static int32
interp_read(interp_t * ip, const char *file_name)
{
FILE *fp;
int32 byteswap, chksum_present;
int32 i;
char eofchk;
float f;
char **argname, **argval;
uint32 chksum;
E_INFO("Reading interpolation weights: %s\n", file_name);
if ((fp = fopen(file_name, "rb")) == NULL)
E_FATAL_SYSTEM("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], INTERP_VERSION) != 0)
E_WARN("Version mismatch(%s): %s, expecting %s\n",
file_name, argval[i], INTERP_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 */
if (bio_fread(&(ip->n_sen), sizeof(int32), 1, fp, byteswap, &chksum) !=
1)
E_FATAL("fread(%s) (arraysize) failed\n", file_name);
if (ip->n_sen <= 0)
E_FATAL("%s: arraysize= %d in header\n", file_name, ip->n_sen);
ip->wt =
(struct interp_wt_s *) ckd_calloc(ip->n_sen,
sizeof(struct interp_wt_s));
for (i = 0; i < ip->n_sen; i++) {
if (bio_fread(&f, sizeof(float32), 1, fp, byteswap, &chksum) != 1)
E_FATAL("fread(%s) (arraydata) failed\n", file_name);
if ((f < 0.0) || (f > 1.0))
E_FATAL("%s: interpolation weight(%d)= %e\n", file_name, i, f);
ip->wt[i].cd = (f == 0.0) ? S3_LOGPROB_ZERO : logs3(ip->logmath, f);
ip->wt[i].ci = (f == 1.0) ? S3_LOGPROB_ZERO : logs3(ip->logmath, 1.0 - f);
}
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 interpolation weights\n", ip->n_sen);
return 1;
}
/*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 SPEC_CPU_WINDOWS
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 SPEC_CPU_WINDOWS
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);
}
}
tmat_t *tmat_init (char *file_name, float64 tpfloor)
{
char tmp;
int32 n_src, n_dst;
FILE *fp;
int32 byteswap, chksum_present;
uint32 chksum;
float32 **tp;
int32 i, j, k, tp_per_tmat;
char **argname, **argval;
tmat_t *t;
E_INFO("Reading HMM transition probability matrices: %s\n", file_name);
t = (tmat_t *) ckd_calloc (1, sizeof(tmat_t));
if ((fp = fopen(file_name, "rb")) == NULL)
E_FATAL_SYSTEM("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], TMAT_PARAM_VERSION) != 0)
E_WARN("Version mismatch(%s): %s, expecting %s\n",
file_name, argval[i], TMAT_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;
/* Read #tmat, #from-states, #to-states, arraysize */
if ((bio_fread (&(t->n_tmat), sizeof(int32), 1, fp, byteswap, &chksum) != 1) ||
(bio_fread (&n_src, sizeof(int32), 1, fp, byteswap, &chksum) != 1) ||
(bio_fread (&n_dst, sizeof(int32), 1, fp, byteswap, &chksum) != 1) ||
(bio_fread (&i, sizeof(int32), 1, fp, byteswap, &chksum) != 1)) {
E_FATAL("bio_fread(%s) (arraysize) failed\n", file_name);
}
if (t->n_tmat >= MAX_S3TMATID)
E_FATAL("%s: #tmat (%d) exceeds limit (%d)\n", file_name, t->n_tmat, MAX_S3TMATID);
if (n_dst != n_src+1)
E_FATAL("%s: #from-states(%d) != #to-states(%d)-1\n", file_name, n_src, n_dst);
t->n_state = n_src;
if (i != t->n_tmat * n_src * n_dst) {
E_FATAL("%s: #float32s(%d) doesn't match dimensions: %d x %d x %d\n",
file_name, i, t->n_tmat, n_src, n_dst);
