/**
* Parse a word dictionary and set the maximum state length per word.
*
* @param winfo [i/o]
*/
void
voca_set_stats(WORD_INFO *winfo)
{
int w,p,n;
int maxwn;
int maxwlen;
int states;
int models;
int trnum;
maxwn = 0;
maxwlen = 0;
states = 0;
models = 0;
trnum = 0;
for (w=0; w<winfo->num; w++) {
models += winfo->wlen[w];
if (maxwlen < winfo->wlen[w]) maxwlen = winfo->wlen[w];
n = 0;
for (p=0; p<winfo->wlen[w]; p++) {
n += hmm_logical_state_num(winfo->wseq[w][p]) - 2;
}
if (maxwn < n) maxwn = n;
states += n;
if (winfo->is_transparent[w]) trnum++;
}
winfo->maxwn = maxwn;
winfo->maxwlen = maxwlen;
winfo->totalstatenum = states;
winfo->totalmodelnum = models;
winfo->totaltransnum = trnum;
}
/**
* Calculate total number of states in a phoneme sequence.
*
* @param hdseq [in] phoneme sequence as given by pointer list of logical %HMM
* @param hdseqlen [in] length of above
* @param has_sp [in] indicates where short-pause insertion is possible
* @param hmminfo [in] HMM definition
*
* @return the total number of states in the sequence.
*/
static int
totalstatelen(HMM_Logical **hdseq, int hdseqlen, boolean *has_sp, HTK_HMM_INFO *hmminfo)
{
int i, len;
len = 0;
for (i=0;i<hdseqlen;i++) {
len += hmm_logical_state_num(hdseq[i]) - 2;
if (has_sp) {
if (has_sp[i]) {
len += hmm_logical_state_num(hmminfo->sp) - 2;
}
}
}
if (hmminfo->multipath) {
/* add count for the initial and final state */
len += 2;
}
return(len);
}
/**
* <JA>
* 構築された factoring 情報を multipath 用に調整する. factoring 情報を,
* モデル全体をスキップする遷移がある場合はその先の音素へコピーする.
* また，(出力を持たない)文頭文法ノードに単語先頭ノードからコピーする.
*
* @param wchmm [in] 木構造化辞書
* </JA>
* <EN>
* Adjust factoring data in tree lexicon for multipath transition handling.
*
* @param wchmm [in] tree lexicon
* </EN>
*
* @callgraph
* @callergraph
*
*/
void
adjust_sc_index(WCHMM_INFO *wchmm)
{
WORD_ID w;
int i,j,k;
HMM_Logical *ltmp;
int ltmp_state_num;
int ato;
LOGPROB prob;
int node, scid;
A_CELL2 *ac;
/* duplicate scid for HMMs with more than one arc from initial state */
for(w=0;w<wchmm->winfo->num;w++) {
for(k=0;k<wchmm->winfo->wlen[w];k++) {
node = wchmm->offset[w][k];
scid = wchmm->state[node].scid;
if (scid == 0) continue;
ltmp = wchmm->winfo->wseq[w][k];
ltmp_state_num = hmm_logical_state_num(ltmp);
if ((hmm_logical_trans(ltmp))->a[0][ltmp_state_num-1] != LOG_ZERO) {
j = k + 1;
if (j == wchmm->winfo->wlen[w]) {
if (wchmm->state[wchmm->wordend[w]].scid == 0) {
jlog("STAT: word %d: factoring node copied for skip phone\n", w);
wchmm->state[wchmm->wordend[w]].scid = scid;
}
} else {
if (wchmm->state[wchmm->offset[w][j]].scid == 0) {
jlog("STAT: word %d: factoring node copied for skip phone\n", w);
wchmm->state[wchmm->offset[w][j]].scid = scid;
}
}
}
for(ato=1;ato<ltmp_state_num;ato++) {
prob = (hmm_logical_trans(ltmp))->a[0][ato];
if (prob != LOG_ZERO) {
wchmm->state[node+ato-1].scid = scid;
}
}
}
}
/* move scid and fscore on the head state to the head grammar state */
for(i=0;i<wchmm->startnum;i++) {
node = wchmm->startnode[i];
if (wchmm->state[node].out.state != NULL) {
j_internal_error("adjust_sc_index: outprob exist in word-head node??\n");
}
if (wchmm->next_a[node] != LOG_ZERO) {
if (wchmm->state[node+1].scid != 0) {
if (wchmm->state[node].scid != 0 && wchmm->state[node].scid != wchmm->state[node+1].scid) {
j_internal_error("adjust_sc_index: different successor list within word-head phone?\n");
}
wchmm->state[node].scid = wchmm->state[node+1].scid;
wchmm->state[node+1].scid = 0;
}
}
for(ac=wchmm->ac[node];ac;ac=ac->next) {
for(k=0;k<ac->n;k++) {
if (wchmm->state[ac->arc[k]].scid != 0) {
if (wchmm->state[node].scid != 0 && wchmm->state[node].scid != wchmm->state[ac->arc[k]].scid) {
j_internal_error("adjust_sc_index: different successor list within word-head phone?\n");
}
wchmm->state[node].scid = wchmm->state[ac->arc[k]].scid;
wchmm->state[ac->arc[k]].scid = 0;
}
}
}
}
}
/**
* <JA>
* 文全体のHMMを構築し，Viterbiアラインメントを実行し，結果を出力する.
