ngram_model_t *
ngram_model_set_init(cmd_ln_t * config,
ngram_model_t ** models,
char **names, const float32 * weights, int32 n_models)
{
ngram_model_set_t *model;
ngram_model_t *base;
logmath_t *lmath;
int32 i, n;
if (n_models == 0) /* WTF */
return NULL;
/* Do consistency checking on the models. They must all use the
* same logbase and shift. */
lmath = models[0]->lmath;
for (i = 1; i < n_models; ++i) {
if (logmath_get_base(models[i]->lmath) != logmath_get_base(lmath)
|| logmath_get_shift(models[i]->lmath) !=
logmath_get_shift(lmath)) {
E_ERROR
("Log-math parameters don't match, will not create LM set\n");
return NULL;
}
}
/* Allocate the combined model, initialize it. */
model = ckd_calloc(1, sizeof(*model));
base = &model->base;
model->n_models = n_models;
model->lms = ckd_calloc(n_models, sizeof(*model->lms));
model->names = ckd_calloc(n_models, sizeof(*model->names));
/* Initialize weights to a uniform distribution */
model->lweights = ckd_calloc(n_models, sizeof(*model->lweights));
{
int32 uniform = logmath_log(lmath, 1.0 / n_models);
for (i = 0; i < n_models; ++i)
model->lweights[i] = uniform;
}
/* Default to interpolate if weights were given. */
if (weights)
model->cur = -1;
n = 0;
for (i = 0; i < n_models; ++i) {
model->lms[i] = ngram_model_retain(models[i]);
model->names[i] = ckd_salloc(names[i]);
if (weights)
model->lweights[i] = logmath_log(lmath, weights[i]);
/* N is the maximum of all merged models. */
if (models[i]->n > n)
n = models[i]->n;
}
/* Allocate the history mapping table. */
model->maphist = ckd_calloc(n - 1, sizeof(*model->maphist));
/* Now build the word-ID mapping and merged vocabulary. */
build_widmap(base, lmath, n);
return base;
}
static fsg_model_t *
jsgf_build_fsg_internal(jsgf_t * grammar, jsgf_rule_t * rule,
logmath_t * lmath, float32 lw, int do_closure)
{
fsg_model_t *fsg;
glist_t nulls;
gnode_t *gn;
int rule_entry, rule_exit;
/* Clear previous links */
for (gn = grammar->links; gn; gn = gnode_next(gn)) {
ckd_free(gnode_ptr(gn));
}
glist_free(grammar->links);
grammar->links = NULL;
grammar->nstate = 0;
/* Create the top-level entry state, and expand the
top-level rule. */
rule_entry = grammar->nstate++;
rule_exit = expand_rule(grammar, rule, rule_entry, NO_NODE);
/* If no exit-state was created, create one. */
if (rule_exit == NO_NODE) {
rule_exit = grammar->nstate++;
jsgf_add_link(grammar, NULL, rule_entry, rule_exit);
}
fsg = fsg_model_init(rule->name, lmath, lw, grammar->nstate);
fsg->start_state = rule_entry;
fsg->final_state = rule_exit;
grammar->links = glist_reverse(grammar->links);
for (gn = grammar->links; gn; gn = gnode_next(gn)) {
jsgf_link_t *link = gnode_ptr(gn);
if (link->atom) {
if (jsgf_atom_is_rule(link->atom)) {
fsg_model_null_trans_add(fsg, link->from, link->to,
logmath_log(lmath,
link->atom->weight));
}
else {
int wid = fsg_model_word_add(fsg, link->atom->name);
fsg_model_trans_add(fsg, link->from, link->to,
logmath_log(lmath, link->atom->weight),
wid);
}
}
else {
fsg_model_null_trans_add(fsg, link->from, link->to, 0);
}
}
if (do_closure) {
nulls = fsg_model_null_trans_closure(fsg, NULL);
glist_free(nulls);
}
return fsg;
}
ngram_model_t *
ngram_model_set_add(ngram_model_t * base,
ngram_model_t * model,
const char *name, float32 weight, int reuse_widmap)
{
ngram_model_set_t *set = (ngram_model_set_t *) base;
float32 fprob;
int32 scale, i;
/* Add it to the array of lms. */
++set->n_models;
set->lms = ckd_realloc(set->lms, set->n_models * sizeof(*set->lms));
set->lms[set->n_models - 1] = model;
set->names =
ckd_realloc(set->names, set->n_models * sizeof(*set->names));
set->names[set->n_models - 1] = ckd_salloc(name);
/* Expand the history mapping table if necessary. */
if (model->n > base->n) {
base->n = model->n;
set->maphist = ckd_realloc(set->maphist,
(model->n - 1) * sizeof(*set->maphist));
}
/* Renormalize the interpolation weights. */
fprob = weight * 1.0f / set->n_models;
set->lweights = ckd_realloc(set->lweights,
set->n_models * sizeof(*set->lweights));
set->lweights[set->n_models - 1] = logmath_log(base->lmath, fprob);
/* Now normalize everything else to fit it in. This is
* accomplished by simply scaling all the other probabilities
* by (1-fprob). */
scale = logmath_log(base->lmath, 1.0 - fprob);
for (i = 0; i < set->n_models - 1; ++i)
set->lweights[i] += scale;
/* Reuse the old word ID mapping if requested. */
if (reuse_widmap) {
int32 **new_widmap;
/* Tack another column onto the widmap array. */
new_widmap = (int32 **) ckd_calloc_2d(base->n_words, set->n_models,
sizeof(**new_widmap));
for (i = 0; i < base->n_words; ++i) {
/* Copy all the existing mappings. */
memcpy(new_widmap[i], set->widmap[i],
(set->n_models - 1) * sizeof(**new_widmap));
/* Create the new mapping. */
new_widmap[i][set->n_models - 1] =
ngram_wid(model, base->word_str[i]);
}
ckd_free_2d((void **) set->widmap);
set->widmap = new_widmap;
}
else {
build_widmap(base, base->lmath, base->n);
}
return model;
}
int32
ngram_model_add_class_word(ngram_model_t * model,
const char *classname,
const char *word, float32 weight)
{
ngram_class_t *lmclass;
