/** * @file wchmm.c * * * @brief 木構造化辞書の構築 * * ここでは，与えられた単語辞書, HMM定義および言語制約から木構造化辞書を * 構築する関数が定義されています. 木構造化辞書は起動時に構築され， * 第1パスの認識に用いられます. 木構造化辞書は状態単位で構成され， * 各状態はHMM出力確率と遷移先の他，および探索のための様々な情報を含みます. * * 開発の経緯上，ソース内では木構造化辞書は wchmm (word-conjunction HMM) と * も表現されています. * * * * * @brief Construction of tree lexicon. * * Functions to build a tree lexicon (or called word-conjunction HMM here) * from word dictionary, HMM and language models are defined here. The * constructed tree lexicon will be used for the recognition of the 1st pass. * The lexicon is composed per HMM state unit, and various informations * about output probabilities, arcs, language model constraints, and others * are assembled in the lexicon. * * Note that the word "wchmm" in the source code is a synonim of * "tree lexicon". * * * @author Akinobu Lee * @date Mon Sep 19 23:39:15 2005 * * $Revision: 1.10 $ * */ /* * Copyright (c) 1991-2012 Kawahara Lab., Kyoto University * Copyright (c) 2000-2005 Shikano Lab., Nara Institute of Science and Technology * Copyright (c) 2005-2012 Julius project team, Nagoya Institute of Technology * All rights reserved */ /* wchmm = word conjunction HMM = lexicon tree */ #include #define WCHMM_SIZE_CHECK ///< If defined, do wchmm size estimation (for debug only) /**************************************************************/ /*********** Initialization of tree lexicon *******************/ /**************************************************************/ /** * * 木構造化辞書構造体を新規に割り付ける. * * @return 新たにメモリ上に割り付けられた木構造化辞書構造体へのポインタを返す. * * * Allocate a new tree lexicon structure. * * @return pointer to the newly allocated tree lexicon structure. * * @callgraph * @callergraph */ WCHMM_INFO * wchmm_new() { WCHMM_INFO *w; w = (WCHMM_INFO *)mymalloc(sizeof(WCHMM_INFO)); w->lmtype = LM_UNDEF; w->lmvar = LM_UNDEF; w->ngram = NULL; w->dfa = NULL; w->winfo = NULL; w->malloc_root = NULL; #ifdef PASS1_IWCD w->lcdset_category_root = NULL; w->lcdset_mroot = NULL; #endif /* PASS1_IWCD */ w->wrk.out_from_len = 0; /* reset user function entry point */ w->uni_prob_user = NULL; w->bi_prob_user = NULL; return w; } /** * * 木構造化辞書の内容を初期化する. * * @param wchmm [out] 木構造化辞書へのポインタ * * * Initialize content of a lexicon tree. * * @param wchmm [out] pointer to the lexicon tree structure * */ static void wchmm_init(WCHMM_INFO *wchmm) { /* the resulting tree size is typically half of total state num */ wchmm->maxwcn = wchmm->winfo->totalstatenum / 2; wchmm->state = (WCHMM_STATE *)mymalloc(sizeof(WCHMM_STATE)*wchmm->maxwcn); wchmm->self_a = (LOGPROB *)mymalloc(sizeof(LOGPROB)*wchmm->maxwcn); wchmm->next_a = (LOGPROB *)mymalloc(sizeof(LOGPROB)*wchmm->maxwcn); wchmm->ac = (A_CELL2 **)mymalloc(sizeof(A_CELL2 *)*wchmm->maxwcn); wchmm->stend = (WORD_ID *)mymalloc(sizeof(WORD_ID)*wchmm->maxwcn); wchmm->offset = (int **)mymalloc(sizeof(int *)*wchmm->winfo->num); wchmm->wordend = (int *)mymalloc(sizeof(int)*wchmm->winfo->num); wchmm->maxstartnum = STARTNODE_STEP; wchmm->startnode = (int *)mymalloc(sizeof(int)*STARTNODE_STEP); wchmm->startnum = 0; if (wchmm->category_tree) { wchmm->start2wid = (WORD_ID *)mymalloc(sizeof(WORD_ID)*STARTNODE_STEP); } if (wchmm->hmminfo->multipath) { wchmm->wordbegin = (int *)mymalloc(sizeof(int)*wchmm->winfo->num); wchmm->wrk.out_from = (int *)mymalloc(sizeof(int) * wchmm->winfo->maxwn); wchmm->wrk.out_from_next = (int *)mymalloc(sizeof(int) * wchmm->winfo->maxwn); wchmm->wrk.out_a = (LOGPROB *)mymalloc(sizeof(LOGPROB) * wchmm->winfo->maxwn); wchmm->wrk.out_a_next = (LOGPROB *)mymalloc(sizeof(LOGPROB) * wchmm->winfo->maxwn); wchmm->wrk.out_from_len = wchmm->winfo->maxwn; } else { wchmm->wordend_a = (LOGPROB *)mymalloc(sizeof(LOGPROB)*wchmm->winfo->num); } #ifdef PASS1_IWCD wchmm->outstyle = (unsigned char *)mymalloc(sizeof(unsigned char)*wchmm->maxwcn); #endif #ifdef UNIGRAM_FACTORING wchmm->start2isolate = NULL; wchmm->isolatenum = 0; #endif if (!wchmm->category_tree) { #ifdef UNIGRAM_FACTORING wchmm->scword = NULL; wchmm->fscore = NULL; #endif wchmm->sclist = NULL; wchmm->sclen = NULL; } wchmm->n = 0; } /** * * 木構造化辞書の状態格納領域を MAXWCNSTEP 分だけ伸長する. * * @param wchmm [i/o] 木構造化辞書 * * * Expand state-related area in a tree lexicon by MAXWCNSTEP. * * @param wchmm [i/o] tree lexicon * */ static void wchmm_extend(WCHMM_INFO *wchmm) { /* practical value! */ wchmm->maxwcn += wchmm->winfo->totalstatenum / 6; wchmm->state = (WCHMM_STATE *)myrealloc(wchmm->state, sizeof(WCHMM_STATE)*wchmm->maxwcn); wchmm->self_a = (LOGPROB *)myrealloc(wchmm->self_a, sizeof(LOGPROB)*wchmm->maxwcn); wchmm->next_a = (LOGPROB *)myrealloc(wchmm->next_a, sizeof(LOGPROB)*wchmm->maxwcn); wchmm->ac = (A_CELL2 **)myrealloc(wchmm->ac, sizeof(A_CELL2 *)*wchmm->maxwcn); wchmm->stend = (WORD_ID *)myrealloc(wchmm->stend, sizeof(WORD_ID)*wchmm->maxwcn); #ifdef PASS1_IWCD wchmm->outstyle = (unsigned char *)myrealloc(wchmm->outstyle, sizeof(unsigned char)*wchmm->maxwcn); #endif } /** * * 木構造化辞書の単語先頭ノード格納領域を STARTNODE_STEP分だけ伸長する. (multipath) * * @param wchmm [i/o] 木構造化辞書 * * * Expand word-start nodes area in a tree lexicon by STARTNODE_STEP. (multipath) * * @param wchmm [i/o] tree lexicon * */ static void wchmm_extend_startnode(WCHMM_INFO *wchmm) { wchmm->maxstartnum += STARTNODE_STEP; wchmm->startnode = (int *)myrealloc(wchmm->startnode, sizeof(int) * wchmm->maxstartnum); if (wchmm->category_tree) { wchmm->start2wid = (WORD_ID *)myrealloc(wchmm->start2wid, sizeof(WORD_ID) * wchmm->maxstartnum); } } /** * * 木構造化辞書およびその内部の割付メモリを全て解放する. * * @param w [in] 木構造化辞書 * * * Free all data in a tree lexicon. * * @param w [in] tree lexicon * * @callgraph * @callergraph */ void wchmm_free(WCHMM_INFO *w) { int i; /* wchmm->state[i].ac malloced by mybmalloc2() */ /* wchmm->offset[][] malloced by mybmalloc2() */ #ifdef PASS1_IWCD /* LRC_INFO, RC_INFO in wchmm->state[i].outsty malloced by mybmalloc2() */ #endif /* wchmm->sclist[][] and wchmm->sclen[] malloced by mybmalloc2() */ /* they all will be freed by a single mybfree2() call */ mybfree2(&(w->malloc_root)); if (!w->category_tree) { #ifdef UNIGRAM_FACTORING if (w->fscore != NULL) free(w->fscore); #endif } #ifdef UNIGRAM_FACTORING if (w->start2isolate != NULL) free(w->start2isolate); #endif #ifdef PASS1_IWCD free(w->outstyle); #endif if (w->hmminfo->multipath) { free(w->wordbegin); } else { free(w->wordend_a); } if (w->category_tree) free(w->start2wid); free(w->startnode); free(w->wordend); free(w->offset); free(w->stend); free(w->ac); free(w->next_a); free(w->self_a); free(w->state); #ifdef PASS1_IWCD if (w->category_tree) lcdset_remove_with_category_all(w); #endif /* PASS1_IWCD */ if (w->wrk.out_from_len != 0) { free(w->wrk.out_from); free(w->wrk.out_from_next); free(w->wrk.out_a); free(w->wrk.out_a_next); w->wrk.out_from_len = 0; } free(w); } /**************************************************************/ /*********** Word sort functions for tree construction ********/ /**************************************************************/ /** * * 単語を音素のならびでソートするqsort_reentrant関数 * * @param widx1 [in] 単語ID 1 へのポインタ * @param widx2 [in] 単語ID 2 へのポインタ * * @return 単語widx2が単語widx1の一部か昇順であれば 1, 単語widx1が単語widx2の一部か昇順であれば -1, 全く同じ音素並びであれば 0 を返す. * * * qsort_reentrant function to sort words by their phoneme sequence. * * @param widx1 [in] pointer to word id #1 * @param widx2 [in] pointer to wrod id #2 * * @return 1 if word[widx2] is part of word[widx1], -1 if word[widx1] is part of word[widx2], or 0 if the two words are equal. * */ static int compare_wseq(WORD_ID *widx1, WORD_ID *widx2, WORD_INFO *winfo) { int len1, len2, n; int p=0; len1 = winfo->wlen[*widx1]; len2 = winfo->wlen[*widx2]; n=0; /* while (n < len1 && n < len2 && (p = (int)winfo->wseq[*widx1][n] - (int)winfo->wseq[*widx2][n]) == 0 ) n++;*/ while (n < len1 && n < len2 && (p = strcmp((winfo->wseq[*widx1][n])->name, (winfo->wseq[*widx2][n])->name)) == 0 ) n++; if (n < len1) { if (n < len2) { /* differ */ return(p); } else { /* 2 is part of 1 */ return(1); } } else { if (n < len2) { /* 1 is part of 2 */ return(-1); } else { /* same */ return(0); } } } /** * * 単語IDの集合 windex[bgn..bgn+len-1] を単語の音素ならびでソートする. * * @param winfo [in] 単語辞書 * @param windex [i/o] 単語IDのインデックス列（内部でソートされる） * @param bgn [in] @a windex のソート開始点 * @param len [in] @a windex の @a bgn からのソートする要素数 * * * Sort word IDs in windex[bgn..bgn+len-1] by their phoneme sequence order. * * @param winfo [in] word lexicon * @param windex [i/o] index sequence of word IDs, (will be sorted in this function) * @param bgn [in] start point to sort in @a windex * @param len [in] length of indexes to be sorted from @a bgn * */ static void wchmm_sort_idx_by_wseq(WORD_INFO *winfo, WORD_ID *windex, WORD_ID bgn, WORD_ID len) { qsort_reentrant(&(windex[bgn]), len, sizeof(WORD_ID), (int (*)(const void *, const void *, void *))compare_wseq, winfo); } /** * * 単語をカテゴリIDでソートするqsort関数. * * @param widx1 [in] 要素1へのポインタ * @param widx2 [in] 要素2へのポインタ * * @return * * * qsort function to sort words by their category ID. * * @param widx1 [in] pointer to element #1 * @param widx2 [in] pointer to element #2 * * @return * */ static int compare_category(WORD_ID *widx1, WORD_ID *widx2, WORD_INFO *winfo) { int c1,c2; c1 = winfo->wton[*widx1]; c2 = winfo->wton[*widx2]; return(c1 - c2); } /** * * 単語ID集合 windex[0..len-1] をカテゴリIDでソートする. * * @param winfo [in] 単語辞書 * @param windex [i/o] 単語IDのインデックス列（内部でソートされる） * @param len [in] @a windex の要素数 * * * Sort word IDs in windex[0..len-1] by their category ID. * * @param winfo [in] tree lexicon * @param windex [i/o] index sequence of word IDs, (will be sorted in this function) * @param len [in] number of elements in @a