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;
}
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;
}
static int
process(sbthread_t *th)
{
FILE *raw;
int16 *buf;
mfcc_t **cepbuf;
size_t nsamps;
fe_t *fe;
long fsize;
int32 nfr;
char outfile[16];
FILE *logfile;
sprintf(outfile, "%03ld.log", (long) sbthread_arg(th));
logfile = fopen(outfile, "w");
pthread_setspecific(logfp_index, (void *)logfile);
if ((fe = fe_init_auto_r(sbthread_config(th))) == NULL)
return -1;
if ((raw = fopen(TESTDATADIR "/chan3.raw", "rb")) == NULL)
return -1;
fseek(raw, 0, SEEK_END);
fsize = ftell(raw);
fseek(raw, 0, SEEK_SET);
buf = ckd_malloc(fsize);
fread(buf, 1, fsize, raw);
nsamps = fsize / 2;
fe_process_utt(fe, buf, nsamps, &cepbuf, &nfr);
E_INFO("nfr = %d\n", nfr);
fe_free_2d(cepbuf);
ckd_free(buf);
fclose(raw);
fe_free(fe);
fclose(logfile);
return 0;
}
int
main(int argc, char *argv[])
{
char const *cfg;
int i;
int16 buf[2048];
if (argc == 2) {
config = cmd_ln_parse_file_r(NULL, cont_args_def, argv[1], TRUE);
}
else {
config = cmd_ln_parse_r(NULL, cont_args_def, argc, argv, FALSE);
}
/* Handle argument file as -argfile. */
if (config && (cfg = cmd_ln_str_r(config, "-argfile")) != NULL) {
config = cmd_ln_parse_file_r(config, cont_args_def, cfg, FALSE);
}
if (config == NULL)
return 1;
singlefile = cmd_ln_boolean_r(config, "-singlefile");
if ((infile_path = cmd_ln_str_r(config, "-infile")) != NULL) {
if ((infile = fopen(infile_path, "rb")) == NULL) {
E_FATAL_SYSTEM("Failed to read audio from '%s'", infile_path);
return 1;
}
read_audio = &read_audio_file;
/* skip wav header */
read_audio(buf, 44);
}
else {
if ((ad = ad_open_dev(cmd_ln_str_r(config, "-adcdev"),
(int) cmd_ln_float32_r(config,
"-samprate"))) ==
NULL) {
E_FATAL("Failed to open audio device\n");
return 1;
}
read_audio = &read_audio_adev;
printf("Start recording ...\n");
fflush(stdout);
if (ad_start_rec(ad) < 0)
E_FATAL("Failed to start recording\n");
/* TODO remove this thing */
for (i = 0; i < 5; i++) {
sleep_msec(200);
read_audio(buf, 2048);
}
printf("You may speak now\n");
fflush(stdout);
}
fe = fe_init_auto_r(config);
if (fe == NULL)
return 1;
segment_audio();
if (ad)
ad_close(ad);
if (infile)
fclose(infile);
fe_free(fe);
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;
}
int ofxSphinxASR::engineInit(ofAsrEngineArgs *e)
{
#if defined TARGET_WIN32
char cfg_filename[] = "sphinx.cfg";
char grammarJSGF_filename[] = "grammar.jsgf";
char grammarFSG_filename[] = "grammar.fsg";
#else
char cfg_filename[] = "/tmp/sphinx.cfg";
char grammarJSGF_filename[] = "/tmp/grammar.jsgf";
char grammarFSG_filename[] = "/tmp/grammar.fsg";
#endif
FILE *cfg_fp = fopen(cfg_filename, "wt");
if (cfg_fp==NULL)
return OFXASR_FAIL_WRITE_CONFIG;
if ( access(e->sphinxmodel_am.c_str(), 0) != 0 )
return OFXASR_FAIL_READ_FILES;
if ( access(e->sphinxmodel_lm.c_str(), 0) != 0 )
return OFXASR_FAIL_READ_FILES;