*
* @param words [in] 文仮説をあらわす単語列
* @param wnum [in] @a words の長さ
* @param param [in] 入力特徴パラメータ列
* @param per_what [in] 単語・音素・状態のどの単位でアラインメントを取るかを指定
* @param align [out] アラインメント結果を格納するSentence構造体
* @param r [i/o] 認識処理インスタンス
* </JA>
* <EN>
* Build sentence HMM, call viterbi_segment() and output result.
*
* @param words [in] word sequence of the sentence
* @param wnum [in] number of words in @a words
* @param param [in] input parameter vector
* @param per_what [in] specify the alignment unit (word / phoneme / state)
* @param s [out] Sentence data area to store the alignment result
* @param r [i/o] recognition process instance
* </EN>
*/
static void
do_align(WORD_ID *words, short wnum, HTK_Param *param, int per_what, SentenceAlign *align, RecogProcess *r)
{
HMM_Logical **phones; /* phoneme sequence */
boolean *has_sp; /* whether phone can follow short pause */
int k;
int phonenum; /* num of above */
HMM *shmm; /* sentence HMM */
int *end_state; /* state number of word ends */
int *end_frame; /* segmented last frame of words */
LOGPROB *end_score; /* normalized score of each words */
LOGPROB allscore; /* total score of this word sequence */
WORD_ID w;
int i, rlen;
int end_num = 0;
int *id_seq, *phloc = NULL, *stloc = NULL;
int j,n,p;
WORD_INFO *winfo;
HTK_HMM_INFO *hmminfo;
boolean enable_iwsp; /* for multipath */
winfo = r->lm->winfo;
hmminfo = r->am->hmminfo;
if (hmminfo->multipath) enable_iwsp = r->lm->config->enable_iwsp;
/* initialize result storage buffer */
switch(per_what) {
case PER_WORD:
jlog("ALIGN: === word alignment begin ===\n");
end_num = wnum;
phloc = (int *)mymalloc(sizeof(int)*wnum);
i = 0;
for(w=0;w<wnum;w++) {
phloc[w] = i;
i += winfo->wlen[words[w]];
}
break;
case PER_PHONEME:
jlog("ALIGN: === phoneme alignment begin ===\n");
end_num = 0;
for(w=0;w<wnum;w++) end_num += winfo->wlen[words[w]];
break;
case PER_STATE:
jlog("ALIGN: === state alignment begin ===\n");
end_num = 0;
for(w=0;w<wnum;w++) {
for (i=0;i<winfo->wlen[words[w]]; i++) {
end_num += hmm_logical_state_num(winfo->wseq[words[w]][i]) - 2;
}
if (hmminfo->multipath && enable_iwsp) {
end_num += hmm_logical_state_num(hmminfo->sp) - 2;
}
}
phloc = (int *)mymalloc(sizeof(int)*end_num);
stloc = (int *)mymalloc(sizeof(int)*end_num);
{
n = 0;
p = 0;
for(w=0;w<wnum;w++) {
for(i=0;i<winfo->wlen[words[w]]; i++) {
for(j=0; j<hmm_logical_state_num(winfo->wseq[words[w]][i]) - 2; j++) {
phloc[n] = p;
stloc[n] = j + 1;
n++;
}
if (hmminfo->multipath && enable_iwsp && i == winfo->wlen[words[w]] - 1) {
for(k=0;k<hmm_logical_state_num(hmminfo->sp)-2;k++) {
phloc[n] = p;
stloc[n] = j + 1 + k + end_num;
n++;
}
}
p++;
}
}
}
break;
}
end_state = (int *)mymalloc(sizeof(int) * end_num);
/* make phoneme sequence word sequence */
phones = make_phseq(words, wnum, &has_sp, &phonenum, &end_state, per_what, r);
/* build the sentence HMMs */
shmm = new_make_word_hmm(hmminfo, phones, phonenum, has_sp);
if (shmm == NULL) {
j_internal_error("Error: failed to make word hmm for alignment\n");
}
/* call viterbi segmentation function */
allscore = viterbi_segment(shmm, param, r->wchmm->hmmwrk, hmminfo->multipath, end_state, end_num, &id_seq, &end_frame, &end_score, &rlen);
/* store result to s */
align->num = rlen;
align->unittype = per_what;
align->begin_frame = (int *)mymalloc(sizeof(int) * rlen);
align->end_frame = (int *)mymalloc(sizeof(int) * rlen);
align->avgscore = (LOGPROB *)mymalloc(sizeof(LOGPROB) * rlen);
for(i=0;i<rlen;i++) {
align->begin_frame[i] = (i == 0) ? 0 : end_frame[i-1] + 1;
align->end_frame[i] = end_frame[i];
align->avgscore[i] = end_score[i];
}
switch(per_what) {
case PER_WORD:
align->w = (WORD_ID *)mymalloc(sizeof(WORD_ID) * rlen);
for(i=0;i<rlen;i++) {
align->w[i] = words[id_seq[i]];
}
break;
case PER_PHONEME:
align->ph = (HMM_Logical **)mymalloc(sizeof(HMM_Logical *) * rlen);
for(i=0;i<rlen;i++) {
align->ph[i] = phones[id_seq[i]];
}
break;
case PER_STATE:
align->ph = (HMM_Logical **)mymalloc(sizeof(HMM_Logical *) * rlen);
align->loc = (short *)mymalloc(sizeof(short) * rlen);
if (hmminfo->multipath) align->is_iwsp = (boolean *)mymalloc(sizeof(boolean) * rlen);
for(i=0;i<rlen;i++) {
align->ph[i] = phones[phloc[id_seq[i]]];
if (hmminfo->multipath) {
if (enable_iwsp && stloc[id_seq[i]] > end_num) {
align->loc[i] = stloc[id_seq[i]] - end_num;
align->is_iwsp[i] = TRUE;
} else {
align->loc[i] = stloc[id_seq[i]];
align->is_iwsp[i] = FALSE;
}
} else {
align->loc[i] = stloc[id_seq[i]];
}
}
break;
}
align->allscore = allscore;
free_hmm(shmm);
free(id_seq);
free(phones);
if (has_sp) free(has_sp);
free(end_score);
free(end_frame);
free(end_state);
switch(per_what) {
case PER_WORD:
free(phloc);
break;
case PER_PHONEME:
break;
case PER_STATE:
free(phloc);
free(stloc);
}
}
/**
* <JA>
* 最後の1単語の前向きトレリスを計算して，文仮説の前向き尤度を更新する.
