ps_seg_t *
ps_seg_iter(ps_decoder_t *ps)
{
ps_seg_t *itor;
ptmr_start(&ps->perf);
itor = ps_search_seg_iter(ps->search);
ptmr_stop(&ps->perf);
return itor;
}
ps_seg_t *
ps_seg_iter(ps_decoder_t *ps, int32 *out_best_score)
{
ps_seg_t *itor;
ptmr_start(&ps->perf);
itor = ps_search_seg_iter(ps->search, out_best_score);
ptmr_stop(&ps->perf);
return itor;
}
int32
ps_get_prob(ps_decoder_t *ps)
{
int32 prob;
ptmr_start(&ps->perf);
prob = ps_search_prob(ps->search);
ptmr_stop(&ps->perf);
return prob;
}
char const *
ps_get_hyp_final(ps_decoder_t *ps, int32 *out_is_final)
{
char const *hyp;
ptmr_start(&ps->perf);
hyp = ps_search_hyp(ps->search, NULL, out_is_final);
ptmr_stop(&ps->perf);
return hyp;
}
char const *
ps_get_hyp(ps_decoder_t *ps, int32 *out_best_score)
{
char const *hyp;
ptmr_start(&ps->perf);
hyp = ps_search_hyp(ps->search, out_best_score, NULL);
ptmr_stop(&ps->perf);
return hyp;
}
static void
utt_livepretend(void *data, utt_res_t * ur, int32 sf, int32 ef,
char *uttid)
{
char fullrawfn[FILENAME_LENGTH];
char *hypstr;
short samples[SAMPLE_BUFFER_LENGTH];
float32 **frames;
kb_t *kb;
FILE *rawfd;
int len, n_frames;
kb = (kb_t *) data;
/* report_utt_res(ur); */
sprintf(fullrawfn, "%s/%s%s", rawdirfn, ur->uttfile, decoder.rawext);
if ((rawfd = fopen(fullrawfn, "rb")) == NULL) {
E_FATAL("Cannnot open raw file %s.\n", fullrawfn);
}
/* temporary hack */
/* fread(waveheader, 1, 44, rawfd); */
if (ur->lmname != NULL)
srch_set_lm((srch_t *) kb->srch, ur->lmname);
if (ur->regmatname != NULL)
kb_setmllr(ur->regmatname, ur->cb2mllrname, kb);
if (s3_decode_begin_utt(&decoder, ur->uttfile) != S3_DECODE_SUCCESS)
E_FATAL("Cannot begin utterance decoding.\n");
len = fread(samples, sizeof(short), SAMPLE_BUFFER_LENGTH, rawfd);
while (len > 0) {
ptmr_start(&(st->tm));
fe_process_utt(fe, samples, len, &frames, &n_frames);
if (frames != NULL) {
s3_decode_process(&decoder, frames, n_frames);
ckd_free_2d((void **)frames);
}
ptmr_stop(&(st->tm));
if (S3_DECODE_SUCCESS ==
s3_decode_hypothesis(&decoder, NULL, &hypstr, NULL))
if (decoder.phypdump)
E_INFO("PARTIAL_HYP: %s\n", hypstr);
len = fread(samples, sizeof(short), SAMPLE_BUFFER_LENGTH, rawfd);
}
fclose(rawfd);
s3_decode_end_utt(&decoder);
}
int32
ps_get_prob(ps_decoder_t *ps, char const **out_uttid)
{
int32 prob;
ptmr_start(&ps->perf);
prob = ps_search_prob(ps->search);
if (out_uttid)
*out_uttid = ps->uttid;
ptmr_stop(&ps->perf);
return prob;
}
char const *
ps_get_hyp(ps_decoder_t *ps, int32 *out_best_score, char const **out_uttid)
{
char const *hyp;
ptmr_start(&ps->perf);
hyp = ps_search_hyp(ps->search, out_best_score);
if (out_uttid)
*out_uttid = ps->uttid;
ptmr_stop(&ps->perf);
return hyp;
}
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);
}
static int
init_state(const char *obsdmp,
const char *obsidx,
uint32 n_density,
uint32 n_stream,
uint32 *veclen,
uint32 blksize,
int reest,
const char *mixwfn,
const char *meanfn,
const char *varfn,
uint32 ts_off,
uint32 ts_cnt,
uint32 n_ts,
uint32 n_d_ts)
{
vector_t ***mean;
vector_t ***var = NULL;
vector_t ****fullvar = NULL;
float32 ***mixw = NULL;
uint32 n_frame;
uint32 ignore = 0;
codew_t *label;
uint32 n_corpus = 0;
float64 sqerr;
float64 tot_sqerr;
segdmp_type_t t;
uint32 i, j, ts, n;
int32 full_covar;
full_covar = cmd_ln_int32("-fullvar");
/* fully-continuous for now */
mean = gauden_alloc_param(ts_cnt, n_stream, n_density, veclen);
if (full_covar)
fullvar = gauden_alloc_param_full(ts_cnt, n_stream, n_density, veclen);
else
var = gauden_alloc_param(ts_cnt, n_stream, n_density, veclen);
if (mixwfn)
mixw = (float32 ***)ckd_calloc_3d(ts_cnt,
n_stream,
n_density,
sizeof(float32));
if (cmd_ln_str("-segidxfn")) {
E_INFO("Multi-class dump\n");
if (segdmp_open_read(cmd_ln_str_list("-segdmpdirs"),
cmd_ln_str("-segdmpfn"),
cmd_ln_str("-segidxfn"),
&n,
&t,
n_stream, veclen, blksize) != S3_SUCCESS) {
E_FATAL("Unable to open dumps\n");
}
if (n != n_d_ts) {
E_FATAL("Expected %u tied-states in dump, but apparently %u\n",
n_d_ts, n);
}
if (t != SEGDMP_TYPE_FEAT) {
E_FATAL("Expected feature dump, but instead saw %u\n", t);
}
multiclass = TRUE;
}
else {
E_INFO("1-class dump file\n");
multiclass = FALSE;
dmp_fp = s3open(cmd_ln_str("-segdmpfn"), "rb",
&dmp_swp);
if (dmp_fp == NULL) {
E_ERROR_SYSTEM("Unable to open dump file %s for reading\n",
cmd_ln_str("-segdmpfn"));
return S3_ERROR;
}
if (s3read(&n_frame, sizeof(uint32), 1, dmp_fp, dmp_swp, &ignore) != 1) {
E_ERROR_SYSTEM("Unable to open dump file %s for reading\n",
cmd_ln_str("-segdmpfn"));
return S3_ERROR;
}
data_offset = ftell(dmp_fp);
}
tot_sqerr = 0;
for (i = 0; i < ts_cnt; i++) {
ts = ts_off + i;
/* stride not accounted for yet */
if (o2d == NULL) {
if (multiclass)
n_frame = segdmp_n_seg(ts);
}
else {
for (j = 0, n_frame = 0; j < n_o2d[ts]; j++) {
n_frame += segdmp_n_seg(o2d[ts][j]);
}
}
E_INFO("Corpus %u: sz==%u frames%s\n",
ts, n_frame,
(n_frame > cmd_ln_int32("-vartiethr") ? "" : " tied var"));
if (n_frame == 0) {
continue;
}
E_INFO("Convergence ratios are abs(cur - prior) / abs(prior)\n");
/* Do some variety of k-means clustering */
ptmr_start(&km_timer);
sqerr = cluster(ts, n_stream, n_frame, veclen, blksize, mean[i], n_density, &label);
ptmr_stop(&km_timer);
if (sqerr < 0) {
E_ERROR("Unable to do k-means for state %u; skipping...\n", ts);
continue;
}
/* Given the k-means and assuming equal prior liklihoods
* compute the variances */
ptmr_start(&var_timer);
if (full_covar)
full_variances(ts, mean[i], fullvar[i], n_density, n_stream, veclen, blksize,
n_frame, label);
else
variances(ts, mean[i], var[i], n_density, n_stream, veclen, blksize, n_frame, label);
ptmr_stop(&var_timer);
if (mixwfn) {
/* initialize the mixing weights by counting # of occurrances
* of the top codeword over the corpus and normalizing */
init_mixw(mixw[i], mean[i], n_density, veclen, n_frame, n_stream, label);
ckd_free(label);
if (reest == TRUE && full_covar)
E_ERROR("EM re-estimation is not yet supported for full covariances\n");
else if (reest == TRUE) {
ptmr_start(&em_timer);
/* Do iterations of EM to estimate the mixture densities */
reest_sum(ts, mean[i], var[i], mixw[i], n_density, n_stream,
n_frame, veclen, blksize,
cmd_ln_int32("-niter"),
FALSE,
cmd_ln_int32("-vartiethr"));
ptmr_stop(&em_timer);
}
}
++n_corpus;
tot_sqerr += sqerr;
E_INFO("sqerr [%u] == %e\n", ts, sqerr);
}
if (n_corpus > 0) {
E_INFO("sqerr = %e tot %e rms\n", tot_sqerr, sqrt(tot_sqerr/n_corpus));
}
if (!multiclass)
s3close(dmp_fp);
if (meanfn) {
if (s3gau_write(meanfn,
(const vector_t ***)mean,
ts_cnt,
n_stream,
n_density,
veclen) != S3_SUCCESS) {
return S3_ERROR;
}
}
else {
E_INFO("No mean file given; none written\n");
}
if (varfn) {
if (full_covar) {
if (s3gau_write_full(varfn,
(const vector_t ****)fullvar,
ts_cnt,
n_stream,
n_density,
veclen) != S3_SUCCESS)
return S3_ERROR;
}
else {
if (s3gau_write(varfn,
(const vector_t ***)var,
ts_cnt,
n_stream,
n_density,
