/**
* Find PHMM node with same senone sequence and tmat id as the given triphone.
* Return ptr to PHMM node if found, NULL otherwise.
*/
static phmm_t *
phmm_lookup(allphone_search_t * allphs, s3pid_t pid)
{
phmm_t *p;
bin_mdef_t *mdef;
phmm_t **ci_phmm;
mdef = ((ps_search_t *) allphs)->acmod->mdef;
ci_phmm = allphs->ci_phmm;
for (p = ci_phmm[bin_mdef_pid2ci(mdef, pid)]; p; p = p->next) {
if (mdef_pid2tmatid(mdef, p->pid) == mdef_pid2tmatid(mdef, pid))
if (mdef_pid2ssid(mdef, p->pid) == mdef_pid2ssid(mdef, pid))
return p;
}
//not found
return NULL;
}
/**
* Build net from phone HMMs
*/
static int
phmm_build(allphone_search_t * allphs)
{
phmm_t *p, **pid2phmm;
bin_mdef_t *mdef;
int32 lrc_size;
uint32 *lc, *rc;
s3pid_t pid;
s3cipid_t ci;
s3cipid_t *filler;
int n_phmm, n_link;
int i, nphone;
mdef = ((ps_search_t *) allphs)->acmod->mdef;
allphs->ci_phmm =
(phmm_t **) ckd_calloc(bin_mdef_n_ciphone(mdef), sizeof(phmm_t *));
pid2phmm =
(phmm_t **) ckd_calloc(bin_mdef_n_phone(mdef), sizeof(phmm_t *));
/* For each unique ciphone/triphone entry in mdef, create a PHMM node */
n_phmm = 0;
nphone = allphs->ci_only ? bin_mdef_n_ciphone(mdef) : bin_mdef_n_phone(mdef);
E_INFO("Building PHMM net of %d phones\n", nphone);
for (pid = 0; pid < nphone; pid++) {
if ((p = phmm_lookup(allphs, pid)) == NULL) {
//not found, should be created
p = (phmm_t *) ckd_calloc(1, sizeof(*p));
hmm_init(allphs->hmmctx, &(p->hmm), FALSE,
mdef_pid2ssid(mdef, pid), mdef->phone[pid].tmat);
p->pid = pid;
p->ci = bin_mdef_pid2ci(mdef, pid);
p->succlist = NULL;
p->next = allphs->ci_phmm[p->ci];
allphs->ci_phmm[p->ci] = p;
n_phmm++;
}
pid2phmm[pid] = p;
}
/* Fill out bitvecs of each PHMM node, alloc continuous memory chunk for context bitvectors */
lrc_size = bitvec_size(bin_mdef_n_ciphone(mdef));
lc = ckd_calloc(n_phmm * 2 * lrc_size, sizeof(bitvec_t));
rc = lc + (n_phmm * lrc_size);
for (ci = 0; ci < mdef->n_ciphone; ci++) {
for (p = allphs->ci_phmm[ci]; p; p = p->next) {
p->lc = lc;
lc += lrc_size;
p->rc = rc;
rc += lrc_size;
}
}
/* Fill out lc and rc bitmaps (remember to map all fillers to each other!!) */
filler =
(s3cipid_t *) ckd_calloc(bin_mdef_n_ciphone(mdef) + 1,
sizeof(s3cipid_t));
/* Connect fillers */
i = 0;
for (ci = 0; ci < bin_mdef_n_ciphone(mdef); ci++) {
p = pid2phmm[ci];
bitvec_set_all(p->lc, bin_mdef_n_ciphone(mdef));
bitvec_set_all(p->rc, bin_mdef_n_ciphone(mdef));
if (mdef->phone[ci].info.ci.filler) {
filler[i++] = ci;
}
}
filler[i] = BAD_S3CIPID;
/* Loop over cdphones only if ci_only is not set */
for (pid = bin_mdef_n_ciphone(mdef); pid < nphone;
pid++) {
p = pid2phmm[pid];
if (mdef->phone[mdef->phone[pid].info.cd.ctx[1]].info.ci.filler) {
for (i = 0; IS_S3CIPID(filler[i]); i++)
bitvec_set(p->lc, filler[i]);
}
else
bitvec_set(p->lc, mdef->phone[pid].info.cd.ctx[1]);
if (mdef->phone[mdef->phone[pid].info.cd.ctx[2]].info.ci.filler) {
for (i = 0; IS_S3CIPID(filler[i]); i++)
bitvec_set(p->rc, filler[i]);
}
else
bitvec_set(p->rc, mdef->phone[pid].info.cd.ctx[2]);
}
ckd_free(pid2phmm);
ckd_free(filler);
/* Create links between PHMM nodes */
n_link = phmm_link(allphs);
E_INFO("%d nodes, %d links\n", n_phmm, n_link);
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;
}