windex * */ static void wchmm_sort_idx_by_category(WORD_INFO *winfo, WORD_ID *windex, WORD_ID len) { qsort_reentrant(windex, len, sizeof(WORD_ID), (int (*)(const void *, const void *, void *))compare_category, winfo); } /**********************************************************************/ /************** Subroutines to link part of words ********************/ /**********************************************************************/ /** * * ２単語間で，単語の先頭から同一で共有可能な音素の数を調べる. * * @param winfo [in] 単語辞書 * @param i [in] 単語１ * @param j [in] 単語２ * * @return 共有可能な先頭からの音素数を返す. * * * Compare two words from word head per phoneme to see how many phones * can be shared among the two. * * @param winfo [in] word dictionary * @param i [in] a word * @param j [in] another word * * @return the number of phonemes to be shared from the head of the words. * */ static int wchmm_check_match(WORD_INFO *winfo, int i, int j) { int k,tmplen; for (tmplen=0,k=0;kwlen[i];k++) { if (k > winfo->wlen[j]-1) break; if (! (strmatch(winfo->wseq[i][k]->name, winfo->wseq[j][k]->name))) break; tmplen++; } return(tmplen); } /** * * Initialize transition information on a node. * * * ノードの遷移情報を初期化する. * * * @param wchmm [i/o] tree lexicon * @param node [in] node id * */ static void acc_init(WCHMM_INFO *wchmm, int node) { wchmm->self_a[node] = LOG_ZERO; wchmm->next_a[node] = LOG_ZERO; wchmm->ac[node] = NULL; } /** * * Add an arc to a node. * This function is for transition other than self and next node. * * * ノードに遷移を追加する. * この関数は自己遷移・隣への遷移以外の場合に使用される. * * * @param wchmm [i/o] tree lexicon * @param node [in] node id * @param a [in] transition probability in log10 * @param arc [in] transition destination node id * */ static void add_ac(WCHMM_INFO *wchmm, int node, LOGPROB a, int arc) { A_CELL2 *ac2; for(ac2=wchmm->ac[node];ac2;ac2=ac2->next) { if (ac2->n < A_CELL2_ALLOC_STEP) break; } if (ac2 == NULL) { ac2 = (A_CELL2 *)mybmalloc2(sizeof(A_CELL2), &(wchmm->malloc_root)); ac2->n = 0; ac2->next = wchmm->ac[node]; wchmm->ac[node] = ac2; } ac2->arc[ac2->n] = arc; ac2->a[ac2->n] = a; ac2->n++; } /** * * 木構造化辞書のあるノードに，別のノードへの遷移を追加する * * @param wchmm [i/o] 木構造化辞書 * @param node [in] ノード番号 * @param a [in] 遷移確率（対数） * @param arc [in] 遷移先のノード番号 * * * Add a transition arc between two nodes on the tree lexicon * * @param wchmm [i/o] tree lexicon * @param node [in] node number of source node * @param a [in] transition probability in log scale * @param arc [in] node number of destination node * */ static void add_wacc(WCHMM_INFO *wchmm, int node, LOGPROB a, int arc) { if (arc == node) { wchmm->self_a[node] = a; } else if (arc == node + 1) { wchmm->next_a[node] = a; } else { add_ac(wchmm, node, a, arc); } } /** * * ある単語のある位置の音素から単語末端の外へ出る遷移のリストを得る. (multipath) * * @param wchmm [in] 木構造化辞書 * @param w [in] 単語ID * @param pos [in] 音素位置 * @param node [out] 音素内の，単語末端外への遷移を持つ状態のリスト * @param a [out] @a node の各要素の遷移確率 * @param num [out] @a node の要素数. 