veclen) != S3_SUCCESS)
return S3_ERROR;
}
}
else {
E_INFO("No variance file given; none written\n");
}
if (mixwfn) {
if (s3mixw_write(mixwfn,
mixw,
ts_cnt,
n_stream,
n_density) != S3_SUCCESS) {
return S3_ERROR;
}
}
else {
E_INFO("No mixing weight file given; none written\n");
}
return S3_SUCCESS;
}
static void
utt_livepretend(void *data, utt_res_t * ur, int32 sf, int32 ef,
char *uttid)
{
char fullrawfn[FILENAME_LENGTH];
char *hypstr;
cont_ad_t *cont_ad;
ad_rec_t bogus_ad;
short samples[SAMPLE_BUFFER_LENGTH];
float32 **frames;
kb_t *kb;
int nread, n_frames, seg_n_frames;
int ts, listening;
kb = (kb_t *) data;
/* report_utt_res(ur); */
sprintf(fullrawfn, "%s/%s%s", rawdirfn, ur->uttfile, decoder.rawext);
if ((rawfd = fopen(fullrawfn, "rb")) == NULL) {
E_FATAL("Cannnot open raw file %s.\n", fullrawfn);
}
if (ur->lmname != NULL)
srch_set_lm((srch_t *) kb->srch, ur->lmname);
if (ur->regmatname != NULL)
kb_setmllr(ur->regmatname, ur->cb2mllrname, kb);
bogus_ad.sps = (int32) cmd_ln_float32_r(kb->kbcore->config, "-samprate");
if ((cont_ad = cont_ad_init(&bogus_ad, ad_file_read)) == NULL) {
E_FATAL("Failed to initialize energy-based endpointer");
}
listening = 0;
ts = 0;
seg_n_frames = 0;
while ((nread = cont_ad_read(cont_ad, samples, SAMPLE_BUFFER_LENGTH))
>= 0) {
if (nread) {
ts = cont_ad->read_ts;
if (!listening) {
char uttid[FILENAME_LENGTH];
sprintf(uttid, "%s_%.3f", ur->uttfile,
(double) ts / bogus_ad.sps);
if (s3_decode_begin_utt(&decoder, uttid) != S3_DECODE_SUCCESS)
E_FATAL("Cannot begin utterance decoding.\n");
listening = 1;
}
ptmr_start(&(st->tm));
fe_process_utt(fe, samples, nread, &frames, &n_frames);
seg_n_frames += n_frames;
if (frames != NULL) {
s3_decode_process(&decoder, frames, n_frames);
ckd_free_2d((void **)frames);
}
ptmr_stop(&(st->tm));
if (s3_decode_hypothesis(&decoder, NULL, &hypstr, NULL) ==
S3_DECODE_SUCCESS) {
if (decoder.phypdump) {
E_INFO("PARTIAL_HYP: %s\n", hypstr);
}
}
/* If the segment is too long, break it. */
if (seg_n_frames > 15000) {
s3_decode_end_utt(&decoder);
listening = 0;
}
}
else {
if (listening && cont_ad->read_ts - ts > 8000) { /* HACK */
s3_decode_end_utt(&decoder);
listening = 0;
}
}
}
fclose(rawfd);
cont_ad_close(cont_ad);
if (listening)
s3_decode_end_utt(&decoder);
}
int
ps_end_utt(ps_decoder_t *ps)
{
int rv, i;
acmod_end_utt(ps->acmod);
/* Search any remaining frames. */
if ((rv = ps_search_forward(ps)) < 0) {
ptmr_stop(&ps->perf);
return rv;
}
/* Finish phone loop search. */
if (ps->phone_loop) {
if ((rv = ps_search_finish(ps->phone_loop)) < 0) {
ptmr_stop(&ps->perf);
return rv;
}
}
/* Search any frames remaining in the lookahead window. */
if (ps->acmod->output_frame >= ps->pl_window) {
for (i = ps->acmod->output_frame - ps->pl_window;
i < ps->acmod->output_frame; ++i)
ps_search_step(ps->search, i);
}
/* Finish main search. */
if ((rv = ps_search_finish(ps->search)) < 0) {
ptmr_stop(&ps->perf);
return rv;
}
ptmr_stop(&ps->perf);
/* Log a backtrace if requested. */
if (cmd_ln_boolean_r(ps->config, "-backtrace")) {
const char* hyp;
ps_seg_t *seg;
int32 score;
hyp = ps_get_hyp(ps, &score);
if (hyp != NULL) {
E_INFO("%s (%d)\n", hyp, score);
E_INFO_NOFN("%-20s %-5s %-5s %-5s %-10s %-10s %-3s\n",
"word", "start", "end", "pprob", "ascr", "lscr", "lback");
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);
E_INFO_NOFN("%-20s %-5d %-5d %-1.3f %-10d %-10d %-3d\n",
word, sf, ef, logmath_exp(ps_get_logmath(ps), post),
ascr, lscr, lback);
}
}
}
return rv;
}
ptmr_t
ctl_process(const char *ctlfile, const char *ctllmfile, const char *ctlmllrfile, int32 nskip,
int32 count, void (*func) (void *kb, utt_res_t * ur, int32 sf,
int32 ef, char *uttid), void *kb)
{
FILE *fp;
FILE *ctllmfp;
FILE *ctlmllrfp;
char uttfile[16384], uttid[4096];
char lmname[4096];
char regmatname[4096], cb2mllrname[4096];
char tmp[4096];
int32 sf, ef;
utt_res_t *ur;
ptmr_t tm;
kb_t *k;
k = (kb_t *) kb;
ctllmfp = NULL;
ctlmllrfp = NULL;
ur = new_utt_res();
if (ctlfile) {
if ((fp = fopen(ctlfile, "r")) == NULL)
E_FATAL_SYSTEM("fopen(%s,r) failed\n", ctlfile);
}
else
fp = stdin;
if (ctllmfile) {
E_INFO("LM is used in this session\n");
if ((ctllmfp = fopen(ctllmfile, "r")) == NULL)
E_FATAL_SYSTEM("fopen(%s,r) failed\n", ctllmfile);
}
if (ctlmllrfile) {
E_INFO("MLLR is used in this session\n");
if ((ctlmllrfp = fopen(ctlmllrfile, "r")) == NULL)
E_FATAL_SYSTEM("fopen(%s,r) failed\n", ctlmllrfile);
}
ptmr_init(&tm);
if (nskip > 0) {
E_INFO("Skipping %d entries at the beginning of %s\n", nskip,
ctlfile);
for (; nskip > 0; --nskip) {
if (ctl_read_entry(fp, uttfile, &sf, &ef, uttid) < 0) {
fclose(fp);
return tm;
}
/*This checks the size of the control file of the lm in batch mode */
if (ctllmfile) {
if (ctl_read_entry(ctllmfp, lmname, &sf, &ef, tmp) < 0) {
fclose(ctllmfp);
E_ERROR
("An LM control file is specified but LM cannot be read when skipping the %d-th sentence\n",
nskip);
return tm;
}
}
/*This checks the size of the control file of the mllr in batch mode */
if (ctlmllrfile) {
if (ctl_read_entry(ctlmllrfp, regmatname, &sf, &ef, tmp) <
0) {
fclose(ctlmllrfp);
E_ERROR
("A MLLR control file is specified but MLLR cannot be read when skipping the %d-th sentence\n",
nskip);
return tm;
}
}
}
}
for (; count > 0; --count) {
int32 tmp1, tmp2;
if (ctl_read_entry(fp, uttfile, &sf, &ef, uttid) < 0)
break;
/*This checks the size of the control file in batch mode */
if (ctllmfile) {
if (ctl_read_entry(ctllmfp, lmname, &tmp1, &tmp2, tmp) < 0) {
fclose(ctllmfp);
E_ERROR
("LM control file is specified but LM cannot be read when counting the %d-th sentence\n",
count);
break;
}
}
if (ctlmllrfile) {
if (ctl_read_entry
(ctlmllrfp, regmatname, &tmp1, &tmp2, cb2mllrname) < 0) {
E_ERROR
("MLLR control file is specified but MLLR cannot be read when counting the %d-th sentence\n",
count);
break;
}
if (tmp2 == -1)
strcpy(cb2mllrname, ".1cls.");
}
/* Process this utterance */
ptmr_start(&tm);
if (func) {
utt_res_set_uttfile(ur, uttfile);
if (ctllmfile)
utt_res_set_lmname(ur, lmname);
if (ctlmllrfile) {
utt_res_set_regmatname(ur, regmatname);
utt_res_set_cb2mllrname(ur, cb2mllrname);
}
(*func) (kb, ur, sf, ef, uttid);
}
ptmr_stop(&tm);
E_INFO
("%s: %6.1f sec CPU, %6.1f sec Clk; TOT: %8.1f sec CPU, %8.1f sec Clk\n\n",
uttid, tm.t_cpu, tm.t_elapsed, tm.t_tot_cpu,
tm.t_tot_elapsed);
ptmr_reset(&tm);
}
if (fp)
fclose(fp);
if (ctllmfp)
fclose(ctllmfp);
if (ctlmllrfp)
fclose(ctlmllrfp);
if (ur)
ckd_free(ur);
return tm;
}
ptmr_t
ctl_process_utt(const char *uttfile, int32 count,
void (*func) (void *kb, utt_res_t * ur, int32 sf, int32 ef,
char *uttid), void *kb)
{
char utterance_file[16384];
char uttid[4096];
const char *base;
int32 i, c;
int32 ts, newts;
ptmr_t tm;
utt_res_t *ur;
ptmr_init(&tm);
ur = new_utt_res();
base = path2basename(uttfile);
/* strip_fileext() copies base to uttid. So, copying uttid to base
* is redundant if strip_fileext() is not called.