}
/* Allocate memory for tmat data */
t->tp = (int32 ***) ckd_calloc_3d (t->n_tmat, n_src, n_dst, sizeof(int32));
/* Temporary structure to read in the float data */
tp = (float32 **) ckd_calloc_2d (n_src, n_dst, sizeof(float32));
/* Read transition matrices, normalize and floor them, and convert to logs3 domain */
tp_per_tmat = n_src * n_dst;
for (i = 0; i < t->n_tmat; i++) {
if (bio_fread (tp[0], sizeof(float32), tp_per_tmat, fp,
byteswap, &chksum) != tp_per_tmat) {
E_FATAL("fread(%s) (arraydata) failed\n", file_name);
}
/* Normalize and floor */
for (j = 0; j < n_src; j++) {
if (vector_sum_norm (tp[j], n_dst) == 0.0)
E_ERROR("Normalization failed for tmat %d from state %d\n", i, j);
vector_nz_floor (tp[j], n_dst, tpfloor);
vector_sum_norm (tp[j], n_dst);
/* Convert to logs3. Take care of special case when tp = 0.0! */
for (k = 0; k < n_dst; k++)
t->tp[i][j][k] = (tp[j][k] == 0.0) ? S3_LOGPROB_ZERO : logs3(tp[j][k]);
}
}
ckd_free_2d ((void **) tp);
if (chksum_present)
bio_verify_chksum (fp, byteswap, chksum);
if (fread (&tmp, 1, 1, fp) == 1)
E_ERROR("Non-empty file beyond end of data\n");
fclose(fp);
E_INFO("Read %d transition matrices of size %dx%d\n",
t->n_tmat, t->n_state, t->n_state+1);
if (tmat_chk_uppertri (t) < 0)
E_FATAL("Tmat not upper triangular\n");
return t;
}
word_fsg_t *
word_fsg_load(s2_fsg_t * fsg,
int use_altpron, int use_filler,
kbcore_t *kbc)
{
float32 silprob = kbc->fillpen->silprob;
float32 fillprob = kbc->fillpen->fillerprob;
float32 lw = kbc->fillpen->lw;
word_fsg_t *word_fsg;
s2_fsg_trans_t *trans;
int32 n_trans, n_null_trans, n_alt_trans, n_filler_trans, n_unk;
int32 wid;
int32 logp;
glist_t nulls;
int32 i, j;
assert(fsg);
/* Some error checking */
if (lw <= 0.0)
E_WARN("Unusual language-weight value: %.3e\n", lw);
if (use_filler && ((silprob < 0.0) || (fillprob < 0.0))) {
E_ERROR("silprob/fillprob must be >= 0\n");
return NULL;
}
if ((fsg->n_state <= 0)
|| ((fsg->start_state < 0) || (fsg->start_state >= fsg->n_state))
|| ((fsg->final_state < 0) || (fsg->final_state >= fsg->n_state))) {
E_ERROR("Bad #states/start_state/final_state values: %d/%d/%d\n",
fsg->n_state, fsg->start_state, fsg->final_state);
return NULL;
}
for (trans = fsg->trans_list; trans; trans = trans->next) {
if ((trans->from_state < 0) || (trans->from_state >= fsg->n_state)
|| (trans->to_state < 0) || (trans->to_state >= fsg->n_state)
|| (trans->prob <= 0) || (trans->prob > 1.0)) {
E_ERROR("Bad transition: P(%d -> %d) = %e\n",
trans->from_state, trans->to_state, trans->prob);
return NULL;
}
}
word_fsg = (word_fsg_t *) ckd_calloc(1, sizeof(word_fsg_t));
word_fsg->name = ckd_salloc(fsg->name ? fsg->name : "");
word_fsg->n_state = fsg->n_state;
word_fsg->start_state = fsg->start_state;
word_fsg->final_state = fsg->final_state;
word_fsg->use_altpron = use_altpron;
word_fsg->use_filler = use_filler;
word_fsg->lw = lw;
word_fsg->lc = NULL;
