/* 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;
}
bin_mdef_t *
bin_mdef_read(cmd_ln_t *config, const char *filename)
{
bin_mdef_t *m;
FILE *fh;
size_t tree_start;
int32 val, i, swap, pos, end;
int32 *sseq_size;
int do_mmap;
/* Try to read it as text first. */
if ((m = bin_mdef_read_text(config, filename)) != NULL)
return m;
E_INFO("Reading binary model definition: %s\n", filename);
if ((fh = fopen(filename, "rb")) == NULL)
return NULL;
if (fread(&val, 4, 1, fh) != 1) {
fclose(fh);
E_ERROR_SYSTEM("Failed to read byte-order marker from %s\n",
filename);
return NULL;
}
swap = 0;
if (val == BIN_MDEF_OTHER_ENDIAN) {
swap = 1;
E_INFO("Must byte-swap %s\n", filename);
}
if (fread(&val, 4, 1, fh) != 1) {
fclose(fh);
E_ERROR_SYSTEM("Failed to read version from %s\n", filename);
return NULL;
}
if (swap)
SWAP_INT32(&val);
if (val > BIN_MDEF_FORMAT_VERSION) {
E_ERROR("File format version %d for %s is newer than library\n",
val, filename);
fclose(fh);
return NULL;
}
if (fread(&val, 4, 1, fh) != 1) {
fclose(fh);
E_ERROR_SYSTEM("Failed to read header length from %s\n", filename);
return NULL;
}
if (swap)
SWAP_INT32(&val);
/* Skip format descriptor. */
fseek(fh, val, SEEK_CUR);
/* Finally allocate it. */
m = ckd_calloc(1, sizeof(*m));
m->refcnt = 1;
/* Check these, to make gcc/glibc shut up. */
#define FREAD_SWAP32_CHK(dest) \
if (fread((dest), 4, 1, fh) != 1) { \
fclose(fh); \
ckd_free(m); \
E_ERROR_SYSTEM("Failed to read %s from %s\n", #dest, filename); \
return NULL; \
} \
if (swap) SWAP_INT32(dest);
FREAD_SWAP32_CHK(&m->n_ciphone);
FREAD_SWAP32_CHK(&m->n_phone);
FREAD_SWAP32_CHK(&m->n_emit_state);
FREAD_SWAP32_CHK(&m->n_ci_sen);
FREAD_SWAP32_CHK(&m->n_sen);
FREAD_SWAP32_CHK(&m->n_tmat);
FREAD_SWAP32_CHK(&m->n_sseq);
FREAD_SWAP32_CHK(&m->n_ctx);
FREAD_SWAP32_CHK(&m->n_cd_tree);
FREAD_SWAP32_CHK(&m->sil);
/* CI names are first in the file. */
m->ciname = ckd_calloc(m->n_ciphone, sizeof(*m->ciname));
/* Decide whether to read in the whole file or mmap it. */
do_mmap = config ? cmd_ln_boolean_r(config, "-mmap") : TRUE;
if (swap) {
E_WARN("-mmap specified, but mdef is other-endian. Will not memory-map.\n");
do_mmap = FALSE;
}
/* Actually try to mmap it. */
if (do_mmap) {
m->filemap = mmio_file_read(filename);
if (m->filemap == NULL)
do_mmap = FALSE;
}
pos = ftell(fh);
if (do_mmap) {
/* Get the base pointer from the memory map. */
m->ciname[0] = (char *)mmio_file_ptr(m->filemap) + pos;
/* Success! */
m->alloc_mode = BIN_MDEF_ON_DISK;
}
else {
/* Read everything into memory. */
m->alloc_mode = BIN_MDEF_IN_MEMORY;
fseek(fh, 0, SEEK_END);
end = ftell(fh);
fseek(fh, pos, SEEK_SET);
m->ciname[0] = ckd_malloc(end - pos);
if (fread(m->ciname[0], 1, end - pos, fh) != end - pos)