*/
/*
strip_fileext (base, uttid);
strcpy (base, uttid);
*/
ts = -1;
for (c = 0; c < count; c++) {
/* Wait for uttfile to change from previous iteration */
for (i = 0;; i++) {
newts = stat_mtime(uttfile);
if ((newts >= 0) && (newts != ts))
break;
if (i == 0)
E_INFO("Waiting for %s, count %d, c %d\n", uttfile, count,
c);
SLEEP_SEC(1);
}
ts = newts;
/* Form uttid */
sprintf(uttid, "%s_%08d", base, c);
strncpy(utterance_file, uttfile, sizeof(utterance_file) - 1);
utterance_file[sizeof(utterance_file) - 1] = 0;
/* Process this utterance */
ptmr_start(&tm);
if (func) {
utt_res_set_uttfile(ur, utterance_file);
(*func) (kb, ur, 0, -1, uttid);
}
ptmr_stop(&tm);
E_INFO
("%s: %6.1f sec CPU, %6.1f sec Clk; TOT: %8.1f sec CPU, %8.1f sec Clk\n\n",
uttid, tm.t_cpu, tm.t_elapsed, tm.t_tot_cpu,
tm.t_tot_elapsed);
ptmr_reset(&tm);
}
if (ur)
free_utt_res(ur);
return tm;
}
float64 vector_vqgen (float32 **data, int32 rows, int32 cols, int32 vqrows,
float64 epsilon, int32 maxiter,
float32 **mean, int32 *map)
{
int32 i, j, r, it;
static uint32 seed = 1;
float64 sqerr, prev_sqerr=0, t;
bitvec_t sel;
int32 *count;
float32 *gmean;
ptmr_t tm;
assert ((rows >= vqrows) && (maxiter >= 0) && (epsilon > 0.0));
sel = bitvec_alloc (rows);
ptmr_init (&tm);
ptmr_start (&tm);
/* Pick a random initial set of centroids */
#ifndef WIN32 /* RAH */
srandom (seed);
seed ^= random();
#else /* RAH */
srand ((unsigned) time(NULL)); /* RAH */
#endif
for (i = 0; i < vqrows; i++) {
/* Find r = a random, previously unselected row from the input */
#ifndef WIN32 /* RAH */
r = (random() & (int32)0x7fffffff) % rows;
#else /* RAH */
r = (rand() & (int32)0x7fffffff) % rows; /* RAH */
#endif /* RAH */
while (bitvec_is_set (sel, r)) { /* BUG: possible infinite loop!! */
if (++r >= rows)
r = 0;
}
bitvec_set (sel, r);
memcpy ((void *)(mean[i]), (void *)(data[r]), cols * sizeof(float32));
/* BUG: What if two randomly selected rows are identical in content?? */
}
bitvec_free (sel);
count = (int32 *) ckd_calloc (vqrows, sizeof(int32));
/* In k-means, unmapped means in any iteration are a problem. Replace them with gmean */
gmean = (float32 *) ckd_calloc (cols, sizeof(float32));
vector_mean (gmean, mean, vqrows, cols);
for (it = 0;; it++) { /* Iterations of k-means algorithm */
/* Find the current data->mean mappings (labels) */
sqerr = 0.0;
for (i = 0; i < rows; i++) {
map[i] = vector_vqlabel (data[i], mean, vqrows, cols, &t);
sqerr += t;
}
ptmr_stop(&tm);
if (it == 0)
E_INFO("Iter %4d: %.1fs CPU; sqerr= %e\n", it, tm.t_cpu, sqerr);
else
E_INFO("Iter %4d: %.1fs CPU; sqerr= %e; delta= %e\n",
it, tm.t_cpu, sqerr, (prev_sqerr-sqerr)/prev_sqerr);
/* Check if exit condition satisfied */
if ((sqerr == 0.0) || (it >= maxiter-1) ||
((it > 0) && ( ((prev_sqerr - sqerr) / prev_sqerr) < epsilon )) )
break;
prev_sqerr = sqerr;
ptmr_start(&tm);
/* Update (reestimate) means */
for (i = 0; i < vqrows; i++) {
for (j = 0; j < cols; j++)
mean[i][j] = 0.0;
count[i] = 0;
}
for (i = 0; i < rows; i++) {
vector_accum (mean[map[i]], data[i], cols);
count[map[i]]++;
}
for (i = 0; i < vqrows; i++) {
if (count[i] > 1) {
t = 1.0 / (float64)(count[i]);
for (j = 0; j < cols; j++)
/* mean[i][j] *= t; */ /* RAH, compiler was complaining about this, */
mean[i][j] = (float32) ((float64) mean[i][j] * (float64) t); /* */
} else if (count[i] == 0) {
E_ERROR("Iter %d: mean[%d] unmapped\n", it, i);
memcpy (mean[i], gmean, cols * sizeof(float32));
}
}
}
ckd_free (count);
ckd_free (gmean);
return sqerr;
}
/* Decode the given mfc file and write result to given directory */
static void
utt_astar(void *data, utt_res_t * ur, int32 sf, int32 ef, char *uttid)
{
char dagfile[1024], nbestfile[1024];
const char *latdir;
const char *latext;
const char *nbestext;
dag_t *dag;
int32 nfrm;
if (ur->lmname)
lmset_set_curlm_wname(lmset, ur->lmname);
latdir = cmd_ln_str_r(config, "-inlatdir");
latext = cmd_ln_str_r(config, "-latext");
nbestext = cmd_ln_str_r(config, "-nbestext");
if (latdir) {
build_output_uttfile(dagfile, latdir, uttid, ur->uttfile);
strcat(dagfile, ".");
strcat(dagfile, latext);
}
else
sprintf(dagfile, "%s.%s", uttid, latext);
ptmr_reset(&tm_utt);
ptmr_start(&tm_utt);
nfrm = 0;
if ((dag = dag_load(dagfile,
cmd_ln_int32_r(config, "-maxedge"),
cmd_ln_float32_r(config, "-logbase"),
cmd_ln_int32_r(config, "-dagfudge"), dict, fpen, config, logmath)) != NULL) {
if (dict_filler_word(dict, dag->end->wid))
dag->end->wid = dict->finishwid;
dag_remove_unreachable(dag);
if (dag_bypass_filler_nodes(dag, 1.0, dict, fpen) < 0) {
E_ERROR("maxedge limit (%d) exceeded\n", dag->maxedge);
goto search_done;
}
dag_compute_hscr(dag, dict, lmset->cur_lm, 1.0);
dag_remove_bypass_links(dag);
E_INFO("%5d frames, %6d nodes, %8d edges, %8d bypass\n",
dag->nfrm, dag->nnode, dag->nlink, dag->nbypass);
nfrm = dag->nfrm;
build_output_uttfile(nbestfile, nbestdir, uttid, ur->uttfile);
strcat(nbestfile, ".");
strcat(nbestfile, nbestext);
nbest_search(dag, nbestfile, uttid, 1.0, dict, lmset->cur_lm, fpen);
lm_cache_stats_dump(lmset->cur_lm);
lm_cache_reset(lmset->cur_lm);
}
else
E_ERROR("Dag load (%s) failed\n", uttid);
search_done:
dag_destroy(dag);
ptmr_stop(&tm_utt);
printf("%s: TMR: %5d Frm", uttid, nfrm);
if (nfrm > 0) {
printf(" %6.2f xEl", tm_utt.t_elapsed * 100.0 / nfrm);
printf(" %6.2f xCPU", tm_utt.t_cpu * 100.0 / nfrm);
}
printf("\n");
fflush(stdout);
}
static void decode_utt (void *data, char *uttfile, int32 sf, int32 ef, char *uttid)
{
kb_t *kb;
acoustic_t *am;
int32 featwin, nfr, min_utt_frames, n_vithist;
char cepfile[4096], latfile[4096];
vithist_t *finalhist;
int32 i, f;
glist_t hyplist;
FILE *latfp;
printf ("\n");
fflush (stdout);
