] =
(float32 *) ckd_calloc(n_density * veclen[i],
sizeof(float32));
if (bio_fread
((*out_cb)[i], sizeof(float32),
n_density * veclen[i], fp,
byteswap, &chksum) != n_density * veclen[i])
E_FATAL("fread(%s, %d) of feat %d failed\n", file_name,
n_density * veclen[i], i);
}
ckd_free(veclen);
if (chksum_present)
bio_verify_chksum(fp, byteswap, chksum);
if (fread(&tmp, 1, 1, fp) == 1)
E_FATAL("%s: More data than expected\n", file_name);
fclose(fp);
E_INFO("%d mixture Gaussians, %d components, %d feature streams, veclen %d\n", n_mgau,
n_density, n_feat, blk);
return n;
}
int
ps_end_utt(ps_decoder_t *ps)
{
int rv, i;
if (ps->acmod->state == ACMOD_ENDED || ps->acmod->state == ACMOD_IDLE) {
E_ERROR("Utterance is not started\n");
return -1;
}
acmod_end_utt(ps->acmod);
/* Search any remaining frames. */
if ((rv = ps_search_forward(ps)) < 0) {
ptmr_stop(&ps->perf);
return rv;
}
/* Finish phone loop search. */
if (ps->phone_loop) {
if ((rv = ps_search_finish(ps->phone_loop)) < 0) {
ptmr_stop(&ps->perf);
return rv;
}
}
/* Search any frames remaining in the lookahead window. */
if (ps->acmod->output_frame >= ps->pl_window) {
for (i = ps->acmod->output_frame - ps->pl_window;
i < ps->acmod->output_frame; ++i)
ps_search_step(ps->search, i);
}
/* Finish main search. */
if ((rv = ps_search_finish(ps->search)) < 0) {
ptmr_stop(&ps->perf);
return rv;
}
ptmr_stop(&ps->perf);
/* Log a backtrace if requested. */
if (cmd_ln_boolean_r(ps->config, "-backtrace")) {
const char* hyp;
ps_seg_t *seg;
int32 score;
hyp = ps_get_hyp(ps, &score);
if (hyp != NULL) {
E_INFO("%s (%d)\n", hyp, score);
E_INFO_NOFN("%-20s %-5s %-5s %-5s %-10s %-10s %-3s\n",
"word", "start", "end", "pprob", "ascr", "lscr", "lback");
for (seg = ps_seg_iter(ps, &score); seg;
seg = ps_seg_next(seg)) {
char const *word;
int sf, ef;
int32 post, lscr, ascr, lback;
word = ps_seg_word(seg);
ps_seg_frames(seg, &sf, &ef);
post = ps_seg_prob(seg, &ascr, &lscr, &lback);
E_INFO_NOFN("%-20s %-5d %-5d %-1.3f %-10d %-10d %-3d\n",
word, sf, ef, logmath_exp(ps_get_logmath(ps), post),
ascr, lscr, lback);
}
}
}
return rv;
}
static int32
read_sendump(s2_semi_mgau_t *s, mdef_t *mdef, char const *file)
{
FILE *fp;
char line[1000];
int32 i, n;
int32 do_swap, do_mmap;
size_t filesize, offset;
int n_clust = 256; /* Number of clusters (if zero, we are just using
* 8-bit quantized weights) */
int r = s->n_density;
int c = mdef_n_sen(mdef);
s->n_sen = c;
do_mmap = cmd_ln_boolean_r(s->config, "-mmap");
if ((fp = fopen(file, "rb")) == NULL)
return -1;
E_INFO("Loading senones from dump file %s\n", file);
/* Read title size, title */
fread(&n, sizeof(int32), 1, fp);
/* This is extremely bogus */
do_swap = 0;
if (n < 1 || n > 999) {
SWAP_INT32(&n);
if (n < 1 || n > 999) {
E_FATAL("Title length %x in dump file %s out of range\n", n, file);
}
do_swap = 1;
}
if (fread(line, sizeof(char), n, fp) != n)
E_FATAL("Cannot read title\n");
if (line[n - 1] != '\0')
E_FATAL("Bad title in dump file\n");
E_INFO("%s\n", line);
/* Read header size, header */
fread(&n, 1, sizeof(n), fp);
if (do_swap) SWAP_INT32(&n);
if (fread(line, sizeof(char), n, fp) != n)
E_FATAL("Cannot read header\n");
if (line[n - 1] != '\0')
E_FATAL("Bad header in dump file\n");
/* Read other header strings until string length = 0 */
for (;;) {
fread(&n, 1, sizeof(n), fp);
if (do_swap) SWAP_INT32(&n);
if (n == 0)
break;
if (fread(line, sizeof(char), n, fp) != n)
E_FATAL("Cannot read header\n");
/* Look for a cluster count, if present */
if (!strncmp(line, "cluster_count ", strlen("cluster_count "))) {
n_clust = atoi(line + strlen("cluster_count "));
}
}
/* Read #codewords, #pdfs */
fread(&r, 1, sizeof(r), fp);
if (do_swap) SWAP_INT32(&r);
fread(&c, 1, sizeof(c), fp);
if (do_swap) SWAP_INT32(&c);
E_INFO("Rows: %d, Columns: %d\n", r, c);
if (n_clust) {
E_ERROR ("Dump file is incompatible with PocketSphinx\n");
fclose(fp);
return -1;
}
if (do_mmap) {
E_INFO("Using memory-mapped I/O for senones\n");
}
offset = ftell(fp);
fseek(fp, 0, SEEK_END);
filesize = ftell(fp);
fseek(fp, offset, SEEK_SET);
/* Allocate memory for pdfs (or memory map them) */
if (do_mmap)
s->sendump_mmap = mmio_file_read(file);
/* Otherwise, set up all pointers, etc. */
if (s->sendump_mmap) {
s->mixw = ckd_calloc(s->n_feat, sizeof(*s->mixw));
for (i = 0; i < s->n_feat; i++) {
/* Pointers into the mmap()ed 2d array */
s->mixw[i] = ckd_calloc(r, sizeof(**s->mixw));
}
for (n = 0; n < s->n_feat; n++) {
for (i = 0; i < r; i++) {
s->mixw[n][i] = ((uint8 *) mmio_file_ptr(s->sendump_mmap)) + offset;
offset += c;
}
}
}
else {
s->mixw = ckd_calloc_3d(s->n_feat, r, c, sizeof(***s->mixw));
/* Read pdf values and ids */
for (n = 0; n < s->n_feat; n++) {
for (i = 0; i < r; i++) {
if (fread(s->mixw[n][i], sizeof(***s->mixw), c, fp) != (size_t) c) {
E_ERROR("Failed to read %d bytes from sendump\n", c);
return -1;
}
}
}
}
fclose(fp);
return 0;
}
static int32
read_mixw(s2_semi_mgau_t * s, char const *file_name, double SmoothMin)
{
char **argname, **argval;
char eofchk;
FILE *fp;
int32 byteswap, chksum_present;
uint32 chksum;
float32 *pdf;
int32 i, f, c, n;
int32 n_sen;
int32 n_feat;
int32 n_comp;
int32 n_err;
E_INFO("Reading mixture weights file '%s'\n", file_name);
if ((fp = fopen(file_name, "rb")) == NULL)
E_FATAL("fopen(%s,rb) failed\n", file_name);
/* Read header, including argument-value info and 32-bit byteorder magic */
if (bio_readhdr(fp, &argname, &argval, &byteswap) < 0)
E_FATAL("bio_readhdr(%s) failed\n", file_name);
/* Parse argument-value list */
chksum_present = 0;
for (i = 0; argname[i]; i++) {
if (strcmp(argname[i], "version") == 0) {
if (strcmp(argval[i], MGAU_MIXW_VERSION) != 0)
E_WARN("Version mismatch(%s): %s, expecting %s\n",
file_name, argval[i], MGAU_MIXW_VERSION);
}
else if (strcmp(argname[i], "chksum0") == 0) {
chksum_present = 1; /* Ignore the associated value */
}
}
bio_hdrarg_free(argname, argval);
argname = argval = NULL;
chksum = 0;
/* Read #senones, #features, #codewords, arraysize */
if ((bio_fread(&n_sen, sizeof(int32), 1, fp, byteswap, &chksum) != 1)
|| (bio_fread(&n_feat, sizeof(int32), 1, fp, byteswap, &chksum) !=
1)
|| (bio_fread(&n_comp, sizeof(int32), 1, fp, byteswap, &chksum) !=
1)
|| (bio_fread(&n, sizeof(int32), 1, fp, byteswap, &chksum) != 1)) {
E_FATAL("bio_fread(%s) (arraysize) failed\n", file_name);
}
if (n_feat != s->n_feat)
E_FATAL("#Features streams(%d) != %d\n", n_feat, s->n_feat);
if (n != n_sen * n_feat * n_comp) {
E_FATAL
("%s: #float32s(%d) doesn't match header dimensions: %d x %d x %d\n",
file_name, i, n_sen, n_feat, n_comp);
}
/* n_sen = number of mixture weights per codeword, which is
* fixed at the number of senones since we have only one codebook.