E_FATAL("Failed to read %d bytes of data from %s\n", end - pos, filename);
}
for (i = 1; i < m->n_ciphone; ++i)
m->ciname[i] = m->ciname[i - 1] + strlen(m->ciname[i - 1]) + 1;
/* Skip past the padding. */
tree_start =
m->ciname[i - 1] + strlen(m->ciname[i - 1]) + 1 - m->ciname[0];
tree_start = (tree_start + 3) & ~3;
m->cd_tree = (cd_tree_t *) (m->ciname[0] + tree_start);
if (swap) {
for (i = 0; i < m->n_cd_tree; ++i) {
SWAP_INT16(&m->cd_tree[i].ctx);
SWAP_INT16(&m->cd_tree[i].n_down);
SWAP_INT32(&m->cd_tree[i].c.down);
}
}
m->phone = (mdef_entry_t *) (m->cd_tree + m->n_cd_tree);
if (swap) {
for (i = 0; i < m->n_phone; ++i) {
SWAP_INT32(&m->phone[i].ssid);
SWAP_INT32(&m->phone[i].tmat);
}
}
sseq_size = (int32 *) (m->phone + m->n_phone);
if (swap)
SWAP_INT32(sseq_size);
m->sseq = ckd_calloc(m->n_sseq, sizeof(*m->sseq));
m->sseq[0] = (uint16 *) (sseq_size + 1);
if (swap) {
for (i = 0; i < *sseq_size; ++i)
SWAP_INT16(m->sseq[0] + i);
}
if (m->n_emit_state) {
for (i = 1; i < m->n_sseq; ++i)
m->sseq[i] = m->sseq[0] + i * m->n_emit_state;
}
else {
m->sseq_len = (uint8 *) (m->sseq[0] + *sseq_size);
for (i = 1; i < m->n_sseq; ++i)
m->sseq[i] = m->sseq[i - 1] + m->sseq_len[i - 1];
}
/* Now build the CD-to-CI mappings using the senone sequences.
* This is the only really accurate way to do it, though it is
* still inaccurate in the case of heterogeneous topologies or
* cross-state tying. */
m->cd2cisen = (int16 *) ckd_malloc(m->n_sen * sizeof(*m->cd2cisen));
m->sen2cimap = (int16 *) ckd_malloc(m->n_sen * sizeof(*m->sen2cimap));
/* Default mappings (identity, none) */
for (i = 0; i < m->n_ci_sen; ++i)
m->cd2cisen[i] = i;
for (; i < m->n_sen; ++i)
m->cd2cisen[i] = -1;
for (i = 0; i < m->n_sen; ++i)
m->sen2cimap[i] = -1;
for (i = 0; i < m->n_phone; ++i) {
int32 j, ssid = m->phone[i].ssid;
for (j = 0; j < bin_mdef_n_emit_state_phone(m, i); ++j) {
int s = bin_mdef_sseq2sen(m, ssid, j);
int ci = bin_mdef_pid2ci(m, i);
/* Take the first one and warn if we have cross-state tying. */
if (m->sen2cimap[s] == -1)
m->sen2cimap[s] = ci;
if (m->sen2cimap[s] != ci)
E_WARN
("Senone %d is shared between multiple base phones\n",
s);
if (j > bin_mdef_n_emit_state_phone(m, ci))
E_WARN("CD phone %d has fewer states than CI phone %d\n",
i, ci);
else
m->cd2cisen[s] =
bin_mdef_sseq2sen(m, m->phone[ci].ssid, j);
}
}
/* Set the silence phone. */
m->sil = bin_mdef_ciphone_id(m, S3_SILENCE_CIPHONE);
E_INFO
("%d CI-phone, %d CD-phone, %d emitstate/phone, %d CI-sen, %d Sen, %d Sen-Seq\n",
m->n_ciphone, m->n_phone - m->n_ciphone, m->n_emit_state,
m->n_ci_sen, m->n_sen, m->n_sseq);
fclose(fh);
return m;
}
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;
}