E_INFO("Utterance %s\n", uttid);
kb = (kb_t *)data;
am = kb->am;
featwin = feat_window_size(am->fcb);
/* Build complete cepfile name and read cepstrum data; check for min length */
ctl_infile (cepfile, cmd_ln_str("-cepdir"), cmd_ln_str("-cepext"), uttfile);
if ((nfr = s2mfc_read (cepfile, sf, ef, featwin, am->mfc, S3_MAX_FRAMES)) < 0) {
E_ERROR("%s: MFC read failed\n", uttid);
return;
}
E_INFO("%s: %d frames\n", uttid, nfr-(featwin<<1));
ptmr_reset (kb->tm);
ptmr_reset (kb->tm_search);
ptmr_start (kb->tm);
min_utt_frames = (featwin<<1) + 1;
if (nfr < min_utt_frames) {
E_ERROR("%s: Utterance shorter than %d frames; ignored\n",
uttid, min_utt_frames, nfr);
return;
}
/* CMN/AGC */
if (strcmp (cmd_ln_str("-cmn"), "current") == 0)
cmn (am->mfc, nfr, feat_cepsize(am->fcb));
if (strcmp (cmd_ln_str("-agc"), "max") == 0)
agc_max (am->mfc, nfr);
/* Process utterance */
lextree_vit_start (kb, uttid);
for (i = featwin, f = 0; i < nfr-featwin; i++, f++) {
am->senscale[f] = acoustic_eval (am, i);
ptmr_start (kb->tm_search);
lextree_vit_frame (kb, f, uttid);
printf (" %d,%d,%d", f, glist_count (kb->vithist[f]), glist_count (kb->lextree_active));
fflush (stdout);
ptmr_stop (kb->tm_search);
}
printf ("\n");
finalhist = lextree_vit_end (kb, f, uttid);
hyplist = vithist_backtrace (finalhist, kb->am->senscale);
hyp_log (stdout, hyplist, _dict_wordstr, (void *)kb->dict);
hyp_myfree (hyplist);
printf ("\n");
/* Log the entire Viterbi word lattice */
sprintf (latfile, "%s.lat", uttid);
if ((latfp = fopen(latfile, "w")) == NULL) {
E_ERROR("fopen(%s,w) failed; using stdout\n", latfile);
latfp = stdout;
}
n_vithist = vithist_log (latfp, kb->vithist, f, _dict_wordstr, (void *)kb->dict);
if (latfp != stdout)
fclose (latfp);
else {
printf ("\n");
fflush (stdout);
}
ptmr_stop (kb->tm);
if (f > 0) {
printf("TMR(%s): %5d frames; %.1fs CPU, %.2f xRT; %.1fs CPU(search), %.2f xRT; %.1fs Elapsed, %.2f xRT\n",
uttid, f,
kb->tm->t_cpu, kb->tm->t_cpu * 100.0 / f,
kb->tm_search->t_cpu, kb->tm_search->t_cpu * 100.0 / f,
kb->tm->t_elapsed, kb->tm->t_elapsed * 100.0 / f);
printf("CTR(%s): %5d frames; %d Sen (%.1f/fr); %d HMM (%.1f/fr); %d Words (%.1f/fr)\n",
uttid, f,
kb->n_sen_eval, ((float64)kb->n_sen_eval) / f,
kb->n_hmm_eval, ((float64)kb->n_hmm_eval) / f,
n_vithist, ((float64) n_vithist) / f);
}
/* Cleanup */
glist_free (kb->lextree_active);
kb->lextree_active = NULL;
for (; f >= -1; --f) { /* I.e., including dummy START_WORD node at frame -1 */
glist_myfree (kb->vithist[f], sizeof(vithist_t));
kb->vithist[f] = NULL;
}
lm_cache_reset (kb->lm);
}
ptmr_t
ctl_process(char *ctlfile, char *ctlmllrfile, int32 nskip, int32 count,
void (*func) (void *kb, char *uttfile, int32 sf, int32 ef,
char *uttid), void *kb)
{
FILE *fp, *mllrfp;
char uttfile[16384], uttid[4096];
char regmatfile[4096], cb2mllrfile[4096];
int32 sf, ef;
ptmr_t tm;
mllrfp = NULL;
E_INFO("Batch mode recognition without dynamic LM\n");
if (ctlfile) {
if ((fp = fopen(ctlfile, "r")) == NULL)
E_FATAL_SYSTEM("fopen(%s,r) failed\n", ctlfile);
}
else
fp = stdin;
if (ctlmllrfile) {
if ((mllrfp = fopen(ctlmllrfile, "r")) == NULL)
E_FATAL_SYSTEM("fopen(%s,r) failed\n", ctlmllrfile);
}
ptmr_init(&tm);
if (nskip > 0) {
E_INFO("Skipping %d entries at the beginning of %s\n", nskip,
ctlfile);
for (; nskip > 0; --nskip) {
if (ctl_read_entry(fp, uttfile, &sf, &ef, uttid) < 0) {
fclose(fp);
return tm;
}
}
if (ctlmllrfile) {
for (; nskip > 0; --nskip) {
if (ctl_read_entry(fp, regmatfile, &sf, &ef, cb2mllrfile) <
0) {
E_ERROR
("MLLR cannot be read when skipping the %d-th sentence\n",
nskip);
fclose(fp);
return tm;
}
}
}
}
for (; count > 0; --count) {
if (ctl_read_entry(fp, uttfile, &sf, &ef, uttid) < 0)
break;
if (ctlmllrfile) {
int32 tmp1, tmp2;
if (ctl_read_entry
(mllrfp, regmatfile, &tmp1, &tmp2, cb2mllrfile) < 0) {
E_ERROR
("MLLR cannot be read when counting the %d-th sentence\n",
count);
break;
}
if (tmp2 == -1)
strcpy(cb2mllrfile, ".1cls.");
}
/* Process this utterance */
ptmr_start(&tm);
if (func) {
if (ctlmllrfile)
kb_setmllr(regmatfile, cb2mllrfile, kb);
(*func) (kb, uttfile, sf, ef, uttid);
}
ptmr_stop(&tm);
E_INFO
("%s: %6.1f sec CPU, %6.1f sec Clk; TOT: %8.1f sec CPU, %8.1f sec Clk\n\n",
uttid, tm.t_cpu, tm.t_elapsed, tm.t_tot_cpu,
tm.t_tot_elapsed);
ptmr_reset(&tm);
}
if (fp)
fclose(fp);
return tm;
}
int
main(int argc, char *argv[])
{
lexicon_t *lex;
model_def_t *omdef;
model_def_t *dmdef;
feat_t *feat;
uint32 n_stream, blksize;
uint32 *veclen;
uint32 ts_off;
uint32 ts_cnt;
FILE *fp;
if (main_initialize(argc, argv, &lex, &omdef, &dmdef, &feat) != S3_SUCCESS) {
return -1;
}
n_stream = feat_dimension1(feat);
veclen = feat_stream_lengths(feat);
blksize = feat_dimension(feat);
if (strcmp(cmd_ln_str("-gthobj"), "state") == 0) {
ts_off = cmd_ln_int32("-tsoff");
if (cmd_ln_str("-tscnt") == NULL) {
ts_cnt = omdef->n_tied_state - ts_off;
}
else {
ts_cnt = cmd_ln_int32("-tscnt");
}
if (ts_off + ts_cnt > omdef->n_tied_state) {
E_FATAL("Too many tied states specified\n");
}
n_tot_frame = 0;
ptmr_reset(&all_timer);
ptmr_reset(&km_timer);
ptmr_reset(&var_timer);
ptmr_reset(&em_timer);
ptmr_start(&all_timer);
if (init_state(cmd_ln_str("-segdmpfn"),
cmd_ln_str("-segidxfn"),
cmd_ln_int32("-ndensity"),
n_stream,
veclen,
blksize,
cmd_ln_int32("-reest"),
cmd_ln_str("-mixwfn"),
cmd_ln_str("-meanfn"),
cmd_ln_str("-varfn"),
ts_off,
ts_cnt,
omdef->n_tied_state,
(dmdef != NULL ? dmdef->n_tied_state : omdef->n_tied_state))
!= S3_SUCCESS) {
E_ERROR("Unable to train [%u %u]\n", ts_off, ts_off+ts_cnt-1);
}
ptmr_stop(&all_timer);
if (n_tot_frame > 0) {
E_INFO("TOTALS:");
E_INFOCONT(" km %4.3fx %4.3e",
km_timer.t_cpu / (n_tot_frame * 0.01),
(km_timer.t_cpu > 0 ?