*
* @param now [i/o] 文仮説
* @param param [in] 入力パラメータ列
* @param r [in] 認識処理インスタンス
*
* </JA>
* <EN>
* Compute the forward viterbi for the last word to update forward scores
* and ready for word connection.
*
* @param now [i/o] hypothesis
* @param param [in] input parameter vectors
* @param r [in] recognition process instance
*
* </EN>
* @callgraph
* @callergraph
*/
void
scan_word(NODE *now, HTK_Param *param, RecogProcess *r)
{
int i,t, j;
HMM *whmm;
A_CELL *ac;
WORD_ID word;
LOGPROB tmpmax, tmptmp, score1;
int startt = 0, endt = 0;
int wordhmmnum;
LOGPROB tmpmax_store, store_point_maxarc; /* multipath */
LOGPROB tmpmax2 = LOG_ZERO;
int phmmlen;
HMM_Logical *ret, *wend;
int store_point;
int crossword_point = 0;
boolean back_rescan = FALSE;
boolean node_exist_p;
int tn=0; ///< Temporal pointer to current buffer
int tl=0; ///< Temporal pointer to previous buffer
/* store global values to local for rapid access */
WORD_INFO *winfo;
HTK_HMM_INFO *hmminfo;
LOGPROB *framemaxscore;
int peseqlen;
boolean ccd_flag;
boolean enable_iwsp;
#ifdef SCAN_BEAM
LOGPROB scan_beam_thres;
#endif
StackDecode *dwrk;
winfo = r->lm->winfo;
hmminfo = r->am->hmminfo;
dwrk = &(r->pass2);
peseqlen = r->peseqlen;
framemaxscore = r->pass2.framemaxscore;
ccd_flag = r->ccd_flag;
enable_iwsp = r->lm->config->enable_iwsp; /* multipath */
#ifdef SCAN_BEAM
scan_beam_thres = r->config->pass2.scan_beam_thres;
#endif
if (hmminfo->multipath) {
store_point = -1;
} else {
store_point = 0;
}
/* ----------------------- prepare HMM ----------------------- */
if (ccd_flag) {
/* 直前の音素があれば，そこまでさかのぼって scan する */
/* if there are any last phone, enable backscan */
if (now->last_ph == NULL) {
/* initial score: now->g[] */
/* scan range: phones in now->seq[now->seqnum-1] */
back_rescan = FALSE;
} else {
/* initial score: now->g_prev[] (1-phone before)*/
/* scan range: phones in now->seq[now->seqnum-1] + now->last_ph */
back_rescan = TRUE;
}
}
#ifdef TCD
if (now->last_ph != NULL) {
jlog("DEBUG: inherited last_ph: %s\n", (now->last_ph)->name);
if (now->last_ph_sp_attached) jlog("DEBUG: (sp attached)\n"); /* multipath */
} else {
jlog("DEBUG: no last_ph inherited\n");
}
#endif
/* scan 範囲分のHMMを準備 */
/* prepare HMM of the scan range */
word = now->seq[now->seqnum-1];
if (ccd_flag) {
if (back_rescan) {
/* scan range: phones in now->seq[now->seqnum-1] + now->last_ph */
phmmlen = winfo->wlen[word] + 1;
if (phmmlen > dwrk->phmmlen_max) {
j_internal_error("scan_word: num of phonemes in a word exceed phmmlenmax (%d) ?\n", dwrk->phmmlen_max);
}
for (i=0;i<phmmlen - 2;i++) dwrk->phmmseq[i] = winfo->wseq[word][i];
if (enable_iwsp && hmminfo->multipath) {
for (i=0;i<phmmlen - 2;i++) dwrk->has_sp[i] = FALSE;
}
/* 最終単語と last_ph 間の単語間triphoneを考慮 */
/* consider cross-word context dependency between the last word and now->last_ph */
wend = winfo->wseq[word][winfo->wlen[word]-1];
ret = get_right_context_HMM(wend, now->last_ph->name, hmminfo);
if (ret == NULL) { /* triphone not found */
/* fallback to the original bi/mono-phone */
/* error if the original is pseudo phone (not explicitly defined
in hmmdefs/hmmlist) */
/* exception: word with 1 phone (triphone may exist in the next expansion */
if (winfo->wlen[word] > 1 && wend->is_pseudo) {
error_missing_right_triphone(wend, now->last_ph->name);
}
dwrk->phmmseq[phmmlen-2] = wend;
} else {
dwrk->phmmseq[phmmlen-2] = ret;
}
ret = get_left_context_HMM(now->last_ph, wend->name, hmminfo);
if (ret == NULL) {
/* fallback to the original bi/mono-phone */
/* error if the original is pseudo phone (not explicitly defined
in hmmdefs/hmmlist) */
