static int
fe_parse_melfb_params(cmd_ln_t *config, fe_t *fe, melfb_t * mel)
{
mel->sampling_rate = fe->sampling_rate;
mel->fft_size = fe->fft_size;
mel->num_cepstra = fe->num_cepstra;
mel->num_filters = cmd_ln_int32_r(config, "-nfilt");
if (fe->log_spec)
fe->feature_dimension = mel->num_filters;
else
fe->feature_dimension = fe->num_cepstra;
mel->upper_filt_freq = cmd_ln_float32_r(config, "-upperf");
mel->lower_filt_freq = cmd_ln_float32_r(config, "-lowerf");
mel->doublewide = cmd_ln_boolean_r(config, "-doublebw");
mel->warp_type = cmd_ln_str_r(config, "-warp_type");
mel->warp_params = cmd_ln_str_r(config, "-warp_params");
mel->lifter_val = cmd_ln_int32_r(config, "-lifter");
mel->unit_area = cmd_ln_boolean_r(config, "-unit_area");
mel->round_filters = cmd_ln_boolean_r(config, "-round_filters");
if (fe_warp_set(mel, mel->warp_type) != FE_SUCCESS) {
E_ERROR("Failed to initialize the warping function.\n");
return -1;
}
fe_warp_set_parameters(mel, mel->warp_params, mel->sampling_rate);
return 0;
}
int32
main(int32 argc, char *argv[])
{
kb_t kb;
stat_t *st;
cmd_ln_t *config;
print_appl_info(argv[0]);
cmd_ln_appl_enter(argc, argv, "default.arg", arg);
unlimit();
config = cmd_ln_get();
kb_init(&kb, config);
st = kb.stat;
fprintf(stdout, "\n");
if (cmd_ln_str_r(config, "-ctl")) {
/* When -ctlfile is speicified, corpus.c will look at -ctl_lm and
-ctl_mllr to get the corresponding LM and MLLR for the utterance */
st->tm = ctl_process(cmd_ln_str_r(config, "-ctl"),
cmd_ln_str_r(config, "-ctl_lm"),
cmd_ln_str_r(config, "-ctl_mllr"),
cmd_ln_int32_r(config, "-ctloffset"),
cmd_ln_int32_r(config, "-ctlcount"), utt_decode, &kb);
}
else if (cmd_ln_str_r(config, "-utt")) {
/* When -utt is specified, corpus.c will wait for the utterance to
change */
st->tm = ctl_process_utt(cmd_ln_str_r(config, "-utt"),
cmd_ln_int32_r(config, "-ctlcount"),
utt_decode, &kb);
}
else {
/* Is error checking good enough?" */
E_FATAL("Both -utt and -ctl are not specified.\n");
}
if (kb.matchsegfp)
fclose(kb.matchsegfp);
if (kb.matchfp)
fclose(kb.matchfp);
stat_report_corpus(kb.stat);
kb_free(&kb);
#if (! WIN32)
#if defined(_SUN4)
system("ps -el | grep sphinx3_decode");
#else
system("ps aguxwww | grep sphinx3_decode");
#endif
#endif
cmd_ln_free_r(config);
exit(0);
}
static int
acmod_init_feat(acmod_t *acmod)
{
acmod->fcb =
feat_init(cmd_ln_str_r(acmod->config, "-feat"),
cmn_type_from_str(cmd_ln_str_r(acmod->config,"-cmn")),
cmd_ln_boolean_r(acmod->config, "-varnorm"),
agc_type_from_str(cmd_ln_str_r(acmod->config, "-agc")),
1, cmd_ln_int32_r(acmod->config, "-ceplen"));
if (acmod->fcb == NULL)
return -1;
if (cmd_ln_str_r(acmod->config, "-lda")) {
E_INFO("Reading linear feature transformation from %s\n",
cmd_ln_str_r(acmod->config, "-lda"));
if (feat_read_lda(acmod->fcb,
cmd_ln_str_r(acmod->config, "-lda"),
cmd_ln_int32_r(acmod->config, "-ldadim")) < 0)
return -1;
}
if (cmd_ln_str_r(acmod->config, "-svspec")) {
int32 **subvecs;
E_INFO("Using subvector specification %s\n",
cmd_ln_str_r(acmod->config, "-svspec"));
if ((subvecs = parse_subvecs(cmd_ln_str_r(acmod->config, "-svspec"))) == NULL)
return -1;
if ((feat_set_subvecs(acmod->fcb, subvecs)) < 0)
return -1;
}
if (cmd_ln_exists_r(acmod->config, "-agcthresh")
&& 0 != strcmp(cmd_ln_str_r(acmod->config, "-agc"), "none")) {
agc_set_threshold(acmod->fcb->agc_struct,
cmd_ln_float32_r(acmod->config, "-agcthresh"));
}
if (acmod->fcb->cmn_struct
&& cmd_ln_exists_r(acmod->config, "-cmninit")) {
char *c, *cc, *vallist;
int32 nvals;
vallist = ckd_salloc(cmd_ln_str_r(acmod->config, "-cmninit"));
c = vallist;
nvals = 0;
while (nvals < acmod->fcb->cmn_struct->veclen
&& (cc = strchr(c, ',')) != NULL) {