int32 classid, tag_wid, wid, i, scale;
float32 fprob;
/* Find the class corresponding to classname. Linear search
* probably okay here since there won't be very many classes, and
* this doesn't have to be fast. */
tag_wid = ngram_wid(model, classname);
if (tag_wid == NGRAM_INVALID_WID) {
E_ERROR("No such word or class tag: %s\n", classname);
return tag_wid;
}
for (classid = 0; classid < model->n_classes; ++classid) {
if (model->classes[classid]->tag_wid == tag_wid)
break;
}
/* Hmm, no such class. It's probably not a good idea to create one. */
if (classid == model->n_classes) {
E_ERROR
("Word %s is not a class tag (call ngram_model_add_class() first)\n",
classname);
return NGRAM_INVALID_WID;
}
lmclass = model->classes[classid];
/* Add this word to the model's set of words. */
wid = ngram_add_word_internal(model, word, classid);
if (wid == NGRAM_INVALID_WID)
return wid;
/* This is the fixed probability of the new word. */
fprob = weight * 1.0f / (lmclass->n_words + lmclass->n_hash_inuse + 1);
/* Now normalize everything else to fit it in. This is
* accomplished by simply scaling all the other probabilities
* by (1-fprob). */
scale = logmath_log(model->lmath, 1.0 - fprob);
for (i = 0; i < lmclass->n_words; ++i)
lmclass->prob1[i] += scale;
for (i = 0; i < lmclass->n_hash; ++i)
if (lmclass->nword_hash[i].wid != -1)
lmclass->nword_hash[i].prob1 += scale;
/* Now add it to the class hash table. */
return ngram_class_add_word(lmclass, wid,
logmath_log(model->lmath, fprob));
}
static int
write_lattice(ps_decoder_t *ps, char const *latdir, char const *uttid)
{
ps_lattice_t *lat;
logmath_t *lmath;
cmd_ln_t *config;
char *outfile;
int32 beam;
if ((lat = ps_get_lattice(ps)) == NULL) {
E_ERROR("Failed to obtain word lattice for utterance %s\n", uttid);
return -1;
}
config = ps_get_config(ps);
outfile = string_join(latdir, "/", uttid,
cmd_ln_str_r(config, "-outlatext"), NULL);
/* Prune lattice. */
lmath = ps_get_logmath(ps);
beam = logmath_log(lmath, cmd_ln_float64_r(config, "-outlatbeam"));
ps_lattice_posterior_prune(lat, beam);
if (0 == strcmp("htk", cmd_ln_str_r(config, "-outlatfmt"))) {
if (ps_lattice_write_htk(lat, outfile) < 0) {
E_ERROR("Failed to write lattice to %s\n", outfile);
return -1;
}
}
else {
if (ps_lattice_write(lat, outfile) < 0) {
E_ERROR("Failed to write lattice to %s\n", outfile);
return -1;
}
}
return 0;
}
ngram_model_t *
ngram_model_set_interp(ngram_model_t * base,
const char **names, const float32 * weights)
{
ngram_model_set_t *set = (ngram_model_set_t *) base;
/* If we have a set of weights here, then set them. */
if (names && weights) {
int32 i, j;
/* We hope there aren't many models. */
for (i = 0; i < set->n_models; ++i) {
for (j = 0; j < set->n_models; ++j)
if (0 == strcmp(names[i], set->names[j]))
break;
if (j == set->n_models) {
E_ERROR("Unknown LM name %s\n", names[i]);
return NULL;
}
set->lweights[j] = logmath_log(base->lmath, weights[i]);
}
}
else if (weights) {
memcpy(set->lweights, weights,
set->n_models * sizeof(*set->lweights));
}
/* Otherwise just enable existing weights. */
set->cur = -1;
return base;
}
static int32
lm_trie_add_ug(ngram_model_t * base, int32 wid, int32 lweight)
{
ngram_model_trie_t *model = (ngram_model_trie_t *) base;
/* This would be very bad if this happened! */
assert(!NGRAM_IS_CLASSWID(wid));
/* Reallocate unigram array. */
model->trie->unigrams =
(unigram_t *) ckd_realloc(model->trie->unigrams,
sizeof(*model->trie->unigrams) *
(base->n_1g_alloc + 1));
memset(model->trie->unigrams + (base->n_counts[0] + 1), 0,
(size_t) (base->n_1g_alloc -
base->n_counts[0]) * sizeof(*model->trie->unigrams));
++base->n_counts[0];
lweight += logmath_log(base->lmath, 1.0 / base->n_counts[0]);
model->trie->unigrams[wid + 1].next = model->trie->unigrams[wid].next;
model->trie->unigrams[wid].prob = (float) lweight;
/* This unigram by definition doesn't participate in any bigrams,
* so its backoff weight is undefined and next pointer same as in finish unigram*/
model->trie->unigrams[wid].bo = 0;
/* Finally, increase the unigram count */
/* FIXME: Note that this can actually be quite bogus due to the
* presence of class words. If wid falls outside the unigram
* count, increase it to compensate, at the cost of no longer
* really knowing how many unigrams we have :( */
if ((uint32) wid >= base->n_counts[0])
base->n_counts[0] = wid + 1;
return (int32) weight_score(base, lweight);
}
int32
ngram_model_add_word(ngram_model_t * model,
const char *word, float32 weight)
{
int32 wid, prob = model->log_zero;
/* If we add word to unwritable model, we need to make it writable */
if (!model->writable) {
E_WARN("Can't add word '%s' to read-only language model. "
"Disable mmap with '-mmap no' to make it writable\n", word);
return -1;
}
wid = ngram_add_word_internal(model, word, -1);
if (wid == NGRAM_INVALID_WID)
return wid;
/* Do what needs to be done to add the word to the unigram. */
if (model->funcs && model->funcs->add_ug)
prob =
(*model->funcs->add_ug) (model, wid,
logmath_log(model->lmath, weight));
if (prob == 0)
return -1;
return wid;
}
