/*
* Find Viterbi alignment.
*/
static void align_utt (char *sent, /* In: Reference transcript */
float32 **mfc, /* In: MFC cepstra for input utterance */
int32 nfr, /* In: #frames of input */
char *ctlspec, /* In: Utt specifiction from control file */
char *uttid) /* In: Utterance id, for logging and other use */
{
static float32 **feat = NULL;
static int32 w;
static int32 topn;
static gauden_dist_t ***dist;
static int32 *senscr;
static s3senid_t *sen_active;
static int8 *mgau_active;
static char *s2stsegdir;
static char *stsegdir;
static char *phsegdir;
static char *wdsegdir;
int32 i, s, sid, gid, n_sen_active, best;
char *arg;
align_stseg_t *stseg;
align_phseg_t *phseg;
align_wdseg_t *wdseg;
if (! feat) {
/* One-time allocation of necessary intermediate variables */
/* Allocate space for a feature vector */
feat = (float32 **) ckd_calloc (n_feat, sizeof(float32 *));
for (i = 0; i < n_feat; i++)
feat[i] = (float32 *) ckd_calloc (featlen[i], sizeof(float32));
/* Allocate space for top-N codeword density values in a codebook */
w = feat_window_size (); /* #MFC vectors needed on either side of current
frame to compute one feature vector */
topn = *((int32 *) cmd_ln_access("-topn"));
if (topn > g->n_density) {
E_ERROR("-topn argument (%d) > #density codewords (%d); set to latter\n",
topn, g->n_density);
topn = g->n_density;
}
dist = (gauden_dist_t ***) ckd_calloc_3d (g->n_mgau, n_feat, topn,
sizeof(gauden_dist_t));
/* Space for one frame of senone scores, and per frame active flags */
senscr = (int32 *) ckd_calloc (sen->n_sen, sizeof(int32));
sen_active = (s3senid_t *) ckd_calloc (sen->n_sen, sizeof(s3senid_t));
mgau_active = (int8 *) ckd_calloc (g->n_mgau, sizeof(int8));
/* Note various output directories */
s2stsegdir = NULL;
stsegdir = NULL;
phsegdir = NULL;
wdsegdir = NULL;
if ((arg = (char *) cmd_ln_access ("-s2stsegdir")) != NULL)
s2stsegdir = (char *) ckd_salloc (arg);
if ((arg = (char *) cmd_ln_access ("-stsegdir")) != NULL)
stsegdir = (char *) ckd_salloc (arg);
if ((arg = (char *) cmd_ln_access ("-phsegdir")) != NULL)
phsegdir = (char *) ckd_salloc (arg);
if ((arg = (char *) cmd_ln_access ("-wdsegdir")) != NULL)
wdsegdir = (char *) ckd_salloc (arg);
}
/* HACK HACKA HACK BHIKSHA
if (nfr <= (w<<1)) {
E_ERROR("Utterance %s < %d frames (%d); ignored\n", uttid, (w<<1)+1, nfr);
return;
}
END HACK HACKA HACK */
cyctimer_reset_all ();
counter_reset_all ();
timing_reset (tm_utt);
timing_start (tm_utt);
cyctimer_resume (tmr_utt);
/* AGC and CMN */
arg = (char *) cmd_ln_access ("-cmn");
if (strcmp (arg, "current") == 0)
norm_mean (mfc-4, nfr+8, cepsize); /* -4 HACKA HACK */
arg = (char *) cmd_ln_access ("-agc");
if (strcmp (arg, "max") == 0)
agc_max (mfc, nfr);
if (align_build_sent_hmm (sent) != 0) {
align_destroy_sent_hmm ();
cyctimer_pause (tmr_utt);
E_ERROR("No sentence HMM; no alignment for %s\n", uttid);
return;
}
align_start_utt (uttid);
/*
* A feature vector for frame f depends on input MFC vectors [f-w..f+w]. Hence
* the feature vector corresponding to the first w and last w input frames is
* undefined. We define them by simply replicating the first and last true
* feature vectors (presumably silence regions).
