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;
}
int
fe_process_frames_ext(fe_t *fe,
int16 const **inout_spch,
size_t *inout_nsamps,
mfcc_t **buf_cep,
int32 *inout_nframes,
int16 **voiced_spch,
int32 *voiced_spch_nsamps,
int32 *out_frameidx)
{
int proc_result;
fe_prespch_extend_pcm(fe->vad_data->prespch_buf, *inout_nframes);
fe->vad_data->store_pcm = TRUE;
proc_result = fe_process_frames(fe, inout_spch, inout_nsamps, buf_cep, inout_nframes, out_frameidx);
fe->vad_data->store_pcm = FALSE;
if (fe->vad_data->global_state)
fe_prespch_read_pcm(fe->vad_data->prespch_buf, voiced_spch, voiced_spch_nsamps);
else
*voiced_spch_nsamps = 0;
return proc_result;
}
int
fe_process_utt(fe_t * fe, int16 const * spch, size_t nsamps,
mfcc_t *** cep_block, int32 * nframes)
{
mfcc_t **cep;
int rv;
/* Figure out how many frames we will need. */
fe_process_frames(fe, NULL, &nsamps, NULL, nframes, NULL);
/* Create the output buffer (it has to exist, even if there are no output frames). */
if (*nframes)
cep = (mfcc_t **)ckd_calloc_2d(*nframes, fe->feature_dimension, sizeof(**cep));
else
cep = (mfcc_t **)ckd_calloc_2d(1, fe->feature_dimension, sizeof(**cep));
/* Now just call fe_process_frames() with the allocated buffer. */
rv = fe_process_frames(fe, &spch, &nsamps, cep, nframes, NULL);
*cep_block = cep;
return rv;
}
/**
* 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;
}
/**
* 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
acmod_process_raw(acmod_t *acmod,
int16 const **inout_raw,
size_t *inout_n_samps,
int full_utt)
{
int32 ncep;
int32 out_frameidx;
int16 const *prev_audio_inptr;
/* If this is a full utterance, process it all at once. */
if (full_utt)
return acmod_process_full_raw(acmod, inout_raw, inout_n_samps);
/* Append MFCCs to the end of any that are previously in there
* (in practice, there will probably be none) */
if (inout_n_samps && *inout_n_samps) {
int inptr;
int32 processed_samples;
prev_audio_inptr = *inout_raw;
/* Total number of frames available. */
ncep = acmod->n_mfc_alloc - acmod->n_mfc_frame;
/* Where to start writing them (circular buffer) */
inptr = (acmod->mfc_outidx + acmod->n_mfc_frame) % acmod->n_mfc_alloc;
/* Write them in two (or more) parts if there is wraparound. */
while (inptr + ncep > acmod->n_mfc_alloc) {
int32 ncep1 = acmod->n_mfc_alloc - inptr;
if (fe_process_frames(acmod->fe, inout_raw, inout_n_samps,
acmod->mfc_buf + inptr, &ncep1, &out_frameidx) < 0)
return -1;
if (out_frameidx > 0)
acmod->utt_start_frame = out_frameidx;
processed_samples = *inout_raw - prev_audio_inptr;
if (processed_samples + acmod->rawdata_pos < acmod->rawdata_size) {
memcpy(acmod->rawdata + acmod->rawdata_pos, prev_audio_inptr, processed_samples * sizeof(int16));
acmod->rawdata_pos += processed_samples;
}
/* Write to logging file if any. */
if (acmod->rawfh) {
fwrite(prev_audio_inptr, sizeof(int16),
processed_samples,
acmod->rawfh);
}
prev_audio_inptr = *inout_raw;
/* ncep1 now contains the number of frames actually
* processed. This is a good thing, but it means we
* actually still might have some room left at the end of
* the buffer, hence the while loop. Unfortunately it
* also means that in the case where we are really
* actually done, we need to get out totally, hence the
* goto. */
acmod->n_mfc_frame += ncep1;
ncep -= ncep1;
inptr += ncep1;
inptr %= acmod->n_mfc_alloc;
if (ncep1 == 0)
goto alldone;
}
assert(inptr + ncep <= acmod->n_mfc_alloc);
if (fe_process_frames(acmod->fe, inout_raw, inout_n_samps,
acmod->mfc_buf + inptr, &ncep, &out_frameidx) < 0)
return -1;
if (out_frameidx > 0)
acmod->utt_start_frame = out_frameidx;
processed_samples = *inout_raw - prev_audio_inptr;
if (processed_samples + acmod->rawdata_pos < acmod->rawdata_size) {
memcpy(acmod->rawdata + acmod->rawdata_pos, prev_audio_inptr, processed_samples * sizeof(int16));
acmod->rawdata_pos += processed_samples;
}
if (acmod->rawfh) {
fwrite(prev_audio_inptr, sizeof(int16),
processed_samples, acmod->rawfh);
}
prev_audio_inptr = *inout_raw;
acmod->n_mfc_frame += ncep;
alldone:
;
}
/* Hand things off to acmod_process_cep. */
return acmod_process_mfcbuf(acmod);
}
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;
}
