int
ps_decoder_test(cmd_ln_t *config, char const *sname, char const *expected)
{
ps_decoder_t *ps;
mfcc_t **cepbuf;
FILE *rawfh;
int16 *buf;
int16 const *bptr;
size_t nread;
size_t nsamps;
int32 nfr, i, score, prob;
char const *hyp;
char const *uttid;
double n_speech, n_cpu, n_wall;
ps_seg_t *seg;
TEST_ASSERT(ps = ps_init(config));
/* Test it first with pocketsphinx_decode_raw() */
TEST_ASSERT(rawfh = fopen(DATADIR "/goforward.raw", "rb"));
ps_decode_raw(ps, rawfh, "goforward", -1);
hyp = ps_get_hyp(ps, &score, &uttid);
prob = ps_get_prob(ps, &uttid);
printf("%s (%s): %s (%d, %d)\n", sname, uttid, hyp, score, prob);
TEST_EQUAL(0, strcmp(hyp, expected));
TEST_ASSERT(prob <= 0);
ps_get_utt_time(ps, &n_speech, &n_cpu, &n_wall);
printf("%.2f seconds speech, %.2f seconds CPU, %.2f seconds wall\n",
n_speech, n_cpu, n_wall);
printf("%.2f xRT (CPU), %.2f xRT (elapsed)\n",
n_cpu / n_speech, n_wall / n_speech);
/* Test it with ps_process_raw() */
clearerr(rawfh);
fseek(rawfh, 0, SEEK_END);
nsamps = ftell(rawfh) / sizeof(*buf);
fseek(rawfh, 0, SEEK_SET);
TEST_EQUAL(0, ps_start_utt(ps, NULL));
nsamps = 2048;
buf = ckd_calloc(nsamps, sizeof(*buf));
while (!feof(rawfh)) {
nread = fread(buf, sizeof(*buf), nsamps, rawfh);
ps_process_raw(ps, buf, nread, FALSE, FALSE);
}
TEST_EQUAL(0, ps_end_utt(ps));
hyp = ps_get_hyp(ps, &score, &uttid);
prob = ps_get_prob(ps, &uttid);
printf("%s (%s): %s (%d, %d)\n", sname, uttid, hyp, score, prob);
TEST_EQUAL(0, strcmp(uttid, "000000000"));
TEST_EQUAL(0, strcmp(hyp, expected));
ps_get_utt_time(ps, &n_speech, &n_cpu, &n_wall);
printf("%.2f seconds speech, %.2f seconds CPU, %.2f seconds wall\n",
n_speech, n_cpu, n_wall);
printf("%.2f xRT (CPU), %.2f xRT (elapsed)\n",
n_cpu / n_speech, n_wall / n_speech);
/* Now read the whole file and produce an MFCC buffer. */
clearerr(rawfh);
fseek(rawfh, 0, SEEK_END);
nsamps = ftell(rawfh) / sizeof(*buf);
fseek(rawfh, 0, SEEK_SET);
bptr = buf = ckd_realloc(buf, nsamps * sizeof(*buf));
TEST_EQUAL(nsamps, fread(buf, sizeof(*buf), nsamps, rawfh));
fe_process_frames(ps->acmod->fe, &bptr, &nsamps, NULL, &nfr);
cepbuf = ckd_calloc_2d(nfr + 1,
fe_get_output_size(ps->acmod->fe),
sizeof(**cepbuf));
fe_start_utt(ps->acmod->fe);
fe_process_frames(ps->acmod->fe, &bptr, &nsamps, cepbuf, &nfr);
fe_end_utt(ps->acmod->fe, cepbuf[nfr], &i);
/* Decode it with process_cep() */
TEST_EQUAL(0, ps_start_utt(ps, NULL));
for (i = 0; i < nfr; ++i) {
ps_process_cep(ps, cepbuf + i, 1, FALSE, FALSE);
}
TEST_EQUAL(0, ps_end_utt(ps));
hyp = ps_get_hyp(ps, &score, &uttid);
prob = ps_get_prob(ps, &uttid);
printf("%s (%s): %s (%d, %d)\n", sname, uttid, hyp, score, prob);
TEST_EQUAL(0, strcmp(uttid, "000000001"));
TEST_EQUAL(0, strcmp(hyp, expected));
TEST_ASSERT(prob <= 0);
for (seg = ps_seg_iter(ps, &score); seg;