if (now->last_ph->is_pseudo) {
error_missing_left_triphone(now->last_ph, wend->name);
}
dwrk->phmmseq[phmmlen-1] = now->last_ph;
} else {
dwrk->phmmseq[phmmlen-1] = ret;
}
if (enable_iwsp && hmminfo->multipath) {
dwrk->has_sp[phmmlen-2] = TRUE;
dwrk->has_sp[phmmlen-1] = now->last_ph_sp_attached;
}
#ifdef TCD
jlog("DEBUG: w=");
for(i=0;i<winfo->wlen[word];i++) {
jlog(" %s",(winfo->wseq[word][i])->name);
if (enable_iwsp && hmminfo->multipath && dwrk->has_sp[i]) jlog("(sp)");
}
jlog(" | %s\n", (now->last_ph)->name);
if (hmminfo->multipath && now->last_ph_sp_attached) jlog("DEBUG: (sp)\n");
jlog("DEBUG: scan for:");
for (i=0;i<phmmlen;i++) {
jlog(" %s", dwrk->phmmseq[i]->name);
if (enable_iwsp && hmminfo->multipath && dwrk->has_sp[i]) jlog("(sp)");
}
jlog("\n");
#endif
/* 単語HMMを作る */
/* make word HMM */
whmm = new_make_word_hmm(hmminfo, dwrk->phmmseq, phmmlen, (enable_iwsp && hmminfo->multipath) ? dwrk->has_sp : NULL);
if (whmm == NULL) {
j_internal_error("Error: failed to make word hmm for word #%d \"%s [%s]\"\n", word, winfo->wname[word], winfo->woutput[word]);
}
/* backscan なので，計算前の g[] 初期値は now->g_prev[] を使用 */
/* As backscan enabled, the initial forward score g[] is set by
now->g_prev[] */
for (t=0;t<peseqlen;t++) {
dwrk->g[t]=now->g_prev[t];
}
/* 次段用のg_prevを格納するノード位置を設定 */
/* set where to store scores as new g_prev[] for the next backscan
in the HMM */
if (hmminfo->multipath) {
store_point = hmm_logical_state_num(dwrk->phmmseq[0]) - 2;
store_point_maxarc = max_out_arc(dwrk->phmmseq[0]);
if (enable_iwsp && dwrk->has_sp[0]) {
store_point += hmm_logical_state_num(hmminfo->sp) - 2;
if (store_point_maxarc < max_out_arc(hmminfo->sp)) {
store_point_maxarc = max_out_arc(hmminfo->sp);
}
}
} else {
store_point = hmm_logical_state_num(dwrk->phmmseq[0]) - 2 - 1;
}
/* scan中に直前単語とこの単語をまたぐ場所を設定 */
/* set where is the connection point of the last word in the HMM */
if (hmminfo->multipath) {
crossword_point = whmm->len - hmm_logical_state_num(dwrk->phmmseq[phmmlen-1]);
if (enable_iwsp && dwrk->has_sp[phmmlen-1]) {
crossword_point -= hmm_logical_state_num(hmminfo->sp) - 2;
}
} else {
crossword_point = whmm->len - (hmm_logical_state_num(dwrk->phmmseq[phmmlen-1]) - 2) - 1;
}
} else { /* not backscan mode */
/* scan range: phones in now->seq[now->seqnum-1] */
#ifdef TCD
jlog("DEBUG: scan(org):");
for (i=0;i<winfo->wlen[word];i++) {
jlog(" %s", (winfo->wseq[word][i])->name);
}
jlog("\n");
#endif
if (enable_iwsp && hmminfo->multipath) {
/* 必要ならばショートポーズを挟み込む位置を指定する */
for(i=0;i<winfo->wlen[word];i++) {
dwrk->has_sp[i] = FALSE;
}
dwrk->has_sp[winfo->wlen[word]-1] = TRUE;
}
/* 単語HMMを作る */
/* make word HMM */
whmm = new_make_word_hmm(hmminfo, winfo->wseq[word], winfo->wlen[word], (enable_iwsp && hmminfo->multipath) ? dwrk->has_sp : NULL);
if (whmm == NULL) {
j_internal_error("Error: failed to make word hmm for word #%d \"%s [%s]\"\n", word, winfo->wname[word], winfo->woutput[word]);
}
/* 計算前の g[] 初期値は now->g[] を使用 */
/* the initial forward score g[] is set by now->g[] */
for (t=0;t<peseqlen;t++) {
dwrk->g[t]=now->g[t];
}
/* 次段用のg_prevを格納するノード位置を設定 */
/* set where to store scores as new g_prev[] for the next backscan
in the HMM */
if (hmminfo->multipath) {
store_point = hmm_logical_state_num(winfo->wseq[word][0]) - 2;
store_point_maxarc = max_out_arc(winfo->wseq[word][0]);
if (enable_iwsp && dwrk->has_sp[0]) {
store_point += hmm_logical_state_num(hmminfo->sp) - 2;
if (store_point_maxarc < max_out_arc(hmminfo->sp)) {
store_point_maxarc = max_out_arc(hmminfo->sp);
}
}
} else {
store_point = hmm_logical_state_num(winfo->wseq[word][0]) - 2 - 1;
}
/* scan中に直前単語とこの単語をまたぐ場所は，なし */
