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 int
acmod_read_senfh_header(acmod_t *acmod)
{
char **name, **val;
int32 swap;
int i;
if (bio_readhdr(acmod->insenfh, &name, &val, &swap) < 0)
goto error_out;
for (i = 0; name[i] != NULL; ++i) {
if (!strcmp(name[i], "n_sen")) {
if (atoi(val[i]) != bin_mdef_n_sen(acmod->mdef)) {
E_ERROR("Number of senones in senone file (%d) does not "
"match mdef (%d)\n", atoi(val[i]),
bin_mdef_n_sen(acmod->mdef));
goto error_out;
}
}
if (!strcmp(name[i], "logbase")) {
if (fabs(atof_c(val[i]) - logmath_get_base(acmod->lmath)) > 0.001) {
E_ERROR("Logbase in senone file (%f) does not match acmod "
"(%f)\n", atof_c(val[i]),
logmath_get_base(acmod->lmath));
goto error_out;
}
}
}
acmod->insen_swap = swap;
bio_hdrarg_free(name, val);
return 0;
error_out:
bio_hdrarg_free(name, val);
return -1;
}
static void
evaluate_string(ngram_model_t *lm, logmath_t *lmath, const char *text)
{
char *textfoo;
char **words;
int32 n, ch, noovs, nccs, lscr;
/* Split it into an array of strings. */
textfoo = ckd_salloc(text);
n = str2words(textfoo, NULL, 0);
if (n < 0)
E_FATAL("str2words(textfoo, NULL, 0) = %d, should not happen\n", n);
if (n == 0) /* Do nothing! */
return;
words = ckd_calloc(n, sizeof(*words));
str2words(textfoo, words, n);
ch = calc_entropy(lm, words, n, &nccs, &noovs, &lscr);
printf("input: %s\n", text);
printf("cross-entropy: %f bits\n",
ch * log(logmath_get_base(lmath)) / log(2));
/* Calculate perplexity pplx = exp CH */
printf("perplexity: %f\n", logmath_exp(lmath, ch));
printf("lm score: %d\n", lscr);
/* Report OOVs and CCs */
printf("%d words evaluated\n", n);
printf("%d OOVs, %d context cues removed\n",
noovs, nccs);
ckd_free(textfoo);
ckd_free(words);
}
static void
evaluate_file(ngram_model_t *lm, logmath_t *lmath, const char *lsnfn)
{
FILE *fh;
lineiter_t *litor;
int32 nccs, noovs, nwords, lscr;
float64 ch, log_to_log2;;
if ((fh = fopen(lsnfn, "r")) == NULL)
E_FATAL_SYSTEM("failed to open transcript file %s", lsnfn);
/* We have to keep ch in floating-point to avoid overflows, so
* we might as well use log2. */
log_to_log2 = log(logmath_get_base(lmath)) / log(2);
lscr = nccs = noovs = nwords = 0;
ch = 0.0;
for (litor = lineiter_start(fh); litor; litor = lineiter_next(litor)) {
char **words;
int32 n, tmp_ch, tmp_noovs, tmp_nccs, tmp_lscr;
n = str2words(litor->buf, NULL, 0);
if (n < 0)
E_FATAL("str2words(line, NULL, 0) = %d, should not happen\n", n);
if (n == 0) /* Do nothing! */
continue;
words = ckd_calloc(n, sizeof(*words));
str2words(litor->buf, words, n);
/* Remove any utterance ID (FIXME: has to be a single "word") */
if (words[n-1][0] == '('
&& words[n-1][strlen(words[n-1])-1] == ')')
n = n - 1;
tmp_ch = calc_entropy(lm, words, n, &tmp_nccs,
&tmp_noovs, &tmp_lscr);
ch += (float64) tmp_ch * (n - tmp_nccs - tmp_noovs) * log_to_log2;
nccs += tmp_nccs;
noovs += tmp_noovs;
lscr += tmp_lscr;
nwords += n;
ckd_free(words);
}
ch /= (nwords - nccs - noovs);
printf("cross-entropy: %f bits\n", ch);
/* Calculate perplexity pplx = exp CH */
printf("perplexity: %f\n", pow(2.0, ch));
printf("lm score: %d\n", lscr);
/* Report OOVs and CCs */
printf("%d words evaluated\n", nwords);
printf("%d OOVs (%.2f%%), %d context cues removed\n",
noovs, (double)noovs / nwords * 100, nccs);
}
int32
subvq_mgau_eval(mgau_model_t * g, subvq_t * vq, int32 m, int32 n,
int32 * active)
{
mgau_t *mgau;
int32 *map;
int32 i, v, sv_id;
int32 c;
int32 *vqdist;
int32 score;
int32 last_active;
float64 f;
f = 1.0 / log(logmath_get_base(g->logmath));
vqdist = vq->vqdist[0];
score = S3_LOGPROB_ZERO;
mgau = &(g->mgau[m]);
map = vq->map[m][0];
if (!active) {
for (i = 0; i < n; i++) {
v = 0;
for (sv_id = 0; sv_id < vq->n_sv; sv_id++) {
v += vqdist[*(map++)];
}
score = logmath_add(g->logmath, score, v + mgau->mixw[i]);
}
}
else {
last_active = 0;
for (i = 0; active[i] >= 0; i++) {
c = active[i];
}
for (i = 0; active[i] >= 0; i++) {
c = active[i];