*cc = '\0';
acmod->fcb->cmn_struct->cmn_mean[nvals] = FLOAT2MFCC(atof_c(c));
c = cc + 1;
++nvals;
}
if (nvals < acmod->fcb->cmn_struct->veclen && *c != '\0') {
acmod->fcb->cmn_struct->cmn_mean[nvals] = FLOAT2MFCC(atof_c(c));
}
ckd_free(vallist);
}
return 0;
}
int
fe_parse_general_params(cmd_ln_t *config, fe_t * fe)
{
int j;
fe->config = config;
fe->sampling_rate = cmd_ln_float32_r(config, "-samprate");
fe->frame_rate = (int16)cmd_ln_int32_r(config, "-frate");
if (cmd_ln_boolean_r(config, "-dither")) {
fe->dither = 1;
fe->seed = cmd_ln_int32_r(config, "-seed");
}
#ifdef WORDS_BIGENDIAN
fe->swap = strcmp("big", cmd_ln_str_r(config, "-input_endian")) == 0 ? 0 : 1;
#else
fe->swap = strcmp("little", cmd_ln_str_r(config, "-input_endian")) == 0 ? 0 : 1;
#endif
fe->window_length = cmd_ln_float32_r(config, "-wlen");
fe->pre_emphasis_alpha = cmd_ln_float32_r(config, "-alpha");
fe->num_cepstra = (uint8)cmd_ln_int32_r(config, "-ncep");
fe->fft_size = (int16)cmd_ln_int32_r(config, "-nfft");
/* Check FFT size, compute FFT order (log_2(n)) */
for (j = fe->fft_size, fe->fft_order = 0; j > 1; j >>= 1, fe->fft_order++) {
if (((j % 2) != 0) || (fe->fft_size <= 0)) {
E_ERROR("fft: number of points must be a power of 2 (is %d)\n",
fe->fft_size);
return -1;
}
}
/* Verify that FFT size is greater or equal to window length. */
if (fe->fft_size < (int)(fe->window_length * fe->sampling_rate)) {
E_ERROR("FFT: Number of points must be greater or equal to frame size (%d samples)\n",
(int)(fe->window_length * fe->sampling_rate));
return -1;
}
fe->remove_dc = cmd_ln_boolean_r(config, "-remove_dc");
if (0 == strcmp(cmd_ln_str_r(config, "-transform"), "dct"))
fe->transform = DCT_II;
else if (0 == strcmp(cmd_ln_str_r(config, "-transform"), "legacy"))
fe->transform = LEGACY_DCT;
else if (0 == strcmp(cmd_ln_str_r(config, "-transform"), "htk"))
fe->transform = DCT_HTK;
else {
E_ERROR("Invalid transform type (values are 'dct', 'legacy', 'htk')\n");
return -1;
}
if (cmd_ln_boolean_r(config, "-logspec"))
fe->log_spec = RAW_LOG_SPEC;
if (cmd_ln_boolean_r(config, "-smoothspec"))
fe->log_spec = SMOOTH_LOG_SPEC;
return 0;
}
glist_t
srch_FLAT_FWD_bestpath_impl(void *srch, /**< A void pointer to a search structure */
dag_t * dag)
{
srch_t *s;
srch_FLAT_FWD_graph_t *fwg;
float32 bestpathlw;
float64 lwf;
srch_hyp_t *tmph, *bph;
glist_t ghyp, rhyp;
s = (srch_t *) srch;
fwg = (srch_FLAT_FWD_graph_t *) s->grh->graph_struct;
assert(fwg->lathist);
bestpathlw = cmd_ln_float32_r(kbcore_config(fwg->kbcore), "-bestpathlw");
lwf = bestpathlw ? (bestpathlw / cmd_ln_float32_r(kbcore_config(fwg->kbcore), "-lw")) : 1.0;
flat_fwd_dag_add_fudge_edges(fwg,
dag,
cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-dagfudge"),
cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-min_endfr"),
(void *) fwg->lathist, s->kbc->dict);
/* Bypass filler nodes */
if (!dag->filler_removed) {
/* If Viterbi search terminated in filler word coerce final DAG node to FINISH_WORD */
if (dict_filler_word(s->kbc->dict, dag->end->wid))
dag->end->wid = s->kbc->dict->finishwid;
if (dag_bypass_filler_nodes(dag, lwf, s->kbc->dict, s->kbc->fillpen) < 0)
E_ERROR("maxedge limit (%d) exceeded\n", dag->maxedge);
else
dag->filler_removed = 1;
}
bph =
dag_search(dag, s->uttid, lwf, dag->end,
s->kbc->dict, s->kbc->lmset->cur_lm, s->kbc->fillpen);
if (bph != NULL) {
ghyp = NULL;
for (tmph = bph; tmph; tmph = tmph->next)
ghyp = glist_add_ptr(ghyp, (void *) tmph);
rhyp = glist_reverse(ghyp);
return rhyp;
}
else {
return NULL;
}
}
/**
* "Detect" Sphinx MFCC files, meaning verify their lousy headers, and
* set up some parameters from the config object.