*/
s->n_sen = n_sen;
/* Quantized mixture weight arrays. */
s->mixw = ckd_calloc_3d(s->n_feat, s->n_density, n_sen, sizeof(***s->mixw));
/* Temporary structure to read in floats before conversion to (int32) logs3 */
pdf = (float32 *) ckd_calloc(n_comp, sizeof(float32));
/* Read senone probs data, normalize, floor, convert to logs3, truncate to 8 bits */
n_err = 0;
for (i = 0; i < n_sen; i++) {
for (f = 0; f < n_feat; f++) {
if (bio_fread((void *) pdf, sizeof(float32),
n_comp, fp, byteswap, &chksum) != n_comp) {
E_FATAL("bio_fread(%s) (arraydata) failed\n", file_name);
}
/* Normalize and floor */
if (vector_sum_norm(pdf, n_comp) <= 0.0)
n_err++;
vector_floor(pdf, n_comp, SmoothMin);
vector_sum_norm(pdf, n_comp);
/* Convert to LOG, quantize, and transpose */
for (c = 0; c < n_comp; c++) {
int32 qscr;
qscr = -logmath_log(s->lmath_8b, pdf[c]);
if ((qscr > MAX_NEG_MIXW) || (qscr < 0))
qscr = MAX_NEG_MIXW;
s->mixw[f][c][i] = qscr;
}
}
}
if (n_err > 0)
E_WARN("Weight normalization failed for %d senones\n", n_err);
ckd_free(pdf);
if (chksum_present)
bio_verify_chksum(fp, byteswap, chksum);
if (fread(&eofchk, 1, 1, fp) == 1)
E_FATAL("More data than expected in %s\n", file_name);
fclose(fp);
E_INFO("Read %d x %d x %d mixture weights\n", n_sen, n_feat, n_comp);
return n_sen;
}
s2_semi_mgau_t *
s2_semi_mgau_init(cmd_ln_t *config, logmath_t *lmath, feat_t *fcb, mdef_t *mdef)
{
s2_semi_mgau_t *s;
char const *sendump_path;
float32 **fgau;
int i;
s = ckd_calloc(1, sizeof(*s));
s->config = config;
s->lmath = logmath_retain(lmath);
/* Log-add table. */
s->lmath_8b = logmath_init(logmath_get_base(lmath), SENSCR_SHIFT, TRUE);
if (s->lmath_8b == NULL) {
s2_semi_mgau_free(s);
return NULL;
}
/* Ensure that it is only 8 bits wide so that fast_logmath_add() works. */
if (logmath_get_width(s->lmath_8b) != 1) {
E_ERROR("Log base %f is too small to represent add table in 8 bits\n",
logmath_get_base(s->lmath_8b));
s2_semi_mgau_free(s);
return NULL;
}
/* Inherit stream dimensions from acmod, will be checked below. */
s->n_feat = feat_dimension1(fcb);
s->veclen = ckd_calloc(s->n_feat, sizeof(int32));
for (i = 0; i < s->n_feat; ++i)
s->veclen[i] = feat_dimension2(fcb, i);
/* Read means and variances. */
if (s3_read_mgau(s, cmd_ln_str_r(s->config, "-mean"), &fgau) < 0) {
s2_semi_mgau_free(s);
return NULL;
}
s->means = (mfcc_t **)fgau;
if (s3_read_mgau(s, cmd_ln_str_r(s->config, "-var"), &fgau) < 0) {
s2_semi_mgau_free(s);
return NULL;
}
s->vars = (mfcc_t **)fgau;
/* Precompute (and fixed-point-ize) means, variances, and determinants. */
s->dets = (mfcc_t **)ckd_calloc_2d(s->n_feat, s->n_density, sizeof(**s->dets));
s3_precomp(s, s->lmath, cmd_ln_float32_r(s->config, "-varfloor"));
/* Read mixture weights */
if ((sendump_path = cmd_ln_str_r(s->config, "-sendump")))
read_sendump(s, mdef, sendump_path);
else
read_mixw(s, cmd_ln_str_r(s->config, "-mixw"),
cmd_ln_float32_r(s->config, "-mixwfloor"));
s->ds_ratio = cmd_ln_int32_r(s->config, "-ds");
/* Determine top-N for each feature */
s->topn_beam = ckd_calloc(s->n_feat, sizeof(*s->topn_beam));
s->max_topn = cmd_ln_int32_r(s->config, "-topn");
split_topn(cmd_ln_str_r(s->config, "-topn_beam"), s->topn_beam, s->n_feat);
E_INFO("Maximum top-N: %d ", s->max_topn);
E_INFOCONT("Top-N beams:");
for (i = 0; i < s->n_feat; ++i) {
E_INFOCONT(" %d", s->topn_beam[i]);
}
E_INFOCONT("\n");
/* Top-N scores from recent frames */
s->n_topn_hist = cmd_ln_int32_r(s->config, "-pl_window") + 2;
s->topn_hist = (vqFeature_t ***)
ckd_calloc_3d(s->n_topn_hist, s->n_feat, s->max_topn,
sizeof(***s->topn_hist));
s->topn_hist_n = ckd_calloc_2d(s->n_topn_hist, s->n_feat,
sizeof(**s->topn_hist_n));
for (i = 0; i < s->n_topn_hist; ++i) {
int j;
for (j = 0; j < s->n_feat; ++j) {
int k;
for (k = 0; k < s->max_topn; ++k) {
s->topn_hist[i][j][k].score = WORST_DIST;
s->topn_hist[i][j][k].codeword = k;
}
}
}
return s;
}
int
parse_cmd_ln(int argc, char *argv[])
{
uint32 isHelp;
uint32 isExample;
const char helpstr[] =
"Description: \n\
(copied from Eric's comments)\n\
* Create a tied-state-to-codebook mapping file for semi-continuous, \n\
* phone dependent or fully continuous Gaussian density tying.";
const char examplestr[]=
"Example: \n\
(By Arthur: Not sure, may be obsolete) \n\
mk_ts2cb -moddeffn semi -ts2cbfn ts2cb";
static arg_t defn[] = {
{ "-help",
ARG_BOOLEAN,
"no",
"Shows the usage of the tool"
},
{ "-example",
ARG_BOOLEAN,
"no",
"Shows example of how to use the tool"
},
{ "-ts2cbfn",
ARG_STRING,
NULL,
"A SPHINX-III tied-state-to-cb file name"
},
{ "-moddeffn",
ARG_STRING,
NULL,
"A SPHINX-III model definition file name"
},
{ "-tyingtype", /* either "semi", "pd", or "cont" */
ARG_STRING,
"semi",
"Output a state parameter def file for fully continuous models"
},
{ "-pclassfn", /* this switch is reference for -tyingtype pd */
ARG_STRING,
NULL,
"A SPHINX-II phone class file name"
},
{ NULL,
0,
NULL,
NULL
}
};
cmd_ln_parse(defn, argc, argv, TRUE);
isHelp = cmd_ln_int32("-help");
isExample = cmd_ln_int32("-example");
if(isHelp) {
printf("%s\n\n",helpstr);
}
if(isExample) {
printf("%s\n\n",examplestr);
}
if(isHelp || isExample) {
E_INFO("User asked for help or example.\n");
exit(0);
}
return 0;
}
int32 live_utt_decode_block (int16 *samples, int32 nsamples,
int32 live_endutt, partialhyp_t **ohyp)
{
static int32 live_begin_new_utt = 1;
static int32 frmno;
static float32 ***live_feat = NULL;
int32 live_nfr, live_nfeatvec;
int32 nwds =0;
float32 **mfcbuf;
/* int i,j;*/
/* 2004/08/27 L Galescu <*****@*****.**> -- added raw audio file saving */
static char uttfn[1024];
static FILE *rawfp = NULL;
int16 block_peak_amplitude;
if(live_feat==NULL)
live_feat = feat_array_alloc (kbcore_fcb(kbcore), LIVEBUFBLOCKSIZE);
if (live_begin_new_utt){
fe_start_utt(fe);
utt_begin (kb);
frmno = 0;
kb->nfr = 0;
kb->utt_hmm_eval = 0;
kb->utt_sen_eval = 0;
kb->utt_gau_eval = 0;
live_begin_new_utt = 0;
sprintf(uttfn, "%s/%s.raw", cmd_ln_str("-outrawdir"), kb->uttid);
rawfp = fopen(uttfn, "wb");
}
/* 10.jan.01 RAH, fe_process_utt now requires ***mfcbuf and it allocates the memory internally) */
mfcbuf = NULL;
/* LG 20080613 */
block_peak_amplitude = get_peak_amplitude(samples, nsamples);
if (block_peak_amplitude > peak_amplitude)
peak_amplitude = block_peak_amplitude;
E_INFO("segment peak %d\n",peak_amplitude);
live_nfr = fe_process_utt(fe, samples, nsamples, &mfcbuf); /**/
if (rawfp != NULL) {
fwrite(samples, sizeof(int16), nsamples, rawfp);
if (live_endutt)
fclose(rawfp);
}
if (live_endutt) {
/* RAH, It seems that we shouldn't throw out this data */
fe_end_utt(fe,dummyframe); /* Flush out the fe */
}
#if 0
E_INFO("Number frame after fe_process_utt %d\n",live_nfr);
for(i=0;i<live_nfr;i++){
printf("%d ",i);
for(j=0;j<13;j++){
printf("%f ",mfcbuf[i][j]);
fflush(stdout);
}
printf("\n");
fflush(stdout);
}
#endif
/* lgalescu 2004/08/22 -- i am under the impression that
* feat_s2mfc2feat_block() needs to be called at the end of utt
* even if no frames need processing
*/
/* lgalescu 2004/10/13 -- rescinded the above */
if(live_nfr>0){
/* Compute feature vectors */
live_nfeatvec = feat_s2mfc2feat_block(kbcore_fcb(kbcore), mfcbuf,
live_nfr, live_begin_new_utt,
live_endutt, live_feat);
#if 0
E_INFO ("live_nfeatvec: %ld\n",live_nfeatvec);
#endif
#if 0
E_INFO("Current frame number %d, Number of frames %d, Number frame after feat_s2mfcfeat_block %d\n",frmno,live_nfr,live_nfeatvec);
for(i=0;i<live_nfeatvec;i++){
printf("%d\n",i);
printf("Cep: ");
fflush(stdout);
for(j=0;j<13;j++){
printf("%f ",live_feat[i][0][j]);
fflush(stdout);
}
printf("\n");
fflush(stdout);
printf("Del: ");
fflush(stdout);
for(j=13;j<26;j++){
printf("%f ",live_feat[i][0][j]);
fflush(stdout);
}
printf("\n");
fflush(stdout);
printf("Acc: ");
fflush(stdout);
for(j=26;j<39;j++){
printf("%f ",live_feat[i][0][j]);
fflush(stdout);
}
printf("\n");
fflush(stdout);
}
#endif
/* decode the block */
utt_decode_block (live_feat, live_nfeatvec, &frmno, kb,
maxwpf, maxhistpf, maxhmmpf, ptranskip, hmmdumpfp);
/* lgalescu 2004/08/21
* moved the following block out of the previous if(){} because we need
* the output even when no feature computation has to be done.