ngram_class_t *
ngram_class_new(ngram_model_t * model, int32 tag_wid, int32 start_wid,
glist_t classwords)
{
ngram_class_t *lmclass;
gnode_t *gn;
float32 tprob;
int i;
lmclass = ckd_calloc(1, sizeof(*lmclass));
lmclass->tag_wid = tag_wid;
/* wid_base is the wid (minus class tag) of the first word in the list. */
lmclass->start_wid = start_wid;
lmclass->n_words = glist_count(classwords);
lmclass->prob1 = ckd_calloc(lmclass->n_words, sizeof(*lmclass->prob1));
lmclass->nword_hash = NULL;
lmclass->n_hash = 0;
tprob = 0.0;
for (gn = classwords; gn; gn = gnode_next(gn)) {
tprob += gnode_float32(gn);
}
if (tprob > 1.1 || tprob < 0.9) {
E_INFO("Total class probability is %f, will normalize\n", tprob);
for (gn = classwords; gn; gn = gnode_next(gn)) {
gn->data.fl /= tprob;
}
}
for (i = 0, gn = classwords; gn; ++i, gn = gnode_next(gn)) {
lmclass->prob1[i] = logmath_log(model->lmath, gnode_float32(gn));
}
return lmclass;
}
int
fsg_model_add_silence(fsg_model_t * fsg, char const *silword,
int state, float32 silprob)
{
int32 logsilp;
int n_trans, silwid, src;
E_INFO("Adding silence transitions for %s to FSG\n", silword);
silwid = fsg_model_word_add(fsg, silword);
logsilp = (int32) (logmath_log(fsg->lmath, silprob) * fsg->lw);
if (fsg->silwords == NULL)
fsg->silwords = bitvec_alloc(fsg->n_word_alloc);
bitvec_set(fsg->silwords, silwid);
n_trans = 0;
if (state == -1) {
for (src = 0; src < fsg->n_state; src++) {
fsg_model_trans_add(fsg, src, src, logsilp, silwid);
++n_trans;
}
}
else {
fsg_model_trans_add(fsg, state, state, logsilp, silwid);
++n_trans;
}
E_INFO("Added %d silence word transitions\n", n_trans);
return n_trans;
}
static fsg_model_t *
jsgf_build_fsg_internal(jsgf_t *grammar, jsgf_rule_t *rule,
logmath_t *lmath, float32 lw, int do_closure)
{
fsg_model_t *fsg;
glist_t nulls;
gnode_t *gn;
/* Clear previous links */
for (gn = grammar->links; gn; gn = gnode_next(gn)) {
ckd_free(gnode_ptr(gn));
}
glist_free(grammar->links);
grammar->links = NULL;
rule->entry = rule->exit = 0;
grammar->nstate = 0;
expand_rule(grammar, rule);
fsg = fsg_model_init(rule->name, lmath, lw, grammar->nstate);
fsg->start_state = rule->entry;
fsg->final_state = rule->exit;
grammar->links = glist_reverse(grammar->links);
for (gn = grammar->links; gn; gn = gnode_next(gn)) {
jsgf_link_t *link = gnode_ptr(gn);
if (link->atom) {
if (jsgf_atom_is_rule(link->atom)) {
fsg_model_null_trans_add(fsg, link->from, link->to,
logmath_log(lmath, link->atom->weight));
}
else {
int wid = fsg_model_word_add(fsg, link->atom->name);
fsg_model_trans_add(fsg, link->from, link->to,
logmath_log(lmath, link->atom->weight), wid);
}
}
else {
fsg_model_null_trans_add(fsg, link->from, link->to, 0);
}
}
if (do_closure) {
nulls = fsg_model_null_trans_closure(fsg, NULL);
glist_free(nulls);
}
return fsg;
}
static int
trie_apply_weights(ngram_model_t * base, float32 lw, float32 wip)
{
//just update weights that are going to be used on score calculation
base->lw = lw;
base->log_wip = logmath_log(base->lmath, wip);
return 0;
}
ngram_model_t *
ngram_model_set_remove(ngram_model_t * base,
const char *name, int reuse_widmap)
{
ngram_model_set_t *set = (ngram_model_set_t *) base;
ngram_model_t *submodel;
int32 lmidx, scale, n, i;
float32 fprob;
for (lmidx = 0; lmidx < set->n_models; ++lmidx)
if (0 == strcmp(name, set->names[lmidx]))
break;
if (lmidx == set->n_models)
return NULL;
submodel = set->lms[lmidx];
/* Renormalize the interpolation weights by scaling them by
* 1/(1-fprob) */
fprob = (float32) logmath_exp(base->lmath, set->lweights[lmidx]);
scale = logmath_log(base->lmath, 1.0 - fprob);
/* Remove it from the array of lms, renormalize remaining weights,
* and recalcluate n. */
--set->n_models;
n = 0;
ckd_free(set->names[lmidx]);
set->names[lmidx] = NULL;
for (i = 0; i < set->n_models; ++i) {
if (i >= lmidx) {
set->lms[i] = set->lms[i + 1];
set->names[i] = set->names[i + 1];
set->lweights[i] = set->lweights[i + 1];
}
set->lweights[i] -= scale;
if (set->lms[i]->n > n)
n = set->lms[i]->n;
}
/* There's no need to shrink these arrays. */
set->lms[set->n_models] = NULL;
set->lweights[set->n_models] = base->log_zero;
/* No need to shrink maphist either. */
/* Reuse the existing word ID mapping if requested. */
if (reuse_widmap) {
/* Just go through and shrink each row. */
for (i = 0; i < base->n_words; ++i) {
memmove(set->widmap[i] + lmidx, set->widmap[i] + lmidx + 1,
(set->n_models - lmidx) * sizeof(**set->widmap));
}
}
else {
build_widmap(base, base->lmath, n);
}
return submodel;
}
static int
phone_loop_search_reinit(ps_search_t *search, dict_t *dict, dict2pid_t *d2p)
{
phone_loop_search_t *pls = (phone_loop_search_t *)search;
cmd_ln_t *config = ps_search_config(search);
acmod_t *acmod = ps_search_acmod(search);
int i;
/* Free old dict2pid, dict, if necessary. */
ps_search_base_reinit(search, dict, d2p);
/* Initialize HMM context. */
if (pls->hmmctx)
hmm_context_free(pls->hmmctx);
pls->hmmctx = hmm_context_init(bin_mdef_n_emit_state(acmod->mdef),
acmod->tmat->tp, NULL, acmod->mdef->sseq);
if (pls->hmmctx == NULL)
return -1;
/* Initialize phone HMMs. */