/* the connection point of the last word is not exist in the HMM */
crossword_point = -1;
}
} else { /* ccd_flag == FALSE */
if (enable_iwsp && hmminfo->multipath) {
/* 必要ならばショートポーズを挟み込む位置を指定する */
for(i=0;i<winfo->wlen[word];i++) {
dwrk->has_sp[i] = FALSE;
}
dwrk->has_sp[winfo->wlen[word]-1] = TRUE;
}
/* 音素環境非依存の場合は単純に最終単語分の HMM を作成 */
/* for monophone: simple make HMM for the last word */
whmm = new_make_word_hmm(hmminfo, winfo->wseq[word], winfo->wlen[word], (enable_iwsp && hmminfo->multipath) ? dwrk->has_sp : NULL);
if (whmm == NULL) {
j_internal_error("Error: failed to make word hmm for word #%d \"%s [%s]\"\n", word, winfo->wname[word], winfo->woutput[word]);
}
/* 計算前の g[] 初期値は now->g[] を使用 */
/* the initial forward score g[] is set by now->g[] */
for (t=0;t<peseqlen;t++) {
dwrk->g[t]=now->g[t];
}
}
#ifdef TCD
jlog("DEBUG: whmm len = %d\n",whmm->len);
jlog("DEBUG: crossword_point = %d\n", crossword_point);
jlog("DEBUG: g[] store point = %d\n", store_point);
#endif
wordhmmnum = whmm->len;
if (wordhmmnum >= winfo->maxwn + 10) {
j_internal_error("scan_word: word too long (>%d)\n", winfo->maxwn + 10);
}
#ifndef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
if (ccd_flag) {
now->tail_g_score = now->g[now->bestt];
}
}
#endif
/* ----------------------- do scan ----------------------- */
/* scan開始点を検索 -> starttへ*/
/* search for the start frame -> set to startt */
for(t = peseqlen-1; t >=0 ; t--) {
if (
#ifdef SCAN_BEAM
dwrk->g[t] > framemaxscore[t] - scan_beam_thres &&
#endif
dwrk->g[t] > LOG_ZERO) {
break;
}
}
if (t < 0) { /* no node has score > LOG_ZERO */
for(t=0;t<peseqlen;t++) {
if (ccd_flag) now->g_prev[t] = LOG_ZERO;
now->g[t] = LOG_ZERO;
}
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
for(t=0;t<peseqlen;t++) {
now->wordend_frame[t] = -1;
now->wordend_gscore[t] = LOG_ZERO;
}
}
#endif
goto end_of_scan;
}
startt = t;
/* clear [startt+1..peseqlen-1] */
for(t=peseqlen-1;t>startt;t--) {
if (ccd_flag) now->g_prev[t] = LOG_ZERO;
now->g[t] = LOG_ZERO;
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
now->wordend_frame[t] = -1;
now->wordend_gscore[t] = LOG_ZERO;
}
#endif
}
/* バッファポインタ初期化 */
tn = 0; tl = 1;
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
for(i=0;i<wordhmmnum;i++) {
dwrk->wend_token_frame[tn][i] = -1;
dwrk->wend_token_gscore[tn][i] = LOG_ZERO;
}
}
#endif
if (! hmminfo->multipath) {
/* Below initialization is not needed on multipath version, since
the actual viterbi will begin at frame 0 in multipath mode in main loop */
/* 時間 [startt] 上の値を初期化 */
/* initialize scores on frame [startt] */
for(i=0;i<wordhmmnum-1;i++) dwrk->wordtrellis[tn][i] = LOG_ZERO;
dwrk->wordtrellis[tn][wordhmmnum-1] = dwrk->g[startt] + outprob(&(r->am->hmmwrk), startt, &(whmm->state[wordhmmnum-1]), param);
if (ccd_flag) {
now->g_prev[startt] = dwrk->wordtrellis[tn][store_point];
}
now->g[startt] = dwrk->wordtrellis[tn][0];
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
if (ccd_flag) {
if (back_rescan) {
if (wordhmmnum-1 == crossword_point) {
dwrk->wend_token_frame[tn][wordhmmnum-1] = startt;
dwrk->wend_token_gscore[tn][wordhmmnum-1] = dwrk->g[startt];
} else {
dwrk->wend_token_frame[tn][wordhmmnum-1] = -1;
dwrk->wend_token_gscore[tn][wordhmmnum-1] = LOG_ZERO;
}
} else {
dwrk->wend_token_frame[tn][wordhmmnum-1] = startt;
dwrk->wend_token_gscore[tn][wordhmmnum-1] = dwrk->g[startt];
}
} else {
dwrk->wend_token_frame[tn][wordhmmnum-1] = startt;
dwrk->wend_token_gscore[tn][wordhmmnum-1] = dwrk->g[startt];
}
now->wordend_frame[startt] = dwrk->wend_token_frame[tn][0];
now->wordend_gscore[startt] = dwrk->wend_token_gscore[tn][0];