*
* @return TRUE, or -1 on error.
*/
static int
detect_sphinx_mfc(sphinx_wave2feat_t *wtf)
{
FILE *fh;
int32 len;
long flen;
if ((fh = fopen(wtf->infile, "rb")) == NULL) {
E_ERROR_SYSTEM("Failed to open %s", wtf->infile);
return -1;
}
if (fread(&len, 4, 1, fh) != 1) {
E_ERROR_SYSTEM("Failed to read header from %s\n", wtf->infile);
fclose(fh);
return -1;
}
fseek(fh, 0, SEEK_END);
flen = ftell(fh);
/* figure out whether to byteswap */
flen = (flen / 4) - 1;
if (flen != len) {
/* First make sure this is an endianness problem, otherwise fail. */
SWAP_INT32(&len);
if (flen != len) {
SWAP_INT32(&len);
E_ERROR("Mismatch in header/file lengths: 0x%08x vs 0x%08x\n",
len, flen);
return -1;
}
/* Set the input endianness to the opposite of the machine endianness... */
cmd_ln_set_str_r(wtf->config, "-input_endian",
(0 == strcmp("big", cmd_ln_str_r(wtf->config, "-mach_endian"))
? "little" : "big"));
}
fseek(fh, 4, SEEK_SET);
wtf->infh = fh;
if (cmd_ln_boolean_r(wtf->config, "-spec2cep")) {
wtf->in_veclen = cmd_ln_int32_r(wtf->config, "-nfilt");
}
else if (cmd_ln_boolean_r(wtf->config, "-cep2spec")) {
wtf->in_veclen = cmd_ln_int32_r(wtf->config, "-ncep");
wtf->veclen = cmd_ln_int32_r(wtf->config, "-nfilt");
}
else {
/* Should not happen. */
E_ERROR("Sphinx MFCC file reading requested but -spec2cep/-cep2spec not given\n");
assert(FALSE);
}
return TRUE;
}
/**
* Process PCM audio from a filehandle. Assume that wtf->infh is
* positioned just after the file header.
*/
static int
decode_pcm(sphinx_wave2feat_t *wtf)
{
size_t nsamp;
int32 n, nfr, nchans, whichchan;
uint32 nfloat;
nchans = cmd_ln_int32_r(wtf->config, "-nchans");
whichchan = cmd_ln_int32_r(wtf->config, "-whichchan");
fe_start_utt(wtf->fe);
nfloat = 0;
while ((nsamp = fread(wtf->audio, 2, wtf->blocksize, wtf->infh)) != 0) {
size_t nvec;
int16 const *inspeech;
/* Byteswap stuff here if necessary. */
if (wtf->byteswap) {
for (n = 0; n < nsamp; ++n)
SWAP_INT16(wtf->audio + n);
}
/* Mix or pick channels. */
if (nchans > 1)
nsamp = mixnpick_channels(wtf->audio, nsamp, nchans, whichchan);
inspeech = wtf->audio;
nvec = wtf->featsize;
/* Consume all samples. */
while (nsamp) {
nfr = nvec;
fe_process_frames(wtf->fe, &inspeech, &nsamp, wtf->feat, &nfr, NULL);
if (nfr) {
if ((n = (*wtf->ot->output_frames)(wtf, wtf->feat, nfr)) < 0)
return -1;
nfloat += n;
}
}
inspeech = wtf->audio;
}
/* Now process any leftover audio frames. */
fe_end_utt(wtf->fe, wtf->feat[0], &nfr);
if (nfr) {
if ((n = (*wtf->ot->output_frames)(wtf, wtf->feat, nfr)) < 0)
return -1;
nfloat += n;
}
if (fclose(wtf->infh) == EOF)
E_ERROR_SYSTEM("Failed to close input file");
wtf->infh = NULL;
return nfloat;
}
int
main(int argc, char *argv[])
{
print_appl_info(argv[0]);
cmd_ln_appl_enter(argc, argv, "default.arg", defn);
unlimit();
config = cmd_ln_get();
logmath = logs3_init(cmd_ln_float64_r(config, "-logbase"), 1,
cmd_ln_int32_r(config, "-log3table"));
E_INFO("Value of base %f \n", cmd_ln_float32_r(config, "-logbase"));
models_init();
ptmr_init(&tm_utt);
if ((inmatchsegfp = fopen(cmd_ln_str_r(config, "-inhypseg"), "r")) == NULL)
E_ERROR("fopen(%s,r) failed\n", cmd_ln_str_r(config, "-inhypseg"));
if ((outconfmatchsegfp = fopen(cmd_ln_str_r(config, "-output"), "w")) == NULL)
E_ERROR("fopen(%s,w) failed\n", cmd_ln_str_r(config, "-output"));
if (cmd_ln_str_r(config, "-ctl")) {
ctl_process(cmd_ln_str_r(config, "-ctl"),
cmd_ln_str_r(config, "-ctl_lm"),
NULL,
cmd_ln_int32_r(config, "-ctloffset"),
cmd_ln_int32_r(config, "-ctlcount"), utt_confidence, NULL);
}
else {
E_FATAL("-ctl is not specified\n");
}
#if (! WIN32)