*/
/* lgalescu 2004/10/13 -- rescinded */
/* Pull out partial hypothesis */
nwds = live_get_partialhyp(live_endutt);
*ohyp = parthyp;
parthyplen = nwds;
}
/* Clean up */
if (live_endutt) {
live_begin_new_utt = 1;
kb->tot_fr += kb->nfr;
utt_end(kb);
}
else {
live_begin_new_utt = 0;
}
/* I'm starting to think that fe_process_utt should not be allocating its
* memory, that or it should allocate some max and just keep on going,
* this idea of constantly allocating freeing memory seems dangerous to me.
*/
/* 20040318 ARCHAN : It sounds extremely dangerous to me and I will
* eliminate it sometime.
*/
/* lgalescu: i second that! the memory issue needs to be investigated: after a run on linux, i noticed some 1.6M of memory having "disappeared"! */
if(live_nfr>0){
ckd_free_2d((void **) mfcbuf); /* RAH, this must be freed since fe_process_utt allocates it */
}
return(parthyplen);
}
/*
* Continuous recognition from a file
*/
static void
recognize_from_file() {
cont_ad_t *cont;
ad_rec_t file_ad = {0};
int16 adbuf[4096];
const char* hyp;
const char* uttid;
int32 k, ts, start;
char waveheader[44];
if ((rawfd = fopen(cmd_ln_str_r(config, "-infile"), "rb")) == NULL) {
E_FATAL_SYSTEM("Failed to open file '%s' for reading",
cmd_ln_str_r(config, "-infile"));
}
fread(waveheader, 1, 44, rawfd);
file_ad.sps = (int32)cmd_ln_float32_r(config, "-samprate");
file_ad.bps = sizeof(int16);
if ((cont = cont_ad_init(&file_ad, ad_file_read)) == NULL) {
E_FATAL("Failed to initialize voice activity detection\n");
}
if (cont_ad_calib(cont) < 0)
E_INFO("Using default voice activity detection\n");
rewind (rawfd);
for (;;) {
while ((k = cont_ad_read(cont, adbuf, 4096)) == 0);
if (k < 0)
break;
if (ps_start_utt(ps, NULL) < 0)
E_FATAL("ps_start_utt() failed\n");
ps_process_raw(ps, adbuf, k, FALSE, FALSE);
hyp = ps_get_hyp(ps, NULL, &uttid);
printf("= partial_hypothese | %s: %s\n", uttid, hyp);
fflush(stdout);
ts = cont->read_ts;
start = ((ts - k) * 100.0) / file_ad.sps;
for (;;) {
if ((k = cont_ad_read(cont, adbuf, 4096)) < 0)
break;
if (k == 0) {
/*
* No speech data available; check current timestamp with most recent
* speech to see if more than 1 sec elapsed. If so, end of utterance.
*/
if ((cont->read_ts - ts) > 2000) {
printf("= utterance_end | %i\n", (cont->read_ts - ts) );
break;
}
} else {
/* New speech data received; note current timestamp */
ts = cont->read_ts;
}
ps_process_raw(ps, adbuf, k, FALSE, FALSE);
hyp = ps_get_hyp(ps, NULL, &uttid);
printf("= partial_hypothese | %s: %s\n", uttid, hyp);
fflush(stdout);
}
ps_end_utt(ps);
if (cmd_ln_boolean_r(config, "-time")) {
print_word_times(start);
printf("= continue\n");
fflush(stdout);
fgetc( stdin );
} else {
hyp = ps_get_hyp(ps, NULL, &uttid);
printf("= final_hypothese | %s: %s\n\n\n", uttid, hyp);
}
fflush(stdout);
}
cont_ad_close(cont);
fclose(rawfd);
}
int
ps_reinit(ps_decoder_t *ps, cmd_ln_t *config)
{
const char *path;
const char *keyphrase;
int32 lw;
if (config && config != ps->config) {
cmd_ln_free_r(ps->config);
ps->config = cmd_ln_retain(config);
}
err_set_debug_level(cmd_ln_int32_r(ps->config, "-debug"));
/* Set up logging. We need to do this earlier because we want to dump
* the information to the configured log, not to the stderr. */
if (config && cmd_ln_str_r(ps->config, "-logfn")) {
if (err_set_logfile(cmd_ln_str_r(ps->config, "-logfn")) < 0) {
E_ERROR("Cannot redirect log output\n");
return -1;
}
}
ps->mfclogdir = cmd_ln_str_r(ps->config, "-mfclogdir");
ps->rawlogdir = cmd_ln_str_r(ps->config, "-rawlogdir");
ps->senlogdir = cmd_ln_str_r(ps->config, "-senlogdir");
/* Fill in some default arguments. */
ps_expand_model_config(ps);
/* Free old searches (do this before other reinit) */
ps_free_searches(ps);
ps->searches = hash_table_new(3, HASH_CASE_YES);
/* Free old acmod. */
acmod_free(ps->acmod);
ps->acmod = NULL;
/* Free old dictionary (must be done after the two things above) */
dict_free(ps->dict);
ps->dict = NULL;
/* Free d2p */
dict2pid_free(ps->d2p);
ps->d2p = NULL;
/* Logmath computation (used in acmod and search) */
if (ps->lmath == NULL
|| (logmath_get_base(ps->lmath) !=
(float64)cmd_ln_float32_r(ps->config, "-logbase"))) {
if (ps->lmath)
logmath_free(ps->lmath);
ps->lmath = logmath_init
((float64)cmd_ln_float32_r(ps->config, "-logbase"), 0,
cmd_ln_boolean_r(ps->config, "-bestpath"));
}
/* Acoustic model (this is basically everything that
* uttproc.c, senscr.c, and others used to do) */
if ((ps->acmod = acmod_init(ps->config, ps->lmath, NULL, NULL)) == NULL)
return -1;
if (cmd_ln_int32_r(ps->config, "-pl_window") > 0) {
/* Initialize an auxiliary phone loop search, which will run in
* "parallel" with FSG or N-Gram search. */
if ((ps->phone_loop =
phone_loop_search_init(ps->config, ps->acmod, ps->dict)) == NULL)
return -1;
hash_table_enter(ps->searches,
ps_search_name(ps->phone_loop),
ps->phone_loop);
}
/* Dictionary and triphone mappings (depends on acmod). */
/* FIXME: pass config, change arguments, implement LTS, etc. */
if ((ps->dict = dict_init(ps->config, ps->acmod->mdef)) == NULL)
return -1;
if ((ps->d2p = dict2pid_build(ps->acmod->mdef, ps->dict)) == NULL)
return -1;
lw = cmd_ln_float32_r(ps->config, "-lw");
/* Determine whether we are starting out in FSG or N-Gram search mode.