word_fsg->rc = NULL;
word_fsg->dict = kbc->dict;
word_fsg->mdef = kbc->mdef;
word_fsg->tmat = kbc->tmat;
word_fsg->n_ciphone = mdef_n_ciphone(kbc->mdef);
/* Allocate non-epsilon transition matrix array */
word_fsg->trans = (glist_t **) ckd_calloc_2d(word_fsg->n_state,
word_fsg->n_state,
sizeof(glist_t));
/* Allocate epsilon transition matrix array */
word_fsg->null_trans = (word_fsglink_t ***)
ckd_calloc_2d(word_fsg->n_state, word_fsg->n_state,
sizeof(word_fsglink_t *));
/* Process transitions */
n_null_trans = 0;
n_alt_trans = 0;
n_filler_trans = 0;
n_unk = 0;
nulls = NULL;
for (trans = fsg->trans_list, n_trans = 0;
trans; trans = trans->next, n_trans++) {
/* Convert prob to logs2prob and apply language weight */
logp = (int32) (logs3(kbcore_logmath(kbc), trans->prob) * lw);
/* Check if word is in dictionary */
if (trans->word) {
wid = dict_wordid(kbc->dict, trans->word);
if (wid < 0) {
E_ERROR("Unknown word '%s'; ignored\n", trans->word);
n_unk++;
}
else if (use_altpron) {
wid = dict_basewid(kbc->dict, wid);
assert(wid >= 0);
}
}
else
wid = -1; /* Null transition */
/* Add transition to word_fsg structure */
i = trans->from_state;
j = trans->to_state;
if (wid < 0) {
if (word_fsg_null_trans_add(word_fsg, i, j, logp) == 1) {
n_null_trans++;
nulls =
glist_add_ptr(nulls,
(void *) word_fsg->null_trans[i][j]);
}
}
else {
word_fsg_trans_add(word_fsg, i, j, logp, wid);
/* Add transitions for alternative pronunciations, if any */
if (use_altpron) {
for (wid = dict_nextalt(kbc->dict, wid);
wid >= 0; wid = dict_nextalt(kbc->dict, wid)) {
word_fsg_trans_add(word_fsg, i, j, logp, wid);
n_alt_trans++;
n_trans++;
}
}
}
}
/* Add silence and noise filler word transitions if specified */
if (use_filler) {
n_filler_trans = word_fsg_add_filler(word_fsg, silprob, fillprob, kbcore_logmath(kbc));
n_trans += n_filler_trans;
}
E_INFO
("FSG: %d states, %d transitions (%d null, %d alt, %d filler, %d unknown)\n",
word_fsg->n_state, n_trans, n_null_trans, n_alt_trans,
n_filler_trans, n_unk);
#if __FSG_DBG__
E_INFO("FSG before NULL closure:\n");
word_fsg_write(word_fsg, stdout);
#endif
/* Null transitions closure */
nulls = word_fsg_null_trans_closure(word_fsg, nulls);
glist_free(nulls);
#if __FSG_DBG__
E_INFO("FSG after NULL closure:\n");
word_fsg_write(word_fsg, stdout);
#endif
/* Compute left and right context CIphone lists for each state */
word_fsg_lc_rc(word_fsg);
#if __FSG_DBG__
E_INFO("FSG after lc/rc:\n");
word_fsg_write(word_fsg, stdout);
#endif
return word_fsg;
}
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;
}
acoustic_t *acoustic_init (feat_t *f, gauden_t *g, senone_t *s,
float64 beam,
int32 maxfr)
{
acoustic_t *am;
int32 i;
if (senone_n_mgau(s) != gauden_n_mgau(g)) {
E_ERROR("#Parent mixture Gaussians mismatch: senone(%d), gauden(%d)\n",
senone_n_mgau(s), gauden_n_mgau(g));
}
if (feat_n_stream(f) != senone_n_stream(s)) {
E_ERROR("#Feature-streams mismatch: feat(%d), senone(%d)\n",