map += (c - last_active) * vq->n_sv;
v = 0;
for (sv_id = 0; sv_id < vq->n_sv; sv_id++) {
v += vqdist[*(map++)];
}
last_active = c + 1;
score = logmath_add(g->logmath, score, v + mgau->mixw[i]);
}
}
if (score == S3_LOGPROB_ZERO) {
E_INFO("Warning!! Score is S3_LOGPROB_ZERO\n");
}
return score;
}
int
acmod_write_senfh_header(acmod_t *acmod, FILE *logfh)
{
char nsenstr[64], logbasestr[64];
sprintf(nsenstr, "%d", bin_mdef_n_sen(acmod->mdef));
sprintf(logbasestr, "%f", logmath_get_base(acmod->lmath));
return bio_writehdr(logfh,
"version", "0.1",
"mdef_file", cmd_ln_str_r(acmod->config, "-mdef"),
"n_sen", nsenstr,
"logbase", logbasestr, NULL);
}
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"));
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_init_defaults(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,
ckd_salloc(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, ps->acmod->lmath)) == NULL)
return -1;
if ((ps->d2p = dict2pid_build(ps->acmod->mdef, ps->dict)) == NULL)
return -1;
lw = cmd_ln_float32_r(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(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(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(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))
return -1;
ps_set_search(ps, PS_DEFAULT_SEARCH);
}
/* Or load a JSGF grammar */
if ((path = cmd_ln_str_r(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_free(lm);
ngram_model_set_iter_free(lmset_it);
return -1;
}
ngram_model_free(lm);
}
name = cmd_ln_str_r(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;
}
int
ps_reinit(ps_decoder_t *ps, cmd_ln_t *config)
{
char const *lmfile, *lmctl = NULL;
if (config && config != ps->config) {
cmd_ln_free_r(ps->config);
ps->config = config;
}
#ifndef _WIN32_WCE
/* Set up logging. */
if (cmd_ln_str_r(ps->config, "-logfn"))
err_set_logfile(cmd_ln_str_r(ps->config, "-logfn"));
#endif
err_set_debug_level(cmd_ln_int32_r(ps->config, "-debug"));
ps->mfclogdir = cmd_ln_str_r(ps->config, "-mfclogdir");
ps->rawlogdir = cmd_ln_str_r(ps->config, "-rawlogdir");
/* Fill in some default arguments. */
ps_init_defaults(ps);
/* Free old searches (do this before other reinit) */
ps_free_searches(ps);
/* 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;
/* 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;
/* Make the acmod's feature buffer growable if we are doing two-pass search. */
if (cmd_ln_boolean_r(ps->config, "-fwdflat")
&& cmd_ln_boolean_r(ps->config, "-fwdtree"))
acmod_set_grow(ps->acmod, TRUE);
if ((ps->pl_window = cmd_ln_int32_r(ps->config, "-pl_window"))) {
/* 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;
ps->searches = glist_add_ptr(ps->searches, 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;
/* Determine whether we are starting out in FSG or N-Gram search mode. */
if (cmd_ln_str_r(ps->config, "-fsg") || cmd_ln_str_r(ps->config, "-jsgf")) {
ps_search_t *fsgs;
if ((ps->d2p = dict2pid_build(ps->acmod->mdef, ps->dict)) == NULL)
return -1;
if ((fsgs = fsg_search_init(ps->config, ps->acmod, ps->dict, ps->d2p)) == NULL)
return -1;
fsgs->pls = ps->phone_loop;
ps->searches = glist_add_ptr(ps->searches, fsgs);
ps->search = fsgs;
}
else if ((lmfile = cmd_ln_str_r(ps->config, "-lm"))
|| (lmctl = cmd_ln_str_r(ps->config, "-lmctl"))) {
ps_search_t *ngs;
if ((ps->d2p = dict2pid_build(ps->acmod->mdef, ps->dict)) == NULL)
return -1;
if ((ngs = ngram_search_init(ps->config, ps->acmod, ps->dict, ps->d2p)) == NULL)
return -1;
ngs->pls = ps->phone_loop;
ps->searches = glist_add_ptr(ps->searches, ngs);
ps->search = ngs;
}
/* Otherwise, we will initialize the search whenever the user
* decides to load an FSG or a language model. */
else {
if ((ps->d2p = dict2pid_build(ps->acmod->mdef, ps->dict)) == NULL)
return -1;
}
/* Initialize performance timer. */
ps->perf.name = "decode";
ptmr_init(&ps->perf);
return 0;
}
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;
}