if (pls->phones) {
for (i = 0; i < pls->n_phones; ++i)
hmm_deinit((hmm_t *)&pls->phones[i]);
ckd_free(pls->phones);
}
pls->n_phones = bin_mdef_n_ciphone(acmod->mdef);
pls->phones = ckd_calloc(pls->n_phones, sizeof(*pls->phones));
for (i = 0; i < pls->n_phones; ++i) {
pls->phones[i].ciphone = i;
hmm_init(pls->hmmctx, (hmm_t *)&pls->phones[i],
FALSE,
bin_mdef_pid2ssid(acmod->mdef, i),
bin_mdef_pid2tmatid(acmod->mdef, i));
}
pls->beam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-pl_beam"));
pls->pbeam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-pl_pbeam"));
pls->pip = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-pip"));
E_INFO("State beam %d Phone exit beam %d Insertion penalty %d\n",
pls->beam, pls->pbeam, pls->pip);
return 0;
}
int32
ngram_model_init(ngram_model_t *base,
ngram_funcs_t *funcs,
logmath_t *lmath,
int32 n, int32 n_unigram)
{
base->refcount = 1;
base->funcs = funcs;
base->n = n;
/* If this was previously initialized... */
if (base->n_counts == NULL)
base->n_counts = ckd_calloc(3, sizeof(*base->n_counts));
/* Don't reset weights if logmath object hasn't changed. */
if (base->lmath != lmath) {
/* Set default values for weights. */
base->lw = 1.0;
base->log_wip = 0; /* i.e. 1.0 */
base->log_uw = 0; /* i.e. 1.0 */
base->log_uniform = logmath_log(lmath, 1.0 / n_unigram);
base->log_uniform_weight = logmath_get_zero(lmath);
base->log_zero = logmath_get_zero(lmath);
base->lmath = lmath;
}
/* Allocate or reallocate space for word strings. */
if (base->word_str) {
/* Free all previous word strings if they were allocated. */
if (base->writable) {
int32 i;
for (i = 0; i < base->n_words; ++i) {
ckd_free(base->word_str[i]);
base->word_str[i] = NULL;
}
}
base->word_str = ckd_realloc(base->word_str, n_unigram * sizeof(char *));
}
else
base->word_str = ckd_calloc(n_unigram, sizeof(char *));
/* NOTE: They are no longer case-insensitive since we are allowing
* other encodings for word strings. Beware. */
if (base->wid)
hash_table_empty(base->wid);
else
base->wid = hash_table_new(n_unigram, FALSE);
base->n_counts[0] = base->n_1g_alloc = base->n_words = n_unigram;
return 0;
}
/*
* Some of the gaussian density computation can be carried out in advance:
* log(determinant) calculation,
* 1/(2*var) in the exponent,
* NOTE; The density computation is performed in log domain.
*/
static int32
gauden_dist_precompute(gauden_t * g, logmath_t *lmath, float32 varfloor)
{
int32 i, m, f, d, flen;
mfcc_t *meanp;
mfcc_t *varp;
mfcc_t *detp;
int32 floored;
floored = 0;
/* Allocate space for determinants */
g->det = (mfcc_t***)ckd_calloc_3d(g->n_mgau, g->n_feat, g->n_density, sizeof(***g->det));
for (m = 0; m < g->n_mgau; m++) {
for (f = 0; f < g->n_feat; f++) {
flen = g->featlen[f];
/* Determinants for all variance vectors in g->[m][f] */
for (d = 0, detp = g->det[m][f]; d < g->n_density; d++, detp++) {
*detp = 0;
for (i = 0, varp = g->var[m][f][d], meanp = g->mean[m][f][d];
i < flen; i++, varp++, meanp++) {
float32 *fvarp = (float32 *)varp;
#ifdef FIXED_POINT
float32 *fmp = (float32 *)meanp;
*meanp = FLOAT2MFCC(*fmp);
#endif
if (*fvarp < varfloor) {
*fvarp = varfloor;
++floored;
}
*detp += (mfcc_t)logmath_log(lmath,
1.0 / sqrt(*fvarp * 2.0 * M_PI));
/* Precompute this part of the exponential */
*varp = (mfcc_t)logmath_ln_to_log(lmath,
(1.0 / (*fvarp * 2.0)));
}
}
}
}
E_INFO("%d variance values floored\n", floored);
return 0;
}
int32
ngram_model_add_word(ngram_model_t *model,
const char *word, float32 weight)
{
int32 wid, prob = model->log_zero;
wid = ngram_add_word_internal(model, word, -1);
if (wid == NGRAM_INVALID_WID)
return wid;
/* Do what needs to be done to add the word to the unigram. */
if (model->funcs && model->funcs->add_ug)
prob = (*model->funcs->add_ug)(model, wid, logmath_log(model->lmath, weight));
if (prob == 0) {
if (model->writable)
ckd_free(model->word_str[wid]);
return -1;
}
return wid;
}
static int32
lm3g_template_raw_score(ngram_model_t *base, int32 wid,
int32 *history, int32 n_hist,
int32 *n_used)
{
NGRAM_MODEL_TYPE *model = (NGRAM_MODEL_TYPE *)base;
int32 score;
switch (n_hist) {
case 0:
/* Access mode: unigram */
*n_used = 1;
/* Undo insertion penalty. */
score = model->lm3g.unigrams[wid].prob1.l - base->log_wip;
/* Undo language weight. */
score = (int32)(score / base->lw);
/* Undo unigram interpolation */
if (strcmp(base->word_str[wid], "<s>") != 0) { /* FIXME: configurable start_sym */
/* This operation is numerically unstable, so try to avoid it
* as possible */
if (base->log_uniform + base->log_uniform_weight > logmath_get_zero(base->lmath)) {
score = logmath_log(base->lmath,
logmath_exp(base->lmath, score)
- logmath_exp(base->lmath,
base->log_uniform + base->log_uniform_weight));
}
}
return score;
case 1:
score = lm3g_bg_score(model, history[0], wid, n_used);
break;
case 2:
default:
/* Anything greater than 2 is the same as a trigram for now. */
score = lm3g_tg_score(model, history[1], history[0], wid, n_used);
break;
}
/* FIXME (maybe): This doesn't undo unigram weighting in backoff cases. */
return (int32)((score - base->log_wip) / base->lw);