if ( access(e->sphinxmodel_dict.c_str(), 0) != 0 )
return OFXASR_FAIL_READ_FILES;
if ( access(e->sphinxmodel_fdict.c_str(), 0) != 0 )
return OFXASR_FAIL_READ_FILES;
char cur_path[1024];
getcwd(cur_path, 1024);
fprintf(cfg_fp, "-samprate %d\n", e->samplerate);
fprintf(cfg_fp, "-hmm %s/%s\n", cur_path, e->sphinxmodel_am.c_str());
fprintf(cfg_fp, "-dict %s/%s\n", cur_path, e->sphinxmodel_dict.c_str());
fprintf(cfg_fp, "-fdict %s/%s\n", cur_path, e->sphinxmodel_fdict.c_str());
fprintf(cfg_fp, "-lm %s/%s\n", cur_path, e->sphinxmodel_lm.c_str());
if (e->sphinx_mode == 2) {
if(e->sphinx_candidate_sentences.size() < 1) {
printf("Warning: The word list is empty! Use mode 4.\n");
e->sphinx_mode = 4;
}
else {
FILE *gram_fp = fopen(grammarJSGF_filename, "wt");
if (gram_fp==NULL)
return OFXASR_FAIL_WRITE_CONFIG;
fprintf(gram_fp,
"#JSGF V1.0;\n\ngrammar cca_gram;\n\npublic <cca_gram> = (\n");
for (int i=0; i<e->sphinx_candidate_sentences.size()-1; i++) {
fprintf(gram_fp, "%s |\n",
e->sphinx_candidate_sentences[i].c_str());
}
fprintf(gram_fp, "%s );\n\n",
e->sphinx_candidate_sentences[e->sphinx_candidate_sentences.size()-1].c_str());
fclose(gram_fp);
}
jsgf_t *jsgf = jsgf_parse_file(grammarJSGF_filename, NULL);
if (jsgf == NULL) {
printf("Bad jsgf file %s.\n", grammarJSGF_filename);
return OFXASR_INVALID_JSGF_GRAMMAR;
}
fsg_model_t *fsg = get_fsg(jsgf, NULL);
fsg_model_writefile(fsg, grammarFSG_filename);
fsg_model_free(fsg);
jsgf_grammar_free(jsgf);
fprintf(cfg_fp, "-fsg %s\n", grammarFSG_filename);
}
fprintf(cfg_fp, "-op_mode %d\n", e->sphinx_mode);
fclose(cfg_fp);
err_set_logfp(NULL); // disable logs
cmd_ln_t *config = NULL;
config = cmd_ln_parse_file_r(config, S3_DECODE_ARG_DEFS, cfg_filename, TRUE);
if (config == NULL) {
return OFXASR_INVALID_CONFIG;
}
decoder = new s3_decode_t;
if (s3_decode_init(decoder, config) != S3_DECODE_SUCCESS) {
return OFXASR_FAIL_INIT_DECODER;
}
fe = fe_init_auto_r(config);
if (fe == NULL) {
return OFXASR_FAIL_INIT_FRONTEND;
}
bEngineInitialed = true;
return OFXASR_SUCCESS;
}
void
kb_init(kb_t * kb, cmd_ln_t *config)
{
kbcore_t *kbcore;
mdef_t *mdef;
dict_t *dict;
dict2pid_t *d2p;
int32 cisencnt;
/* STRUCTURE: Initialize the kb structure to zero, just in case */
memset(kb, 0, sizeof(*kb));
kb->kbcore = kbcore_init(config);
if (kb->kbcore == NULL)
E_FATAL("Initialization of kb failed\n");
kbcore = kb->kbcore;
mdef = kbcore_mdef(kbcore);
dict = kbcore_dict(kbcore);
d2p = kbcore_dict2pid(kbcore);
err_set_debug_level(cmd_ln_int32_r(config, "-debug"));
/* STRUCTURE INITIALIZATION: Initialize the beam data structure */
if (cmd_ln_exists_r(config, "-ptranskip")) {
kb->beam = beam_init(cmd_ln_float64_r(config, "-beam"),
cmd_ln_float64_r(config, "-pbeam"),
cmd_ln_float64_r(config, "-wbeam"),
cmd_ln_float64_r(config, "-wend_beam"),
cmd_ln_int32_r(config, "-ptranskip"), mdef_n_ciphone(mdef),