*/
for (i = 0; i < nfr; i++) {
cyctimer_resume (tmr_utt);
/* Compute feature vector for current frame from input speech cepstra */
/* HACK HACKA HACK BHIKSHA
if (i < w)
feat_cep2feat (mfc+w, feat);
else if (i >= nfr-w)
feat_cep2feat (mfc+(nfr-w-1), feat);
else
END HACK HACKA HACK */
feat_cep2feat (mfc+i, feat);
/*
* Evaluate gaussian density codebooks and senone scores for input codeword.
* Evaluate only active codebooks and senones.
*/
/* Obtain active senone flags */
cyctimer_resume (tmr_senone);
align_sen_active (sen_active, sen->n_sen);
/* Flag all CI senones to active if interpolating */
if (interp) {
for (s = 0; s < mdef->n_ci_sen; s++)
sen_active[s] = 1;
}
/* Turn active flags into list (for faster access) */
n_sen_active = 0;
for (s = 0; s < mdef->n_sen; s++) {
if (sen_active[s])
sen_active[n_sen_active++] = s;
}
cyctimer_pause (tmr_senone);
/* Flag all active mixture-gaussian codebooks */
cyctimer_resume (tmr_gauden);
for (gid = 0; gid < g->n_mgau; gid++)
mgau_active[gid] = 0;
for (s = 0; s < n_sen_active; s++) {
sid = sen_active[s];
mgau_active[sen->mgau[sid]] = 1;
}
/* Compute topn gaussian density values (for active codebooks) */
for (gid = 0; gid < g->n_mgau; gid++)
if (mgau_active[gid])
gauden_dist (g, gid, topn, feat, dist[gid]);
cyctimer_pause (tmr_gauden);
/* Evaluate active senones */
cyctimer_resume (tmr_senone);
best = (int32) 0x80000000;
for (s = 0; s < n_sen_active; s++) {
sid = sen_active[s];
senscr[sid] = senone_eval (sen, sid, dist[sen->mgau[sid]], topn);
if (best < senscr[sid])
best = senscr[sid];
}
if (interp) {
for (s = 0; s < n_sen_active; s++) {
if ((sid = sen_active[s]) >= mdef->n_ci_sen)
interp_cd_ci (interp, senscr, sid, mdef->cd2cisen[sid]);
}
}
/* Normalize senone scores (interpolation above can only lower best score) */
for (s = 0; s < n_sen_active; s++) {
sid = sen_active[s];
senscr[sid] -= best;
}
senscale[i] = best;
cyctimer_pause (tmr_senone);
/* Step alignment one frame forward */
cyctimer_resume (tmr_align);
align_frame (senscr);
cyctimer_pause (tmr_align);
cyctimer_pause (tmr_utt);
}
timing_stop (tm_utt);
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);
if (stsegdir)
write_stseg (stsegdir, stseg, uttid, ctlspec);
if (phsegdir)
write_phseg (phsegdir, phseg, uttid, ctlspec);
if (wdsegdir)
write_wdseg (wdsegdir, wdseg, uttid, ctlspec);
if (outsentfp)
write_outsent (outsentfp, wdseg, uttid);
}
align_destroy_sent_hmm ();
cyctimer_print_all_norm (stdout, nfr*0.01, tmr_utt);
counter_print_all (stdout);
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
ms_cont_mgau_frame_eval(ps_mgau_t * mg,
int16 *senscr,
uint8 *senone_active,
int32 n_senone_active,
mfcc_t ** feat,
int32 frame,
int32 compallsen)
{
ms_mgau_model_t *msg = (ms_mgau_model_t *)mg;
int32 gid;
int32 topn;
int32 best;
gauden_t *g;
senone_t *sen;
topn = ms_mgau_topn(msg);
g = ms_mgau_gauden(msg);
sen = ms_mgau_senone(msg);
if (compallsen) {
int32 s;
for (gid = 0; gid < g->n_mgau; gid++)
gauden_dist(g, gid, topn, feat, msg->dist[gid]);
best = (int32) 0x7fffffff;
for (s = 0; s <(int) sen->n_sen; s++) {