}
static int32
s3_precomp(s2_semi_mgau_t *s, logmath_t *lmath, float32 vFloor)
{
int feat;
for (feat = 0; feat < s->n_feat; ++feat) {
float32 *fmp;
mfcc_t *mp;
mfcc_t *vp, *dp;
int32 vecLen, i;
vecLen = s->veclen[feat];
fmp = (float32 *) s->means[feat];
mp = s->means[feat];
vp = s->vars[feat];
dp = s->dets[feat];
for (i = 0; i < s->n_density; ++i) {
mfcc_t d;
int32 j;
d = 0;
for (j = 0; j < vecLen; ++j, ++vp, ++mp, ++fmp) {
float64 fvar;
*mp = FLOAT2MFCC(*fmp);
/* Always do these pre-calculations in floating point */
fvar = *(float32 *) vp;
if (fvar < vFloor)
fvar = vFloor;
d += (mfcc_t)logmath_log(lmath, 1 / sqrt(fvar * 2.0 * M_PI));
*vp = (mfcc_t)logmath_ln_to_log(lmath, 1.0 / (2.0 * fvar));
}
*dp++ = d;
}
}
return 0;
}
static int32 senone_mixw_read(logmath_t * logmath, senone_t *s, const char *file_name, float64 mixwfloor)
{
FILE *fp;
char **argname, **argval;
int32 byteswap, chksum_present;
uint32 chksum;
float32 *pdf;
int32 i, j, f, m, c, p, n_sen, n_err, n_cw, nval;
char eofchk;
mixw_t *fw;
E_INFO("Reading senone mixture 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], MIXW_PARAM_VERSION) != 0)
E_WARN("Version mismatch(%s): %s, expecting %s\n",
file_name, argval[i], MIXW_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 #senones, #features, #codewords, arraysize */
n_sen = s->n_sen;
if ((bio_fread (&(s->n_sen), sizeof(int32), 1, fp, byteswap, &chksum) != 1) ||
(bio_fread (&(s->n_feat), sizeof(int32), 1, fp, byteswap, &chksum) != 1) ||
(bio_fread (&(n_cw), sizeof(int32), 1, fp, byteswap, &chksum) != 1) ||
(bio_fread (&nval, sizeof(int32), 1, fp, byteswap, &chksum) != 1)) {
E_FATAL("bio_fread(%s) (arraysize) failed\n", file_name);
}
if ((n_sen != 0) && (s->n_sen != n_sen))
E_FATAL("#senones(%d) conflict with mapping file(%d)\n", s->n_sen, n_sen);
if (s->n_sen >= MAX_SENID)
E_FATAL("%s: #senones (%d) exceeds limit (%d)\n", file_name, s->n_sen, MAX_SENID);
if (s->n_feat <= 0)
E_FATAL("Bad #features: %d\n", s->n_feat);
if (n_cw <= 0)
E_FATAL("Bad #mixing-wts/senone: %d\n", n_cw);
/* Allocate sen2mgau map if not yet done so (i.e. no explicit mapping file given */
if (! s->sen2mgau) {
assert ((s->n_mgau == 0) || (s->n_mgau == 1));
s->sen2mgau = (uint32 *) ckd_calloc (s->n_sen, sizeof(int32));
if (s->n_mgau == 1) {
/* Semicontinuous mode; all senones map to single, shared gaussian: 0 */
for (i = 0; i < s->n_sen; i++)
s->sen2mgau[i] = 0;
} else {
/* Fully continuous mode; each senone maps to own parent gaussian */
s->n_mgau = s->n_sen;
for (i = 0; i < s->n_sen; i++)
s->sen2mgau[i] = i;
}
} else
assert (s->n_mgau != 0);
if (s->n_mgau >= MAX_MGAUID)
E_FATAL("%s: #gauden (%d) exceeds limit (%d)\n", file_name, s->n_mgau, MAX_MGAUID);
if (nval != s->n_sen * s->n_feat * n_cw) {
E_FATAL("%s: #float32 values(%d) doesn't match dimensions: %d x %d x %d\n",
file_name, nval, s->n_sen, s->n_feat, n_cw);
}
/*
* Compute #LSB bits to be dropped to represent mixwfloor with 8 bits.
* All PDF values will be truncated (in the LSB positions) by these many bits.
*/
if ((mixwfloor <= 0.0) || (mixwfloor >= 1.0))
E_FATAL("mixwfloor (%e) not in range (0, 1)\n", mixwfloor);
/* Allocate memory for s->mgau2sen and senone PDF data */
build_mgau2sen (s, n_cw);
/* Temporary structure to read in floats */
pdf = (float32 *) ckd_calloc (n_cw, sizeof(float32));
/* Read senone probs data, normalize, floor, convert to logs3, truncate to 8 bits */
n_err = 0;
for (i = 0; i < s->n_sen; i++) {
m = s->sen2mgau[i]; /* Parent mgau */
j = s->mgau2sen_idx[i]; /* Index of senone i within list of senones for mgau m */
fw = s->mgau2sen[m].feat_mixw;
for (f = 0; f < s->n_feat; f++) {
if (bio_fread((void *)pdf, sizeof(float32), n_cw, fp, byteswap, &chksum)
!= n_cw) {
E_FATAL("bio_fread(%s) (arraydata) failed\n", file_name);
}
/* Normalize and floor */
if (vector_sum_norm (pdf, n_cw) == 0.0)
n_err++;
vector_floor (pdf, n_cw, mixwfloor);
vector_sum_norm (pdf, n_cw);
/* Convert to logs3, truncate to 8 bits, and store in s->pdf */
for (c = 0; c < n_cw; c++) {
p = -logmath_log(logmath, pdf[c]);
printf ("%f %d\n", pdf[c], p);
fw[f].prob[j][c] = p;
}
}
}
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 mixture weights for %d senones: %d features x %d codewords\n",
s->n_sen, s->n_feat, n_cw);
return 0;
}
fsg_model_t *
fsg_model_read(FILE * fp, logmath_t * lmath, float32 lw)
{
fsg_model_t *fsg;
hash_table_t *vocab;
hash_iter_t *itor;
int32 lastwid;
char **wordptr;
char *lineptr;
char *fsgname;
int32 lineno;
int32 n, i, j;
int n_state, n_trans, n_null_trans;
glist_t nulls;
float32 p;
lineno = 0;
vocab = hash_table_new(32, FALSE);
wordptr = NULL;
lineptr = NULL;
nulls = NULL;
fsgname = NULL;
fsg = NULL;
/* Scan upto FSG_BEGIN header */
for (;;) {
n = nextline_str2words(fp, &lineno, &lineptr, &wordptr);
if (n < 0) {
E_ERROR("%s declaration missing\n", FSG_MODEL_BEGIN_DECL);
goto parse_error;
}
if ((strcmp(wordptr[0], FSG_MODEL_BEGIN_DECL) == 0)) {
if (n > 2) {
E_ERROR("Line[%d]: malformed FSG_BEGIN declaration\n",
lineno);
goto parse_error;
}
break;
}
}
/* Save FSG name, or it will get clobbered below :(.