* If neither is used skip search initialization. */
/* Load KWS if one was specified in config */
if ((keyphrase = cmd_ln_str_r(ps->config, "-keyphrase"))) {
if (ps_set_keyphrase(ps, PS_DEFAULT_SEARCH, keyphrase))
return -1;
ps_set_search(ps, PS_DEFAULT_SEARCH);
}
if ((path = cmd_ln_str_r(ps->config, "-kws"))) {
if (ps_set_kws(ps, PS_DEFAULT_SEARCH, path))
return -1;
ps_set_search(ps, PS_DEFAULT_SEARCH);
}
/* Load an FSG if one was specified in config */
if ((path = cmd_ln_str_r(ps->config, "-fsg"))) {
fsg_model_t *fsg = fsg_model_readfile(path, ps->lmath, lw);
if (!fsg)
return -1;
if (ps_set_fsg(ps, PS_DEFAULT_SEARCH, fsg)) {
fsg_model_free(fsg);
return -1;
}
fsg_model_free(fsg);
ps_set_search(ps, PS_DEFAULT_SEARCH);
}
/* Or load a JSGF grammar */
if ((path = cmd_ln_str_r(ps->config, "-jsgf"))) {
if (ps_set_jsgf_file(ps, PS_DEFAULT_SEARCH, path)
|| ps_set_search(ps, PS_DEFAULT_SEARCH))
return -1;
}
if ((path = cmd_ln_str_r(ps->config, "-allphone"))) {
if (ps_set_allphone_file(ps, PS_DEFAULT_SEARCH, path)
|| ps_set_search(ps, PS_DEFAULT_SEARCH))
return -1;
}
if ((path = cmd_ln_str_r(ps->config, "-lm")) &&
!cmd_ln_boolean_r(ps->config, "-allphone")) {
if (ps_set_lm_file(ps, PS_DEFAULT_SEARCH, path)
|| ps_set_search(ps, PS_DEFAULT_SEARCH))
return -1;
}
if ((path = cmd_ln_str_r(ps->config, "-lmctl"))) {
const char *name;
ngram_model_t *lmset;
ngram_model_set_iter_t *lmset_it;
if (!(lmset = ngram_model_set_read(ps->config, path, ps->lmath))) {
E_ERROR("Failed to read language model control file: %s\n", path);
return -1;
}
for(lmset_it = ngram_model_set_iter(lmset);
lmset_it; lmset_it = ngram_model_set_iter_next(lmset_it)) {
ngram_model_t *lm = ngram_model_set_iter_model(lmset_it, &name);
E_INFO("adding search %s\n", name);
if (ps_set_lm(ps, name, lm)) {
ngram_model_set_iter_free(lmset_it);
ngram_model_free(lmset);
return -1;
}
}
ngram_model_free(lmset);
name = cmd_ln_str_r(ps->config, "-lmname");
if (name)
ps_set_search(ps, name);
else {
E_ERROR("No default LM name (-lmname) for `-lmctl'\n");
return -1;
}
}
/* Initialize performance timer. */
ps->perf.name = "decode";
ptmr_init(&ps->perf);
return 0;
}
static int batch_decoder_decode_adc(batch_decoder_t *bd, FILE *infh, int sf,
int ef, alignment_t *al)
{
featbuf_t *fb = search_factory_featbuf(bd->sf);
float32 samprate = cmd_ln_float32_r(bd->config, "-samprate");
int32 frate = cmd_ln_int32_r(bd->config, "-frate");
int16 buf[512];
if (ef != -1) {
ef = (int32) (((ef - sf) * samprate / frate) + (samprate
* cmd_ln_float32_r(bd->config, "-wlen")));
}
sf = (int32) (sf * (samprate / frate));
fseek(infh, cmd_ln_int32_r(bd->config, "-adchdr") + sf * sizeof(int16),
SEEK_SET);
if (al) {
alignment_iter_t *itor;
double starttime = 0.0;
for (itor = alignment_words(al); itor; itor = alignment_iter_next(itor)) {
alignment_entry_t *ent = alignment_iter_get(itor);
double nsec = (double) ent->duration / frate;
double endtime = starttime + nsec;
size_t nsamp = (size_t) (nsec * samprate);
E_INFO("Processing %d samples for %s (%f seconds ending %f)\n",
nsamp, dict_wordstr(search_factory_d2p(bd->sf)->dict, ent->id.wid),
nsec, endtime);
E_INFO("Woke up at delta %f\n", get_time_delta(bd));
while (nsamp > 0) {
size_t nread = 512;
if (nread > nsamp) nread = nsamp;
nread = fread(buf, sizeof(int16), nread, infh);
if (nread == 0)
break;
featbuf_producer_process_raw(fb, buf, nread, FALSE);
nsamp -= nread;
starttime += (nread / samprate);
double delta = get_time_delta(bd);
if (starttime > delta) {
E_INFO("Sleeping until next start time (%f seconds)\n",
starttime - delta);
usleep((int)((starttime - delta) * 1000000));
}
}
double delta = get_time_delta(bd);
if (endtime > delta) {
E_INFO("Sleeping until end time (%f seconds)\n",
endtime - delta);
usleep((int)((endtime - delta) * 1000000));
}
}
}
else {
while (ef == -1 || sf < ef) {
size_t nread = 512;
if (ef != -1 && nread > ef - sf)
nread = ef - sf;
nread = fread(buf, sizeof(int16), nread, infh);
if (nread == 0)
break;
featbuf_producer_process_raw(fb, buf, nread, FALSE);
//usleep((int)((double)nread / 16000 * 1000000));
sf += nread;
}
}
return 0;
}
int
ps_start_utt(ps_decoder_t *ps)
{
int rv;
char uttid[16];
if (ps->acmod->state == ACMOD_STARTED || ps->acmod->state == ACMOD_PROCESSING) {
E_ERROR("Utterance already started\n");
return -1;
}
if (ps->search == NULL) {
E_ERROR("No search module is selected, did you forget to "
"specify a language model or grammar?\n");
return -1;
}
ptmr_reset(&ps->perf);
ptmr_start(&ps->perf);
sprintf(uttid, "%09u", ps->uttno);
++ps->uttno;
/* Remove any residual word lattice and hypothesis. */
ps_lattice_free(ps->search->dag);
ps->search->dag = NULL;
ps->search->last_link = NULL;
ps->search->post = 0;
ckd_free(ps->search->hyp_str);
ps->search->hyp_str = NULL;
if ((rv = acmod_start_utt(ps->acmod)) < 0)
return rv;
/* Start logging features and audio if requested. */
if (ps->mfclogdir) {
char *logfn = string_join(ps->mfclogdir, "/",
uttid, ".mfc", NULL);
FILE *mfcfh;
E_INFO("Writing MFCC log file: %s\n", logfn);
if ((mfcfh = fopen(logfn, "wb")) == NULL) {
E_ERROR_SYSTEM("Failed to open MFCC log file %s", logfn);
ckd_free(logfn);
return -1;
}
ckd_free(logfn);
acmod_set_mfcfh(ps->acmod, mfcfh);
}
if (ps->rawlogdir) {
char *logfn = string_join(ps->rawlogdir, "/",
uttid, ".raw", NULL);
FILE *rawfh;
E_INFO("Writing raw audio log file: %s\n", logfn);
if ((rawfh = fopen(logfn, "wb")) == NULL) {
E_ERROR_SYSTEM("Failed to open raw audio log file %s", logfn);
ckd_free(logfn);
return -1;
}
ckd_free(logfn);
acmod_set_rawfh(ps->acmod, rawfh);
}
if (ps->senlogdir) {
char *logfn = string_join(ps->senlogdir, "/",
uttid, ".sen", NULL);
FILE *senfh;
E_INFO("Writing senone score log file: %s\n", logfn);
if ((senfh = fopen(logfn, "wb")) == NULL) {
E_ERROR_SYSTEM("Failed to open senone score log file %s", logfn);
ckd_free(logfn);
return -1;
}
ckd_free(logfn);
acmod_set_senfh(ps->acmod, senfh);
}
/* Start auxiliary phone loop search. */
if (ps->phone_loop)
ps_search_start(ps->phone_loop);
return ps_search_start(ps->search);
}
/**
* 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);
}
static int
cp_parm()
{
FILE *fp;
uint32 i, o;
uint32 max=0;
/* Open the file first to see whether command-line parameters
match
*/
if(cmd_ln_str("-cpopsfn")==NULL) {
E_INFO("Please specify -cpopsfn\n");
return S3_ERROR;
}
fp = fopen(cmd_ln_str("-cpopsfn"), "r");
if (fp == NULL) {
E_INFO("Unable to open cpops file\n");
return S3_ERROR;
}
while (fscanf(fp, "%u %u", &o, &i) == 2) {
if(o+1>max) {
max=o+1;
}
}
if (omixw) {
if(max != n_mixw_o) {
E_INFO("Mismatch between cp operation file (max out %d) and -nmixout (%d)\n",max, n_mixw_o);
return S3_ERROR;
}
}
if (ogau) {
if(max != n_cb_o) {
E_INFO("Mismatch between cp operation file (max out %d) and -ncbout (%d)\n",max, n_cb_o);
return S3_ERROR;
}
}
if (ogau_full) {
if(max != n_cb_o) {
E_INFO("Mismatch between cp operation file (max out %d) and -ncbout (%d)\n",max, n_cb_o);
return S3_ERROR;
}
}
if (otmat) {
if(max != n_tmat_o) {
E_INFO("Mismatch between cp operation file (max out %d) and -ntmatout (%d)\n",max, n_tmat_o);
return S3_ERROR;
}
}
fclose(fp);
fp = fopen(cmd_ln_str("-cpopsfn"), "r");
while (fscanf(fp, "%u %u", &o, &i) == 2) {
if (omixw) {
cp_mixw(o, i);
}
if (ogau) {
cp_gau(o, i);
}
if (ogau_full) {
cp_gau_full(o, i);
}
if (otmat) {
cp_tmat(o, i);
}
}
fclose(fp);
return S3_SUCCESS;
}
void
utt_decode(void *data, utt_res_t * ur, int32 sf, int32 ef, char *uttid)
{
kb_t *kb;
kbcore_t *kbcore;
cmd_ln_t *config;
int32 num_decode_frame;
int32 total_frame;
stat_t *st;
srch_t *s;
num_decode_frame = 0;
E_INFO("Processing: %s\n", uttid);
kb = (kb_t *) data;
kbcore = kb->kbcore;
config = kbcore_config(kbcore);
kb_set_uttid(uttid, ur->uttfile, kb);
st = kb->stat;
/* Convert input file to cepstra if waveform input is selected */
if (cmd_ln_boolean_r(config, "-adcin")) {
int16 *adcdata;
int32 nsamps = 0;
if ((adcdata = bio_read_wavfile(cmd_ln_str_r(config, "-cepdir"),
ur->uttfile,
cmd_ln_str_r(config, "-cepext"),
cmd_ln_int32_r(config, "-adchdr"),
strcmp(cmd_ln_str_r(config, "-input_endian"), "big"),
&nsamps)) == NULL) {
E_FATAL("Cannot read file %s\n", ur->uttfile);
}
if (kb->mfcc) {
ckd_free_2d((void **)kb->mfcc);
}
fe_start_utt(kb->fe);
if (fe_process_utt(kb->fe, adcdata, nsamps, &kb->mfcc, &total_frame) < 0) {
E_FATAL("MFCC calculation failed\n", ur->uttfile);
}
ckd_free(adcdata);
if (total_frame > S3_MAX_FRAMES) {
E_FATAL("Maximum number of frames (%d) exceeded\n", S3_MAX_FRAMES);
}
if ((total_frame = feat_s2mfc2feat_live(kbcore_fcb(kbcore),
kb->mfcc,
&total_frame,
TRUE, TRUE,
kb->feat)) < 0) {
E_FATAL("Feature computation failed\n");
}
}
else {
/* Read mfc file and build feature vectors for entire utterance */
if ((total_frame = feat_s2mfc2feat(kbcore_fcb(kbcore), ur->uttfile,
cmd_ln_str_r(config, "-cepdir"),
cmd_ln_str_r(config, "-cepext"), sf, ef,
kb->feat, S3_MAX_FRAMES)) < 0) {
E_FATAL("Cannot read file %s. Forced exit\n", ur->uttfile);
}
}
/* Also need to make sure we don't set resource if it is the same. Well, this mechanism
could be provided inside the following function.
