ps_search_t *
state_align_search_init(const char *name,
cmd_ln_t *config,
acmod_t *acmod,
ps_alignment_t *al)
{
state_align_search_t *sas;
ps_alignment_iter_t *itor;
hmm_t *hmm;
sas = ckd_calloc(1, sizeof(*sas));
ps_search_init(ps_search_base(sas), &state_align_search_funcs,
PS_SEARCH_TYPE_STATE_ALIGN, name,
config, acmod, al->d2p->dict, al->d2p);
sas->hmmctx = hmm_context_init(bin_mdef_n_emit_state(acmod->mdef),
acmod->tmat->tp, NULL, acmod->mdef->sseq);
if (sas->hmmctx == NULL) {
ckd_free(sas);
return NULL;
}
sas->al = al;
/* Generate HMM vector from phone level of alignment. */
sas->n_phones = ps_alignment_n_phones(al);
sas->n_emit_state = ps_alignment_n_states(al);
sas->hmms = ckd_calloc(sas->n_phones, sizeof(*sas->hmms));
for (hmm = sas->hmms, itor = ps_alignment_phones(al); itor;
++hmm, itor = ps_alignment_iter_next(itor)) {
ps_alignment_entry_t *ent = ps_alignment_iter_get(itor);
hmm_init(sas->hmmctx, hmm, FALSE,
ent->id.pid.ssid, ent->id.pid.tmatid);
}
return ps_search_base(sas);
}
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 penalty storage */
pls->n_phones = bin_mdef_n_ciphone(acmod->mdef);
pls->window = cmd_ln_int32_r(config, "-pl_window");
if (pls->penalties)
ckd_free(pls->penalties);
pls->penalties = (int32 *)ckd_calloc(pls->n_phones, sizeof(*pls->penalties));
if (pls->pen_buf)
ckd_free_2d(pls->pen_buf);
pls->pen_buf = (int32 **)ckd_calloc_2d(pls->window, pls->n_phones, sizeof(**pls->pen_buf));
/* Initialize phone HMMs. */
if (pls->hmms) {
for (i = 0; i < pls->n_phones; ++i)
hmm_deinit((hmm_t *)&pls->hmms[i]);
ckd_free(pls->hmms);
}
pls->hmms = (hmm_t *)ckd_calloc(pls->n_phones, sizeof(*pls->hmms));
for (i = 0; i < pls->n_phones; ++i) {
hmm_init(pls->hmmctx, (hmm_t *)&pls->hmms[i],
FALSE,
bin_mdef_pid2ssid(acmod->mdef, i),
bin_mdef_pid2tmatid(acmod->mdef, i));
}
pls->penalty_weight = cmd_ln_float64_r(config, "-pl_weight");
pls->beam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-pl_beam")) >> SENSCR_SHIFT;
pls->pbeam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-pl_pbeam")) >> SENSCR_SHIFT;
pls->pip = logmath_log(acmod->lmath, cmd_ln_float32_r(config, "-pl_pip")) >> SENSCR_SHIFT;
E_INFO("State beam %d Phone exit beam %d Insertion penalty %d\n",
pls->beam, pls->pbeam, pls->pip);
return 0;
}
void
dict2pid_dump(FILE * fp, dict2pid_t * d2p)
{
int32 w, p, pronlen;
int32 i, j, b, l, r;
bin_mdef_t *mdef = d2p->mdef;
dict_t *dict = d2p->dict;
fprintf(fp, "# INTERNAL (wd comssid ssid ssid ... ssid comssid)\n");
for (w = 0; w < dict_size(dict); w++) {
fprintf(fp, "%30s ", dict_wordstr(dict, w));
pronlen = dict_pronlen(dict, w);
for (p = 0; p < pronlen; p++)
fprintf(fp, " %5d", dict2pid_internal(d2p, w, p));
fprintf(fp, "\n");
}
fprintf(fp, "#\n");
fprintf(fp, "# LDIPH_LC (b r l ssid)\n");
for (b = 0; b < bin_mdef_n_ciphone(mdef); b++) {
for (r = 0; r < bin_mdef_n_ciphone(mdef); r++) {
for (l = 0; l < bin_mdef_n_ciphone(mdef); l++) {
if (IS_S3SSID(d2p->ldiph_lc[b][r][l]))
fprintf(fp, "%6s %6s %6s %5d\n", bin_mdef_ciphone_str(mdef, (s3cipid_t) b), bin_mdef_ciphone_str(mdef, (s3cipid_t) r), bin_mdef_ciphone_str(mdef, (s3cipid_t) l), d2p->ldiph_lc[b][r][l]); /* RAH, ldiph_lc is returning an int32, %d expects an int16 */