* If name is missing, try the default.
*/
if (n == 2) {
fsgname = ckd_salloc(wordptr[1]);
}
else {
E_WARN("FSG name is missing\n");
fsgname = ckd_salloc("unknown");
}
/* Read #states */
n = nextline_str2words(fp, &lineno, &lineptr, &wordptr);
if ((n != 2)
|| ((strcmp(wordptr[0], FSG_MODEL_N_DECL) != 0)
&& (strcmp(wordptr[0], FSG_MODEL_NUM_STATES_DECL) != 0))
|| (sscanf(wordptr[1], "%d", &n_state) != 1)
|| (n_state <= 0)) {
E_ERROR
("Line[%d]: #states declaration line missing or malformed\n",
lineno);
goto parse_error;
}
/* Now create the FSG. */
fsg = fsg_model_init(fsgname, lmath, lw, n_state);
ckd_free(fsgname);
fsgname = NULL;
/* Read start state */
n = nextline_str2words(fp, &lineno, &lineptr, &wordptr);
if ((n != 2)
|| ((strcmp(wordptr[0], FSG_MODEL_S_DECL) != 0)
&& (strcmp(wordptr[0], FSG_MODEL_START_STATE_DECL) != 0))
|| (sscanf(wordptr[1], "%d", &(fsg->start_state)) != 1)
|| (fsg->start_state < 0)
|| (fsg->start_state >= fsg->n_state)) {
E_ERROR
("Line[%d]: start state declaration line missing or malformed\n",
lineno);
goto parse_error;
}
/* Read final state */
n = nextline_str2words(fp, &lineno, &lineptr, &wordptr);
if ((n != 2)
|| ((strcmp(wordptr[0], FSG_MODEL_F_DECL) != 0)
&& (strcmp(wordptr[0], FSG_MODEL_FINAL_STATE_DECL) != 0))
|| (sscanf(wordptr[1], "%d", &(fsg->final_state)) != 1)
|| (fsg->final_state < 0)
|| (fsg->final_state >= fsg->n_state)) {
E_ERROR
("Line[%d]: final state declaration line missing or malformed\n",
lineno);
goto parse_error;
}
/* Read transitions */
lastwid = 0;
n_trans = n_null_trans = 0;
for (;;) {
int32 wid, tprob;
n = nextline_str2words(fp, &lineno, &lineptr, &wordptr);
if (n <= 0) {
E_ERROR("Line[%d]: transition or FSG_END statement expected\n",
lineno);
goto parse_error;
}
if ((strcmp(wordptr[0], FSG_MODEL_END_DECL) == 0)) {
break;
}
if ((strcmp(wordptr[0], FSG_MODEL_T_DECL) == 0)
|| (strcmp(wordptr[0], FSG_MODEL_TRANSITION_DECL) == 0)) {
if (((n != 4) && (n != 5))
|| (sscanf(wordptr[1], "%d", &i) != 1)
|| (sscanf(wordptr[2], "%d", &j) != 1)
|| (i < 0) || (i >= fsg->n_state)
|| (j < 0) || (j >= fsg->n_state)) {
E_ERROR
("Line[%d]: transition spec malformed; Expecting: from-state to-state trans-prob [word]\n",
lineno);
goto parse_error;
}
p = atof_c(wordptr[3]);
if ((p <= 0.0) || (p > 1.0)) {
E_ERROR
("Line[%d]: transition spec malformed; Expecting float as transition probability\n",
lineno);
goto parse_error;
}
}
else {
E_ERROR("Line[%d]: transition or FSG_END statement expected\n",
lineno);
goto parse_error;
}
tprob = (int32) (logmath_log(lmath, p) * fsg->lw);
/* Add word to "dictionary". */
if (n > 4) {
if (hash_table_lookup_int32(vocab, wordptr[4], &wid) < 0) {
(void) hash_table_enter_int32(vocab,
ckd_salloc(wordptr[4]),
lastwid);
wid = lastwid;
++lastwid;
}
fsg_model_trans_add(fsg, i, j, tprob, wid);
++n_trans;
}
else {
if (fsg_model_null_trans_add(fsg, i, j, tprob) == 1) {
++n_null_trans;
nulls =
glist_add_ptr(nulls, fsg_model_null_trans(fsg, i, j));
}
}
}
E_INFO("FSG: %d states, %d unique words, %d transitions (%d null)\n",
fsg->n_state, hash_table_inuse(vocab), n_trans, n_null_trans);
/* Now create a string table from the "dictionary" */
fsg->n_word = hash_table_inuse(vocab);
fsg->n_word_alloc = fsg->n_word + 10; /* Pad it a bit. */
fsg->vocab = ckd_calloc(fsg->n_word_alloc, sizeof(*fsg->vocab));
for (itor = hash_table_iter(vocab); itor;
itor = hash_table_iter_next(itor)) {
char const *word = hash_entry_key(itor->ent);
int32 wid = (int32) (long) hash_entry_val(itor->ent);
fsg->vocab[wid] = (char *) word;
}
hash_table_free(vocab);
/* Do transitive closure on null transitions */
nulls = fsg_model_null_trans_closure(fsg, nulls);
glist_free(nulls);
ckd_free(lineptr);
ckd_free(wordptr);
return fsg;
parse_error:
for (itor = hash_table_iter(vocab); itor;
itor = hash_table_iter_next(itor))
ckd_free((char *) hash_entry_key(itor->ent));
glist_free(nulls);
hash_table_free(vocab);
ckd_free(fsgname);
ckd_free(lineptr);
ckd_free(wordptr);
fsg_model_free(fsg);
return NULL;
}
int
main(int argc, char *argv[])
{
logmath_t *lmath;
int32 rv;
lmath = logmath_init(1.0001, 0, 1);
TEST_ASSERT(lmath);
printf("log(1e-150) = %d\n", logmath_log(lmath, 1e-150));
TEST_EQUAL_LOG(logmath_log(lmath, 1e-150), -3454050);
printf("exp(log(1e-150)) = %e\n",logmath_exp(lmath, logmath_log(lmath, 1e-150)));
TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 1e-150)), 1e-150);
