int
ps_load_dict(ps_decoder_t *ps, char const *dictfile,
char const *fdictfile, char const *format)
{
cmd_ln_t *newconfig;
dict2pid_t *d2p;
dict_t *dict;
hash_iter_t *search_it;
/* Create a new scratch config to load this dict (so existing one
* won't be affected if it fails) */
newconfig = cmd_ln_init(NULL, ps_args(), TRUE, NULL);
cmd_ln_set_boolean_r(newconfig, "-dictcase",
cmd_ln_boolean_r(ps->config, "-dictcase"));
cmd_ln_set_str_r(newconfig, "-dict", dictfile);
if (fdictfile)
cmd_ln_set_str_r(newconfig, "-fdict", fdictfile);
else
cmd_ln_set_str_r(newconfig, "-fdict",
cmd_ln_str_r(ps->config, "-fdict"));
/* Try to load it. */
if ((dict = dict_init(newconfig, ps->acmod->mdef, ps->acmod->lmath)) == NULL) {
cmd_ln_free_r(newconfig);
return -1;
}
/* Reinit the dict2pid. */
if ((d2p = dict2pid_build(ps->acmod->mdef, dict)) == NULL) {
cmd_ln_free_r(newconfig);
return -1;
}
/* Success! Update the existing config to reflect new dicts and
* drop everything into place. */
cmd_ln_free_r(newconfig);
cmd_ln_set_str_r(ps->config, "-dict", dictfile);
if (fdictfile)
cmd_ln_set_str_r(ps->config, "-fdict", fdictfile);
dict_free(ps->dict);
ps->dict = dict;
dict2pid_free(ps->d2p);
ps->d2p = d2p;
/* And tell all searches to reconfigure themselves. */
for (search_it = hash_table_iter(ps->searches); search_it;
search_it = hash_table_iter_next(search_it)) {
if (ps_search_reinit(hash_entry_val(search_it->ent), dict, d2p) < 0) {
hash_table_iter_free(search_it);
return -1;
}
}
return 0;
}
int
ps_load_dict(ps_decoder_t *ps, char const *dictfile,
char const *fdictfile, char const *format)
{
cmd_ln_t *newconfig;
dict2pid_t *d2p;
dict_t *dict;
gnode_t *gn;
int rv;
/* Create a new scratch config to load this dict (so existing one
* won't be affected if it fails) */
newconfig = cmd_ln_init(NULL, ps_args(), TRUE, NULL);
cmd_ln_set_boolean_r(newconfig, "-dictcase",
cmd_ln_boolean_r(ps->config, "-dictcase"));
cmd_ln_set_str_r(newconfig, "-dict", dictfile);
if (fdictfile)
cmd_ln_set_str_r(newconfig, "-fdict", fdictfile);
else
cmd_ln_set_str_r(newconfig, "-fdict",
cmd_ln_str_r(ps->config, "-fdict"));
/* Try to load it. */
if ((dict = dict_init(newconfig, ps->acmod->mdef)) == NULL) {
cmd_ln_free_r(newconfig);
return -1;
}
/* Reinit the dict2pid. */
if ((d2p = dict2pid_build(ps->acmod->mdef, dict)) == NULL) {
cmd_ln_free_r(newconfig);
return -1;
}
/* Success! Update the existing config to reflect new dicts and
* drop everything into place. */
cmd_ln_free_r(newconfig);
cmd_ln_set_str_r(ps->config, "-dict", dictfile);
if (fdictfile)
cmd_ln_set_str_r(ps->config, "-fdict", fdictfile);
dict_free(ps->dict);
ps->dict = dict;
dict2pid_free(ps->d2p);
ps->d2p = d2p;
/* And tell all searches to reconfigure themselves. */
for (gn = ps->searches; gn; gn = gnode_next(gn)) {
ps_search_t *search = gnode_ptr(gn);
if ((rv = ps_search_reinit(search, dict, d2p)) < 0)
return rv;
}
return 0;
}
main (int32 argc, char *argv[])
{
mdef_t *mdef;
dict_t *dict;
dict2pid_t *d2p;
if (argc != 4)
usage(argv[0]);
logs3_init (1.0003);
mdef = mdef_init (argv[1]);
dict = dict_init (mdef, argv[2], argv[3], 0);
d2p = dict2pid_build (mdef, dict);
dict2pid_dump (stdout, d2p, mdef, dict);
}
int
main(int argc, char *argv[])
{
bin_mdef_t *mdef;
dict_t *dict;
dict2pid_t *d2p;
cmd_ln_t *config;
TEST_ASSERT(config = cmd_ln_init(NULL, NULL, FALSE,
"-dict", MODELDIR "/lm/en_US/cmu07a.dic",
"-fdict", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/noisedict",
NULL));
TEST_ASSERT(mdef = bin_mdef_read(NULL, MODELDIR "/hmm/en_US/hub4wsj_sc_8k/mdef"));
TEST_ASSERT(dict = dict_init(config, mdef));
TEST_ASSERT(d2p = dict2pid_build(mdef, dict));
dict_free(dict);
dict2pid_free(d2p);
bin_mdef_free(mdef);
cmd_ln_free_r(config);
return 0;
}
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;
}