km_timer.t_elapsed / km_timer.t_cpu : 0.0));
E_INFOCONT(" var %4.3fx %4.3e",
var_timer.t_cpu / (n_tot_frame * 0.01),
(var_timer.t_cpu > 0 ?
var_timer.t_elapsed / var_timer.t_cpu : 0.0));
E_INFOCONT(" em %4.3fx %4.3e",
em_timer.t_cpu / (n_tot_frame * 0.01),
(em_timer.t_cpu > 0 ?
em_timer.t_elapsed / em_timer.t_cpu : 0.0));
E_INFOCONT(" all %4.3fx %4.3e",
all_timer.t_cpu / (n_tot_frame * 0.01),
(all_timer.t_cpu > 0 ?
all_timer.t_elapsed / all_timer.t_cpu : 0.0));
E_INFOCONT("\n");
}
if (cmd_ln_str("-tsrngfn") != NULL) {
fp = fopen(cmd_ln_str("-tsrngfn"),
"w");
if (fp == NULL) {
E_FATAL_SYSTEM("Unable to open %s for reading",
cmd_ln_str("-tsrngfn"));
}
fprintf(fp, "%d %d\n", ts_off, ts_cnt);
}
else if (ts_cnt != omdef->n_tied_state) {
E_WARN("Subset of tied states specified, but no -tsrngfn arg");
}
}
else if (strcmp(cmd_ln_str("-gthobj"), "single") == 0) {
n_tot_frame = 0;
ptmr_reset(&all_timer);
ptmr_reset(&km_timer);
ptmr_reset(&var_timer);
ptmr_reset(&em_timer);
ptmr_start(&all_timer);
if (init_state(cmd_ln_str("-segdmpfn"),
NULL, /* No index -> single class dump file */
cmd_ln_int32("-ndensity"),
n_stream,
veclen,
blksize,
cmd_ln_int32("-reest"),
cmd_ln_str("-mixwfn"),
cmd_ln_str("-meanfn"),
cmd_ln_str("-varfn"),
0,
1,
1,
1) != S3_SUCCESS) {
E_ERROR("Unable to train\n");
}
ptmr_stop(&all_timer);
if (n_tot_frame > 0) {
E_INFO("TOTALS:");
E_INFOCONT(" km %4.3fx %4.3e",
km_timer.t_cpu / (n_tot_frame * 0.01),
(km_timer.t_cpu > 0 ?
km_timer.t_elapsed / km_timer.t_cpu : 0.0));
E_INFOCONT(" var %4.3fx %4.3e",
var_timer.t_cpu / (n_tot_frame * 0.01),
(var_timer.t_cpu > 0 ?
var_timer.t_elapsed / var_timer.t_cpu : 0.0));
E_INFOCONT(" em %4.3fx %4.3e",
em_timer.t_cpu / (n_tot_frame * 0.01),
(em_timer.t_cpu > 0 ?
em_timer.t_elapsed / em_timer.t_cpu : 0.0));
E_INFOCONT(" all %4.3fx %4.3e",
all_timer.t_cpu / (n_tot_frame * 0.01),
(all_timer.t_cpu > 0 ?
all_timer.t_elapsed / all_timer.t_cpu : 0.0));
E_INFOCONT("\n");
}
}
return 0;
}
/*
* Find Viterbi alignment.
*/
static void
align_utt(char *sent, /* In: Reference transcript */
int32 nfr, /* In: #frames of input */
char *ctlspec, /* In: Utt specifiction from control file */
char *uttid)
{ /* In: Utterance id, for logging and other use */
int32 i;
align_stseg_t *stseg;
align_phseg_t *phseg;
align_wdseg_t *wdseg;
int32 w;
w = feat_window_size(kbcore_fcb(kbc)); /* #MFC vectors needed on either side of current
frame to compute one feature vector */
if (nfr <= (w << 1)) {
E_ERROR("Utterance %s < %d frames (%d); ignored\n", uttid,
(w << 1) + 1, nfr);
return;
}
ptmr_reset_all(timers);
ptmr_reset(&tm_utt);
ptmr_start(&tm_utt);
ptmr_reset(&tm_ovrhd);
ptmr_start(&tm_ovrhd);
ptmr_start(timers + tmr_utt);
if (align_build_sent_hmm(sent, cmd_ln_int32_r(kbc->config, "-insert_sil")) != 0) {
align_destroy_sent_hmm();
ptmr_stop(timers + tmr_utt);
E_ERROR("No sentence HMM; no alignment for %s\n", uttid);
return;
}
align_start_utt(uttid);
for (i = 0; i < nfr; i++) {
ptmr_start(timers + tmr_utt);
/* Obtain active senone flags */
ptmr_start(timers + tmr_gauden);
ptmr_start(timers + tmr_senone);
align_sen_active(ascr->sen_active, ascr->n_sen);
/* Bah, there ought to be a function for this. */
if (kbc->ms_mgau) {
ms_cont_mgau_frame_eval(ascr,
kbc->ms_mgau,
kbc->mdef, feat[i], i);
}
else if (kbc->s2_mgau) {
s2_semi_mgau_frame_eval(kbc->s2_mgau,
ascr, fastgmm, feat[i],
i);
}
else if (kbc->mgau) {
approx_cont_mgau_ci_eval(kbcore_svq(kbc),
kbcore_gs(kbc),
kbcore_mgau(kbc),
fastgmm,
kbc->mdef,
feat[i][0],
ascr->cache_ci_senscr[0],
&(ascr->cache_best_list[0]), i,
kbcore_logmath(kbc));
approx_cont_mgau_frame_eval(kbcore_mdef(kbc),
kbcore_svq(kbc),
kbcore_gs(kbc),
kbcore_mgau(kbc),
fastgmm, ascr,
feat[i][0], i,
ascr->
cache_ci_senscr[0],
&tm_ovrhd,
kbcore_logmath(kbc));
}
ptmr_stop(timers + tmr_gauden);
ptmr_stop(timers + tmr_senone);
/* Step alignment one frame forward */
ptmr_start(timers + tmr_align);
align_frame(ascr->senscr);
ptmr_stop(timers + tmr_align);
ptmr_stop(timers + tmr_utt);
}
ptmr_stop(&tm_utt);
ptmr_stop(&tm_ovrhd);
printf("\n");
/* Wind up alignment for this utterance */
if (align_end_utt(&stseg, &phseg, &wdseg) < 0)
E_ERROR("Final state not reached; no alignment for %s\n\n", uttid);
else {
if (s2stsegdir)
write_s2stseg(s2stsegdir, stseg, uttid, ctlspec, cmd_ln_boolean_r(kbc->config, "-s2cdsen"));
if (stsegdir)
write_stseg(stsegdir, stseg, uttid, ctlspec);
if (phsegdir)
write_phseg(phsegdir, phseg, uttid, ctlspec);
if (phlabdir)
write_phlab(phlabdir, phseg, uttid, ctlspec, cmd_ln_int32_r(kbc->config, "-frate"));
if (wdsegdir)
write_wdseg(wdsegdir, wdseg, uttid, ctlspec);
if (outsentfp)
write_outsent(outsentfp, wdseg, uttid);
if (outctlfp)
write_outctl(outctlfp, ctlspec);
}
align_destroy_sent_hmm();
ptmr_print_all(stdout, timers, nfr * 0.1);
printf
("EXECTIME: %5d frames, %7.2f sec CPU, %6.2f xRT; %7.2f sec elapsed, %6.2f xRT\n",
nfr, tm_utt.t_cpu, tm_utt.t_cpu * 100.0 / nfr, tm_utt.t_elapsed,
tm_utt.t_elapsed * 100.0 / nfr);
tot_nfr += nfr;
}
int32
baum_welch_update(float64 *log_forw_prob,
vector_t **feature,
uint32 n_obs,
state_t *state,
uint32 n_state,
model_inventory_t *inv,
float64 a_beam,
float64 b_beam,
float32 spthresh,
s3phseg_t *phseg,
int32 mixw_reest,
int32 tmat_reest,
int32 mean_reest,
int32 var_reest,
int32 pass2var,
int32 var_is_full,
FILE *pdumpfh,
bw_timers_t *timers,
feat_t *fcb)
{
float64 *scale = NULL;
float64 **dscale = NULL;
float64 **active_alpha;
uint32 **active_astate;
uint32 **bp;
uint32 *n_active_astate;
float64 log_fp; /* accumulator for the log of the probability
* of observing the input given the model */
uint32 t; /* time */
int ret;
uint32 i,j;
/* caller must ensure that there is some non-zero amount
of work to be done here */
assert(n_obs > 0);
assert(n_state > 0);
scale = (float64 *)ckd_calloc(n_obs, sizeof(float64));
dscale = (float64 **)ckd_calloc(n_obs, sizeof(float64 *));
n_active_astate = (uint32 *)ckd_calloc(n_obs, sizeof(uint32));
active_alpha = (float64 **)ckd_calloc(n_obs, sizeof(float64 *));
active_astate = (uint32 **)ckd_calloc(n_obs, sizeof(uint32 *));
bp = (uint32 **)ckd_calloc(n_obs, sizeof(uint32 *));
/* Compute the scaled alpha variable and scale factors
* for all states and time subject to the pruning constraints */
if (timers)
ptmr_start(&timers->fwd_timer);
/*
* Debug?