printf("log(1e-48) = %d\n", logmath_log(lmath, 1e-48));
printf("exp(log(1e-48)) = %e\n",logmath_exp(lmath, logmath_log(lmath, 1e-48)));
TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 1e-48)), 1e-48);
printf("log(42) = %d\n", logmath_log(lmath, 42));
TEST_EQUAL_LOG(logmath_log(lmath, 42), 37378);
printf("exp(log(42)) = %f\n",logmath_exp(lmath, logmath_log(lmath, 42)));
TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 42)), 42);
printf("log(1e-3 + 5e-3) = %d l+ %d = %d\n",
logmath_log(lmath, 1e-3),
logmath_log(lmath, 5e-3),
logmath_add(lmath, logmath_log(lmath, 1e-3),
logmath_log(lmath, 5e-3)));
printf("log(1e-3 + 5e-3) = %e + %e = %e\n",
logmath_exp(lmath, logmath_log(lmath, 1e-3)),
logmath_exp(lmath, logmath_log(lmath, 5e-3)),
logmath_exp(lmath, logmath_add(lmath, logmath_log(lmath, 1e-3),
logmath_log(lmath, 5e-3))));
TEST_EQUAL_LOG(logmath_add(lmath, logmath_log(lmath, 1e-48),
logmath_log(lmath, 5e-48)),
logmath_log(lmath, 6e-48));
TEST_EQUAL_LOG(logmath_add(lmath, logmath_log(lmath, 1e-48),
logmath_log(lmath, 42)),
logmath_log(lmath, 42));
rv = logmath_write(lmath, "tmp.logadd");
TEST_EQUAL(rv, 0);
logmath_free(lmath);
lmath = logmath_read("tmp.logadd");
TEST_ASSERT(lmath);
printf("log(1e-150) = %d\n", logmath_log(lmath, 1e-150));
TEST_EQUAL_LOG(logmath_log(lmath, 1e-150), -3454050);
printf("exp(log(1e-150)) = %e\n",logmath_exp(lmath, logmath_log(lmath, 1e-150)));
TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 1e-150)), 1e-150);
printf("log(1e-48) = %d\n", logmath_log(lmath, 1e-48));
printf("exp(log(1e-48)) = %e\n",logmath_exp(lmath, logmath_log(lmath, 1e-48)));
TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 1e-48)), 1e-48);
printf("log(42) = %d\n", logmath_log(lmath, 42));
TEST_EQUAL_LOG(logmath_log(lmath, 42), 37378);
printf("exp(log(42)) = %f\n",logmath_exp(lmath, logmath_log(lmath, 42)));
TEST_EQUAL_FLOAT(logmath_exp(lmath, logmath_log(lmath, 42)), 41.99);
printf("log(1e-3 + 5e-3) = %d l+ %d = %d\n",
logmath_log(lmath, 1e-3),
logmath_log(lmath, 5e-3),
logmath_add(lmath, logmath_log(lmath, 1e-3),
logmath_log(lmath, 5e-3)));
printf("log(1e-3 + 5e-3) = %e + %e = %e\n",
logmath_exp(lmath, logmath_log(lmath, 1e-3)),
logmath_exp(lmath, logmath_log(lmath, 5e-3)),
logmath_exp(lmath, logmath_add(lmath, logmath_log(lmath, 1e-3),
logmath_log(lmath, 5e-3))));
TEST_EQUAL_LOG(logmath_add(lmath, logmath_log(lmath, 1e-48),
logmath_log(lmath, 5e-48)),
logmath_log(lmath, 6e-48));
TEST_EQUAL_LOG(logmath_add(lmath, logmath_log(lmath, 1e-48),
logmath_log(lmath, 42)),
logmath_log(lmath, 42));
return 0;
}
static int32
senone_mixw_read(senone_t * s, char const *file_name, logmath_t *lmath)
{
char eofchk;
FILE *fp;
int32 byteswap, chksum_present;
uint32 chksum;
float32 *pdf;
int32 i, f, c, p, n_err;
char **argname, **argval;
E_INFO("Reading senone mixture weights: %s\n", file_name);
if ((fp = fopen(file_name, "rb")) == NULL)
E_FATAL_SYSTEM("Failed to open mixture weights file '%s' for reading", file_name);
/* Read header, including argument-value info and 32-bit byteorder magic */
if (bio_readhdr(fp, &argname, &argval, &byteswap) < 0)
E_FATAL("Failed to read header from file '%s'\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], MIXW_PARAM_VERSION) != 0)
E_WARN("Version mismatch(%s): %s, expecting %s\n",
file_name, argval[i], MIXW_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 #senones, #features, #codewords, arraysize */
if ((bio_fread(&(s->n_sen), sizeof(int32), 1, fp, byteswap, &chksum) !=
1)
||
(bio_fread(&(s->n_feat), sizeof(int32), 1, fp, byteswap, &chksum)
!= 1)
|| (bio_fread(&(s->n_cw), 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 (i != s->n_sen * s->n_feat * s->n_cw) {
E_FATAL
("%s: #float32s(%d) doesn't match dimensions: %d x %d x %d\n",
file_name, i, s->n_sen, s->n_feat, s->n_cw);
}
/*
* Compute #LSB bits to be dropped to represent mixwfloor with 8 bits.
* All PDF values will be truncated (in the LSB positions) by these many bits.
*/
if ((s->mixwfloor <= 0.0) || (s->mixwfloor >= 1.0))
E_FATAL("mixwfloor (%e) not in range (0, 1)\n", s->mixwfloor);
/* Use a fixed shift for compatibility with everything else. */
E_INFO("Truncating senone logs3(pdf) values by %d bits\n", SENSCR_SHIFT);
/*
* Allocate memory for senone PDF data. Organize normally or transposed depending on
* s->n_gauden.