system("ps auxwww | grep s3dag");
#endif
fclose(outconfmatchsegfp);
fclose(inmatchsegfp);
models_free();
logmath_free(logmath);
cmd_ln_free_r(config);
return 0;
}
glist_t
srch_FLAT_FWD_nbest_impl(void *srch, /**< A void pointer to a search structure */
dag_t * dag)
{
srch_t *s;
srch_FLAT_FWD_graph_t *fwg;
float32 bestpathlw;
float64 lwf;
char str[2000];
s = (srch_t *) srch;
fwg = (srch_FLAT_FWD_graph_t *) s->grh->graph_struct;
assert(fwg->lathist);
if (!(cmd_ln_exists_r(kbcore_config(fwg->kbcore), "-nbestdir")
&& cmd_ln_str_r(kbcore_config(fwg->kbcore), "-nbestdir")))
return NULL;
ctl_outfile(str, cmd_ln_str_r(kbcore_config(fwg->kbcore), "-nbestdir"),
cmd_ln_str_r(kbcore_config(fwg->kbcore), "-nbestext"),
(s->uttfile ? s->uttfile : s->uttid), s->uttid,
cmd_ln_boolean_r(kbcore_config(fwg->kbcore), "-build_outdirs"));
bestpathlw = cmd_ln_float32_r(kbcore_config(fwg->kbcore), "-bestpathlw");
lwf = bestpathlw ? (bestpathlw / cmd_ln_float32_r(kbcore_config(fwg->kbcore), "-lw")) : 1.0;
flat_fwd_dag_add_fudge_edges(fwg,
dag,
cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-dagfudge"),
cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-min_endfr"),
(void *) fwg->lathist, s->kbc->dict);
/* Bypass filler nodes */
if (!dag->filler_removed) {
/* If Viterbi search terminated in filler word coerce final DAG node to FINISH_WORD */
if (dict_filler_word(s->kbc->dict, dag->end->wid))
dag->end->wid = s->kbc->dict->finishwid;
dag_remove_unreachable(dag);
if (dag_bypass_filler_nodes(dag, lwf, s->kbc->dict, s->kbc->fillpen) < 0)
E_ERROR("maxedge limit (%d) exceeded\n", dag->maxedge);
}
dag_compute_hscr(dag, kbcore_dict(s->kbc), kbcore_lm(s->kbc), lwf);
dag_remove_bypass_links(dag);
dag->filler_removed = 0;
nbest_search(dag, str, s->uttid, lwf,
kbcore_dict(s->kbc),
kbcore_lm(s->kbc), kbcore_fillpen(s->kbc)
);
return NULL;
}
int
main(int _argc, char **_argv)
{
char *ctrlfn;
char *cfgfn;
cmd_ln_t *config = NULL;
print_appl_info(_argv[0]);
if (_argc != 4) {
printf("\nUSAGE: %s <ctrlfile> <rawdir> <cfgfile>\n", _argv[0]);
return -1;
}
ctrlfn = _argv[1];
rawdirfn = _argv[2];
cfgfn = _argv[3];
if ((config = cmd_ln_parse_file_r(config, S3_DECODE_ARG_DEFS, cfgfn, TRUE)) == NULL)
E_FATAL("Bad configuration file %s.\n", cfgfn);
if (s3_decode_init(&decoder, config) != S3_DECODE_SUCCESS)
E_FATAL("Failed to initialize live-decoder.\n");
fe = fe_init_auto_r(config);
st = decoder.kb.stat;
ptmr_init(&(st->tm));
if (ctrlfn) {
/* When -ctlfile is speicified, corpus.c will look at -ctl_lm and
-ctl_mllr to get the corresponding LM and MLLR for the utterance */
st->tm = ctl_process(ctrlfn,
cmd_ln_str_r(config, "-ctl_lm"),
cmd_ln_str_r(config, "-ctl_mllr"),
cmd_ln_int32_r(config, "-ctloffset"),
cmd_ln_int32_r(config, "-ctlcount"),
utt_livepretend, &(decoder.kb));
}
else {
E_FATAL("control file is not specified.\n");
}
stat_report_corpus(decoder.kb.stat);
s3_decode_close(&decoder);
fe_free(fe);
return 0;
}
int
acmod_fe_mismatch(acmod_t *acmod, fe_t *fe)
{
/* Output vector dimension needs to be the same. */
if (cmd_ln_int32_r(acmod->config, "-ceplen") != fe_get_output_size(fe)) {
E_ERROR("Configured feature length %d doesn't match feature extraction output size %d\n",
cmd_ln_int32_r(acmod->config, "-ceplen"),
fe_get_output_size(fe));
return TRUE;
}
/* Feature parameters need to be the same. */
/* ... */
return FALSE;
}
static int
write_nbest(ps_decoder_t *ps, char const *nbestdir, char const *uttid)
{
cmd_ln_t *config;
char *outfile;
FILE *fh;
ps_nbest_t *nbest;
int32 i, n, score;
const char* hyp;
config = ps_get_config(ps);
outfile = string_join(nbestdir, "/", uttid,