* E_INFO("Before Forward search\n");
*/
ret = forward(active_alpha, active_astate, n_active_astate, bp,
scale, dscale,
feature, n_obs, state, n_state,
inv, a_beam, phseg, timers, 0);
#if BW_DEBUG
for (i=0 ; i < n_obs; i++) {
E_INFO("Number of active states %d at time %d\n",n_active_astate[i],i);
E_INFO("Scale of time %d is %e \n",i,scale[i]);
for(j=0 ; j < n_active_astate[i]; j++) {
E_INFO("Active state: %d Active alpha: %e\n",active_astate[i][j], active_alpha[i][j]);
}
}
i=0;
j=0;
#endif
/* Dump a phoneme segmentation if requested */
if (cmd_ln_str("-outphsegdir")) {
const char *phsegdir;
char *segfn, *uttid;
phsegdir = cmd_ln_str("-outphsegdir");
uttid = (cmd_ln_int32("-outputfullpath")
? corpus_utt_full_name() : corpus_utt());
segfn = ckd_calloc(strlen(phsegdir) + 1
+ strlen(uttid)
+ strlen(".phseg") + 1, 1);
strcpy(segfn, phsegdir);
strcat(segfn, "/");
strcat(segfn, uttid);
strcat(segfn, ".phseg");
write_phseg(segfn, inv, state, active_astate, n_active_astate,
n_state, n_obs, active_alpha, scale, bp);
ckd_free(segfn);
}
if (timers)
ptmr_stop(&timers->fwd_timer);
if (ret != S3_SUCCESS) {
/* Some problem with the utterance, release per utterance storage and
* forget about adding the utterance accumulators to the global accumulators */
goto error;
}
/* Compute the scaled beta variable and update the reestimation
* sums */
if (timers)
ptmr_start(&timers->bwd_timer);
#if BW_DEBUG
E_INFO("Before Backward search\n");
#endif
ret = backward_update(active_alpha, active_astate, n_active_astate, scale, dscale,
feature, n_obs,
state, n_state,
inv, b_beam, spthresh,
mixw_reest, tmat_reest, mean_reest, var_reest, pass2var,
var_is_full, pdumpfh, timers, fcb);
if (timers)
ptmr_stop(&timers->bwd_timer);
if (ret != S3_SUCCESS) {
/* Some problem with the utterance, release per utterance storage and
* forget about adding the utterance accumulators to the global accumulators */
goto error;
}
#if BW_DEBUG
E_INFO("Before Global Accumulation\n");
#endif
/* If no error was found in the forward or backward procedures,
* add the resulting utterance reestimation accumulators to the
* global reestimation accumulators */
if (timers)
ptmr_start(&timers->rstu_timer);
accum_global(inv, state, n_state,
mixw_reest, tmat_reest, mean_reest, var_reest,
var_is_full);
if (timers)
ptmr_stop(&timers->rstu_timer);
for (i = 0; i < n_active_astate[n_obs-1] && active_astate[n_obs-1][i] != (n_state-1); i++);
assert(i < n_active_astate[n_obs-1]);
/* Calculate log[ p( O | \lambda ) ] */
assert(active_alpha[n_obs-1][i] > 0);
log_fp = log(active_alpha[n_obs-1][i]);
for (t = 0; t < n_obs; t++) {
assert(scale[t] > 0);
log_fp -= log(scale[t]);
for (j = 0; j < inv->gauden->n_feat; j++) {
log_fp += dscale[t][j];
}
}
*log_forw_prob = log_fp;
ckd_free((void *)scale);
ckd_free(n_active_astate);
for (i = 0; i < n_obs; i++) {
ckd_free((void *)active_alpha[i]);
ckd_free((void *)active_astate[i]);
ckd_free((void *)dscale[i]);
ckd_free((void *)bp[i]);
}
ckd_free((void *)active_alpha);
ckd_free((void *)active_astate);
ckd_free((void **)dscale);
ckd_free(bp);
return S3_SUCCESS;
error:
ckd_free((void *)scale);
for (i = 0; i < n_obs; i++) {
if (dscale[i])
ckd_free((void *)dscale[i]);
}
ckd_free((void **)dscale);
ckd_free(n_active_astate);
for (i = 0; i < n_obs; i++) {
ckd_free((void *)active_alpha[i]);
ckd_free((void *)active_astate[i]);
ckd_free((void *)bp[i]);
}
ckd_free((void *)active_alpha);
ckd_free((void *)active_astate);
ckd_free(bp);
E_ERROR("%s ignored\n", corpus_utt_brief_name());
return S3_ERROR;
}
static void process_reffile (char *reffile)
{
FILE *rfp, *hfp;
char line[16384], uttid[4096], file[4096], lc_uttid[4096];
int32 i, k;
dagnode_t ref[MAX_UTT_LEN];
int32 nref, noov, nhyp;
int32 tot_err, tot_ref, tot_corr, tot_oov, tot_hyp;
dag_t *dag;
dpnode_t retval;
ptmr_t *tm;
char *latdir, *hypfile;
if ((rfp = fopen(reffile, "r")) == NULL)
E_FATAL("fopen(%s,r) failed\n", reffile);
latdir = (char *) cmd_ln_access ("-latdir");
hypfile = (char *) cmd_ln_access ("-hyp");
if ((! latdir) && (! hypfile))
E_FATAL("Both -latdir and -hyp arguments missing\n");
if (latdir && hypfile)
E_FATAL("-latdir and -hyp arguments are mutually exclusive\n");
hfp = NULL;
if (hypfile) {
if ((hfp = fopen(hypfile, "r")) == NULL)
E_FATAL("fopen(%s,r) failed\n", hypfile);
}
tot_err = 0;
tot_ref = 0;
tot_hyp = 0;
tot_corr = 0;
tot_oov = 0;
tm = (ptmr_t *) ckd_calloc (1, sizeof(ptmr_t));
while (fgets(line, sizeof(line), rfp) != NULL) {
ptmr_reset (tm);
ptmr_start (tm);
if ((nref = refline2wds (line, ref, &noov, uttid)) < 0)
E_FATAL("Bad line in file %s: %s\n", reffile, line);
/* Read lattice or hypfile, whichever is specified */
if (latdir) {
sprintf (file, "%s/%s.lat", latdir, uttid);
dag = dag_load (file);
if (! dag) {
/* Try lower casing uttid */
strcpy (lc_uttid, uttid);
lcase (lc_uttid);
sprintf (file, "%s/%s.lat", latdir, lc_uttid);
dag = dag_load (file);
}
} else {
if (fgets(line, sizeof(line), hfp) == NULL)
E_FATAL("Premature EOF(%s) at uttid %s\n", hypfile, uttid);
dag = hypline2dag (uttid, line);
}
if (dag) {
/* Append sentinel silwid node to end of DAG */
dag_append_sentinel (dag, silwid);
/* Find best path (returns #errors/#correct and updates *nhyp) */
retval = dp (uttid, dict, oovbegin, ref, nref, dag, &nhyp, 0, 1);
dag_destroy (dag);
} else {
retval.c = 0;
retval.e = nref-1;
nhyp = 0;
}
ptmr_stop (tm);
tot_ref += nref-1;
tot_hyp += nhyp;
tot_err += retval.e;
tot_corr += retval.c;
tot_oov += noov;
printf("(%s) << %d ref; %d %.1f%% oov; %d hyp; %d %.1f%% corr; %d %.1f%% err; %.1fs CPU >>\n",
uttid, nref-1,
noov, (nref > 1) ? (noov * 100.0) / (nref-1) : 0.0,
nhyp,
retval.c, (nref > 1) ? (retval.c * 100.0) / (nref-1) : 0.0,
retval.e, (nref > 1) ? (retval.e * 100.0) / (nref-1) : 0.0,
tm->t_cpu);