seg = ps_seg_next(seg)) {
char const *word;
int sf, ef;
int32 post, lscr, ascr, lback;
word = ps_seg_word(seg);
ps_seg_frames(seg, &sf, &ef);
post = ps_seg_prob(seg, &ascr, &lscr, &lback);
printf("%s (%d:%d) P(w|o) = %f ascr = %d lscr = %d lback = %d\n", word, sf, ef,
logmath_exp(ps_get_logmath(ps), post), ascr, lscr, lback);
TEST_ASSERT(post <= 2); // Due to numerical errors with float it sometimes could go out of 0
}
ps_get_utt_time(ps, &n_speech, &n_cpu, &n_wall);
printf("%.2f seconds speech, %.2f seconds CPU, %.2f seconds wall\n",
n_speech, n_cpu, n_wall);
printf("%.2f xRT (CPU), %.2f xRT (elapsed)\n",
n_cpu / n_speech, n_wall / n_speech);
ps_get_all_time(ps, &n_speech, &n_cpu, &n_wall);
printf("TOTAL: %.2f seconds speech, %.2f seconds CPU, %.2f seconds wall\n",
n_speech, n_cpu, n_wall);
printf("TOTAL: %.2f xRT (CPU), %.2f xRT (elapsed)\n",
n_cpu / n_speech, n_wall / n_speech);
fclose(rawfh);
ps_free(ps);
cmd_ln_free_r(config);
ckd_free_2d(cepbuf);
ckd_free(buf);
return 0;
}
static void
process_ctl(ps_decoder_t *ps, cmd_ln_t *config, FILE *ctlfh)
{
int32 ctloffset, ctlcount, ctlincr;
int32 i;
char *line;
size_t len;
FILE *hypfh = NULL, *hypsegfh = NULL, *ctmfh = NULL;
FILE *mllrfh = NULL, *lmfh = NULL, *fsgfh = NULL;
double n_speech, n_cpu, n_wall;
char const *outlatdir;
char const *nbestdir;
char const *str;
int frate;
ctloffset = cmd_ln_int32_r(config, "-ctloffset");
ctlcount = cmd_ln_int32_r(config, "-ctlcount");
ctlincr = cmd_ln_int32_r(config, "-ctlincr");
outlatdir = cmd_ln_str_r(config, "-outlatdir");
nbestdir = cmd_ln_str_r(config, "-nbestdir");
frate = cmd_ln_int32_r(config, "-frate");
if ((str = cmd_ln_str_r(config, "-mllrctl"))) {
mllrfh = fopen(str, "r");
if (mllrfh == NULL) {
E_ERROR_SYSTEM("Failed to open MLLR control file file %s", str);
goto done;
}
}
if ((str = cmd_ln_str_r(config, "-fsgctl"))) {
fsgfh = fopen(str, "r");
if (fsgfh == NULL) {
E_ERROR_SYSTEM("Failed to open FSG control file file %s", str);
goto done;
}
}
if ((str = cmd_ln_str_r(config, "-lmnamectl"))) {
lmfh = fopen(str, "r");
if (lmfh == NULL) {
E_ERROR_SYSTEM("Failed to open LM name control file file %s", str);
goto done;
}
}
if ((str = cmd_ln_str_r(config, "-hyp"))) {
hypfh = fopen(str, "w");
if (hypfh == NULL) {
E_ERROR_SYSTEM("Failed to open hypothesis file %s for writing", str);
goto done;
}
setbuf(hypfh, NULL);
}
if ((str = cmd_ln_str_r(config, "-hypseg"))) {
hypsegfh = fopen(str, "w");
if (hypsegfh == NULL) {
E_ERROR_SYSTEM("Failed to open hypothesis file %s for writing", str);
goto done;
}
setbuf(hypsegfh, NULL);
}
if ((str = cmd_ln_str_r(config, "-ctm"))) {
ctmfh = fopen(str, "w");
if (ctmfh == NULL) {
E_ERROR_SYSTEM("Failed to open hypothesis file %s for writing", str);
goto done;
}
setbuf(ctmfh, NULL);
}
i = 0;
while ((line = fread_line(ctlfh, &len))) {