cmd_ln_str_r(config, "-nbestext"), NULL);
n = cmd_ln_int32_r(config, "-nbest");
fh = fopen(outfile, "w");
if (fh == NULL) {
E_ERROR_SYSTEM("Failed to write a lattice to file %s\n", outfile);
return -1;
}
nbest = ps_nbest(ps, 0, -1, NULL, NULL);
for (i = 0; i < n && nbest && (nbest = ps_nbest_next(nbest)); i++) {
hyp = ps_nbest_hyp(nbest, &score);
fprintf(fh, "%s %d\n", hyp, score);
}
if (nbest)
ps_nbest_free(nbest);
fclose(fh);
return 0;
}
int filter_buffer_create(context_t *ctx)
{
options_t *opts;
decoder_set_t *decset;
decoder_t *dec;
cmd_ln_t *cfg;
uint32_t rate;
int32 fps;
int32_t frlen;
filter_buf_t *filtbuf;
if (!ctx || !(opts = ctx->opts) ||
!(decset = ctx->decset) || !(dec = decset->decs) ||
!(cfg = dec->cfg) || !(filtbuf = mrp_allocz(sizeof(filter_buf_t))))
return -1;
rate = opts->rate;
fps = cmd_ln_int32_r(cfg, "-frate");
frlen = rate / (double)fps;
filtbuf->len = 0;
filtbuf->frlen = frlen;
filtbuf->fdrec = open_file_for_recording(opts->audio);
ctx->filtbuf = filtbuf;
return 0;
}
/**
* Process PCM audio from a libsndfile file. FIXME: looks a lot like
* decode_pcm! Also needs stereo support (as does decode_pcm).
*/
static int
decode_sndfile(sphinx_wave2feat_t *wtf)
{
size_t nsamp;
int32 nfr, nchans, whichchan;
int nfloat, n;
nchans = cmd_ln_int32_r(wtf->config, "-nchans");
whichchan = cmd_ln_int32_r(wtf->config, "-whichchan");
fe_start_utt(wtf->fe);
nfloat = 0;
while ((nsamp = sf_read_short(wtf->insfh,
wtf->audio,
wtf->blocksize)) != 0) {
int16 const *inspeech;
size_t nvec;
/* Mix or pick channels. */
if (nchans > 1)
nsamp = mixnpick_channels(wtf->audio, nsamp, nchans, whichchan);
inspeech = wtf->audio;
nvec = wtf->featsize;
/* Consume all samples. */
while (nsamp) {
nfr = nvec;
fe_process_frames(wtf->fe, &inspeech, &nsamp, wtf->feat, &nfr, NULL);
if (nfr) {
if ((n = (*wtf->ot->output_frames)(wtf, wtf->feat, nfr)) < 0)
return -1;
nfloat += n;
}
}
inspeech = wtf->audio;
}
/* Now process any leftover audio frames. */
fe_end_utt(wtf->fe, wtf->feat[0], &nfr);
if (nfr) {
if ((n = (*wtf->ot->output_frames)(wtf, wtf->feat, nfr)) < 0)
return -1;
nfloat += n;
}
sf_close(wtf->insfh);
wtf->insfh = NULL;
return nfloat;
}
int
main(int argc, char *argv[])
{
cmd_ln_t *config;
config = cmd_ln_parse_r(NULL, defs, argc, argv, TRUE);
if (config == NULL)
return 1;
printf("%d %s %d %f\n",
cmd_ln_int32_r(config, "-a"),
cmd_ln_str_r(config, "-b") ? cmd_ln_str_r(config, "-b") : "(null)",
cmd_ln_boolean_r(config, "-c"),
cmd_ln_float64_r(config, "-d"));
cmd_ln_free_r(config);
config = cmd_ln_init(NULL, NULL, FALSE,
"-b", "foobie", NULL);
if (config == NULL)
return 1;
cmd_ln_free_r(config);
config = cmd_ln_init(NULL, defs, TRUE,
"-b", "foobie", NULL);
if (config == NULL)
return 1;
printf("%d %s %d %f\n",
cmd_ln_int32_r(config, "-a"),
cmd_ln_str_r(config, "-b") ? cmd_ln_str_r(config, "-b") : "(null)",
cmd_ln_boolean_r(config, "-c"),
cmd_ln_float64_r(config, "-d"));
cmd_ln_free_r(config);
config = cmd_ln_init(NULL, NULL, FALSE,
"-b", "foobie", NULL);
if (config == NULL)
return 1;
printf("%s\n",
cmd_ln_str_r(config, "-b") ? cmd_ln_str_r(config, "-b") : "(null)");
cmd_ln_set_str_r(config, "-b", "blatz");
printf("%s\n",
cmd_ln_str_r(config, "-b") ? cmd_ln_str_r(config, "-b") : "(null)");
cmd_ln_free_r(config);
return 0;
}
int
main(int32 argc, char *argv[])
{
/* kb_t kb;
ptmr_t tm; */
cmd_ln_appl_enter(argc, argv, "default.arg", defn);
config = cmd_ln_get();
logmath = logs3_init(cmd_ln_float64_r(config, "-logbase"), 1,
cmd_ln_int32_r(config, "-log3table"));
/* Read in input databases */
models_init();
ptmr_init(&tm_utt);