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;
}
kbcore_t *kbcore_init (float64 logbase,
char *feattype,
char *cmn,
char *varnorm,
char *agc,
char *mdeffile,
char *dictfile,
char *fdictfile,
char *compsep,
char *lmfile,
char *fillpenfile,
float64 silprob,
float64 fillprob,
float64 langwt,
float64 inspen,
float64 uw,
char *meanfile,
char *varfile,
float64 varfloor,
char *mixwfile,
float64 mixwfloor,
char *subvqfile,
char *tmatfile,
float64 tmatfloor)
{
kbcore_t *kb;
E_INFO("Initializing core models:\n");
kb = (kbcore_t *) ckd_calloc (1, sizeof(kbcore_t));
kb->fcb = NULL;
kb->mdef = NULL;
kb->dict = NULL;
kb->dict2pid = NULL;
kb->lm = NULL;
kb->fillpen = NULL;
kb->dict2lmwid = NULL;
kb->mgau = NULL;
kb->svq = NULL;
kb->tmat = NULL;
logs3_init (logbase);
if (feattype) {
if ((kb->fcb = feat_init (feattype, cmn, varnorm, agc)) == NULL)
E_FATAL("feat_init(%s) failed\n", feattype);
if (feat_n_stream(kb->fcb) != 1)
E_FATAL("#Feature streams(%d) != 1\n", feat_n_stream(kb->fcb));
}
if (mdeffile) {
if ((kb->mdef = mdef_init (mdeffile)) == NULL)
E_FATAL("mdef_init(%s) failed\n", mdeffile);
}
if (dictfile) {
if (! compsep)
compsep = "";
else if ((compsep[0] != '\0') && (compsep[1] != '\0')) {
E_FATAL("Compound word separator(%s) must be empty or single character string\n",
compsep);
}
if ((kb->dict = dict_init (kb->mdef, dictfile, fdictfile, compsep[0])) == NULL)
E_FATAL("dict_init(%s,%s,%s) failed\n", dictfile,
fdictfile ? fdictfile : "", compsep);
}
if (lmfile) {
if ((kb->lm = lm_read (lmfile, langwt, inspen, uw)) == NULL)
E_FATAL("lm_read(%s, %e, %e, %e) failed\n", lmfile, langwt, inspen, uw);
}
if (fillpenfile || (lmfile && kb->dict)) {
if (! kb->dict) /* Sic */
E_FATAL("No dictionary for associating filler penalty file(%s)\n", fillpenfile);
if ((kb->fillpen = fillpen_init (kb->dict, fillpenfile, silprob, fillprob,
langwt, inspen)) == NULL)
E_FATAL("fillpen_init(%s) failed\n", fillpenfile);
}
if (meanfile) {
if ((! varfile) || (! mixwfile))
E_FATAL("Varfile or mixwfile not specified along with meanfile(%s)\n", meanfile);
kb->mgau = mgau_init (meanfile, varfile, varfloor, mixwfile, mixwfloor, TRUE);
if (kb->mgau == NULL)
E_FATAL("gauden_init(%s, %s, %e) failed\n", meanfile, varfile, varfloor);
if (subvqfile) {
if ((kb->svq = subvq_init (subvqfile, varfloor, -1, kb->mgau)) == NULL)
E_FATAL("subvq_init (%s, %e, -1) failed\n", subvqfile, varfloor);
}
}
if (tmatfile) {
if ((kb->tmat = tmat_init (tmatfile, tmatfloor)) == NULL)
E_FATAL("tmat_init (%s, %e) failed\n", tmatfile, tmatfloor);
}
if (kb->dict && kb->lm) { /* Initialize dict2lmwid */
if ((kb->dict2lmwid = wid_dict_lm_map (kb->dict, kb->lm)) == NULL)
E_FATAL("Dict/LM word-id mapping failed\n");
}
if (kb->mdef && kb->dict) { /* Initialize dict2pid */
kb->dict2pid = dict2pid_build (kb->mdef, kb->dict);
}
/* ***************** Verifications ***************** */
E_INFO("Verifying models consistency:\n");
if (kb->fcb && kb->mgau) {
/* Verify feature streams against gauden codebooks */
if (feat_stream_len(kb->fcb, 0) != mgau_veclen(kb->mgau))
E_FATAL("Feature streamlen(%d) != mgau streamlen(%d)\n",
feat_stream_len(kb->fcb, 0), mgau_veclen(kb->mgau));
}
if (kb->mdef && kb->mgau) {
/* Verify senone parameters against model definition parameters */
if (kb->mdef->n_sen != mgau_n_mgau(kb->mgau))
E_FATAL("Mdef #senones(%d) != mgau #senones(%d)\n",
kb->mdef->n_sen, mgau_n_mgau(kb->mgau));
}
if (kb->mdef && kb->tmat) {
/* Verify transition matrices parameters against model definition parameters */
if (kb->mdef->n_tmat != kb->tmat->n_tmat)
E_FATAL("Mdef #tmat(%d) != tmatfile(%d)\n", kb->mdef->n_tmat, kb->tmat->n_tmat);
if (kb->mdef->n_emit_state != kb->tmat->n_state)
E_FATAL("Mdef #states(%d) != tmat #states(%d)\n",
kb->mdef->n_emit_state, kb->tmat->n_state);
}
return kb;
}