senscr[s] = senone_eval(sen, s, msg->dist[sen->mgau[s]], topn);
if (best > senscr[s]) {
best = senscr[s];
}
}
/* Normalize senone scores */
for (s = 0; s <(int) sen->n_sen; s++) {
int32 bs = senscr[s] - best;
if (bs > 32767)
bs = 32767;
if (bs < -32768)
bs = -32768;
senscr[s] = bs;
}
}
else {
int32 i, n;
/* Flag all active mixture-gaussian codebooks */
for (gid = 0; gid < g->n_mgau; gid++)
msg->mgau_active[gid] = 0;
n = 0;
for (i = 0; i < n_senone_active; i++) {
/* senone_active consists of deltas. */
int32 s = senone_active[i] + n;
msg->mgau_active[sen->mgau[s]] = 1;
n = s;
}
/* Compute topn gaussian density values (for active codebooks) */
for (gid = 0; gid < g->n_mgau; gid++) {
if (msg->mgau_active[gid])
gauden_dist(g, gid, topn, feat, msg->dist[gid]);
}
best = (int32) 0x7fffffff;
n = 0;
for (i = 0; i < n_senone_active; i++) {
int32 s = senone_active[i] + n;
senscr[s] = senone_eval(sen, s, msg->dist[sen->mgau[s]], topn);
if (best > senscr[s]) {
best = senscr[s];
}
n = s;
}
/* Normalize senone scores */
n = 0;
for (i = 0; i < n_senone_active; i++) {
int32 s = senone_active[i] + n;
int32 bs = senscr[s] - best;
if (bs > 32767)
bs = 32767;
if (bs < -32768)
bs = -32768;
senscr[s] = bs;
n = s;
}
}
return 0;
}
int32 acoustic_eval (acoustic_t *am, int32 frm)
{
senone_t *sen;
gauden_t *gau;
int32 m, f, s, best, bestgau;
int32 i, j, k;
float32 **fv;
float32 **mfc;
sen = am->sen;
gau = am->gau;
if (am->mfc) {
mfc = am->mfc + frm;
am->fcb->compute_feat(am->fcb, mfc, am->feat[0]);
fv = am->feat[0];
} else {
fv = am->feat[frm];
}
/* Identify the active mixture Gaussians */
if (senactive_to_mgauactive(am) == 0)
E_FATAL("No states active\n");
/* Since we're going to accumulate into senscr for each feature stream */
memset (am->senscr, 0, sen->n_sen * sizeof(int32));
/* Evaluate all senones; FOR NOW NOT YET OPTIMIZED TO JUST THE ACTIVE ONES */
best = MAX_NEG_INT32;
for (m = 0; m < gau->n_mgau; m++) {
if (bitvec_is_set (am->gauden_active, m)) {
for (f = 0; f < gau->n_feat; f++) {
k = gauden_dist (gau, m, f, fv[f], am->dist);
if (am->dist_valid) {
/* Determine set of active mgau components based on pruning beam */
bestgau = MAX_NEG_INT32;
for (i = 0; i < k; i++)
if (am->dist[i] > bestgau)
bestgau = am->dist[i];
j = 0;
for (i = 0; i < k; i++) {
if (am->dist[i] >= bestgau + am->mgaubeam) {
am->dist_valid[j++] = i;
}
}
am->n_dist_valid = j;
k = j;
am->tot_dist_valid += j;
am->tot_mgau_eval += 1;
}
#if 1
for (i = 0; i < sen->mgau2sen[m].n_sen; i++) {
s = sen->mgau2sen[m].sen[i];
if (bitvec_is_set (am->sen_active, s))
am->senscr[s] += senone_eval (sen, s, f, am->dist,
am->dist_valid, k);
}
#else
senone_eval_all (sen, m, f, am->dist, am->dist_valid, k, am->senscr);
#endif
}
/* Find the best senone score so far */
for (i = 0; i < sen->mgau2sen[m].n_sen; i++) {
s = sen->mgau2sen[m].sen[i];
if (bitvec_is_set (am->sen_active, s) && (am->senscr[s] > best))
best = am->senscr[s];
}
}
}
/* CD-CI likelihood-interpolation (later) */
/* Normalize senone score by subtracting the best */
for (s = 0; s < sen->n_sen; s++)
if (bitvec_is_set (am->sen_active, s))
am->senscr[s] -= best;
return best;
}