}
}
}
fprintf(fp, "#\n");
fprintf(fp, "# SSEQ %d (senid senid ...)\n", mdef->n_sseq);
for (i = 0; i < mdef->n_sseq; i++) {
fprintf(fp, "%5d ", i);
for (j = 0; j < bin_mdef_n_emit_state(mdef); j++)
fprintf(fp, " %5d", mdef->sseq[i][j]);
fprintf(fp, "\n");
}
fprintf(fp, "#\n");
fprintf(fp, "# END\n");
fflush(fp);
}
int
main(int argc, char *argv[])
{
dict_t *dict;
dict2pid_t *d2p;
bin_mdef_t *mdef;
glextree_t *tree;
hmm_context_t *ctx;
logmath_t *lmath;
tmat_t *tmat;
int i;
TEST_ASSERT(mdef = bin_mdef_read(NULL, MODELDIR "/hmm/en_US/hub4wsj_sc_8k/mdef"));
TEST_ASSERT(dict = dict_init(cmd_ln_init(NULL, NULL, FALSE,
"-dict", DATADIR "/turtle.dic",
"-dictcase", "no", NULL),
mdef));
TEST_ASSERT(d2p = dict2pid_build(mdef, dict));
lmath = logmath_init(1.0001, 0, TRUE);
tmat = tmat_init(MODELDIR "/hmm/en_US/hub4wsj_sc_8k/transition_matrices",
lmath, 1e-5, TRUE);
ctx = hmm_context_init(bin_mdef_n_emit_state(mdef), tmat->tp, NULL, mdef->sseq);
TEST_ASSERT(tree = glextree_build(ctx, dict, d2p, NULL, NULL));
/* Check that a path exists for all dictionary words. */
for (i = 0; i < dict_size(dict); ++i)
TEST_ASSERT(glextree_has_word(tree, i));
dict_free(dict);
dict2pid_free(d2p);
bin_mdef_free(mdef);
tmat_free(tmat);
hmm_context_free(ctx);
glextree_free(tree);
return 0;
}
/* FIXME: Somewhat the same as the above function, needs refactoring */
int
ps_alignment_populate_ci(ps_alignment_t *al)
{
dict2pid_t *d2p;
dict_t *dict;
bin_mdef_t *mdef;
int i;
/* Clear phone and state sequences. */
ps_alignment_vector_empty(&al->sseq);
ps_alignment_vector_empty(&al->state);
/* For each word, expand to phones/senone sequences. */
d2p = al->d2p;
dict = d2p->dict;
mdef = d2p->mdef;
for (i = 0; i < al->word.n_ent; ++i) {
ps_alignment_entry_t *went = al->word.seq + i;
ps_alignment_entry_t *sent;
int wid = went->id.wid;
int len = dict_pronlen(dict, wid);
int j;
for (j = 0; j < len; ++j) {
if ((sent = ps_alignment_vector_grow_one(&al->sseq)) == NULL) {
E_ERROR("Failed to add phone entry!\n");
return -1;
}
sent->id.pid.cipid = dict_pron(dict, wid, j);
sent->id.pid.tmatid = bin_mdef_pid2tmatid(mdef, sent->id.pid.cipid);
sent->id.pid.ssid = bin_mdef_pid2ssid(mdef, sent->id.pid.cipid);
oe_assert(sent->id.pid.ssid != BAD_SSID);
sent->start = went->start;
sent->duration = went->duration;
sent->parent = i;
}
}
/* For each senone sequence, expand to senones. (we could do this
* nested above but this makes it more clear and easier to
* refactor) */
for (i = 0; i < al->sseq.n_ent; ++i) {
ps_alignment_entry_t *pent = al->sseq.seq + i;
ps_alignment_entry_t *sent;
int j;
for (j = 0; j < bin_mdef_n_emit_state(mdef); ++j) {
if ((sent = ps_alignment_vector_grow_one(&al->state)) == NULL) {
E_ERROR("Failed to add state entry!\n");
return -1;
}
sent->id.senid = bin_mdef_sseq2sen(mdef, pent->id.pid.ssid, j);
oe_assert(sent->id.senid != BAD_SENID);
sent->start = pent->start;
sent->duration = pent->duration;
sent->parent = i;
if (j == 0)
pent->child = (uint16)(sent - al->state.seq);
}
}
return 0;
}
int