}
#endif
} /* end of hmminfo->multipath */
endt = startt;
/* メインループ: startt から始まり 0 に向かって Viterbi 計算 */
/* main loop: start from [startt], and compute Viterbi toward [0] */
for(t = hmminfo->multipath ? startt : startt - 1; t >= 0; t--) {
/* wordtrellisのワークエリアをスワップ */
i = tn; tn = tl; tl = i;
node_exist_p = FALSE; /* TRUE if there is at least 1 survived node in this frame */
if (hmminfo->multipath) {
/* 端のノード [t][wordhmmnum-1]は g[] を参照する */
/* the edge node [t][wordhmmnum-1] is equal to g[] */
/* ノード [t][wordhmmnum-2..0] についてトレリスを計算 */
/* expand trellis for node [t][wordhmmnum-2..0] */
tmpmax_store = LOG_ZERO;
} else {
/* 端のノード [t][wordhmmnum-1]は，内部遷移 か g[]の高い方になる */
/* the edge node [t][wordhmmnum-1] is either internal transitin or g[] */
tmptmp = LOG_ZERO;
for (ac=whmm->state[wordhmmnum-1].ac;ac;ac=ac->next) {
score1 = dwrk->wordtrellis[tl][ac->arc] + ac->a;
if (tmptmp < score1) {
j = ac->arc;
tmptmp = score1;
}
}
if (dwrk->g[t] > tmptmp) {
tmpmax = dwrk->g[t];
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
if (!back_rescan || wordhmmnum-1 == crossword_point) {
dwrk->wend_token_frame[tn][wordhmmnum-1] = t;
dwrk->wend_token_gscore[tn][wordhmmnum-1] = dwrk->g[t];
} else {
dwrk->wend_token_frame[tn][wordhmmnum-1] = dwrk->wend_token_frame[tl][j];
dwrk->wend_token_gscore[tn][wordhmmnum-1] = dwrk->wend_token_gscore[tl][j];
}
}
#endif
} else {
tmpmax = tmptmp;
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
dwrk->wend_token_frame[tn][wordhmmnum-1] = dwrk->wend_token_frame[tl][j];
dwrk->wend_token_gscore[tn][wordhmmnum-1] = dwrk->wend_token_gscore[tl][j];
}
#endif
}
/* 端のノードのスコアエンベロープチェック: 一定幅外なら落とす */
/* check if the edge node is within score envelope */
if (
#ifdef SCAN_BEAM
tmpmax <= framemaxscore[t] - scan_beam_thres ||
#endif
tmpmax <= LOG_ZERO
) {
dwrk->wordtrellis[tn][wordhmmnum-1] = LOG_ZERO;
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
dwrk->wend_token_frame[tn][wordhmmnum-1] = -1;
dwrk->wend_token_gscore[tn][wordhmmnum-1] = LOG_ZERO;
}
#endif
} else {
node_exist_p = TRUE;
dwrk->wordtrellis[tn][wordhmmnum-1] = tmpmax + outprob(&(r->am->hmmwrk), t, &(whmm->state[wordhmmnum-1]), param);
}
} /* end of ~multipath */
/* ノード [t][wordhmmnum-2..0] についてトレリスを計算 */
/* expand trellis for node [t][wordhmmnum-2..0] */
for(i=wordhmmnum-2;i>=0;i--) {
if (ccd_flag) {
/* 最尤パスと最尤スコア tmpmax を見つける */
/* tmpmax2 は次回用 g_prev[] のための最大値(自己遷移を除いた最大値) */
/* find most likely path and the max score 'tmpmax' */
/* 'tmpmax2' is max score excluding self transition, for next g_prev[] */
if (! hmminfo->multipath) {
if (i == store_point) {
tmpmax2 = LOG_ZERO;
}
}
tmpmax = LOG_ZERO;
for (ac=whmm->state[i].ac;ac;ac=ac->next) {
if (hmminfo->multipath) {
if (ac->arc == wordhmmnum-1) score1 = dwrk->g[t];
else if (t + 1 > startt) score1 = LOG_ZERO;
else score1 = dwrk->wordtrellis[tl][ac->arc];
score1 += ac->a;
} else {
score1 = dwrk->wordtrellis[tl][ac->arc] + ac->a;
}
if (i <= crossword_point && ac->arc > crossword_point) {
/* これは単語を越える遷移 (backscan 実行時) */
/* this is a transition across word (when backscan is enabled) */
score1 += now->lscore; /* add LM score */
}
if (hmminfo->multipath) {
if (i <= store_point && ac->arc > store_point) {
if (tmpmax_store < score1) tmpmax_store = score1;
}
} else {
if (i == store_point && i != ac->arc) {
if (tmpmax2 < score1) tmpmax2 = score1;
}
}
if (tmpmax < score1) {
tmpmax = score1;
j = ac->arc;
}
}
/* スコアエンベロープチェック: 一定幅外なら落とす */
/* check if score of this node is within the score envelope */
if (
#ifdef SCAN_BEAM
tmpmax <= framemaxscore[t] - scan_beam_thres ||
#endif