printf("== %7d ref; %5d %5.1f%% oov; %7d hyp; %7d %5.1f%% corr; %6d %5.1f%% err; %5.1fs CPU; %s\n",
tot_ref,
tot_oov, (tot_ref > 0) ? (tot_oov * 100.0) / tot_ref : 0.0,
tot_hyp,
tot_corr, (tot_ref > 0) ? (tot_corr * 100.0) / tot_ref : 0.0,
tot_err, (tot_ref > 0) ? (tot_err * 100.0) / tot_ref : 0.0,
tm->t_tot_cpu,
uttid);
fflush (stderr);
fflush (stdout);
}
fclose (rfp);
if (hfp)
fclose (hfp);
printf("SUMMARY: %d ref; %d %.3f%% oov; %d hyp; %d %.3f%% corr; %d %.3f%% err; %.1fs CPU\n",
tot_ref,
tot_oov, (tot_ref > 0) ? (tot_oov * 100.0) / tot_ref : 0.0,
tot_hyp,
tot_corr, (tot_ref > 0) ? (tot_corr * 100.0) / tot_ref : 0.0,
tot_err, (tot_ref > 0) ? (tot_err * 100.0) / tot_ref : 0.0,
tm->t_tot_cpu);
}
int32
viterbi_update(float64 *log_forw_prob,
vector_t **feature,
uint32 n_obs,
state_t *state_seq,
uint32 n_state,
model_inventory_t *inv,
float64 a_beam,
float32 spthresh,
s3phseg_t *phseg,
int32 mixw_reest,
int32 tmat_reest,
int32 mean_reest,
int32 var_reest,
int32 pass2var,
int32 var_is_full,
FILE *pdumpfh,
bw_timers_t *timers,
feat_t *fcb)
{
float64 *scale = NULL;
float64 **dscale = NULL;
float64 **active_alpha;
uint32 **active_astate;
uint32 **bp;
uint32 *n_active_astate;
gauden_t *g; /* Gaussian density parameters and
reestimation sums */
float32 ***mixw; /* all mixing weights */
float64 ***now_den = NULL; /* Short for den[t] */
uint32 ***now_den_idx = NULL;/* Short for den_idx[t] */
uint32 *active_cb;
uint32 n_active_cb;
float32 **tacc; /* Transition matrix reestimation sum accumulators
for the utterance. */
float32 ***wacc; /* mixing weight reestimation sum accumulators
for the utterance. */
float32 ***denacc = NULL; /* mean/var reestimation accumulators for time t */
size_t denacc_size; /* Total size of data references in denacc. Allows
for quick clears between time frames */
uint32 n_lcl_cb;
uint32 *cb_inv;
uint32 i, j, q;
int32 t;
uint32 n_feat;
uint32 n_density;
uint32 n_top;
int ret;
float64 log_fp; /* accumulator for the log of the probability
* of observing the input given the model */
uint32 max_n_next = 0;
uint32 n_cb;
static float64 *p_op = NULL;
static float64 *p_ci_op = NULL;
static float64 **d_term = NULL;
static float64 **d_term_ci = NULL;
/* caller must ensure that there is some non-zero amount
of work to be done here */
assert(n_obs > 0);
assert(n_state > 0);
g = inv->gauden;
n_feat = gauden_n_feat(g);
n_density = gauden_n_density(g);
n_top = gauden_n_top(g);
n_cb = gauden_n_mgau(g);
if (p_op == NULL) {
p_op = ckd_calloc(n_feat, sizeof(float64));
p_ci_op = ckd_calloc(n_feat, sizeof(float64));
}
if (d_term == NULL) {
d_term = (float64 **)ckd_calloc_2d(n_feat, n_top, sizeof(float64));
d_term_ci = (float64 **)ckd_calloc_2d(n_feat, n_top, sizeof(float64));
}
scale = (float64 *)ckd_calloc(n_obs, sizeof(float64));
dscale = (float64 **)ckd_calloc(n_obs, sizeof(float64 *));
n_active_astate = (uint32 *)ckd_calloc(n_obs, sizeof(uint32));
active_alpha = (float64 **)ckd_calloc(n_obs, sizeof(float64 *));
active_astate = (uint32 **)ckd_calloc(n_obs, sizeof(uint32 *));
active_cb = ckd_calloc(2*n_state, sizeof(uint32));
bp = (uint32 **)ckd_calloc(n_obs, sizeof(uint32 *));
/* Run forward algorithm, which has embedded Viterbi. */
if (timers)
ptmr_start(&timers->fwd_timer);
ret = forward(active_alpha, active_astate, n_active_astate, bp,
scale, dscale,
feature, n_obs, state_seq, n_state,
inv, a_beam, phseg, timers, 0);
/* Dump a phoneme segmentation if requested */
if (cmd_ln_str("-outphsegdir")) {
const char *phsegdir;
char *segfn, *uttid;
phsegdir = cmd_ln_str("-outphsegdir");
uttid = (cmd_ln_int32("-outputfullpath")
? corpus_utt_full_name() : corpus_utt());
segfn = ckd_calloc(strlen(phsegdir) + 1
+ strlen(uttid)
+ strlen(".phseg") + 1, 1);
strcpy(segfn, phsegdir);
strcat(segfn, "/");
strcat(segfn, uttid);
strcat(segfn, ".phseg");
write_phseg(segfn, inv, state_seq, active_astate, n_active_astate,
n_state, n_obs, active_alpha, scale, bp);
ckd_free(segfn);
}
if (timers)
ptmr_stop(&timers->fwd_timer);
if (ret != S3_SUCCESS) {
/* Some problem with the utterance, release per utterance storage and
* forget about adding the utterance accumulators to the global accumulators */
goto all_done;
}
mixw = inv->mixw;
if (mixw_reest) {
/* Need to reallocate mixing accumulators for utt */
if (inv->l_mixw_acc) {
ckd_free_3d((void ***)inv->l_mixw_acc);
inv->l_mixw_acc = NULL;
}
inv->l_mixw_acc = (float32 ***)ckd_calloc_3d(inv->n_mixw_inverse,
n_feat,
n_density,
sizeof(float32));
}
wacc = inv->l_mixw_acc;
n_lcl_cb = inv->n_cb_inverse;
cb_inv = inv->cb_inverse;
/* Allocate local accumulators for mean, variance reestimation
sums if necessary */
gauden_alloc_l_acc(g, n_lcl_cb,
mean_reest, var_reest,
var_is_full);
if (tmat_reest) {
if (inv->l_tmat_acc) {
ckd_free_2d((void **)inv->l_tmat_acc);
inv->l_tmat_acc = NULL;
}
for (i = 0; i < n_state; i++) {
if (state_seq[i].n_next > max_n_next)
max_n_next = state_seq[i].n_next;
}
inv->l_tmat_acc = (float32 **)ckd_calloc_2d(n_state,
max_n_next,
sizeof(float32));
}
/* transition matrix reestimation sum accumulators
for the utterance */
tacc = inv->l_tmat_acc;
n_active_cb = 0;
now_den = (float64 ***)ckd_calloc_3d(n_lcl_cb,
n_feat,
n_top,
sizeof(float64));
now_den_idx = (uint32 ***)ckd_calloc_3d(n_lcl_cb,
n_feat,
n_top,
sizeof(uint32));
if (mean_reest || var_reest) {
/* allocate space for the per frame density counts */
denacc = (float32 ***)ckd_calloc_3d(n_lcl_cb,
n_feat,
n_density,
sizeof(float32));
/* # of bytes required to store all weighted vectors */
denacc_size = n_lcl_cb * n_feat * n_density * sizeof(float32);
}
else {
denacc = NULL;
denacc_size = 0;
}
/* Okay now run through the backtrace and accumulate counts. */
/* Find the non-emitting ending state */
for (q = 0; q < n_active_astate[n_obs-1]; ++q) {
