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;
}
sphinx_wave2feat_t *
sphinx_wave2feat_init(cmd_ln_t *config)
{
sphinx_wave2feat_t *wtf;
int i;
wtf = (sphinx_wave2feat_t *)ckd_calloc(1, sizeof(*wtf));
wtf->refcount = 1;
wtf->config = cmd_ln_retain(config);
wtf->fe = fe_init_auto_r(wtf->config);
if (!wtf->fe) {
E_FATAL("Failed to create feature extraction\n");
}
wtf->ot = outtypes; /* Default (sphinx) type. */
for (i = 0; i < nouttypes; ++i) {
output_type_t const *otype = &outtypes[i];
if (0 == strcmp(cmd_ln_str_r(config, "-ofmt"), otype->name)) {
wtf->ot = otype;
break;
}
}
if (i == nouttypes) {
E_ERROR("Unknown output type: '%s'\n",
cmd_ln_str_r(config, "-ofmt"));
sphinx_wave2feat_free(wtf);
return NULL;
}
return wtf;
}
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;
}
int
main(int argc, char *argv[])
{
sbthread_t *threads[10];
cmd_ln_t *config;
int i;
E_INFO("Processing chan3.raw in 10 threads\n");
if ((config = cmd_ln_parse_r(NULL, fe_args, 0, NULL, FALSE)) == NULL)
return -1;
err_set_callback(err_threaded_cb, NULL);
pthread_key_create(&logfp_index, NULL);
pthread_setspecific(logfp_index, (void*)stderr);
for (i = 0; i < 10; ++i) {
config = cmd_ln_retain(config);
threads[i] = sbthread_start(config, process, (void *)(long)i);
}
for (i = 0; i < 10; ++i) {
int rv;
rv = sbthread_wait(threads[i]);
E_INFO("Thread %d exited with status %d\n", i, rv);
sbthread_free(threads[i]);
}
/* Now check to make sure they all created logfiles with the
* correct contents. */
for (i = 0; i < 10; ++i) {
char logfile[16], line[256];
FILE *logfh;
sprintf(logfile, "%03d.log", i);
TEST_ASSERT(logfh = fopen(logfile, "r"));
while (fgets(line, sizeof(line), logfh)) {
string_trim(line, STRING_BOTH);
printf("%s: |%s|\n", logfile, line);
/* total number of frames in audio file is 1436, but there are only 1290 voiced */
TEST_EQUAL(0, strcmp(line, "INFO: test_tls_log.c(61): nfr = 1290"));
}
fclose(logfh);
}
cmd_ln_free_r(config);
return 0;
}
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;
}
SWIGINTERN Config *Decoder_getConfig(Decoder *self){
return cmd_ln_retain(ps_get_config(self));
}
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;
}
fe_t *
fe_init_auto_r(cmd_ln_t *config)
{
fe_t *fe;
int prespch_frame_len;
fe = (fe_t*)ckd_calloc(1, sizeof(*fe));
fe->refcount = 1;
/* transfer params to front end */
if (fe_parse_general_params(cmd_ln_retain(config), fe) < 0) {
fe_free(fe);
return NULL;
}
/* compute remaining fe parameters */
/* We add 0.5 so approximate the float with the closest
* integer. E.g., 2.3 is truncate to 2, whereas 3.7 becomes 4
*/
fe->frame_shift = (int32) (fe->sampling_rate / fe->frame_rate + 0.5);
fe->frame_size = (int32) (fe->window_length * fe->sampling_rate + 0.5);
fe->prior = 0;
fe_start_stream(fe);
assert (fe->frame_shift > 1);
if (fe->frame_size < fe->frame_shift) {
E_ERROR
("Frame size %d (-wlen) must be greater than frame shift %d (-frate)\n",
fe->frame_size, fe->frame_shift);
fe_free(fe);
return NULL;
}
if (fe->frame_size > (fe->fft_size)) {
E_ERROR
("Number of FFT points has to be a power of 2 higher than %d, it is %d\n",
fe->frame_size, fe->fft_size);
fe_free(fe);
return NULL;
}
if (fe->dither)
fe_init_dither(fe->seed);
/* establish buffers for overflow samps and hamming window */
fe->overflow_samps = ckd_calloc(fe->frame_size, sizeof(int16));
fe->hamming_window = ckd_calloc(fe->frame_size/2, sizeof(window_t));
/* create hamming window */
fe_create_hamming(fe->hamming_window, fe->frame_size);
/* init and fill appropriate filter structure */
fe->mel_fb = ckd_calloc(1, sizeof(*fe->mel_fb));
/* transfer params to mel fb */
fe_parse_melfb_params(config, fe, fe->mel_fb);
if (fe->mel_fb->upper_filt_freq > fe->sampling_rate / 2 + 1.0) {