nbestdir = cmd_ln_str_r(config, "-nbestdir");
if (cmd_ln_str_r(config, "-ctl")) {
ctl_process(cmd_ln_str_r(config, "-ctl"),
cmd_ln_str_r(config, "-ctl_lm"),
NULL,
cmd_ln_int32_r(config, "-ctloffset"),
cmd_ln_int32_r(config, "-ctlcount"), utt_astar, NULL);
}
else {
E_FATAL("-ctl is not specified\n");
}
models_free();
logmath_free(logmath);
#if (! WIN32)
system("ps aguxwww | grep s3astar");
#endif
cmd_ln_free_r(config);
return 0;
}
void
ps_get_all_time(ps_decoder_t *ps, double *out_nspeech,
double *out_ncpu, double *out_nwall)
{
int32 frate;
frate = cmd_ln_int32_r(ps->config, "-frate");
*out_nspeech = (double)ps->n_frame / frate;
*out_ncpu = ps->perf.t_tot_cpu;
*out_nwall = ps->perf.t_tot_elapsed;
}
int
acmod_feat_mismatch(acmod_t *acmod, feat_t *fcb)
{
/* Feature type needs to be the same. */
if (0 != strcmp(cmd_ln_str_r(acmod->config, "-feat"), feat_name(fcb)))
return TRUE;
/* Input vector dimension needs to be the same. */
if (cmd_ln_int32_r(acmod->config, "-ceplen") != feat_cepsize(fcb))
return TRUE;
/* FIXME: Need to check LDA and stuff too. */
return FALSE;
}
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;
}
static int search_cb(search_t *search, search_event_t *evt, void *udata)
{
batch_decoder_t *bd = (batch_decoder_t *) udata;
dict_t *d = search_factory_d2p(bd->sf)->dict;
double delta = get_time_delta(bd);
double frate = cmd_ln_int32_r(search_config(search), "-frate");
FILE *hypfh = NULL;
void *val;
if (hash_table_lookup(bd->hypfiles, search_name(search), &val) == 0)
hypfh = val;
else
hypfh = stdout;
fprintf(hypfh, "time delta %f ", delta);
switch (evt->event) {
case SEARCH_PARTIAL_RESULT: {
int32 score;
seg_iter_t *seg = search_seg_iter(search, &score);
fprintf(hypfh, "partial: ");
for (; seg; seg = seg_iter_next(seg)) {
int sf, ef;
seg_iter_times(seg, &sf, &ef);
fprintf(hypfh, "%s:%.3f ", dict_basestr(d, seg_iter_wid(seg)),
(double) ef / frate);
}
fprintf(hypfh, "(%s)\n", search_uttid(search));
break;
}
case SEARCH_START_UTT:
fprintf(hypfh, "start %s\n", search_uttid(search));
break;
case SEARCH_END_UTT:
fprintf(hypfh, "end %s\n", search_uttid(search));
break;
case SEARCH_FINAL_RESULT: {
int32 score;
seg_iter_t *seg = search_seg_iter(search, &score);
fprintf(hypfh, "full: ");
for (; seg; seg = seg_iter_next(seg)) {
int sf, ef;
seg_iter_times(seg, &sf, &ef);
fprintf(hypfh, "%s:%.3f ", dict_basestr(d, seg_iter_wid(seg)),
(double) ef / frate);
}
fprintf(hypfh, "(%s)\n", search_uttid(search));
break;
}
}
return 0;
}
void
ngram_fwdflat_init(ngram_search_t *ngs)
{
int n_words;
n_words = ps_search_n_words(ngs);
ngs->fwdflat_wordlist = ckd_calloc(n_words + 1, sizeof(*ngs->fwdflat_wordlist));
ngs->expand_word_flag = bitvec_alloc(n_words);
ngs->expand_word_list = ckd_calloc(n_words + 1, sizeof(*ngs->expand_word_list));
ngs->frm_wordlist = ckd_calloc(ngs->n_frame_alloc, sizeof(*ngs->frm_wordlist));
ngs->min_ef_width = cmd_ln_int32_r(ps_search_config(ngs), "-fwdflatefwid");
ngs->max_sf_win = cmd_ln_int32_r(ps_search_config(ngs), "-fwdflatsfwin");
E_INFO("fwdflat: min_ef_width = %d, max_sf_win = %d\n",
ngs->min_ef_width, ngs->max_sf_win);
/* No tree-search; pre-build the expansion list, including all LM words. */
if (!ngs->fwdtree) {
/* Build full expansion list from LM words. */
ngram_fwdflat_expand_all(ngs);
/* Allocate single phone words. */
ngram_fwdflat_allocate_1ph(ngs);
}
}
static int batch_decoder_decode_mfc(batch_decoder_t *bd, FILE *infh, int sf,
int ef, alignment_t *al)
{
featbuf_t *fb = search_factory_featbuf(bd->sf);
mfcc_t **mfcs;
int nfr, rv;
if (NULL == (mfcs = read_mfc_file(infh, sf, ef, &nfr,
cmd_ln_int32_r(bd->config, "-ceplen"))))
