acmod_t *
acmod_copy(acmod_t *other)
{
acmod_t *acmod;
acmod = ckd_calloc(1, sizeof(*acmod));
acmod->refcount = 1;
acmod->config = cmd_ln_retain(other->config);
acmod->lmath = logmath_retain(other->lmath);
acmod->mdef = bin_mdef_retain(other->mdef);
acmod->tmat = tmat_retain(other->tmat);
acmod->mgau = ps_mgau_copy(other->mgau);
acmod->fb = featbuf_retain(other->fb);
acmod->fcb = other->fcb; /* Implicitly retained with fb, I think */
/* 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(acmod->config, "-compallsen");
acmod->feat_buf = feat_array_alloc(acmod->fcb, 1);
return acmod;
}
acmod_t *
acmod_init(cmd_ln_t *config, logmath_t *lmath, featbuf_t *fb)
{
acmod_t *acmod;
acmod = ckd_calloc(1, sizeof(*acmod));
acmod->refcount = 1;
acmod->config = cmd_ln_retain(config);
acmod->lmath = logmath_retain(lmath);
acmod->fb = featbuf_retain(fb);
acmod->fcb = featbuf_get_fcb(acmod->fb);
/* Load acoustic model parameters. */
if (acmod_init_am(acmod) < 0)
goto error_out;
/* 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");
acmod->feat_buf = feat_array_alloc(acmod->fcb, 1);
return acmod;
error_out:
acmod_free(acmod);
return NULL;
}
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;
}
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);
}
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;
}
int
test_decode(ps_decoder_t *ps)
{
FILE *rawfh;
int16 buf[2048];
size_t nread;
int16 const *bptr;
int nfr;
ps_lattice_t *dag;
acmod_t *acmod;
ngram_search_t *ngs;
int i, j;
ps_latlink_t *link;
ps_latnode_t *node;
latlink_list_t *x;
int32 norm, post;
ngs = (ngram_search_t *)ps->search;
acmod = ps->acmod;
/* Decode stuff and build a DAG. */
TEST_ASSERT(rawfh = fopen(DATADIR "/goforward.raw", "rb"));
TEST_EQUAL(0, acmod_start_utt(acmod));
ngram_fwdtree_start(ngs);
while (!feof(rawfh)) {
nread = fread(buf, sizeof(*buf), 2048, rawfh);
bptr = buf;
while ((nfr = acmod_process_raw(acmod, &bptr, &nread, FALSE)) > 0) {
while (acmod->n_feat_frame > 0) {
ngram_fwdtree_search(ngs, acmod->output_frame);
acmod_advance(acmod);
}
}
}
ngram_fwdtree_finish(ngs);
printf("FWDTREE: %s\n",
ngram_search_bp_hyp(ngs, ngram_search_find_exit(ngs, -1, NULL, NULL)));
TEST_ASSERT(acmod_end_utt(acmod) >= 0);
fclose(rawfh);
dag = ngram_search_lattice(ps->search);
if (dag == NULL) {
E_ERROR("Failed to build DAG!\n");
return -1;
}
/* Write lattice to disk. */
TEST_EQUAL(0, ps_lattice_write(dag, "test_posterior.lat"));
/* Do a bunch of checks on the DAG generation and traversal code: */
/* Verify that forward and backward iteration give the same number of edges. */
i = j = 0;
for (link = ps_lattice_traverse_edges(dag, NULL, NULL);
link; link = ps_lattice_traverse_next(dag, NULL)) {
++i;
}
for (link = ps_lattice_reverse_edges(dag, NULL, NULL);
link; link = ps_lattice_reverse_next(dag, NULL)) {
++j;
}
printf("%d forward edges, %d reverse edges\n", i, j);
TEST_EQUAL(i,j);
/* Verify that the same links are reachable via entries and exits. */
for (node = dag->nodes; node; node = node->next) {
for (x = node->exits; x; x = x->next)
x->link->alpha = -42;
}
for (node = dag->nodes; node; node = node->next) {
for (x = node->entries; x; x = x->next)
TEST_EQUAL(x->link->alpha, -42);
}
/* Verify that forward iteration is properly ordered. */
for (link = ps_lattice_traverse_edges(dag, NULL, NULL);
link; link = ps_lattice_traverse_next(dag, NULL)) {
link->alpha = 0;
for (x = link->from->entries; x; x = x->next) {
TEST_EQUAL(x->link->alpha, 0);
}
}
/* Verify that backward iteration is properly ordered. */
for (node = dag->nodes; node; node = node->next) {
for (x = node->exits; x; x = x->next)
x->link->alpha = -42;
}
for (link = ps_lattice_reverse_edges(dag, NULL, NULL);
link; link = ps_lattice_reverse_next(dag, NULL)) {
link->alpha = 0;
for (x = link->to->exits; x; x = x->next) {
TEST_EQUAL(x->link->alpha, 0);
}
}
/* Find and print best path. */
link = ps_lattice_bestpath(dag, ngs->lmset, 1.0, 1.0/20.0);
printf("BESTPATH: %s\n", ps_lattice_hyp(dag, link));
/* Calculate betas. */
post = ps_lattice_posterior(dag, ngs->lmset, 1.0/20.0);
printf("Best path score: %d\n",
link->path_scr + dag->final_node_ascr);
printf("P(S|O) = %d\n", post);
/* Verify that sum of final alphas and initial alphas+betas is
* sufficiently similar. */
norm = logmath_get_zero(acmod->lmath);
for (x = dag->start->exits; x; x = x->next)
norm = logmath_add(acmod->lmath, norm, x->link->beta + x->link->alpha);
E_INFO("Sum of final alphas+betas = %d\n", dag->norm);
E_INFO("Sum of initial alphas+betas = %d\n", norm);
TEST_EQUAL_LOG(dag->norm, norm);
/* Print posterior probabilities for each link in best path. */
while (link) {
printf("P(%s,%d) = %d = %f\n",
dict_wordstr(ps->search->dict, link->from->wid),
link->ef,
link->alpha + link->beta - dag->norm,
logmath_exp(acmod->lmath, link->alpha + link->beta - dag->norm));
link = link->best_prev;
}
return 0;
}
