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,
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;
timing_t *fwd_timer = NULL;
timing_t *rstu_timer = NULL;
timing_t *gau_timer = NULL;
timing_t *rsts_timer = NULL;
timing_t *rstf_timer = NULL;
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);
/* Get the forward estimation CPU timer */
fwd_timer = timing_get("fwd");
/* Get the per utterance reestimation CPU timer */
rstu_timer = timing_get("rstu");
/* Get the Gaussian density evaluation CPU timer */
gau_timer = timing_get("gau");
/* Get the per state reestimation CPU timer */
rsts_timer = timing_get("rsts");
/* Get the per frame reestimation CPU timer */
rstf_timer = timing_get("rstf");
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 (fwd_timer)
timing_start(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, 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 (fwd_timer)
timing_stop(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 (gau_timer)
timing_start(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 (gau_timer)
timing_stop(gau_timer);
if (rsts_timer)
timing_start(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 (rsts_timer)
timing_stop(rsts_timer);
/* Note that there is only one state/frame so this is kind of
redundant */
if (rstf_timer)
timing_start(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 (rstf_timer)
timing_stop(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 (rstu_timer)
timing_start(rstu_timer);
accum_global(inv, state_seq, n_state,
mixw_reest, tmat_reest, mean_reest, var_reest,
var_is_full);
if (rstu_timer)
timing_stop(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;
}
int32
mmi_viterbi_update(vector_t **feature,
uint32 n_obs,
state_t *state_seq,
uint32 n_state,
model_inventory_t *inv,
float64 a_beam,
int32 mean_reest,
int32 var_reest,
float64 arc_gamma,
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 ***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;
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. */
ret = forward(active_alpha, active_astate, n_active_astate, bp,
scale, dscale,
feature, n_obs, state_seq, n_state,
inv, a_beam, NULL, 1);
if (cmd_ln_str("-outphsegdir")) {
E_FATAL("current MMI implementation don't support -outphsegdir\n");
}
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;
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,
FALSE);
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]];
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;
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;
}
ret = gauden_scale_densities_bwd(now_den, now_den_idx,
&dscale[t],
active_cb, n_active_cb, g);
if (ret != S3_SUCCESS)
goto all_done;
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);
/* 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
* 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);
}
}
/* Note that there is only one state/frame so this is kind of
redundant */
if (mean_reest || var_reest) {
/* Update the mean and variance reestimation accumulators */
mmi_accum_gauden(denacc,
cb_inv,
n_lcl_cb,
feature[t],
now_den_idx,
g,
mean_reest,
var_reest,
arc_gamma,
fcb);
memset(&denacc[0][0][0], 0, denacc_size);
}
if (t > 0) {
prev = active_astate[t-1][bp[t][q]];
q = bp[t][q];
j = prev;
}
}
/* If no error was found, add the resulting utterance reestimation
* accumulators to the global reestimation accumulators */
accum_global(inv, state_seq, n_state,
FALSE, FALSE, mean_reest, var_reest,
FALSE);
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);
ckd_free((void **)bp);
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("viterbi update error in sentence %s\n", corpus_utt_brief_name());
return ret;
}
int
accum_gauden(float32 ***denacc,
uint32 *lcl2gbl,
uint32 n_lcl2gbl,
vector_t *frame,
uint32 ***den_idx,
gauden_t *g,
int32 mean_reest,
int32 var_reest,
int32 pass2var,
float32 ***wacc,
int32 var_is_full,
FILE *pdumpfh,
float32 ***lda)
{
uint32 g_i, i, j, k, kk, l;
vector_t ***macc = g->l_macc; /* local to utt */
vector_t m;
vector_t ***mean = g->mean;
vector_t pm = NULL;
vector_t ***vacc = g->l_vacc;
vector_t v = NULL;
vector_t ****fullvacc = g->l_fullvacc;
vector_t *fv = NULL;
vector_t *cov = NULL;
vector_t dvec = NULL;
float32 ***dnom = g->l_dnom;
float32 diff;
float32 obs_cnt;
vector_t feat = NULL;
/* Apply LDA if desired. */
if (lda) {
/* Note that we ignore -ldadim here, because it's rather
* complicated to change the length of veclen for the
* output only. */
lda_transform(&frame, 1, lda, g->veclen[0], g->veclen[0]);
}
/* for each density family found in the utterance */
for (i = 0; i < n_lcl2gbl; i++) {
g_i = lcl2gbl[i];
/* for each feature */
for (j = 0; j < gauden_n_feat(g); j++) {
feat = frame[j];
if (var_is_full) {
ckd_free_2d((void **)cov);
cov = (vector_t *)ckd_calloc_2d(g->veclen[j], g->veclen[j], sizeof(float32));
ckd_free(dvec);
dvec = ckd_calloc(g->veclen[j], sizeof(float32));
}
/* for each density in the mixture density */
for (kk = 0; kk < gauden_n_top(g); kk++) {
k = den_idx[i][j][kk]; /* i.e. density k is one of the n_top densities */
obs_cnt = denacc[i][j][k]; /* observation count for density (k) at this time frame
given the model */
/* don't bother adding a bunch of
essentially zero values */
if (obs_cnt <= MIN_POS_FLOAT32)
continue;
#ifdef ACCUM_VERBOSE
printf("denacc[%u][%u][%3u] == %.3e\n", i, j, k, obs_cnt);
#endif
if (pdumpfh)
fprintf(pdumpfh, "%u %u %u %g\n", lcl2gbl[i], j, k, obs_cnt);
m = macc[i][j][k]; /* the vector accumulator for mean (i,j,k) */
if (pass2var) {
g_i = lcl2gbl[i];
pm = mean[g_i][j][k]; /* the mean (i, j, k) */
}
if (var_reest) {
if (var_is_full)
fv = fullvacc[i][j][k];
else
v = vacc[i][j][k]; /* the vector accumulator for variance (i,j,k) */
}
if (var_reest && var_is_full) {
if (!pass2var)
outerproduct(cov, feat, feat, g->veclen[j]);
else {
for (l = 0; l < g->veclen[j]; ++l)
dvec[l] = feat[l] - pm[l];
outerproduct(cov, dvec, dvec, g->veclen[j]);
}
scalarmultiply(cov, obs_cnt, g->veclen[j], g->veclen[j]);
matrixadd(fv, cov, g->veclen[j], g->veclen[j]);
}
for (l = 0; l < g->veclen[j]; l++) {
if (mean_reest) {
m[l] += obs_cnt * feat[l];
}
if (var_reest && !var_is_full) {
/* Always reest vars on untransformed features for now */
/* This does NOT work with -2passvar no (for pretty obvious reasons) */
if (!pass2var)
v[l] += obs_cnt * feat[l] * feat[l];
else {
diff = feat[l] - pm[l];
diff *= diff;
v[l] += obs_cnt * diff;
}
}
}
/* accumulate observation count for all densities */
dnom[i][j][k] += obs_cnt;
}
if (var_is_full) {
ckd_free_2d((void **)cov);
cov = NULL;
ckd_free(dvec);
dvec = NULL;
}
}
}
if (pdumpfh)
fputs("\n", pdumpfh);
return S3_SUCCESS;
}