kbcore->logmath
);
/* REPORT : Report the parameters in the beam data structure */
if (REPORT_KB)
beam_report(kb->beam);
}
/* STRUCTURE INITIALIZATION: Initialize the fast GMM computation data structure */
if (cmd_ln_exists_r(config, "-ci_pbeam")) {
kb->fastgmm = fast_gmm_init(cmd_ln_int32_r(config, "-ds"),
cmd_ln_int32_r(config, "-cond_ds"),
cmd_ln_int32_r(config, "-dist_ds"),
cmd_ln_int32_r(config, "-gs4gs"),
cmd_ln_int32_r(config, "-svq4svq"),
cmd_ln_float64_r(config, "-subvqbeam"),
cmd_ln_float64_r(config, "-ci_pbeam"),
cmd_ln_float64_r(config, "-tighten_factor"),
cmd_ln_int32_r(config, "-maxcdsenpf"),
mdef->n_ci_sen,
kbcore->logmath);
/* REPORT : Report the parameters in the fast_gmm_t data struture */
if (REPORT_KB)
fast_gmm_report(kb->fastgmm);
}
/* STRUCTURE INITIALIZATION: Initialize the phoneme lookahead data structure */
if (cmd_ln_exists_r(config, "-pl_beam")) {
kb->pl = pl_init(cmd_ln_int32_r(config, "-pheurtype"),
cmd_ln_float64_r(config, "-pl_beam"), mdef_n_ciphone(mdef),
kbcore->logmath
);
/* REPORT : Report the parameters in the pl_t data struture */
if (REPORT_KB)
pl_report(kb->pl);
}
/* STRUCTURE INITIALIZATION: Initialize the acoustic score data structure */
{
int32 pl_window = 1;
if (cmd_ln_exists_r(config, "-pl_window"))
pl_window = cmd_ln_int32_r(config, "-pl_window");
for (cisencnt = 0; cisencnt == mdef->cd2cisen[cisencnt]; cisencnt++) ;
kb->ascr = ascr_init(kbcore_n_mgau(kbcore),
kb->kbcore->dict2pid->n_comstate,
mdef_n_sseq(mdef),
dict2pid_n_comsseq(d2p),
pl_window, cisencnt);
if (REPORT_KB)
ascr_report(kb->ascr);
}
/* Initialize the front end if -adcin is specified */
if (cmd_ln_exists_r(config, "-adcin") && cmd_ln_boolean_r(config, "-adcin")) {
if ((kb->fe = fe_init_auto_r(config)) == NULL) {
E_FATAL("fe_init_auto_r() failed\n");
}
}
/* STRUCTURE INITIALIZATION : The feature vector */
if ((kb->feat =
feat_array_alloc(kbcore_fcb(kbcore), S3_MAX_FRAMES)) == NULL)
E_FATAL("feat_array_alloc() failed\n");
/* STRUCTURE INITIALIZATION : The statistics for the search */
kb->stat = stat_init();
/* STRUCTURE INITIALIZATION : The adaptation routines of the search */
kb->adapt_am = adapt_am_init();
if (cmd_ln_str_r(config, "-mllr")) {
kb_setmllr(cmd_ln_str_r(config, "-mllr"), cmd_ln_str_r(config, "-cb2mllr"), kb);
}
/* CHECK: make sure when (-cond_ds) is specified, a Gaussian map is also specified */
if (cmd_ln_int32_r(config, "-cond_ds") > 0 && kb->kbcore->gs == NULL)
E_FATAL
("Conditional Down Sampling require the use of Gaussian Selection map\n");
/* MEMORY ALLOCATION : Word best score and exit */
/* Open hypseg file if specified */
kb->matchsegfp = kb->matchfp = NULL;
kb->matchsegfp = file_open(cmd_ln_str_r(config, "-hypseg"));
kb->matchfp = file_open(cmd_ln_str_r(config, "-hyp"));
if (cmd_ln_exists_r(config, "-hmmdump"))
kb->hmmdumpfp = cmd_ln_int32_r(config, "-hmmdump") ? stderr : NULL;