E_ERROR("Upper frequency %.1f is higher than samprate/2 (%.1f)\n",
fe->mel_fb->upper_filt_freq, fe->sampling_rate / 2);
fe_free(fe);
return NULL;
}
fe_build_melfilters(fe->mel_fb);
fe_compute_melcosine(fe->mel_fb);
if (fe->remove_noise || fe->remove_silence)
fe->noise_stats = fe_init_noisestats(fe->mel_fb->num_filters);
fe->vad_data = (vad_data_t*)ckd_calloc(1, sizeof(*fe->vad_data));
prespch_frame_len = fe->log_spec != RAW_LOG_SPEC ? fe->num_cepstra : fe->mel_fb->num_filters;
fe->vad_data->prespch_buf = fe_prespch_init(fe->pre_speech + 1, prespch_frame_len, fe->frame_shift);
/* Create temporary FFT, spectrum and mel-spectrum buffers. */
/* FIXME: Gosh there are a lot of these. */
fe->spch = ckd_calloc(fe->frame_size, sizeof(*fe->spch));
fe->frame = ckd_calloc(fe->fft_size, sizeof(*fe->frame));
fe->spec = ckd_calloc(fe->fft_size, sizeof(*fe->spec));
fe->mfspec = ckd_calloc(fe->mel_fb->num_filters, sizeof(*fe->mfspec));
/* create twiddle factors */
fe->ccc = ckd_calloc(fe->fft_size / 4, sizeof(*fe->ccc));
fe->sss = ckd_calloc(fe->fft_size / 4, sizeof(*fe->sss));
fe_create_twiddle(fe);
if (cmd_ln_boolean_r(config, "-verbose")) {
fe_print_current(fe);
}
/*** Initialize the overflow buffers ***/
fe_start_utt(fe);
return fe;
}
fe_t *
fe_init_auto_r(cmd_ln_t *config)
{
fe_t *fe;
fe = ckd_calloc(1, sizeof(*fe));
fe->refcount = 1;
/* transfer params to front end */
if (fe_parse_general_params(cmd_ln_retain(config), fe) < 0) {
fe_free(fe);
return NULL;
}
/* compute remaining fe parameters */
/* We add 0.5 so approximate the float with the closest
* integer. E.g., 2.3 is truncate to 2, whereas 3.7 becomes 4
*/
fe->frame_shift = (int32) (fe->sampling_rate / fe->frame_rate + 0.5);
fe->frame_size = (int32) (fe->window_length * fe->sampling_rate + 0.5);
fe->prior = 0;
fe->frame_counter = 0;
assert (fe->frame_shift > 1);
if (fe->frame_size > (fe->fft_size)) {
E_WARN
("Number of FFT points has to be a power of 2 higher than %d\n",
(fe->frame_size));
fe_free(fe);
return (NULL);
}
if (fe->dither)
fe_init_dither(fe->seed);
/* establish buffers for overflow samps and hamming window */
fe->overflow_samps = ckd_calloc(fe->frame_size, sizeof(int16));
fe->hamming_window = ckd_calloc(fe->frame_size/2, sizeof(window_t));
/* create hamming window */
fe_create_hamming(fe->hamming_window, fe->frame_size);
/* init and fill appropriate filter structure */
fe->mel_fb = ckd_calloc(1, sizeof(*fe->mel_fb));
/* transfer params to mel fb */
fe_parse_melfb_params(config, fe, fe->mel_fb);
fe_build_melfilters(fe->mel_fb);
fe_compute_melcosine(fe->mel_fb);
/* Create temporary FFT, spectrum and mel-spectrum buffers. */
/* FIXME: Gosh there are a lot of these. */
fe->spch = ckd_calloc(fe->frame_size, sizeof(*fe->spch));
fe->frame = ckd_calloc(fe->fft_size, sizeof(*fe->frame));
fe->spec = ckd_calloc(fe->fft_size, sizeof(*fe->spec));
fe->mfspec = ckd_calloc(fe->mel_fb->num_filters, sizeof(*fe->mfspec));
/* create twiddle factors */
fe->ccc = ckd_calloc(fe->fft_size / 4, sizeof(*fe->ccc));
fe->sss = ckd_calloc(fe->fft_size / 4, sizeof(*fe->sss));
fe_create_twiddle(fe);
if (cmd_ln_boolean_r(config, "-verbose")) {
fe_print_current(fe);
}
/*** Z.A.B. ***/
/*** Initialize the overflow buffers ***/
fe_start_utt(fe);
return fe;
}
SWIGINTERN Config *ps_decoder_s_getConfig(struct ps_decoder_s *self){
return cmd_ln_retain(ps_get_config(self));
}
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 = 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);
/* 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 ((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;
/* 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->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;
}
fe_t *
fe_init_auto()
{
return fe_init_auto_r(cmd_ln_retain(cmd_ln_get()));
}