*/
if (s->n_gauden > 1) {
E_INFO("Not transposing mixture weights in memory\n");
s->pdf =
(senprob_t ***) ckd_calloc_3d(s->n_sen, s->n_feat, s->n_cw,
sizeof(senprob_t));
}
else {
E_INFO("Transposing mixture weights in memory\n");
s->pdf =
(senprob_t ***) ckd_calloc_3d(s->n_feat, s->n_cw, s->n_sen,
sizeof(senprob_t));
}
/* Temporary structure to read in floats */
pdf = (float32 *) ckd_calloc(s->n_cw, sizeof(float32));
/* Read senone probs data, normalize, floor, convert to logs3, truncate to 8 bits */
n_err = 0;
for (i = 0; i < s->n_sen; i++) {
for (f = 0; f < s->n_feat; f++) {
if (bio_fread
((void *) pdf, sizeof(float32), s->n_cw, fp, byteswap,
&chksum)
!= s->n_cw) {
E_FATAL("bio_fread(%s) (arraydata) failed\n", file_name);
}
/* Normalize and floor */
if (vector_sum_norm(pdf, s->n_cw) <= 0.0)
n_err++;
vector_floor(pdf, s->n_cw, s->mixwfloor);
vector_sum_norm(pdf, s->n_cw);
/* Convert to logs3, truncate to 8 bits, and store in s->pdf */
for (c = 0; c < s->n_cw; c++) {
p = -(logmath_log(lmath, pdf[c]));
p += (1 << (SENSCR_SHIFT - 1)) - 1; /* Rounding before truncation */
if (s->n_gauden > 1)
s->pdf[i][f][c] =
(p < (255 << SENSCR_SHIFT)) ? (p >> SENSCR_SHIFT) : 255;
else
s->pdf[f][c][i] =
(p < (255 << SENSCR_SHIFT)) ? (p >> SENSCR_SHIFT) : 255;
}
}
}
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;
}
void
run_tests(logmath_t *lmath, ngram_model_t *model)
{
int32 rv, i;
TEST_ASSERT(model);
TEST_EQUAL(ngram_wid(model, "scylla"), 285);
TEST_EQUAL(strcmp(ngram_word(model, 285), "scylla"), 0);
rv = ngram_model_read_classdef(model, LMDIR "/100.probdef");
TEST_EQUAL(rv, 0);
/* Verify that class word IDs remain the same. */
TEST_EQUAL(ngram_wid(model, "scylla"), 285);
TEST_EQUAL(strcmp(ngram_word(model, 285), "scylla"), 0);
/* Verify in-class word IDs. */
TEST_EQUAL(ngram_wid(model, "scylla:scylla"), 0x80000000 | 400);
/* Verify in-class and out-class unigram scores. */
TEST_EQUAL_LOG(ngram_score(model, "scylla:scylla", NULL),
logmath_log10_to_log(lmath, -2.7884) + logmath_log(lmath, 0.4));
TEST_EQUAL_LOG(ngram_score(model, "scooby:scylla", NULL),
logmath_log10_to_log(lmath, -2.7884) + logmath_log(lmath, 0.1));
TEST_EQUAL_LOG(ngram_score(model, "scylla", NULL),
logmath_log10_to_log(lmath, -2.7884));
TEST_EQUAL_LOG(ngram_score(model, "oh:zero", NULL),
logmath_log10_to_log(lmath, -1.9038) + logmath_log(lmath, 0.7));
TEST_EQUAL_LOG(ngram_score(model, "zero", NULL),
logmath_log10_to_log(lmath, -1.9038));
/* Verify class bigram scores. */
TEST_EQUAL_LOG(ngram_score(model, "scylla", "on", NULL),
logmath_log10_to_log(lmath, -1.2642));
TEST_EQUAL_LOG(ngram_score(model, "scylla:scylla", "on", NULL),
logmath_log10_to_log(lmath, -1.2642) + logmath_log(lmath, 0.4));
TEST_EQUAL_LOG(ngram_score(model, "apparently", "scylla", NULL),
logmath_log10_to_log(lmath, -0.5172));
TEST_EQUAL_LOG(ngram_score(model, "apparently", "karybdis:scylla", NULL),
logmath_log10_to_log(lmath, -0.5172));
TEST_EQUAL_LOG(ngram_score(model, "apparently", "scooby:scylla", NULL),
logmath_log10_to_log(lmath, -0.5172));
/* Verify class trigram scores. */
TEST_EQUAL_LOG(ngram_score(model, "zero", "be", "will", NULL),
logmath_log10_to_log(lmath, -0.5725));
TEST_EQUAL_LOG(ngram_score(model, "oh:zero", "be", "will", NULL),
logmath_log10_to_log(lmath, -0.5725) + logmath_log(lmath, 0.7));
TEST_EQUAL_LOG(ngram_score(model, "should", "variance", "zero", NULL),
logmath_log10_to_log(lmath, -0.9404));
TEST_EQUAL_LOG(ngram_score(model, "should", "variance", "zero:zero", NULL),
logmath_log10_to_log(lmath, -0.9404));
/* Add words to classes. */
rv = ngram_model_add_class_word(model, "scylla", "scrappy:scylla", 1.0);
TEST_ASSERT(rv >= 0);
TEST_EQUAL(ngram_wid(model, "scrappy:scylla"), 0x80000196);
TEST_EQUAL_LOG(ngram_score(model, "scrappy:scylla", NULL),
logmath_log10_to_log(lmath, -2.7884) + logmath_log(lmath, 0.2));
printf("scrappy:scylla %08x %d %f\n",
ngram_wid(model, "scrappy:scylla"),
ngram_score(model, "scrappy:scylla", NULL),
logmath_exp(lmath, ngram_score(model, "scrappy:scylla", NULL)));
/* Add a lot of words to a class. */
for (i = 0; i < 129; ++i) {
char word[32];
sprintf(word, "%d:scylla", i);
rv = ngram_model_add_class_word(model, "scylla", word, 1.0);
printf("%s %08x %d %f\n", word,
ngram_wid(model, word),
ngram_score(model, word, NULL),
logmath_exp(lmath, ngram_score(model, word, NULL)));
TEST_ASSERT(rv >= 0);
TEST_EQUAL(ngram_wid(model, word), 0x80000197 + i);
}
/* Add a new class. */
{
const char *words[] = { "blatz:foobie", "hurf:foobie" };
float32 weights[] = { 0.6, 0.4 };
int32 foobie_prob;
rv = ngram_model_add_class(model, "[foobie]", 1.0,
words, weights, 2);
TEST_ASSERT(rv >= 0);
foobie_prob = ngram_score(model, "[foobie]", NULL);
TEST_EQUAL_LOG(ngram_score(model, "blatz:foobie", NULL),
foobie_prob + logmath_log(lmath, 0.6));
TEST_EQUAL_LOG(ngram_score(model, "hurf:foobie", NULL),
foobie_prob + logmath_log(lmath, 0.4));
}
}