/* STRUCTURE INITIALIZATION : The search data structure, done only
after kb is initialized kb is acted as a clipboard. */
if (cmd_ln_exists_r(config, "-op_mode")) {
/* -op_mode, if set (i.e. not -1), takes precedence over -mode. */
if (cmd_ln_int32_r(config, "-op_mode") != -1)
kb->op_mode = cmd_ln_int32_r(config, "-op_mode");
else
kb->op_mode = srch_mode_str_to_index(cmd_ln_str_r(config, "-mode"));
E_INFO("SEARCH MODE INDEX %d\n", kb->op_mode);
if ((kb->srch = (srch_t *) srch_init(kb, kb->op_mode)) == NULL) {
E_FATAL("Search initialization failed. Forced exit\n");
}
if (REPORT_KB) {
srch_report(kb->srch);
}
}
}
int
main(int argc, char *argv[])
{
static const arg_t fe_args[] = {
waveform_to_cepstral_command_line_macro(),
{ NULL, 0, NULL, NULL }
};
FILE *raw;
cmd_ln_t *config;
fe_t *fe;
int16 buf[1024];
int16 const *inptr;
int32 frame_shift, frame_size;
mfcc_t **cepbuf1, **cepbuf2, **cptr;
int32 nfr, i;
size_t nsamp;
TEST_ASSERT(config = cmd_ln_parse_r(NULL, fe_args, argc, argv, FALSE));
TEST_ASSERT(fe = fe_init_auto_r(config));
TEST_EQUAL(fe_get_output_size(fe), DEFAULT_NUM_CEPSTRA);
fe_get_input_size(fe, &frame_shift, &frame_size);
TEST_EQUAL(frame_shift, DEFAULT_FRAME_SHIFT);
TEST_EQUAL(frame_size, (int)(DEFAULT_WINDOW_LENGTH*DEFAULT_SAMPLING_RATE));
TEST_ASSERT(raw = fopen(TESTDATADIR "/chan3.raw", "rb"));
TEST_EQUAL(0, fe_start_utt(fe));
TEST_EQUAL(1024, fread(buf, sizeof(int16), 1024, raw));
nsamp = 1024;
TEST_ASSERT(fe_process_frames(fe, NULL, &nsamp, NULL, &nfr) >= 0);
TEST_EQUAL(1024, nsamp);
TEST_EQUAL(4, nfr);
cepbuf1 = ckd_calloc_2d(5, DEFAULT_NUM_CEPSTRA, sizeof(**cepbuf1));
inptr = &buf[0];
nfr = 1;
printf("frame_size %d frame_shift %d\n", frame_size, frame_shift);
/* Process the first frame. */
TEST_ASSERT(fe_process_frames(fe, &inptr, &nsamp, &cepbuf1[0], &nfr) >= 0);
printf("inptr %d nsamp %d nfr %d\n", inptr - buf, nsamp, nfr);
TEST_EQUAL(nfr, 1);
/* Note that this next one won't actually consume any frames
* of input, because it already got sufficient overflow
* samples last time around. This is implementation-dependent
* so we shouldn't actually test for it. */
TEST_ASSERT(fe_process_frames(fe, &inptr, &nsamp, &cepbuf1[1], &nfr) >= 0);
printf("inptr %d nsamp %d nfr %d\n", inptr - buf, nsamp, nfr);
TEST_EQUAL(nfr, 1);
TEST_ASSERT(fe_process_frames(fe, &inptr, &nsamp, &cepbuf1[2], &nfr) >= 0);
printf("inptr %d nsamp %d nfr %d\n", inptr - buf, nsamp, nfr);
TEST_EQUAL(nfr, 1);
TEST_ASSERT(fe_process_frames(fe, &inptr, &nsamp, &cepbuf1[3], &nfr) >= 0);
printf("inptr %d nsamp %d nfr %d\n", inptr - buf, nsamp, nfr);
TEST_EQUAL(nfr, 1);
TEST_ASSERT(fe_end_utt(fe, cepbuf1[4], &nfr) >= 0);
printf("nfr %d\n", nfr);
TEST_EQUAL(nfr, 1);
/* What we *should* test is that the output we get by
* processing one frame at a time is exactly the same as what
* we get from doing them all at once. So let's do that */