*/
s = kb->srch;
if (ur->lmname != NULL)
srch_set_lm(s, ur->lmname);
if (ur->regmatname != NULL)
kb_setmllr(ur->regmatname, ur->cb2mllrname, kb);
/* These are necessary! */
s->uttid = kb->uttid;
s->uttfile = kb->uttfile;
utt_begin(kb);
utt_decode_block(kb->feat, total_frame, &num_decode_frame, kb);
utt_end(kb);
st->tot_fr += st->nfr;
}
int
read_seno_dtree_file(dtree_t **out_dt,
const char *file_name)
{
uint32 n_base;
uint32 n_cd;
char tree_id[64];
float32 ent;
uint32 key, l_key, r_key, n_key = 0, *k2q = NULL;
char q_str[1024], *rem_q_str;
FILE *fp;
int i, j, n_quest;
dtree_t *dt;
bt_node_t *node;
bt_t *tree;
comp_quest_t *q;
float32 *q2ent;
*out_dt = dt = (dtree_t *)ckd_calloc(1, sizeof(dtree_t));
dt->tree = tree = bt_new();
fp = fopen(file_name, "r");
if (fp == NULL) {
E_WARN_SYSTEM("Unable to open %s for reading", file_name);
return S3_ERROR;
}
for (n_quest = 0; fgets(q_str, 1024, fp) != NULL; n_quest++);
--n_quest; /* account for header line */
dt->n_quest = n_quest;
rewind(fp);
read_header_line(&n_base, tree_id, &n_cd, fp);
for (i = 0; read_node_line(&ent, &key, &l_key, &r_key, q_str, fp) > 0; i++) {
if (n_key < l_key)
n_key = l_key;
if (n_key < r_key)
n_key = r_key;
}
++n_key;
dt->n_key = n_key;
rewind(fp);
read_header_line(&n_base, tree_id, &n_cd, fp);
E_INFO("Reading tree %s (%u base phones, %u CD phones, %u quest)\n",
tree_id, n_base, n_cd, n_quest);
dt->quest = q = (comp_quest_t *)ckd_calloc(n_quest,
sizeof(comp_quest_t));
dt->k2q = k2q = (uint32 *)ckd_calloc(n_key, sizeof(uint32));
dt->q2ent = q2ent = (float32 *)ckd_calloc(n_quest, sizeof(float32));
for (i = 0; i < n_key; i++) {
k2q[i] = NO_MAP;
}
for (i = 0; read_node_line(&ent, &key, &l_key, &r_key, q_str, fp) > 0; i++) {
if (tree->root) {
E_INFO("%u\n", key);
node = bt_find_node(tree, key);
if (node) {
/* grow left and right children */
bt_add_left(node, l_key);
bt_add_right(node, r_key);
}
else {
E_FATAL("Find node w/ key %u failed\n", key);
}
}
else {
E_INFO("root %u\n", key);
tree->root = bt_new_node(key);
bt_add_left(tree->root, l_key);
bt_add_right(tree->root, r_key);
}
k2q[key] = i;
q2ent[i] = ent;
parse_compound_q(&q[i], q_str);
}
assert(i == n_quest);
return S3_SUCCESS;
}
static void
ps_init_defaults(ps_decoder_t *ps)
{
char const *hmmdir, *lmfile, *dictfile;
/* Disable memory mapping on Blackfin (FIXME: should be uClinux in general). */
#ifdef __ADSPBLACKFIN__
E_INFO("Will not use mmap() on uClinux/Blackfin.");
cmd_ln_set_boolean_r(ps->config, "-mmap", FALSE);
#endif
#ifdef MODELDIR
/* Set default acoustic and language models. */
hmmdir = cmd_ln_str_r(ps->config, "-hmm");
lmfile = cmd_ln_str_r(ps->config, "-lm");
dictfile = cmd_ln_str_r(ps->config, "-dict");
if (hmmdir == NULL && hmmdir_exists(MODELDIR "/hmm/en_US/hub4wsj_sc_8k")) {
hmmdir = MODELDIR "/hmm/en_US/hub4wsj_sc_8k";
cmd_ln_set_str_r(ps->config, "-hmm", hmmdir);
}
if (lmfile == NULL && !cmd_ln_str_r(ps->config, "-fsg")
&& !cmd_ln_str_r(ps->config, "-jsgf")
&& file_exists(MODELDIR "/lm/en_US/hub4.5000.DMP")) {
lmfile = MODELDIR "/lm/en_US/hub4.5000.DMP";
cmd_ln_set_str_r(ps->config, "-lm", lmfile);
}
if (dictfile == NULL && file_exists(MODELDIR "/lm/en_US/cmu07a.dic")) {
dictfile = MODELDIR "/lm/en_US/cmu07a.dic";
cmd_ln_set_str_r(ps->config, "-dict", dictfile);
}
/* Expand acoustic and language model filenames relative to installation path. */
if (hmmdir && !path_is_absolute(hmmdir) && !hmmdir_exists(hmmdir)) {
char *tmphmm = string_join(MODELDIR "/hmm/", hmmdir, NULL);
cmd_ln_set_str_r(ps->config, "-hmm", tmphmm);
ckd_free(tmphmm);
}
if (lmfile && !path_is_absolute(lmfile) && !file_exists(lmfile)) {
char *tmplm = string_join(MODELDIR "/lm/", lmfile, NULL);
cmd_ln_set_str_r(ps->config, "-lm", tmplm);
ckd_free(tmplm);
}
if (dictfile && !path_is_absolute(dictfile) && !file_exists(dictfile)) {
char *tmpdict = string_join(MODELDIR "/lm/", dictfile, NULL);
cmd_ln_set_str_r(ps->config, "-dict", tmpdict);
ckd_free(tmpdict);
}
#endif
/* Get acoustic model filenames and add them to the command-line */
if ((hmmdir = cmd_ln_str_r(ps->config, "-hmm")) != NULL) {
ps_add_file(ps, "-mdef", hmmdir, "mdef");
ps_add_file(ps, "-mean", hmmdir, "means");
ps_add_file(ps, "-var", hmmdir, "variances");
ps_add_file(ps, "-tmat", hmmdir, "transition_matrices");
ps_add_file(ps, "-mixw", hmmdir, "mixture_weights");
ps_add_file(ps, "-sendump", hmmdir, "sendump");
ps_add_file(ps, "-fdict", hmmdir, "noisedict");
ps_add_file(ps, "-lda", hmmdir, "feature_transform");
ps_add_file(ps, "-featparams", hmmdir, "feat.params");
ps_add_file(ps, "-senmgau", hmmdir, "senmgau");
}
}
/**
* Build net from phone HMMs
*/
static int
phmm_build(allphone_search_t * allphs)
{
phmm_t *p, **pid2phmm;
bin_mdef_t *mdef;
int32 lrc_size;
uint32 *lc, *rc;
s3pid_t pid;
s3cipid_t ci;
s3cipid_t *filler;
int n_phmm, n_link;
int i, nphone;
mdef = ((ps_search_t *) allphs)->acmod->mdef;
allphs->ci_phmm =
(phmm_t **) ckd_calloc(bin_mdef_n_ciphone(mdef), sizeof(phmm_t *));
pid2phmm =
(phmm_t **) ckd_calloc(bin_mdef_n_phone(mdef), sizeof(phmm_t *));