return -1;
rv = featbuf_producer_process_cep(fb, mfcs, nfr, TRUE);
ckd_free_2d(mfcs);
return rv;
}
int
ps_set_search(ps_decoder_t *ps, const char *name)
{
ps_search_t *search = ps_find_search(ps, name);
if (search)
ps->search = search;
/* Set pl window depending on the search */
if (!strcmp(PS_SEARCH_NGRAM, ps_search_name(search))) {
ps->pl_window = cmd_ln_int32_r(ps->config, "-pl_window");
} else {
ps->pl_window = 0;
}
return search ? 0 : -1;
}
static void
print_word_times()
{
int frame_rate = cmd_ln_int32_r(config, "-frate");
ps_seg_t *iter = ps_seg_iter(ps);
while (iter != NULL) {
int32 sf, ef, pprob;
float conf;
ps_seg_frames(iter, &sf, &ef);
pprob = ps_seg_prob(iter, NULL, NULL, NULL);
conf = logmath_exp(ps_get_logmath(ps), pprob);
printf("%s %.3f %.3f %f\n", ps_seg_word(iter), ((float)sf / frame_rate),
((float) ef / frame_rate), conf);
iter = ps_seg_next(iter);
}
}
void ofApp::process_result()
{
int frame_rate = cmd_ln_int32_r(config, "-frate");
ps_seg_t *iter = ps_seg_iter(ps, NULL);
printf("\n\n");
while (iter != NULL)
{
int32 sf, ef, pprob;
float conf;
ps_seg_frames(iter, &sf, &ef);
pprob = ps_seg_prob(iter, NULL, NULL, NULL);
conf = logmath_exp(ps_get_logmath(ps), pprob);
//here is where we process the word
new_utterance new_utt;
new_utt.conf = conf;
new_utt.sf = sf;
new_utt.st = (float)sf / frame_rate;
new_utt.ef = ef;
new_utt.et = (float) ef / frame_rate;
new_utt.utt = ps_seg_word(iter);
printf("Recognised: %s %.3f %.3f %f\n", ps_seg_word(iter), new_utt.st,
new_utt.et, new_utt.conf);
std::string word = ps_seg_word(iter);
result.push_back(new_utt);
iter = ps_seg_next(iter);
}
printf("\n\n");
engineExit();
}
static void
test_set_search()
{
cmd_ln_t *config = default_config();
ps_decoder_t *ps = ps_init(config);
ps_search_iter_t *itor;
jsgf_t *jsgf = jsgf_parse_file(DATADIR "/goforward.gram", NULL);
fsg_model_t *fsg = jsgf_build_fsg(jsgf,
jsgf_get_rule(jsgf, "goforward.move"),
ps->lmath, cmd_ln_int32_r(config, "-lw"));
TEST_ASSERT(!ps_set_fsg(ps, "goforward", fsg));
fsg_model_free(fsg);
TEST_ASSERT(!ps_set_jsgf_file(ps, "goforward_other", DATADIR "/goforward.gram"));
ngram_model_t *lm = ngram_model_read(config, DATADIR "/tidigits/lm/tidigits.lm.dmp",
NGRAM_AUTO, ps->lmath);
TEST_ASSERT(!ps_set_lm(ps, "tidigits", lm));
ngram_model_free(lm);
TEST_ASSERT(!ps_set_search(ps, "tidigits"));
TEST_ASSERT(!ps_set_search(ps, "goforward"));
itor = ps_search_iter(ps);
TEST_EQUAL(0, strcmp("goforward_other", ps_search_iter_val(itor)));
itor = ps_search_iter_next(itor);
TEST_EQUAL(0, strcmp("tidigits", ps_search_iter_val(itor)));
itor = ps_search_iter_next(itor);
TEST_EQUAL(0, strcmp("goforward", ps_search_iter_val(itor)));
itor = ps_search_iter_next(itor);
TEST_EQUAL(0, strcmp("phone_loop", ps_search_iter_val(itor)));
itor = ps_search_iter_next(itor);
TEST_EQUAL(NULL, itor);
TEST_ASSERT(!ps_start_utt(ps));
TEST_ASSERT(!ps_end_utt(ps));
ps_free(ps);
cmd_ln_free_r(config);
}
static fwd_dbg_t *
init_fwd_dbg(srch_FLAT_FWD_graph_t * fwg)
{
const char *tmpstr;
fwd_dbg_t *fd;
fd = (fwd_dbg_t *) ckd_calloc(1, sizeof(fwd_dbg_t));
assert(fd);
/* Word to be traced in detail */
if ((tmpstr = cmd_ln_str_r(kbcore_config(fwg->kbcore), "-tracewhmm")) != NULL) {
fd->trace_wid = dict_wordid(fwg->kbcore->dict, tmpstr);
if (NOT_S3WID(fd->trace_wid))
E_ERROR("%s not in dictionary; cannot be traced\n", tmpstr);
}
else
fd->trace_wid = BAD_S3WID;
/* Active words to be dumped for debugging after and before the given frame nos, if any */
fd->word_dump_sf = (int32) 0x7ffffff0;
if (cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-worddumpsf"))