cepbuf2 = ckd_calloc_2d(5, DEFAULT_NUM_CEPSTRA, sizeof(**cepbuf2));
inptr = &buf[0];
nfr = 5;
nsamp = 1024;
TEST_EQUAL(0, fe_start_utt(fe));
TEST_ASSERT(fe_process_frames(fe, &inptr, &nsamp, cepbuf2, &nfr) >= 0);
printf("nfr %d\n", nfr);
TEST_EQUAL(nfr, 4);
nfr = 1;
TEST_ASSERT(fe_end_utt(fe, cepbuf2[4], &nfr) >= 0);
printf("nfr %d\n", nfr);
TEST_EQUAL(nfr, 1);
for (i = 0; i < 5; ++i) {
int j;
printf("%d: ", i);
for (j = 0; j < DEFAULT_NUM_CEPSTRA; ++j) {
printf("%.2f,%.2f ",
MFCC2FLOAT(cepbuf1[i][j]),
MFCC2FLOAT(cepbuf2[i][j]));
TEST_EQUAL_FLOAT(cepbuf1[i][j], cepbuf2[i][j]);
}
printf("\n");
}
/* Now, also test to make sure that even if we feed data in
* little tiny bits we can still make things work. */
memset(cepbuf2[0], 0, 5 * DEFAULT_NUM_CEPSTRA * sizeof(**cepbuf2));
inptr = &buf[0];
cptr = &cepbuf2[0];
nfr = 5;
i = 5;
nsamp = 256;
TEST_EQUAL(0, fe_start_utt(fe));
TEST_ASSERT(fe_process_frames(fe, &inptr, &nsamp, cptr, &i) >= 0);
printf("inptr %d nsamp %d nfr %d\n", inptr - buf, nsamp, i);
cptr += i;
nfr -= i;
i = nfr;
nsamp = 256;
TEST_ASSERT(fe_process_frames(fe, &inptr, &nsamp, cptr, &i) >= 0);
printf("inptr %d nsamp %d nfr %d\n", inptr - buf, nsamp, i);
cptr += i;
nfr -= i;
i = nfr;
nsamp = 256;
TEST_ASSERT(fe_process_frames(fe, &inptr, &nsamp, cptr, &i) >= 0);
printf("inptr %d nsamp %d nfr %d\n", inptr - buf, nsamp, i);
cptr += i;
nfr -= i;
i = nfr;
nsamp = 256;
TEST_ASSERT(fe_process_frames(fe, &inptr, &nsamp, cptr, &i) >= 0);
printf("inptr %d nsamp %d nfr %d\n", inptr - buf, nsamp, i);
cptr += i;
nfr -= i;
TEST_ASSERT(fe_end_utt(fe, *cptr, &nfr) >= 0);
printf("nfr %d\n", nfr);
TEST_EQUAL(nfr, 1);
for (i = 0; i < 5; ++i) {
int j;
printf("%d: ", i);
for (j = 0; j < DEFAULT_NUM_CEPSTRA; ++j) {
printf("%.2f,%.2f ",
MFCC2FLOAT(cepbuf1[i][j]),
MFCC2FLOAT(cepbuf2[i][j]));
TEST_EQUAL_FLOAT(cepbuf1[i][j], cepbuf2[i][j]);
}
printf("\n");
}
/* And now, finally, test fe_process_utt() */
inptr = &buf[0];
i = 0;
TEST_EQUAL(0, fe_start_utt(fe));
TEST_ASSERT(fe_process_utt(fe, inptr, 256, &cptr, &nfr) >= 0);
printf("i %d nfr %d\n", i, nfr);
if (nfr)
memcpy(cepbuf2[i], cptr[0], nfr * DEFAULT_NUM_CEPSTRA * sizeof(**cptr));
ckd_free_2d(cptr);
i += nfr;
inptr += 256;
TEST_ASSERT(fe_process_utt(fe, inptr, 256, &cptr, &nfr) >= 0);
printf("i %d nfr %d\n", i, nfr);
if (nfr)
memcpy(cepbuf2[i], cptr[0], nfr * DEFAULT_NUM_CEPSTRA * sizeof(**cptr));
ckd_free_2d(cptr);
i += nfr;
inptr += 256;
TEST_ASSERT(fe_process_utt(fe, inptr, 256, &cptr, &nfr) >= 0);
printf("i %d nfr %d\n", i, nfr);
if (nfr)
memcpy(cepbuf2[i], cptr[0], nfr * DEFAULT_NUM_CEPSTRA * sizeof(**cptr));
ckd_free_2d(cptr);
i += nfr;
inptr += 256;
TEST_ASSERT(fe_process_utt(fe, inptr, 256, &cptr, &nfr) >= 0);
printf("i %d nfr %d\n", i, nfr);
if (nfr)
memcpy(cepbuf2[i], cptr[0], nfr * DEFAULT_NUM_CEPSTRA * sizeof(**cptr));
ckd_free_2d(cptr);
i += nfr;
inptr += 256;