ps_alignment_populate(ps_alignment_t *al)
{
dict2pid_t *d2p;
dict_t *dict;
bin_mdef_t *mdef;
int i, lc;
/* Clear phone and state sequences. */
ps_alignment_vector_empty(&al->sseq);
ps_alignment_vector_empty(&al->state);
/* For each word, expand to phones/senone sequences. */
d2p = al->d2p;
dict = d2p->dict;
mdef = d2p->mdef;
lc = bin_mdef_silphone(mdef);
for (i = 0; i < al->word.n_ent; ++i) {
ps_alignment_entry_t *went = al->word.seq + i;
ps_alignment_entry_t *sent;
int wid = went->id.wid;
int len = dict_pronlen(dict, wid);
int j, rc;
if (i < al->word.n_ent - 1)
rc = dict_first_phone(dict, al->word.seq[i+1].id.wid);
else
rc = bin_mdef_silphone(mdef);
/* First phone. */
if ((sent = ps_alignment_vector_grow_one(&al->sseq)) == NULL) {
E_ERROR("Failed to add phone entry!\n");
return -1;
}
sent->id.pid.cipid = dict_first_phone(dict, wid);
sent->id.pid.tmatid = bin_mdef_pid2tmatid(mdef, sent->id.pid.cipid);
sent->start = went->start;
sent->duration = went->duration;
sent->parent = i;
went->child = (uint16)(sent - al->sseq.seq);
if (len == 1)
sent->id.pid.ssid
= dict2pid_lrdiph_rc(d2p, sent->id.pid.cipid, lc, rc);
else
sent->id.pid.ssid
= dict2pid_ldiph_lc(d2p, sent->id.pid.cipid,
dict_second_phone(dict, wid), lc);
oe_assert(sent->id.pid.ssid != BAD_SSID);
/* Internal phones. */
for (j = 1; j < len - 1; ++j) {
if ((sent = ps_alignment_vector_grow_one(&al->sseq)) == NULL) {
E_ERROR("Failed to add phone entry!\n");
return -1;
}
sent->id.pid.cipid = dict_pron(dict, wid, j);
sent->id.pid.tmatid = bin_mdef_pid2tmatid(mdef, sent->id.pid.cipid);
sent->id.pid.ssid = dict2pid_internal(d2p, wid, j);
oe_assert(sent->id.pid.ssid != BAD_SSID);
sent->start = went->start;
sent->duration = went->duration;
sent->parent = i;
}
/* Last phone. */
if (j < len) {
xwdssid_t *rssid;
oe_assert(j == len - 1);
if ((sent = ps_alignment_vector_grow_one(&al->sseq)) == NULL) {
E_ERROR("Failed to add phone entry!\n");
return -1;
}
sent->id.pid.cipid = dict_last_phone(dict, wid);
sent->id.pid.tmatid = bin_mdef_pid2tmatid(mdef, sent->id.pid.cipid);
rssid = dict2pid_rssid(d2p, sent->id.pid.cipid,
dict_second_last_phone(dict, wid));
sent->id.pid.ssid = rssid->ssid[rssid->cimap[rc]];
oe_assert(sent->id.pid.ssid != BAD_SSID);
sent->start = went->start;
sent->duration = went->duration;
sent->parent = i;
}
/* Update lc. Could just use sent->id.pid.cipid here but that
* seems needlessly obscure. */
lc = dict_last_phone(dict, wid);
}
/* For each senone sequence, expand to senones. (we could do this
* nested above but this makes it more clear and easier to
* refactor) */
for (i = 0; i < al->sseq.n_ent; ++i) {
ps_alignment_entry_t *pent = al->sseq.seq + i;
ps_alignment_entry_t *sent;
int j;
for (j = 0; j < bin_mdef_n_emit_state(mdef); ++j) {
if ((sent = ps_alignment_vector_grow_one(&al->state)) == NULL) {
E_ERROR("Failed to add state entry!\n");
return -1;
}
sent->id.senid = bin_mdef_sseq2sen(mdef, pent->id.pid.ssid, j);
oe_assert(sent->id.senid != BAD_SENID);
sent->start = pent->start;
sent->duration = pent->duration;
sent->parent = i;
if (j == 0)
pent->child = (uint16)(sent - al->state.seq);
}
}
return 0;
}