feat_n_stream(f), senone_n_stream(s));
}
if (feat_n_stream(f) != gauden_n_stream(g)) {
E_ERROR("#Feature-streams mismatch: feat(%d), gauden(%d)\n",
feat_n_stream(f), gauden_n_stream(g));
return NULL;
}
for (i = 0; i < feat_n_stream(f); i++) {
if (feat_stream_len(f, i) != gauden_stream_len(g, i)) {
E_ERROR("Feature stream(%d) length mismatch: feat(%d), gauden(%d)\n",
feat_stream_len(f, i), gauden_stream_len(g, i));
return NULL;
}
}
if (beam > 1.0) {
E_ERROR("mgaubeam > 1.0 (%e)\n", beam);
return NULL;
}
am = (acoustic_t *) ckd_calloc (1, sizeof(acoustic_t));
am->fcb = f;
am->gau = g;
am->sen = s;
am->mgaubeam = (beam == 0.0) ? LOGPROB_ZERO : logs3(beam);
if (am->mgaubeam > 0)
am->mgaubeam = 0;
am->tot_mgau_eval = 0;
am->tot_dist_valid = 0.0;
am->dist_valid = (am->mgaubeam <= LOGPROB_ZERO) ? NULL :
(int32 *) ckd_calloc (g->max_n_mean, sizeof(int32));
if (f->compute_feat) {
/* Input is MFC cepstra; feature vectors computed from that */
am->mfc = (float32 **) ckd_calloc_2d (maxfr, feat_cepsize(am->fcb),
sizeof(float32));
am->feat = feat_array_alloc (f, 1);
} else {
/* Input is directly feature vectors */
am->mfc = NULL;
am->feat = feat_array_alloc (f, maxfr);
}
am->dist = (int32 *) ckd_calloc (g->max_n_mean, sizeof(int32));
am->gauden_active = bitvec_alloc (g->n_mgau);
am->senscr = (int32 *) ckd_calloc (s->n_sen, sizeof(int32));
am->senscale = (int32 *) ckd_calloc (maxfr, sizeof(int32));
am->sen_active = bitvec_alloc (s->n_sen);
return am;
}
main (int32 argc, char *argv[])
{
kb_t kb;
kbcore_t *kbcore;
bitvec_t active;
int32 w;
cmd_ln_parse (arglist, argc, argv);
unlimit();
kbcore = kbcore_init (cmd_ln_float32("-logbase"),
cmd_ln_str("-feat"),
cmd_ln_str("-mdef"),
cmd_ln_str("-dict"),
cmd_ln_str("-fdict"),
cmd_ln_str("-compsep"),
cmd_ln_str("-lm"),
cmd_ln_str("-fillpen"),
cmd_ln_float32("-silprob"),
cmd_ln_float32("-fillprob"),
cmd_ln_float32("-lw"),
cmd_ln_float32("-wip"),
cmd_ln_str("-mean"),
cmd_ln_str("-var"),
cmd_ln_float32("-varfloor"),
cmd_ln_str("-senmgau"),
cmd_ln_str("-mixw"),
cmd_ln_float32("-mixwfloor"),
cmd_ln_str("-tmat"),
cmd_ln_float32("-tmatfloor"));
/* Here's the perfect candidate for inheritance */
kb.mdef = kbcore->mdef;
kb.dict = kbcore->dict;
kb.lm = kbcore->lm;
kb.fillpen = kbcore->fillpen;
kb.tmat = kbcore->tmat;
kb.dict2lmwid = kbcore->dict2lmwid;
if ((kb.am = acoustic_init (kbcore->fcb, kbcore->gau, kbcore->sen, cmd_ln_float32("-mgaubeam"),
S3_MAX_FRAMES)) == NULL) {
E_FATAL("Acoustic models initialization failed\n");
}
kb.beam = logs3 (cmd_ln_float64("-beam"));
kb.wordbeam = logs3 (cmd_ln_float64("-wordbeam"));
kb.wordmax = cmd_ln_int32("-wordmax");
/* Mark the active words and build lextree */
active = bitvec_alloc (dict_size (kb.dict));
bitvec_clear_all (active, dict_size(kb.dict));
for (w = 0; w < dict_size(kb.dict); w++) {
if (IS_LMWID(kb.dict2lmwid[w]) || dict_filler_word (kb.dict, w))