TEST_ASSERT(fe_end_utt(fe, cepbuf2[i], &nfr) >= 0);
printf("i %d nfr %d\n", i, nfr);
TEST_EQUAL(nfr, 1);
for (i = 0; i < 5; ++i) {
int j;
printf("%d: ", i);
for (j = 0; j < DEFAULT_NUM_CEPSTRA; ++j) {
printf("%.2f,%.2f ",
MFCC2FLOAT(cepbuf1[i][j]),
MFCC2FLOAT(cepbuf2[i][j]));
TEST_EQUAL_FLOAT(cepbuf1[i][j], cepbuf2[i][j]);
}
printf("\n");
}
ckd_free_2d(cepbuf1);
ckd_free_2d(cepbuf2);
fclose(raw);
fe_free(fe);
return 0;
}
fe_t *
fe_init_auto()
{
return fe_init_auto_r(cmd_ln_get());
}
int
sphinx_wave2feat_convert_file(sphinx_wave2feat_t *wtf,
char const *infile, char const *outfile)
{
int nchans, minfft, nfft, nfloat, veclen;
audio_type_t const *atype;
int fshift, fsize;
if (cmd_ln_boolean_r(wtf->config, "-verbose"))
E_INFO("Converting %s to %s\n", infile, outfile);
wtf->infile = ckd_salloc(infile);
/* Detect input file type. */
if ((atype = detect_audio_type(wtf)) == NULL)
return -1;
/* Determine whether to byteswap input. */
wtf->byteswap = strcmp(cmd_ln_str_r(wtf->config, "-mach_endian"),
cmd_ln_str_r(wtf->config, "-input_endian"));
/* Make sure the FFT size is sufficiently large. */
minfft = (int)(cmd_ln_float32_r(wtf->config, "-samprate")
* cmd_ln_float32_r(wtf->config, "-wlen") + 0.5);
for (nfft = 1; nfft < minfft; nfft <<= 1)
;
if (nfft > cmd_ln_int32_r(wtf->config, "-nfft")) {
E_WARN("Value of -nfft = %d is too small, increasing to %d\n",
cmd_ln_int32_r(wtf->config, "-nfft"), nfft);
cmd_ln_set_int32_r(wtf->config, "-nfft", nfft);
fe_free(wtf->fe);
wtf->fe = fe_init_auto_r(wtf->config);
}
/* Get the output frame size (if not already set). */
if (wtf->veclen == 0)
wtf->veclen = fe_get_output_size(wtf->fe);
/* Set up the input and output buffers. */
fe_get_input_size(wtf->fe, &fshift, &fsize);
/* Want to get at least a whole frame plus shift in here. Also we
will either pick or mix multiple channels so we need to read
them all at once. */
nchans = cmd_ln_int32_r(wtf->config, "-nchans");
wtf->blocksize = cmd_ln_int32_r(wtf->config, "-blocksize") * nchans;
if (wtf->blocksize < (fsize + fshift) * nchans) {
E_INFO("Block size of %d too small, increasing to %d\n",
wtf->blocksize,
(fsize + fshift) * nchans);
wtf->blocksize = (fsize + fshift) * nchans;
}
wtf->audio = (short *)ckd_calloc(wtf->blocksize, sizeof(*wtf->audio));
wtf->featsize = (wtf->blocksize / nchans - fsize) / fshift;
/* Use the maximum of the input and output frame sizes to allocate this. */
veclen = wtf->veclen;
if (wtf->in_veclen > veclen) veclen = wtf->in_veclen;
wtf->feat = (mfcc_t**)ckd_calloc_2d(wtf->featsize, veclen, sizeof(**wtf->feat));
/* Let's go! */
if ((wtf->outfh = fopen(outfile, "wb")) == NULL) {
E_ERROR_SYSTEM("Failed to open %s for writing", outfile);
return -1;
}
/* Write an empty header, which we'll fill in later. */
if (wtf->ot->output_header &&
(*wtf->ot->output_header)(wtf, 0) < 0) {
E_ERROR_SYSTEM("Failed to write empty header to %s\n", outfile);