/* For each unique ciphone/triphone entry in mdef, create a PHMM node */
n_phmm = 0;
nphone = allphs->ci_only ? bin_mdef_n_ciphone(mdef) : bin_mdef_n_phone(mdef);
E_INFO("Building PHMM net of %d phones\n", nphone);
for (pid = 0; pid < nphone; pid++) {
if ((p = phmm_lookup(allphs, pid)) == NULL) {
//not found, should be created
p = (phmm_t *) ckd_calloc(1, sizeof(*p));
hmm_init(allphs->hmmctx, &(p->hmm), FALSE,
mdef_pid2ssid(mdef, pid), mdef->phone[pid].tmat);
p->pid = pid;
p->ci = bin_mdef_pid2ci(mdef, pid);
p->succlist = NULL;
p->next = allphs->ci_phmm[p->ci];
allphs->ci_phmm[p->ci] = p;
n_phmm++;
}
pid2phmm[pid] = p;
}
/* Fill out bitvecs of each PHMM node, alloc continuous memory chunk for context bitvectors */
lrc_size = bitvec_size(bin_mdef_n_ciphone(mdef));
lc = ckd_calloc(n_phmm * 2 * lrc_size, sizeof(bitvec_t));
rc = lc + (n_phmm * lrc_size);
for (ci = 0; ci < mdef->n_ciphone; ci++) {
for (p = allphs->ci_phmm[ci]; p; p = p->next) {
p->lc = lc;
lc += lrc_size;
p->rc = rc;
rc += lrc_size;
}
}
/* Fill out lc and rc bitmaps (remember to map all fillers to each other!!) */
filler =
(s3cipid_t *) ckd_calloc(bin_mdef_n_ciphone(mdef) + 1,
sizeof(s3cipid_t));
/* Connect fillers */
i = 0;
for (ci = 0; ci < bin_mdef_n_ciphone(mdef); ci++) {
p = pid2phmm[ci];
bitvec_set_all(p->lc, bin_mdef_n_ciphone(mdef));
bitvec_set_all(p->rc, bin_mdef_n_ciphone(mdef));
if (mdef->phone[ci].info.ci.filler) {
filler[i++] = ci;
}
}
filler[i] = BAD_S3CIPID;
/* Loop over cdphones only if ci_only is not set */
for (pid = bin_mdef_n_ciphone(mdef); pid < nphone;
pid++) {
p = pid2phmm[pid];
if (mdef->phone[mdef->phone[pid].info.cd.ctx[1]].info.ci.filler) {
for (i = 0; IS_S3CIPID(filler[i]); i++)
bitvec_set(p->lc, filler[i]);
}
else
bitvec_set(p->lc, mdef->phone[pid].info.cd.ctx[1]);
if (mdef->phone[mdef->phone[pid].info.cd.ctx[2]].info.ci.filler) {
for (i = 0; IS_S3CIPID(filler[i]); i++)
bitvec_set(p->rc, filler[i]);
}
else
bitvec_set(p->rc, mdef->phone[pid].info.cd.ctx[2]);
}
ckd_free(pid2phmm);
ckd_free(filler);
/* Create links between PHMM nodes */
n_link = phmm_link(allphs);
E_INFO("%d nodes, %d links\n", n_phmm, n_link);
return 0;
}
static int
acmod_init_am(acmod_t *acmod)
{
char const *mdeffn, *tmatfn, *mllrfn, *hmmdir;
/* Read model definition. */
if ((mdeffn = cmd_ln_str_r(acmod->config, "-mdef")) == NULL) {
if ((hmmdir = cmd_ln_str_r(acmod->config, "-hmm")) == NULL)
E_ERROR("Acoustic model definition is not specified either "
"with -mdef option or with -hmm\n");
else
E_ERROR("Folder '%s' does not contain acoustic model "
"definition 'mdef'\n", hmmdir);
return -1;
}
if ((acmod->mdef = bin_mdef_read(acmod->config, mdeffn)) == NULL) {
E_ERROR("Failed to read acoustic model definition from %s\n", mdeffn);
return -1;
}
/* Read transition matrices. */
if ((tmatfn = cmd_ln_str_r(acmod->config, "-tmat")) == NULL) {
E_ERROR("No tmat file specified\n");
return -1;
}
acmod->tmat = tmat_init(tmatfn, acmod->lmath,
cmd_ln_float32_r(acmod->config, "-tmatfloor"),
TRUE);
/* Read the acoustic models. */
if ((cmd_ln_str_r(acmod->config, "-mean") == NULL)
|| (cmd_ln_str_r(acmod->config, "-var") == NULL)
|| (cmd_ln_str_r(acmod->config, "-tmat") == NULL)) {
E_ERROR("No mean/var/tmat files specified\n");
return -1;
}
if (cmd_ln_str_r(acmod->config, "-senmgau")) {
E_INFO("Using general multi-stream GMM computation\n");
acmod->mgau = ms_mgau_init(acmod, acmod->lmath, acmod->mdef);
if (acmod->mgau == NULL)
return -1;
}
else {
E_INFO("Attempting to use SCHMM computation module\n");
if ((acmod->mgau = s2_semi_mgau_init(acmod)) == NULL) {
E_INFO("Attempting to use PTHMM computation module\n");
if ((acmod->mgau = ptm_mgau_init(acmod, acmod->mdef)) == NULL) {
E_INFO("Falling back to general multi-stream GMM computation\n");
acmod->mgau = ms_mgau_init(acmod, acmod->lmath, acmod->mdef);
if (acmod->mgau == NULL)
return -1;
}
}
}
/* If there is an MLLR transform, apply it. */
if ((mllrfn = cmd_ln_str_r(acmod->config, "-mllr"))) {
ps_mllr_t *mllr = ps_mllr_read(mllrfn);
if (mllr == NULL)
return -1;
acmod_update_mllr(acmod, mllr);
}
return 0;
}
dict_t *dict_init (mdef_t *mdef, char *dictfile, char *fillerfile, char comp_sep)
{
FILE *fp, *fp2;
int32 n ;
char line[1024];
dict_t *d;
if (! dictfile)
E_FATAL("No dictionary file\n");
/*
* First obtain #words in dictionary (for hash table allocation).
* Reason: The PC NT system doesn't like to grow memory gradually. Better to allocate
* all the required memory in one go.
*/
if ((fp = fopen(dictfile, "r")) == NULL)
E_FATAL_SYSTEM("fopen(%s,r) failed\n", dictfile);
n = 0;
while (fgets (line, sizeof(line), fp) != NULL) {
if (line[0] != '#')
n++;
}
rewind (fp);
if (fillerfile) {
if ((fp2 = fopen(fillerfile, "r")) == NULL)
E_FATAL_SYSTEM("fopen(%s,r) failed\n", fillerfile);
while (fgets (line, sizeof(line), fp2) != NULL) {
if (line[0] != '#')
n++;
}
rewind (fp2);
}
/*
* Allocate dict entries. HACK!! Allow some extra entries for words not in file.
* Also check for type size restrictions.