fd->word_dump_sf = cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-worddumpsf");
fd->word_dump_ef = (int32) 0x7ffffff0;
if (cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-worddumpef"))
fd->word_dump_ef = cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-worddumpef");
/* Active HMMs to be dumped for debugging after and before the given frame nos, if any */
fd->hmm_dump_sf = (int32) 0x7ffffff0;
if (cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-hmmdumpsf"))
fd->hmm_dump_sf = cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-hmmdumpsf");
fd->hmm_dump_ef = (int32) 0x7ffffff0;
if (cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-hmmdumpef"))
fd->hmm_dump_ef = cmd_ln_int32_r(kbcore_config(fwg->kbcore), "-hmmdumpef");
return fd;
}
void
ngram_fwdflat_deinit(ngram_search_t *ngs)
{
double n_speech = (double)ngs->n_tot_frame
/ cmd_ln_int32_r(ps_search_config(ngs), "-frate");
E_INFO("TOTAL fwdflat %.2f CPU %.3f xRT\n",
ngs->fwdflat_perf.t_tot_cpu,
ngs->fwdflat_perf.t_tot_cpu / n_speech);
E_INFO("TOTAL fwdflat %.2f wall %.3f xRT\n",
ngs->fwdflat_perf.t_tot_elapsed,
ngs->fwdflat_perf.t_tot_elapsed / n_speech);
/* Free single-phone words if we allocated them. */
if (!ngs->fwdtree) {
ngram_fwdflat_free_1ph(ngs);
}
ckd_free(ngs->fwdflat_wordlist);
bitvec_free(ngs->expand_word_flag);
ckd_free(ngs->expand_word_list);
ckd_free(ngs->frm_wordlist);
}
int
s3_decode_init(s3_decode_t * _decode, cmd_ln_t *_config)
{
if (_decode == NULL)
return S3_DECODE_ERROR_NULL_POINTER;
/* capture decoder parameters */
kb_init(&_decode->kb, _config);
/* initialize decoder variables */
_decode->kbcore = _decode->kb.kbcore;
_decode->hyp_frame_num = -1;
_decode->uttid = NULL;
_decode->state = S3_DECODE_STATE_IDLE;
_decode->hyp_str = NULL;
_decode->hyp_segs = NULL;
_decode->swap =
strcmp(cmd_ln_str_r(_config,"-machine_endian"),
cmd_ln_str_r(_config,"-input_endian"));
if (_decode->swap)
E_INFO("Input data WILL be byte swapped\n");
else
E_INFO("Input data will NOT be byte swapped\n");
_decode->phypdump = (cmd_ln_int32_r(_config, "-phypdump"));
if (_decode->phypdump)
E_INFO("Partial hypothesis WILL be dumped\n");
else
E_INFO("Partial hypothesis will NOT be dumped\n");
_decode->rawext = (cmd_ln_str_r(_config, "-rawext"));
return S3_DECODE_SUCCESS;
}
void
ngram_fwdflat_finish(ngram_search_t *ngs)
{
int32 cf;
destroy_fwdflat_chan(ngs);
destroy_fwdflat_wordlist(ngs);
bitvec_clear_all(ngs->word_active, ps_search_n_words(ngs));
/* This is the number of frames processed. */
cf = ps_search_acmod(ngs)->output_frame;
/* Add a mark in the backpointer table for one past the final frame. */
ngram_search_mark_bptable(ngs, cf);
ptmr_stop(&ngs->fwdflat_perf);
/* Print out some statistics. */
if (cf > 0) {
double n_speech = (double)(cf + 1)
/ cmd_ln_int32_r(ps_search_config(ngs), "-frate");
E_INFO("%8d words recognized (%d/fr)\n",
ngs->bpidx, (ngs->bpidx + (cf >> 1)) / (cf + 1));
E_INFO("%8d senones evaluated (%d/fr)\n", ngs->st.n_senone_active_utt,
(ngs->st.n_senone_active_utt + (cf >> 1)) / (cf + 1));
E_INFO("%8d channels searched (%d/fr)\n",
ngs->st.n_fwdflat_chan, ngs->st.n_fwdflat_chan / (cf + 1));
E_INFO("%8d words searched (%d/fr)\n",
ngs->st.n_fwdflat_words, ngs->st.n_fwdflat_words / (cf + 1));
E_INFO("%8d word transitions (%d/fr)\n",
ngs->st.n_fwdflat_word_transition,
ngs->st.n_fwdflat_word_transition / (cf + 1));
E_INFO("fwdflat %.2f CPU %.3f xRT\n",
ngs->fwdflat_perf.t_cpu,
ngs->fwdflat_perf.t_cpu / n_speech);
E_INFO("fwdflat %.2f wall %.3f xRT\n",
ngs->fwdflat_perf.t_elapsed,
ngs->fwdflat_perf.t_elapsed / n_speech);
}