char *wptr[4];
int32 nf, sf, ef;
char *mllrline = NULL, *lmline = NULL, *fsgline = NULL;
char *fsgfile = NULL, *lmname = NULL, *mllrfile = NULL;
if (mllrfh) {
mllrline = fread_line(mllrfh, &len);
if (mllrline == NULL) {
E_ERROR("File size mismatch between control and MLLR control\n");
ckd_free(line);
ckd_free(mllrline);
goto done;
}
mllrfile = string_trim(mllrline, STRING_BOTH);
}
if (lmfh) {
lmline = fread_line(lmfh, &len);
if (lmline == NULL) {
E_ERROR("File size mismatch between control and LM control\n");
ckd_free(line);
ckd_free(lmline);
goto done;
}
lmname = string_trim(lmline, STRING_BOTH);
}
if (fsgfh) {
fsgline = fread_line(fsgfh, &len);
if (fsgline == NULL) {
E_ERROR("File size mismatch between control and FSG control\n");
ckd_free(line);
ckd_free(fsgline);
goto done;
}
fsgfile = string_trim(fsgline, STRING_BOTH);
}
if (i < ctloffset) {
i += ctlincr;
goto nextline;
}
if (ctlcount != -1 && i >= ctloffset + ctlcount) {
goto nextline;
}
sf = 0;
ef = -1;
nf = str2words(line, wptr, 4);
if (nf == 0) {
/* Do nothing. */
}
else if (nf < 0) {
E_ERROR("Unexpected extra data in control file at line %d\n", i);
}
else {
char const *hyp, *file, *uttid;
int32 score;
file = wptr[0];
uttid = NULL;
if (nf > 1)
sf = atoi(wptr[1]);
if (nf > 2)
ef = atoi(wptr[2]);
if (nf > 3)
uttid = wptr[3];
E_INFO("Decoding '%s'\n", uttid ? uttid : file);
/* Do actual decoding. */
if(process_mllrctl_line(ps, config, mllrfile) < 0)
continue;
if(process_lmnamectl_line(ps, config, lmname) < 0)
continue;
if(process_fsgctl_line(ps, config, fsgfile) < 0)
continue;
if(process_ctl_line(ps, config, file, uttid, sf, ef) < 0)
continue;
hyp = ps_get_hyp(ps, &score, &uttid);
/* Write out results and such. */
if (hypfh) {
fprintf(hypfh, "%s (%s %d)\n", hyp ? hyp : "", uttid, score);
}
if (hypsegfh) {
write_hypseg(hypsegfh, ps, uttid);
}
if (ctmfh) {
ps_seg_t *itor = ps_seg_iter(ps, &score);
write_ctm(ctmfh, ps, itor, uttid, frate);
}
if (outlatdir) {
write_lattice(ps, outlatdir, uttid);
}
if (nbestdir) {
write_nbest(ps, nbestdir, uttid);
}
ps_get_utt_time(ps, &n_speech, &n_cpu, &n_wall);
E_INFO("%s: %.2f seconds speech, %.2f seconds CPU, %.2f seconds wall\n",
uttid, n_speech, n_cpu, n_wall);
E_INFO("%s: %.2f xRT (CPU), %.2f xRT (elapsed)\n",
uttid, n_cpu / n_speech, n_wall / n_speech);
E_INFO_NOFN("%s (%s %d)\n", hyp ? hyp : "", uttid, score);
}
i += ctlincr;
nextline:
ckd_free(mllrline);
ckd_free(fsgline);
ckd_free(lmline);
ckd_free(line);
}
ps_get_all_time(ps, &n_speech, &n_cpu, &n_wall);
E_INFO("TOTAL %.2f seconds speech, %.2f seconds CPU, %.2f seconds wall\n",
n_speech, n_cpu, n_wall);
E_INFO("AVERAGE %.2f xRT (CPU), %.2f xRT (elapsed)\n",
n_cpu / n_speech, n_wall / n_speech);
done:
if (hypfh)
fclose(hypfh);
if (hypsegfh)
fclose(hypsegfh);
if (ctmfh)
fclose(ctmfh);
}