*/
d = (dict_t *) ckd_calloc (1, sizeof(dict_t));
d->max_words = (n+1024 < MAX_WID) ? n+1024 : MAX_WID;
if (n >= MAX_WID)
E_FATAL("#Words in dictionaries (%d) exceeds limit (%d)\n", n, MAX_WID);
d->word = (dictword_t *) ckd_calloc (d->max_words, sizeof(dictword_t));
d->n_word = 0;
d->mdef = mdef;
if (mdef) {
d->pht = NULL;
d->ciphone_str = NULL;
} else {
d->pht = hash_new (DEFAULT_NUM_PHONE, 1 /* No case */);
d->ciphone_str = (char **) ckd_calloc (DEFAULT_NUM_PHONE, sizeof(char *));
}
d->n_ciphone = 0;
/* Create new hash table for word strings; case-insensitive word strings */
d->ht = hash_new (d->max_words, 1 /* no-case */);
/* Initialize with no compound words */
d->comp_head = NULL;
/* Digest main dictionary file */
E_INFO("Reading main dictionary: %s\n", dictfile);
dict_read (fp, d);
fclose (fp);
E_INFO("%d words read\n", d->n_word);
/* Now the filler dictionary file, if it exists */
d->filler_start = d->n_word;
if (fillerfile) {
E_INFO("Reading filler dictionary: %s\n", fillerfile);
dict_read (fp2, d);
fclose (fp2);
E_INFO("%d words read\n", d->n_word - d->filler_start);
}
d->filler_end = d->n_word-1;
/* Initialize distinguished word-ids */
d->startwid = dict_wordid (d, START_WORD);
d->finishwid = dict_wordid (d, FINISH_WORD);
d->silwid = dict_wordid (d, SILENCE_WORD);
if (NOT_WID(d->startwid))
E_WARN("%s not in dictionary\n", START_WORD);
if (NOT_WID(d->finishwid))
E_WARN("%s not in dictionary\n", FINISH_WORD);
if (NOT_WID(d->silwid))
E_WARN("%s not in dictionary\n", SILENCE_WORD);
/* Identify compound words if indicated */
if (comp_sep) {
E_INFO("Building compound words (separator = '%c')\n", comp_sep);
n = dict_build_comp (d, comp_sep);
E_INFO("%d compound words\n", n);
}
return d;
}
dtree_t *
mk_tree(float32 ****mixw,
/* ADDITION FOR CONTINUOUS_TREES, 20 May 98 */
float32 ****means,
float32 ****vars,
uint32 *veclen,
/* END ADDITION FOR CONTINUOUS_TREES */
uint32 n_model,
uint32 n_state,
uint32 n_stream,
uint32 n_density,
float32 *stwt,
uint32 *id,
uint32 n_id,
quest_t *all_q,
uint32 n_all_q,
pset_t *pset,
uint32 **dfeat,
uint32 n_dfeat,
uint32 split_min,
uint32 split_max,
float32 split_thr,
float32 mwfloor)
{
dtree_t *s_tree;
uint32 i;
dtree_node_t *b_n, *root;
s_tree = ckd_calloc(1, sizeof(dtree_t));
s_tree->node = ckd_calloc(2*split_max + 1, sizeof(dtree_node_t));
s_tree->n_node = 0;
s_tree->node[0].node_id = 0;
s_tree->n_node = 1;
root = &s_tree->node[0];
mk_node(root, 0,
id, n_id,
mixw,
/* ADDITION FOR CONTINUOUS_TREES; 20 May 98; passing means, vars, veclen */
means,
vars,
veclen,
/* END ADDITION FOR CONTINUOUS_TREES */
n_model,
n_state,
n_stream,
n_density,
stwt,
mwfloor);
set_best_quest(root,
mixw,
/* ADDITION FOR CONTINUOUS_TREES; 20 May 98; passing means, vars and veclen */
means,
vars,
veclen,
/* END ADDITION FOR CONTINUOUS_TREES */
n_model,
n_state,
n_stream,
n_density,
stwt,
all_q,
n_all_q,
pset,
dfeat, n_dfeat,
mwfloor);
if (root->q == NULL) {
/* No question found that is able to split node;
can't go any further */
free_tree(s_tree);
return NULL;
}
for (i = 0; i < split_max; i++) {
b_n = best_leaf_node(root);
if (b_n == NULL) {
E_INFO("stop. leaf nodes are specific\n");
break;
}
/* DDDDDBUG The following criteria will fail if we use only likelihood and no
likelihood increase */
if (b_n->wt_ent_dec <= 0) {
E_INFO("stop. b_n->wt_ent_dec (%.3e) <= 0\n",
b_n->wt_ent_dec);
break;
}
if ((i > split_min) &&
(b_n->wt_ent_dec < split_thr * b_n->wt_ent)) {
E_INFO("stop. b_n->wt_ent_dec (%.3e) < split_thr * b_n->wt_ent (%.3e)\n",
b_n->wt_ent_dec, b_n->wt_ent * split_thr);
break;
}
split_node(s_tree, b_n->node_id,
mixw,
/* ADDITION FOR CONTINUOUS_TREES; 20 May 98; passing means, vars and veclen */
means,
vars,
veclen,
/* END ADDITION FOR CONTINUOUS_TREES */
n_model,
n_state,
n_stream,
n_density,
stwt,
all_q, n_all_q, pset,
dfeat, n_dfeat,
mwfloor);
}
#if 1
E_INFO("Final simple tree\n");
print_tree(stderr, "|", root, pset, 0);
fprintf(stderr, "\n");
#endif
return s_tree;
}
s3wid_t
dict_add_word(dict_t * d, char *word, s3cipid_t * p, int32 np)
{
int32 len;
dictword_t *wordp;
s3wid_t newwid;
if (d->n_word >= d->max_words) {
E_INFO
("Dictionary max size (%d) exceeded; reallocate another entries %d \n",
d->max_words, DICT_INC_SZ);
d->word =
(dictword_t *) ckd_realloc(d->word,
(d->max_words +
DICT_INC_SZ) * sizeof(dictword_t));
d->max_words = d->max_words + DICT_INC_SZ;
return (BAD_S3WID);
}
wordp = d->word + d->n_word;
wordp->word = (char *) ckd_salloc(word); /* Freed in dict_free */
/* Associate word string with d->n_word in hash table */
if (hash_table_enter(d->ht, wordp->word, (void *)(long)d->n_word) != (void *)(long)d->n_word) {
ckd_free(wordp->word);
return (BAD_S3WID);
}
/* Fill in word entry, and set defaults */
if (p && (np > 0)) {
wordp->ciphone = (s3cipid_t *) ckd_malloc(np * sizeof(s3cipid_t)); /* Freed in dict_free */
memcpy(wordp->ciphone, p, np * sizeof(s3cipid_t));
wordp->pronlen = np;
}
else {
wordp->ciphone = NULL;
wordp->pronlen = 0;
}
wordp->alt = BAD_S3WID;
wordp->basewid = d->n_word;
wordp->n_comp = 0;
wordp->comp = NULL;
/* Determine base/alt wids */
if ((len = dict_word2basestr(word)) > 0) {
void *val;
s3wid_t w;
/* Truncated to a baseword string; find its ID */
if (hash_table_lookup(d->ht, word, &val) < 0) {
word[len] = '('; /* Get back the original word */
E_FATAL("Missing base word for: %s\n", word);
}
else
word[len] = '('; /* Get back the original word */
/* Link into alt list */
w = (s3wid_t)(long)val;
wordp->basewid = w;
wordp->alt = d->word[w].alt;
d->word[w].alt = d->n_word;
}
newwid = d->n_word++;
return (newwid);
}
int
main(int argc, char *argv[])
{
acmod_t *acmod;
logmath_t *lmath;
cmd_ln_t *config;
FILE *rawfh;
int16 *buf;
int16 const *bptr;
mfcc_t **cepbuf, **cptr;
size_t nread, nsamps;
int nfr;
int frame_counter;
int bestsen1[270];
lmath = logmath_init(1.0001, 0, 0);
config = cmd_ln_init(NULL, ps_args(), TRUE,
"-mdef", MODELDIR "/en-us/en-us/mdef",
"-mean", MODELDIR "/en-us/en-us/means",
"-var", MODELDIR "/en-us/en-us/variances",
"-tmat", MODELDIR "/en-us/en-us/transition_matrices",
"-sendump", MODELDIR "/en-us/en-us/sendump",
"-compallsen", "true",
"-cmn", "prior",
"-tmatfloor", "0.0001",
"-mixwfloor", "0.001",
"-varfloor", "0.0001",
"-mmap", "no",
"-topn", "4",
"-ds", "1",
"-input_endian", "little",
"-samprate", "16000", NULL);
TEST_ASSERT(config);
cmd_ln_parse_file_r(config, ps_args(), MODELDIR "/en-us/en-us/feat.params", FALSE);
TEST_ASSERT(acmod = acmod_init(config, lmath, NULL, NULL));
cmn_prior_set(acmod->fcb->cmn_struct, prior);
nsamps = 2048;
frame_counter = 0;
buf = ckd_calloc(nsamps, sizeof(*buf));
TEST_ASSERT(rawfh = fopen(DATADIR "/goforward.raw", "rb"));
TEST_EQUAL(0, acmod_start_utt(acmod));
E_INFO("Incremental(2048):\n");
while (!feof(rawfh)) {
nread = fread(buf, sizeof(*buf), nsamps, rawfh);
bptr = buf;
while ((nfr = acmod_process_raw(acmod, &bptr, &nread, FALSE)) > 0 || nread > 0) {
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
TEST_EQUAL(frame_counter, frame_idx);
if (frame_counter < 190)
bestsen1[frame_counter] = best_score;
++frame_counter;
frame_idx = -1;
}
}
}
TEST_EQUAL(0, acmod_end_utt(acmod));
nread = 0;
{
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
bestsen1[frame_counter] = best_score;
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
/* Now try to process the whole thing at once. */
E_INFO("Whole utterance:\n");
cmn_prior_set(acmod->fcb->cmn_struct, prior);
nsamps = ftell(rawfh) / sizeof(*buf);
clearerr(rawfh);
fseek(rawfh, 0, SEEK_SET);
buf = ckd_realloc(buf, nsamps * sizeof(*buf));
TEST_EQUAL(nsamps, fread(buf, sizeof(*buf), nsamps, rawfh));
bptr = buf;
TEST_EQUAL(0, acmod_start_utt(acmod));
acmod_process_raw(acmod, &bptr, &nsamps, TRUE);
TEST_EQUAL(0, acmod_end_utt(acmod));
{
int16 best_score;
int frame_idx = -1, best_senid;
frame_counter = 0;
while (acmod->n_feat_frame > 0) {
acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
/* Now process MFCCs and make sure we get the same results. */
cepbuf = ckd_calloc_2d(frame_counter,
fe_get_output_size(acmod->fe),
sizeof(**cepbuf));
fe_start_utt(acmod->fe);
nsamps = ftell(rawfh) / sizeof(*buf);
bptr = buf;
nfr = frame_counter;
fe_process_frames(acmod->fe, &bptr, &nsamps, cepbuf, &nfr, NULL);
fe_end_utt(acmod->fe, cepbuf[frame_counter-1], &nfr);
E_INFO("Incremental(MFCC):\n");
cmn_prior_set(acmod->fcb->cmn_struct, prior);
TEST_EQUAL(0, acmod_start_utt(acmod));
cptr = cepbuf;
nfr = frame_counter;
frame_counter = 0;
while ((acmod_process_cep(acmod, &cptr, &nfr, FALSE)) > 0) {
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
TEST_EQUAL(frame_counter, frame_idx);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
++frame_counter;
frame_idx = -1;
}
}
TEST_EQUAL(0, acmod_end_utt(acmod));
nfr = 0;
acmod_process_cep(acmod, &cptr, &nfr, FALSE);
{
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
TEST_EQUAL(frame_counter, frame_idx);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
++frame_counter;
frame_idx = -1;
}
}
/* Note that we have to process the whole thing again because
* !#@$@ s2mfc2feat modifies its argument (not for long) */
fe_start_utt(acmod->fe);
nsamps = ftell(rawfh) / sizeof(*buf);
bptr = buf;
nfr = frame_counter;
fe_process_frames(acmod->fe, &bptr, &nsamps, cepbuf, &nfr, NULL);
fe_end_utt(acmod->fe, cepbuf[frame_counter-1], &nfr);
E_INFO("Whole utterance (MFCC):\n");
cmn_prior_set(acmod->fcb->cmn_struct, prior);
TEST_EQUAL(0, acmod_start_utt(acmod));
cptr = cepbuf;
nfr = frame_counter;
acmod_process_cep(acmod, &cptr, &nfr, TRUE);
TEST_EQUAL(0, acmod_end_utt(acmod));
{
int16 best_score;
int frame_idx = -1, best_senid;
frame_counter = 0;
while (acmod->n_feat_frame > 0) {
acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
E_INFO("Rewound (MFCC):\n");
TEST_EQUAL(0, acmod_rewind(acmod));
{
int16 best_score;
int frame_idx = -1, best_senid;
frame_counter = 0;
while (acmod->n_feat_frame > 0) {
acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
/* Clean up, go home. */
ckd_free_2d(cepbuf);
fclose(rawfh);
ckd_free(buf);
acmod_free(acmod);
logmath_free(lmath);
cmd_ln_free_r(config);
return 0;
}
/* Update kb w/ new dictionary and new LM.
