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
acmod_process_full_raw(acmod_t *acmod,
int16 const **inout_raw,
size_t *inout_n_samps)
{
int32 nfr, ntail;
mfcc_t **cepptr;
/* Write to logging file if any. */
if (acmod->rawfh)
fwrite(*inout_raw, 2, *inout_n_samps, acmod->rawfh);
/* Resize mfc_buf to fit. */
if (fe_process_frames(acmod->fe, NULL, inout_n_samps, NULL, &nfr) < 0)
return -1;
if (acmod->n_mfc_alloc < nfr + 1) {
ckd_free_2d(acmod->mfc_buf);
acmod->mfc_buf = ckd_calloc_2d(nfr + 1, fe_get_output_size(acmod->fe),
sizeof(**acmod->mfc_buf));
acmod->n_mfc_alloc = nfr + 1;
}
acmod->n_mfc_frame = 0;
acmod->mfc_outidx = 0;
fe_start_utt(acmod->fe);
if (fe_process_frames(acmod->fe, inout_raw, inout_n_samps,
acmod->mfc_buf, &nfr) < 0)
return -1;
fe_end_utt(acmod->fe, acmod->mfc_buf[nfr], &ntail);
nfr += ntail;
cepptr = acmod->mfc_buf;
nfr = acmod_process_full_cep(acmod, &cepptr, &nfr);
acmod->n_mfc_frame = 0;
return nfr;
}
void
segment_audio()
{
FILE *file;
int16 pcm_buf[BLOCKSIZE];
mfcc_t **cep_buf;
int16 voiced_buf = NULL;
int32 voiced_nsamps, out_frameidx, uttstart = 0;
char file_name[1024];
uint8 cur_vad_state, vad_state, writing;
int uttno, uttlen, sample_rate;
int32 nframes, nframes_tmp;
int16 frame_size, frame_shift, frame_rate;
size_t k;
sample_rate = (int) cmd_ln_float32_r(config, "-samprate");
frame_rate = cmd_ln_int32_r(config, "-frate");
frame_size =
(int32) (cmd_ln_float32_r(config, "-wlen") * sample_rate + 0.5);
frame_shift =
(int32) (sample_rate / cmd_ln_int32_r(config, "-frate") + 0.5);
nframes = (BLOCKSIZE - frame_size) / frame_shift;
cep_buf =
(mfcc_t **) ckd_calloc_2d(nframes, fe_get_output_size(fe),
sizeof(mfcc_t));
uttno = 0;
uttlen = 0;
cur_vad_state = 0;
voiced_nsamps = 0;
writing = 0;
file = NULL;
fe_start_stream(fe);
fe_start_utt(fe);
while ((k = read_audio(pcm_buf, BLOCKSIZE)) > 0) {
int16 const *pcm_buf_tmp;
pcm_buf_tmp = &pcm_buf[0];
while (k) {
nframes_tmp = nframes;
fe_process_frames_ext(fe, &pcm_buf_tmp, &k, cep_buf,
&nframes_tmp, voiced_buf,
&voiced_nsamps, &out_frameidx);
if (out_frameidx > 0) {
uttstart = out_frameidx;
}
vad_state = fe_get_vad_state(fe);
if (!cur_vad_state && vad_state) {
/* silence->speech transition, time to start new file */
uttno++;
if (!singlefile) {
sprintf(file_name, "%s%04d.raw", infile_path, uttno);
if ((file = fopen(file_name, "wb")) == NULL)
E_FATAL_SYSTEM("Failed to open '%s' for writing",
file_name);
} else {
sprintf(file_name, "%s.raw", infile_path);
if ((file = fopen(file_name, "ab")) == NULL)
E_FATAL_SYSTEM("Failed to open '%s' for writing",
file_name);
}
writing = 1;
}
if (writing && file && voiced_nsamps > 0) {
fwrite(voiced_buf, sizeof(int16), voiced_nsamps, file);
uttlen += voiced_nsamps;
}
if (cur_vad_state && !vad_state) {
/* speech -> silence transition, time to finish file */
fclose(file);
printf("Utterance %04d: file %s start %.1f sec length %d samples ( %.2f sec )\n",
uttno,
file_name,
((double) uttstart) / frame_rate,
uttlen,
((double) uttlen) / sample_rate);
fflush(stdout);
fe_end_utt(fe, cep_buf[0], &nframes_tmp);
writing = 0;
uttlen = 0;
voiced_nsamps = 0;
fe_start_utt(fe);
}
cur_vad_state = vad_state;
}
}
if (writing) {
fclose(file);
printf("Utterance %04d: file %s start %.1f sec length %d samples ( %.2f sec )\n",
uttno,
file_name,
((double) uttstart) / frame_rate,
uttlen,
((double) uttlen) / sample_rate);
fflush(stdout);
}
fe_end_utt(fe, cep_buf[0], &nframes);
ckd_free_2d(cep_buf);
}
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;
}
int
main(int argc, char *argv[])
{
acmod_t *acmod;
logmath_t *lmath;
cmd_ln_t *config;
FILE *rawfh;
int16 *buf;
int16 const *bptr;
mfcc_t **cepbuf, **cptr;
size_t nread, nsamps;
int nfr;
int frame_counter;
int bestsen1[270];
lmath = logmath_init(1.0001, 0, 0);
config = cmd_ln_init(NULL, ps_args(), TRUE,
"-featparams", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/feat.params",
"-mdef", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/mdef",
"-mean", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/means",
"-var", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/variances",