goto error_out;
}
wtf->outfile = ckd_salloc(outfile);
if ((nfloat = (*atype->decode)(wtf)) < 0) {
E_ERROR("Failed to convert");
goto error_out;
}
if (wtf->ot->output_header) {
if (fseek(wtf->outfh, 0, SEEK_SET) < 0) {
E_ERROR_SYSTEM("Failed to seek to beginning of %s\n", outfile);
goto error_out;
}
if ((*wtf->ot->output_header)(wtf, nfloat) < 0) {
E_ERROR_SYSTEM("Failed to write header to %s\n", outfile);
goto error_out;
}
}
if (wtf->audio)
ckd_free(wtf->audio);
if (wtf->feat)
ckd_free_2d(wtf->feat);
if (wtf->infile)
ckd_free(wtf->infile);
if (wtf->outfile)
ckd_free(wtf->outfile);
wtf->audio = NULL;
wtf->infile = NULL;
wtf->feat = NULL;
wtf->outfile = NULL;
if (wtf->outfh)
if (fclose(wtf->outfh) == EOF)
E_ERROR_SYSTEM("Failed to close output file");
wtf->outfh = NULL;
return 0;
error_out:
if (wtf->audio)
ckd_free(wtf->audio);
if (wtf->feat)
ckd_free_2d(wtf->feat);
if (wtf->infile)
ckd_free(wtf->infile);
if (wtf->outfile)
ckd_free(wtf->outfile);
wtf->audio = NULL;
wtf->infile = NULL;
wtf->feat = NULL;
wtf->outfile = NULL;
if (wtf->outfh)
if (fclose(wtf->outfh) == EOF)
E_ERROR_SYSTEM("Failed to close output file");
wtf->outfh = NULL;
return -1;
}
static void
models_init(cmd_ln_t *config)
{
int32 cisencnt;
kbc = New_kbcore(config);
kbc->logmath = logs3_init(cmd_ln_float64_r(config, "-logbase"), 1,
cmd_ln_int32_r(config, "-log3table"));
/* Initialize feaure stream type */
kbc->fcb = feat_init(cmd_ln_str_r(config, "-feat"),
cmn_type_from_str(cmd_ln_str_r(config, "-cmn")),
cmd_ln_boolean_r(config, "-varnorm"),
agc_type_from_str(cmd_ln_str_r(config, "-agc")), 1,
cmd_ln_int32_r(config, "-ceplen"));
s3_am_init(kbc);
/* Initialize the front end if -adcin is specified */
if (cmd_ln_exists_r(config, "-adcin") && cmd_ln_boolean_r(config, "-adcin")) {
if ((fe = fe_init_auto_r(config)) == NULL) {
E_FATAL("fe_init_auto_r() failed\n");
}
}
assert(kbc);
assert(kbc->mdef);
assert(kbc->tmat);
/* Dictionary */
dict = dict_init(kbc->mdef, cmd_ln_str_r(config, "-dict"),
cmd_ln_str_r(config, "-fdict"),
cmd_ln_int32_r(config, "-lts_mismatch"),
cmd_ln_boolean_r(config, "-mdef_fillers"),
/* Never do mdef filler phones. */
FALSE,
TRUE);
for (cisencnt = 0; cisencnt == kbc->mdef->cd2cisen[cisencnt];
cisencnt++);
ascr = ascr_init(kbc->mdef->n_sen, 0, /* No composite senone */
mdef_n_sseq(kbc->mdef), 0, /* No composite senone sequence */
1, /* Phoneme lookahead window =1. Not enabled phoneme lookahead at this moment */
cisencnt);
fastgmm = fast_gmm_init(cmd_ln_int32_r(config, "-ds"),
cmd_ln_int32_r(config, "-cond_ds"),
cmd_ln_int32_r(config, "-dist_ds"),
cmd_ln_int32_r(config, "-gs4gs"),
cmd_ln_int32_r(config, "-svq4svq"),
cmd_ln_float64_r(config, "-subvqbeam"),
cmd_ln_float64_r(config, "-ci_pbeam"),
cmd_ln_float64_r(config, "-tighten_factor"),
cmd_ln_int32_r(config, "-maxcdsenpf"),
kbc->mdef->n_ci_sen,
kbc->logmath);
adapt_am = adapt_am_init();
}