* assumes: single-LM kbcore (before & after)
* requires: updating kbcore
* Lucian Galescu, 08/11/2005
*/
void kb_update_lm(kb_t *kb, char *dictfile, char *lmfile)
{
kbcore_t *kbcore;
mdef_t *mdef;
dict_t *dict;
dict2pid_t *d2p;
lm_t *lm;
s3cipid_t ci;
s3wid_t w;
int32 i, n, n_lc;
wordprob_t *wp;
s3cipid_t *lc;
bitvec_t lc_active;
/*** clean up ***/
vithist_t *vithist = kb->vithist;
if (kb->fillertree)
ckd_free ((void *)kb->fillertree);
if (kb->hmm_hist)
ckd_free ((void *)kb->hmm_hist);
/* vithist */
if (vithist) {
ckd_free ((void *) vithist->entry);
ckd_free ((void *) vithist->frame_start);
ckd_free ((void *) vithist->bestscore);
ckd_free ((void *) vithist->bestvh);
ckd_free ((void *) vithist->lms2vh_root);
ckd_free ((void *) kb->vithist);
}
/*** re-initialize ***/
kb->kbcore = kbcore_update_lm(kb->kbcore,
dictfile,
cmd_ln_str("-fdict"),
"", /* Hack!! Hardwired constant for -compsep argument */
lmfile,
cmd_ln_str("-fillpen"),
cmd_ln_float32("-silprob"),
cmd_ln_float32("-fillprob"),
cmd_ln_float32("-lw"),
cmd_ln_float32("-wip"),
cmd_ln_float32("-uw"));
if(kb->kbcore==NULL){
E_FATAL("Updating kbcore failed\n");
}
kbcore = kb->kbcore;
mdef = kbcore_mdef(kbcore);
dict = kbcore_dict(kbcore);
lm = kbcore_lm(kbcore);
d2p = kbcore_dict2pid(kbcore);
if (NOT_S3WID(dict_startwid(dict)) || NOT_S3WID(dict_finishwid(dict)))
E_FATAL("%s or %s not in dictionary\n", S3_START_WORD, S3_FINISH_WORD);
if(lm){
if (NOT_S3LMWID(lm_startwid(lm)) || NOT_S3LMWID(lm_finishwid(lm)))
E_FATAL("%s or %s not in LM\n", S3_START_WORD, S3_FINISH_WORD);
}
/*
* Unlink <s> and </s> between dictionary and LM, to prevent their
* recognition. They are merely dummy words (anchors) at the beginning
* and end of each utterance.
*/
if(lm){
lm_lmwid2dictwid(lm, lm_startwid(lm)) = BAD_S3WID;
lm_lmwid2dictwid(lm, lm_finishwid(lm)) = BAD_S3WID;
for (w = dict_startwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
kbcore->dict2lmwid[w] = BAD_S3LMWID;
for (w = dict_finishwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
kbcore->dict2lmwid[w] = BAD_S3LMWID;
}
/* Build set of all possible left contexts */
lc = (s3cipid_t *) ckd_calloc (mdef_n_ciphone(mdef) + 1, sizeof(s3cipid_t));
lc_active = bitvec_alloc (mdef_n_ciphone (mdef));
for (w = 0; w < dict_size (dict); w++) {
ci = dict_pron (dict, w, dict_pronlen(dict, w) - 1);
if (! mdef_is_fillerphone (mdef, (int)ci))
bitvec_set (lc_active, ci);
}
ci = mdef_silphone(mdef);
bitvec_set (lc_active, ci);
for (ci = 0, n_lc = 0; ci < mdef_n_ciphone(mdef); ci++) {
if (bitvec_is_set (lc_active, ci))
lc[n_lc++] = ci;
}
lc[n_lc] = BAD_S3CIPID;
E_INFO("Building lextrees\n");
/* Get the number of lexical tree*/
kb->n_lextree = cmd_ln_int32 ("-Nlextree");
if (kb->n_lextree < 1) {
E_ERROR("No. of ugtrees specified: %d; will instantiate 1 ugtree\n",
kb->n_lextree);
kb->n_lextree = 1;
}
/* ARCHAN: This code was rearranged in s3.4 implementation of dynamic LM */
/* Build active word list */
wp = (wordprob_t *) ckd_calloc (dict_size(dict), sizeof(wordprob_t));
if (lm) {
E_INFO("Creating Unigram Table\n");
n=0;
n = lm_ug_wordprob (lm, dict, MAX_NEG_INT32, wp);
E_INFO("Size of word table after unigram + words in class: %d\n",n);
if (n < 1)
E_FATAL("%d active words\n", n);
n = wid_wordprob2alt (dict, wp, n); /* Add alternative pronunciations */
/* Retain or remove unigram probs from lextree, depending on option */
if (cmd_ln_int32("-treeugprob") == 0) {
for (i = 0; i < n; i++)
wp[i].prob = -1; /* Flatten all initial probabilities */
}
/* Create the desired no. of unigram lextrees */
kb->ugtree = (lextree_t **) ckd_calloc (kb->n_lextree, sizeof(lextree_t *));
for (i = 0; i < kb->n_lextree; i++) {
kb->ugtree[i] = lextree_build (kbcore, wp, n, lc);
lextree_type (kb->ugtree[i]) = 0;
}
E_INFO("Lextrees(%d), %d nodes(ug)\n",
kb->n_lextree, lextree_n_node(kb->ugtree[0]));
}
/* Create filler lextrees */
/* ARCHAN : only one filler tree is supposed to be build even for dynamic LMs */
n = 0;
for (i = dict_filler_start(dict); i <= dict_filler_end(dict); i++) {
if (dict_filler_word(dict, i)) {
wp[n].wid = i;
wp[n].prob = fillpen (kbcore->fillpen, i);
n++;
}
}
kb->fillertree = (lextree_t **)ckd_calloc(kb->n_lextree,sizeof(lextree_t*));
for (i = 0; i < kb->n_lextree; i++) {
kb->fillertree[i] = lextree_build (kbcore, wp, n, NULL);
lextree_type (kb->fillertree[i]) = -1;
}
ckd_free ((void *) wp);
ckd_free ((void *) lc);
bitvec_free (lc_active);
E_INFO("Lextrees(%d), %d nodes(filler)\n",
kb->n_lextree,
lextree_n_node(kb->fillertree[0]));
if (cmd_ln_int32("-lextreedump")) {
for (i = 0; i < kb->n_lextree; i++) {
fprintf (stderr, "UGTREE %d\n", i);
lextree_dump (kb->ugtree[i], dict, stderr);
}
for (i = 0; i < kb->n_lextree; i++) {
fprintf (stderr, "FILLERTREE %d\n", i);
lextree_dump (kb->fillertree[i], dict, stderr);
}
fflush (stderr);
}
kb->ascr = ascr_init (mgau_n_mgau(kbcore_mgau(kbcore)),
kbcore->dict2pid->n_comstate);
kb->vithist = vithist_init(kbcore, kb->beam->word, cmd_ln_int32("-bghist"));
kb->hmm_hist_binsize = cmd_ln_int32("-hmmhistbinsize");
n = ((kb->ugtree[0]->n_node) + (kb->fillertree[0]->n_node)) * kb->n_lextree;
n /= kb->hmm_hist_binsize;
kb->hmm_hist_bins = n+1;
kb->hmm_hist = (int32 *) ckd_calloc (n+1, sizeof(int32)); /* Really no need for +1 */
}