"-tmat", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/transition_matrices",
"-sendump", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/sendump",
"-compallsen", "true",
"-cmn", "prior",
"-tmatfloor", "0.0001",
"-mixwfloor", "0.001",
"-varfloor", "0.0001",
"-mmap", "no",
"-topn", "4",
"-ds", "1",
"-input_endian", "little",
"-samprate", "16000", NULL);
TEST_ASSERT(config);
TEST_ASSERT(acmod = acmod_init(config, lmath, NULL, NULL));
cmn_prior_set(acmod->fcb->cmn_struct, prior);
nsamps = 2048;
frame_counter = 0;
buf = ckd_calloc(nsamps, sizeof(*buf));
TEST_ASSERT(rawfh = fopen(DATADIR "/goforward.raw", "rb"));
TEST_EQUAL(0, acmod_start_utt(acmod));
E_INFO("Incremental(2048):\n");
while (!feof(rawfh)) {
nread = fread(buf, sizeof(*buf), nsamps, rawfh);
bptr = buf;
while ((nfr = acmod_process_raw(acmod, &bptr, &nread, FALSE)) > 0 || nread > 0) {
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
TEST_EQUAL(frame_counter, frame_idx);
if (frame_counter < 190)
bestsen1[frame_counter] = best_score;
++frame_counter;
frame_idx = -1;
}
}
}
TEST_EQUAL(0, acmod_end_utt(acmod));
nread = 0;
{
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
bestsen1[frame_counter] = best_score;
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
/* Now try to process the whole thing at once. */
E_INFO("Whole utterance:\n");
cmn_prior_set(acmod->fcb->cmn_struct, prior);
nsamps = ftell(rawfh) / sizeof(*buf);
clearerr(rawfh);
fseek(rawfh, 0, SEEK_SET);
buf = ckd_realloc(buf, nsamps * sizeof(*buf));
TEST_EQUAL(nsamps, fread(buf, sizeof(*buf), nsamps, rawfh));
bptr = buf;
TEST_EQUAL(0, acmod_start_utt(acmod));
acmod_process_raw(acmod, &bptr, &nsamps, TRUE);
TEST_EQUAL(0, acmod_end_utt(acmod));
{
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
frame_counter = 0;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
/* Now process MFCCs and make sure we get the same results. */
cepbuf = ckd_calloc_2d(frame_counter,
fe_get_output_size(acmod->fe),
sizeof(**cepbuf));
fe_start_utt(acmod->fe);
nsamps = ftell(rawfh) / sizeof(*buf);
bptr = buf;
nfr = frame_counter;
fe_process_frames(acmod->fe, &bptr, &nsamps, cepbuf, &nfr);
fe_end_utt(acmod->fe, cepbuf[frame_counter-1], &nfr);
E_INFO("Incremental(MFCC):\n");
cmn_prior_set(acmod->fcb->cmn_struct, prior);
TEST_EQUAL(0, acmod_start_utt(acmod));
cptr = cepbuf;
nfr = frame_counter;
frame_counter = 0;
while ((acmod_process_cep(acmod, &cptr, &nfr, FALSE)) > 0) {
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
TEST_EQUAL(frame_counter, frame_idx);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
++frame_counter;
frame_idx = -1;
}
}
TEST_EQUAL(0, acmod_end_utt(acmod));
nfr = 0;
acmod_process_cep(acmod, &cptr, &nfr, FALSE);
{
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
TEST_EQUAL(frame_counter, frame_idx);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
++frame_counter;
frame_idx = -1;
}
}
/* Note that we have to process the whole thing again because
* !#@$@ s2mfc2feat modifies its argument (not for long) */
fe_start_utt(acmod->fe);
nsamps = ftell(rawfh) / sizeof(*buf);
bptr = buf;
nfr = frame_counter;
fe_process_frames(acmod->fe, &bptr, &nsamps, cepbuf, &nfr);
fe_end_utt(acmod->fe, cepbuf[frame_counter-1], &nfr);
E_INFO("Whole utterance (MFCC):\n");
cmn_prior_set(acmod->fcb->cmn_struct, prior);
TEST_EQUAL(0, acmod_start_utt(acmod));
cptr = cepbuf;
nfr = frame_counter;
acmod_process_cep(acmod, &cptr, &nfr, TRUE);
TEST_EQUAL(0, acmod_end_utt(acmod));
{
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
frame_counter = 0;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
E_INFO("Rewound (MFCC):\n");
TEST_EQUAL(0, acmod_rewind(acmod));
{
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
frame_counter = 0;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
/* Clean up, go home. */
ckd_free_2d(cepbuf);
fclose(rawfh);
ckd_free(buf);
acmod_free(acmod);
logmath_free(lmath);
cmd_ln_free_r(config);
return 0;
}
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;
}
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;
}