tmpmax <= LOG_ZERO
) { /* invalid node */
dwrk->wordtrellis[tn][i] = LOG_ZERO;
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
dwrk->wend_token_frame[tn][i] = -1;
dwrk->wend_token_gscore[tn][i] = LOG_ZERO;
}
#endif
if (! hmminfo->multipath) {
if (i == store_point) now->g_prev[t] = LOG_ZERO;
}
} else { /* survived node */
if (! hmminfo->multipath) {
if (i == store_point) now->g_prev[t] = tmpmax2;
}
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
if (hmminfo->multipath) {
if ((back_rescan && i <= crossword_point && j > crossword_point)
|| j == wordhmmnum-1) {
dwrk->wend_token_frame[tn][i] = t;
dwrk->wend_token_gscore[tn][i] = tmpmax;
} else {
dwrk->wend_token_frame[tn][i] = dwrk->wend_token_frame[tl][j];
dwrk->wend_token_gscore[tn][i] = dwrk->wend_token_gscore[tl][j];
}
} else {
if (i <= crossword_point && j > crossword_point) {
dwrk->wend_token_frame[tn][i] = t;
dwrk->wend_token_gscore[tn][i] = tmpmax;
} else {
dwrk->wend_token_frame[tn][i] = dwrk->wend_token_frame[tl][j];
dwrk->wend_token_gscore[tn][i] = dwrk->wend_token_gscore[tl][j];
}
}
}
#endif
node_exist_p = TRUE; /* at least one node survive in this frame */
dwrk->wordtrellis[tn][i] = tmpmax;
if (! hmminfo->multipath || i > 0) {
/* compute output probability */
dwrk->wordtrellis[tn][i] += outprob(&(r->am->hmmwrk), t, &(whmm->state[i]), param);
}
}
} else { /* not triphone */
/* backscan 無し: store_point, crossword_point は無関係 */
/* no backscan: store_point, crossword_point ignored */
tmpmax = LOG_ZERO;
if (hmminfo->multipath) {
for (ac=whmm->state[i].ac;ac;ac=ac->next) {
if (ac->arc == wordhmmnum-1) score1 = dwrk->g[t];
else if (t + 1 > startt) score1 = LOG_ZERO;
else score1 = dwrk->wordtrellis[tl][ac->arc];
score1 += ac->a;
if (tmpmax < score1) {
tmpmax = score1;
j = ac->arc;
}
}
} else {
for (ac=whmm->state[i].ac;ac;ac=ac->next) {
score1 = dwrk->wordtrellis[tl][ac->arc] + ac->a;
if (tmpmax < score1) {
tmpmax = score1;
j = ac->arc;
}
}
}
/* スコアエンベロープチェック: 一定幅外なら落とす */
/* check if score of this node is within the score envelope */
if (
#ifdef SCAN_BEAM
tmpmax <= framemaxscore[t] - scan_beam_thres ||
#endif
tmpmax <= LOG_ZERO
) {
/* invalid node */
dwrk->wordtrellis[tn][i] = LOG_ZERO;
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
dwrk->wend_token_frame[tn][i] = -1;
dwrk->wend_token_gscore[tn][i] = LOG_ZERO;
}
#endif
} else {
/* survived node */
node_exist_p = TRUE;
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
if (hmminfo->multipath) {
if (j == wordhmmnum-1) {
dwrk->wend_token_frame[tn][i] = t;
dwrk->wend_token_gscore[tn][i] = tmpmax;
} else {
dwrk->wend_token_frame[tn][i] = dwrk->wend_token_frame[tl][j];
dwrk->wend_token_gscore[tn][i] = dwrk->wend_token_gscore[tl][j];
}
} else {
dwrk->wend_token_frame[tn][i] = dwrk->wend_token_frame[tl][j];
dwrk->wend_token_gscore[tn][i] = dwrk->wend_token_gscore[tl][j];
}
}
#endif
/* score of node [t][i] has been determined here */
dwrk->wordtrellis[tn][i] = tmpmax;
if (! hmminfo->multipath || i > 0) {
dwrk->wordtrellis[tn][i] += outprob(&(r->am->hmmwrk), t, &(whmm->state[i]), param);
}
}
}
} /* end of node loop */
/* 時間 t のViterbi計算終了. 前向きスコアはscanした単語の始端 */
/* Viterbi end for frame [t]. the forward score is the score of word
beginning scanned */
now->g[t] = dwrk->wordtrellis[tn][0];
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
now->wordend_frame[t] = dwrk->wend_token_frame[tn][0];
now->wordend_gscore[t] = dwrk->wend_token_gscore[tn][0];
}
#endif
if (hmminfo->multipath) {
/* triphone 時, 次段のために store_point のデータをg_prevに保存 */
/* store the scores crossing the store_point to g_prev, for next scan */