bitvec_set (active, w);
}
kb.lextree_root = lextree_build (kb.dict, kb.mdef, active, cmd_ln_int32("-flatdepth"));
kb.vithist = (glist_t *) ckd_calloc (S3_MAX_FRAMES+2, sizeof(glist_t));
kb.vithist++; /* Allow for dummy frame -1 for start word */
kb.lextree_active = NULL;
kb.wd_last_sf = (int32 *) ckd_calloc (dict_size(kb.dict), sizeof(int32));
kb.tm = (ptmr_t *) ckd_calloc (1, sizeof(ptmr_t));
kb.tm_search = (ptmr_t *) ckd_calloc (1, sizeof(ptmr_t));
ctl_process (cmd_ln_str("-ctl"), cmd_ln_int32("-ctloffset"), cmd_ln_int32("-ctlcount"),
decode_utt, &kb);
exit(0);
}
/* RAH 4.16.01 This code has several leaks that must be fixed */
dict2pid_t *dict2pid_build (mdef_t *mdef, dict_t *dict)
{
dict2pid_t *dict2pid;
s3ssid_t *internal, **ldiph, **rdiph, *single;
int32 pronlen;
hash_table_t *hs, *hp;
glist_t g;
gnode_t *gn;
s3senid_t *sen;
hash_entry_t *he;
int32 *cslen;
int32 i, j, b, l, r, w, n, p;
E_INFO("Building PID tables for dictionary\n");
dict2pid = (dict2pid_t *) ckd_calloc (1, sizeof(dict2pid_t));
dict2pid->internal = (s3ssid_t **) ckd_calloc (dict_size(dict), sizeof(s3ssid_t *));
dict2pid->ldiph_lc = (s3ssid_t ***) ckd_calloc_3d (mdef->n_ciphone,
mdef->n_ciphone,
mdef->n_ciphone,
sizeof(s3ssid_t));
dict2pid->single_lc = (s3ssid_t **) ckd_calloc_2d (mdef->n_ciphone,
mdef->n_ciphone,
sizeof(s3ssid_t));
dict2pid->n_comstate = 0;
dict2pid->n_comsseq = 0;
hs = hash_new (mdef->n_ciphone * mdef->n_ciphone * mdef->n_emit_state, HASH_CASE_YES);
hp = hash_new (mdef->n_ciphone * mdef->n_ciphone, HASH_CASE_YES);
for (w = 0, n = 0; w < dict_size(dict); w++) {
pronlen = dict_pronlen(dict, w);
if (pronlen < 0)
E_FATAL("Pronunciation-length(%s)= %d\n", dict_wordstr(dict, w), pronlen);
n += pronlen;
}
internal = (s3ssid_t *) ckd_calloc (n, sizeof(s3ssid_t));
/* Temporary */
ldiph = (s3ssid_t **) ckd_calloc_2d (mdef->n_ciphone, mdef->n_ciphone, sizeof(s3ssid_t));
rdiph = (s3ssid_t **) ckd_calloc_2d (mdef->n_ciphone, mdef->n_ciphone, sizeof(s3ssid_t));
single = (s3ssid_t *) ckd_calloc (mdef->n_ciphone, sizeof(s3ssid_t));
for (b = 0; b < mdef->n_ciphone; b++) {
for (l = 0; l < mdef->n_ciphone; l++) {
for (r = 0; r < mdef->n_ciphone; r++)
dict2pid->ldiph_lc[b][r][l] = BAD_S3SSID;
dict2pid->single_lc[b][l] = BAD_S3SSID;
ldiph[b][l] = BAD_S3SSID;
rdiph[b][l] = BAD_S3SSID;
}
single[b] = BAD_S3SSID;
}
for (w = 0; w < dict_size(dict); w++) {
dict2pid->internal[w] = internal;
pronlen = dict_pronlen(dict,w);
if (pronlen >= 2) {
b = dict_pron(dict, w, 0);
r = dict_pron(dict, w, 1);
if (NOT_S3SSID(ldiph[b][r])) {
g = ldiph_comsseq(mdef, b, r);
ldiph[b][r] = ssidlist2comsseq (g, mdef, dict2pid, hs, hp);
glist_free (g);
for (l = 0; l < mdef_n_ciphone(mdef); l++) {
p = mdef_phone_id_nearest (mdef, (s3cipid_t)b, (s3cipid_t)l, (s3cipid_t)r, WORD_POSN_BEGIN);
dict2pid->ldiph_lc[b][r][l] = mdef_pid2ssid(mdef, p);
}
}
internal[0] = ldiph[b][r];
for (i = 1; i < pronlen-1; i++) {