if (active_astate[n_obs-1][q] == n_state-1)
break;
}
if (q == n_active_astate[n_obs-1]) {
E_ERROR("Failed to align audio to trancript: final state of the search is not reached\n");
ret = S3_ERROR;
goto all_done;
}
for (t = n_obs-1; t >= 0; --t) {
uint32 l_cb;
uint32 l_ci_cb;
float64 op, p_reest_term;
uint32 prev;
j = active_astate[t][q];
/* Follow any non-emitting states at time t first. */
while (state_seq[j].mixw == TYING_NON_EMITTING) {
prev = active_astate[t][bp[t][q]];
#if VITERBI_DEBUG
printf("Following non-emitting state at time %d, %u => %u\n",
t, j, prev);
#endif
/* Backtrace and accumulate transition counts. */
if (tmat_reest) {
assert(tacc != NULL);
tacc[prev][j - prev] += 1.0;
}
q = bp[t][q];
j = prev;
}
/* Now accumulate statistics for the real state. */
l_cb = state_seq[j].l_cb;
l_ci_cb = state_seq[j].l_ci_cb;
n_active_cb = 0;
if (timers)
ptmr_start(&timers->gau_timer);
gauden_compute_log(now_den[l_cb],
now_den_idx[l_cb],
feature[t],
g,
state_seq[j].cb,
NULL);
active_cb[n_active_cb++] = l_cb;
if (l_cb != l_ci_cb) {
gauden_compute_log(now_den[l_ci_cb],
now_den_idx[l_ci_cb],
feature[t],
g,
state_seq[j].ci_cb,
NULL);
active_cb[n_active_cb++] = l_ci_cb;
}
gauden_scale_densities_bwd(now_den, now_den_idx,
&dscale[t],
active_cb, n_active_cb, g);
assert(state_seq[j].mixw != TYING_NON_EMITTING);
/* Now calculate mixture densities. */
/* This is the normalizer sum_m c_{jm} p(o_t|\lambda_{jm}) */
op = gauden_mixture(now_den[l_cb], now_den_idx[l_cb],
mixw[state_seq[j].mixw], g);
if (timers)
ptmr_stop(&timers->gau_timer);
if (timers)
ptmr_start(&timers->rsts_timer);
/* Make up this bogus value to be consistent with backward.c */
p_reest_term = 1.0 / op;
/* Compute the output probability excluding the contribution
* of each feature stream. i.e. p_op[0] is the output
* probability excluding feature stream 0 */
partial_op(p_op,
op,
now_den[l_cb],
now_den_idx[l_cb],
mixw[state_seq[j].mixw],
n_feat,
n_top);
/* compute the probability of each (of possibly topn) density */
den_terms(d_term,
p_reest_term,
p_op,
now_den[l_cb],
now_den_idx[l_cb],
mixw[state_seq[j].mixw],
n_feat,
n_top);
if (l_cb != l_ci_cb) {
/* For each feature stream f, compute:
* sum_k(mixw[f][k] den[f][k])
* and store the results in p_ci_op */
partial_ci_op(p_ci_op,
now_den[l_ci_cb],
now_den_idx[l_ci_cb],
mixw[state_seq[j].ci_mixw],
n_feat,
n_top);
/* For each feature stream and density compute the terms:
* w[f][k] den[f][k] / sum_k(w[f][k] den[f][k]) * post_j
* and store results in d_term_ci */
den_terms_ci(d_term_ci,
1.0, /* post_j = 1.0 */
p_ci_op,
now_den[l_ci_cb],
now_den_idx[l_ci_cb],
mixw[state_seq[j].ci_mixw],
n_feat,
n_top);
}
/* accumulate the probability for each density in the mixing
* weight reestimation accumulators */
if (mixw_reest) {
accum_den_terms(wacc[state_seq[j].l_mixw], d_term,
now_den_idx[l_cb], n_feat, n_top);
/* check if mixw and ci_mixw are different to avoid
* doubling the EM counts in a CI run. */
if (state_seq[j].mixw != state_seq[j].ci_mixw) {
if (n_cb < inv->n_mixw) {
/* semi-continuous, tied mixture, and discrete case */
accum_den_terms(wacc[state_seq[j].l_ci_mixw], d_term,
now_den_idx[l_cb], n_feat, n_top);
}
else {
/* continuous case */
accum_den_terms(wacc[state_seq[j].l_ci_mixw], d_term_ci,
now_den_idx[l_ci_cb], n_feat, n_top);
}
}
}
/* accumulate the probability for each density in the
* density reestimation accumulators */
if (mean_reest || var_reest) {
accum_den_terms(denacc[l_cb], d_term,
now_den_idx[l_cb], n_feat, n_top);
if (l_cb != l_ci_cb) {
accum_den_terms(denacc[l_ci_cb], d_term_ci,
now_den_idx[l_ci_cb], n_feat, n_top);
}
}
if (timers)
ptmr_stop(&timers->rsts_timer);
/* Note that there is only one state/frame so this is kind of
redundant */
if (timers)
ptmr_start(&timers->rstf_timer);
if (mean_reest || var_reest) {
/* Update the mean and variance reestimation accumulators */
if (pdumpfh)
fprintf(pdumpfh, "time %d:\n", t);
accum_gauden(denacc,
cb_inv,
n_lcl_cb,
feature[t],
now_den_idx,
g,
mean_reest,
var_reest,
pass2var,
inv->l_mixw_acc,
var_is_full,
pdumpfh,
fcb);
memset(&denacc[0][0][0], 0, denacc_size);
}
if (timers)
ptmr_stop(&timers->rstf_timer);
if (t > 0) {
prev = active_astate[t-1][bp[t][q]];
#if VITERBI_DEBUG
printf("Backtrace at time %d, %u => %u\n",
t, j, prev);
#endif
/* Backtrace and accumulate transition counts. */
if (tmat_reest) {
assert(tacc != NULL);
tacc[prev][j-prev] += 1.0;
}
q = bp[t][q];
j = prev;
}
}
/* If no error was found, add the resulting utterance reestimation
* accumulators to the global reestimation accumulators */
if (timers)
ptmr_start(&timers->rstu_timer);
accum_global(inv, state_seq, n_state,
mixw_reest, tmat_reest, mean_reest, var_reest,
var_is_full);
if (timers)
ptmr_stop(&timers->rstu_timer);
/* Find the final state */
for (i = 0; i < n_active_astate[n_obs-1]; ++i) {
if (active_astate[n_obs-1][i] == n_state-1)
break;
}
/* Calculate log[ p( O | \lambda ) ] */
assert(active_alpha[n_obs-1][i] > 0);
log_fp = log(active_alpha[n_obs-1][i]);
for (t = 0; t < n_obs; t++) {
assert(scale[t] > 0);
log_fp -= log(scale[t]);
for (j = 0; j < inv->gauden->n_feat; j++) {
log_fp += dscale[t][j];
}
}
*log_forw_prob = log_fp;
all_done:
ckd_free((void *)scale);
for (i = 0; i < n_obs; i++) {
if (dscale[i])
ckd_free((void *)dscale[i]);
}
ckd_free((void **)dscale);
ckd_free(n_active_astate);
for (i = 0; i < n_obs; i++) {
ckd_free((void *)active_alpha[i]);
ckd_free((void *)active_astate[i]);
ckd_free((void *)bp[i]);
}
ckd_free((void *)active_alpha);
ckd_free((void *)active_astate);
ckd_free((void *)active_cb);
if (denacc)
ckd_free_3d((void ***)denacc);
if (now_den)
ckd_free_3d((void ***)now_den);
if (now_den_idx)
ckd_free_3d((void ***)now_den_idx);
if (ret != S3_SUCCESS)
E_ERROR("%s ignored\n", corpus_utt_brief_name());
return ret;
}