if (ccd_flag) {
/* the max arc crossing the store_point always selected as tmpmax_score */
tmpmax_store -= store_point_maxarc;
if (tmpmax_store < LOG_ZERO) tmpmax_store = LOG_ZERO;
now->g_prev[t] = tmpmax_store;
}
}
/* store the number of last computed frame */
if (node_exist_p) endt = t;
/* scanした単語の第１パスでの始端時刻より先まで t が進んでおり，かつ
この t においてスコアエンベロープによって生き残ったノードが一つも
無かったならば,このフレームで計算を打ち切りそれ以上先([0..t-1])は
計算しない */
/* if frame 't' already reached the beginning frame of scanned word
in 1st pass and no node was survived in this frame (all nodes pruned
by score envelope), terminate computation at this frame and
do not computer further frame ([0..t-1]). */
if (t < now->estimated_next_t && (!node_exist_p)) {
/* clear the rest scores */
for (i=t-1;i>=0;i--) {
now->g[i] = LOG_ZERO;
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
now->wordend_frame[i] = -1;
now->wordend_gscore[i] = LOG_ZERO;
}
#endif
if (ccd_flag) now->g_prev[i] = LOG_ZERO;
}
/* terminate loop */
break;
}
} /* end of time loop */
if (debug2_flag) jlog("DEBUG: scanned: [%3d-%3d]\n", endt, startt);
end_of_scan:
if (hmminfo->multipath) {
/* 前向きスコアの最終値を計算 (状態 0 から時間 0 への遷移) */
/* compute the total forward score (transition from state 0 to frame 0 */
if (endt == 0) {
tmpmax = LOG_ZERO;
for(ac=whmm->state[0].ac;ac;ac=ac->next) {
score1 = dwrk->wordtrellis[tn][ac->arc] + ac->a;
if (tmpmax < score1) tmpmax = score1;
}
now->final_g = score1;
} else {
now->final_g = LOG_ZERO;
}
}
/* 次回 backscan のための情報格納 */
/* store data for next backscan */
if (ccd_flag) {
if (store_point == (hmminfo->multipath ? wordhmmnum - 2 : wordhmmnum - 1)) {
/* last_ph無し，かつ単語の音素長=1の場合、次回の scan_word() で
単語全体がもう一度再計算される. この場合,
g_prev は，このscan_wordを開始する前のスコアを入れておく必要がある */
/* if there was no 'last_ph' and the scanned word consists of only
1 phone, the whole word should be re-computed in the future scan_word().
So the next 'g_prev[]' should be the initial forward scores
before we begin Viterbi (= g[t]). */
for (t = startt; t>=0; t--) {
now->g_prev[t] = dwrk->g[t];
}
}
#ifndef GRAPHOUT_PRECISE_BOUNDARY
if (r->graphout) {
if (now->tail_g_score != LOG_ZERO) {
if (now->prevgraph != NULL) {
(now->prevgraph)->leftscore = now->tail_g_score;
}
}
}
#endif
/* 次回のために now->last_ph を更新 */
/* update 'now->last_ph' for future scan_word() */
if (back_rescan) {
now->last_ph = dwrk->phmmseq[0];
} else {
now->last_ph = winfo->wseq[word][0];
}
if (enable_iwsp && hmminfo->multipath) {
now->last_ph_sp_attached = dwrk->has_sp[0];
}
}
#ifdef GRAPHOUT_PRECISE_BOUNDARY
if (! hmminfo->multipath) {
if (r->graphout) {
/* 次回の next_word 用に境界情報を調整 */
/* proceed word boundary for one step for next_word */
now->wordend_frame[peseqlen-1] = now->wordend_frame[0];
now->wordend_gscore[peseqlen-1] = now->wordend_gscore[0];
for (t=0;t<peseqlen-1;t++) {
now->wordend_frame[t] = now->wordend_frame[t+1];
now->wordend_gscore[t] = now->wordend_gscore[t+1];
}
}
}
#endif
/* free work area */
free_hmm(whmm);
#ifdef TCD
if (hmminfo->multipath) {
if (ccd_flag) {
jlog("DEBUG: last_ph = %s", (now->last_ph)->name);
if (now->last_ph_sp_attached) jlog(" (sp attached)");
jlog("\n");
}
} else {
jlog("DEBUG: last_ph = %s\n", (now->last_ph)->name);
}
#endif
}
/**
* <JA>
* 音素の出力状態への遷移確率の最大値を求める. (multipath)
*
* @param l [in] 音素
*
* @return 出力状態への遷移確率の最大値を返す.
* </JA>
* <EN>
* Get the maximum transition log probability outside a phone. (multipath)
*
* @param l [in] phone
*
* @return the maximum transition log probability outside a phone.
* </EN>
*/
static LOGPROB
max_out_arc(HMM_Logical *l)
{
return(get_max_out_arc(hmm_logical_trans(l), hmm_logical_state_num(l)));
}