l = b;
b = r;
r = dict_pron(dict, w, i+1);
p = mdef_phone_id_nearest(mdef, (s3cipid_t)b, (s3cipid_t)l, (s3cipid_t)r, WORD_POSN_INTERNAL);
internal[i] = mdef_pid2ssid(mdef, p);
}
l = b;
b = r;
if (NOT_S3SSID(rdiph[b][l])) {
g = rdiph_comsseq(mdef, b, l);
rdiph[b][l] = ssidlist2comsseq (g, mdef, dict2pid, hs, hp);
glist_free (g);
}
internal[pronlen-1] = rdiph[b][l];
} else if (pronlen == 1) {
b = dict_pron(dict, w, 0);
if (NOT_S3SSID(single[b])) {
g = single_comsseq(mdef, b);
single[b] = ssidlist2comsseq (g, mdef, dict2pid, hs, hp);
glist_free (g);
for (l = 0; l < mdef_n_ciphone(mdef); l++) {
g = single_lc_comsseq(mdef, b, l);
dict2pid->single_lc[b][l] = ssidlist2comsseq (g, mdef, dict2pid, hs, hp);
glist_free (g);
}
}
internal[0] = single[b];
}
internal += pronlen;
}
ckd_free_2d ((void **) ldiph);
ckd_free_2d ((void **) rdiph);
ckd_free ((void *) single);
/* Allocate space for composite state table */
cslen = (int32 *) ckd_calloc (dict2pid->n_comstate, sizeof(int32));
g = hash_tolist(hs, &n);
assert (n == dict2pid->n_comstate);
n = 0;
for (gn = g; gn; gn = gnode_next(gn)) {
he = (hash_entry_t *) gnode_ptr (gn);
sen = (s3senid_t *) hash_entry_key(he);
for (i = 0; IS_S3SENID(sen[i]); i++);
cslen[hash_entry_val(he)] = i+1; /* +1 for terminating sentinel */
n += (i+1);
}
dict2pid->comstate = (s3senid_t **) ckd_calloc (dict2pid->n_comstate, sizeof(s3senid_t *));
sen = (s3senid_t *) ckd_calloc (n, sizeof(s3senid_t));
for (i = 0; i < dict2pid->n_comstate; i++) {
dict2pid->comstate[i] = sen;
sen += cslen[i];
}
/* Build composite state table from hash table hs */
for (gn = g; gn; gn = gnode_next(gn)) {
he = (hash_entry_t *) gnode_ptr (gn);
sen = (s3senid_t *) hash_entry_key(he);
i = hash_entry_val(he);
for (j = 0; j < cslen[i]; j++)
dict2pid->comstate[i][j] = sen[j];
assert (sen[j-1] == BAD_S3SENID);
ckd_free ((void *)sen);
}
ckd_free (cslen);
glist_free (g);
hash_free (hs);
/* Allocate space for composite sseq table */
dict2pid->comsseq = (s3senid_t **) ckd_calloc (dict2pid->n_comsseq, sizeof(s3senid_t *));
g = hash_tolist (hp, &n);
assert (n == dict2pid->n_comsseq);
/* Build composite sseq table */
for (gn = g; gn; gn = gnode_next(gn)) {
he = (hash_entry_t *) gnode_ptr (gn);
i = hash_entry_val(he);
dict2pid->comsseq[i] = (s3senid_t *) hash_entry_key(he);
}
glist_free (g);
hash_free (hp);
/* Weight for each composite state */
dict2pid->comwt = (int32 *) ckd_calloc (dict2pid->n_comstate, sizeof(int32));
for (i = 0; i < dict2pid->n_comstate; i++) {
sen = dict2pid->comstate[i];
for (j = 0; IS_S3SENID(sen[j]); j++);
#if 0
/* if comstate i has N states, its weight= (1/N^2) (Major Hack!!) */
dict2pid->comwt[i] = - (logs3 ((float64)j) << 1);
#else
/* if comstate i has N states, its weight= 1/N */
dict2pid->comwt[i] = - logs3 ((float64)j);
#endif
}
E_INFO("%d composite states; %d composite sseq\n",
dict2pid->n_comstate, dict2pid->n_comsseq);
return dict2pid;
}