発見数だけ増加される. * @param maxnum [in] @a node の格納可能な最大数 * @param insert_sp [in] 単語終端での sp 挟み込みを考慮するならTRUE * * * Make outgoing transition list for given phone position of a word. (multipath) * * @param wchmm [in] tree lexicon * @param w [in] word ID * @param pos [in] location of target phone to be inspected in the word @a w * @param node [out] list of wchmm states that possibly has outgoing transition * @param a [out] transition probabilities of the outgoing transitions in @a node * @param num [out] number of elements in @a out (found num will be added) * @param maxnum [in] maximum number of elements that can be stored in @a node * @param insert_sp [in] TRUE if consider short-pause insertion on word end * */ static void get_outtrans_list(WCHMM_INFO *wchmm, WORD_ID w, int pos, int *node, LOGPROB *a, int *num, int maxnum, boolean insert_sp) { HMM_Logical *ltmp; int states; int k; LOGPROB prob; int oldnum; if (pos < 0) { /* set the word-beginning node, and return */ node[*num] = wchmm->wordbegin[w]; a[*num] = 0.0; (*num)++; } else { ltmp = wchmm->winfo->wseq[w][pos]; states = hmm_logical_state_num(ltmp); /* check initial->final state */ if ((hmm_logical_trans(ltmp))->a[0][states-1] != LOG_ZERO) { /* recursive call for previous phone */ oldnum = *num; get_outtrans_list(wchmm, w, pos-1, node, a, num, maxnum, FALSE); /* previous phone should not be an sp-inserted phone */ /* add probability of the skip transition to all the previous ones */ for(k=oldnum;k<*num;k++) { a[k] += (hmm_logical_trans(ltmp))->a[0][states-1]; } } /* add to list the arcs from output state to final state */ for (k = 1; k < states - 1; k++) { prob = (hmm_logical_trans(ltmp))->a[k][states-1]; if (prob != LOG_ZERO) { if (*num >= maxnum) { j_internal_error("get_outtrans_list: maximum outtrans list num exceeded %d\n", maxnum); } node[*num] = wchmm->offset[w][pos] + k - 1; a[*num] = prob; (*num)++; } } /* for -iwsp, add outgoing arc from the tail sp model only if need_sp == TRUE. need_sp should be TRUE only when the connecting [pos] phone is also an end phone of the to-be-added word (i.e. homophone word) */ /* */ if (insert_sp) { /* consider sp */ for (k = 1; k < hmm_logical_state_num(wchmm->hmminfo->sp) - 1; k++) { prob = hmm_logical_trans(wchmm->hmminfo->sp)->a[k][hmm_logical_state_num(wchmm->hmminfo->sp)-1]; if (prob != LOG_ZERO) { if (*num >= maxnum) { j_internal_error("get_outtrans_list: maximum outtrans list num exceeded %d\n", maxnum); } node[*num] = wchmm->offset[w][pos] + (states - 2) + k - 1; a[*num] = prob; (*num)++; } } } } /*printf(" %d(%s)-%d:\"%s\", num=%d\n", w, wchmm->winfo->woutput[w], pos, (pos < 0) ? "BGN" : wchmm->winfo->wseq[w][pos]->name, *num);*/ return; } /** * * ある音素の末尾の状態から，ある音素の先頭状態への遷移を追加する. * * @param wchmm [i/o] 木構造化辞書 * @param from_node [in] ある音素の末尾の状態 * @param to_node [in] ある音素の先頭状態 * @param tinfo [in] @a from_node の属する音素HMMの遷移確率行列 * * * Add a transition from end node of a phone to start node of another phone. * * @param wchmm [i/o] tree lexicon * @param from_node [in] end node of a phone * @param to_node [in] start node of a phone * @param tinfo [in] transition prob. matrix of the @a from_node phone. * */ static void wchmm_link_hmm(WCHMM_INFO *wchmm, int from_node, int to_node, HTK_HMM_Trans *tinfo) { A_CELL2 *actmp; LOGPROB a; int i, j; boolean tflag; /* get transition probability to outer state in tinfo */ for(i = tinfo->statenum - 2; i >= 0; i--) { if ((a = tinfo->a[i][tinfo->statenum-1]) != LOG_ZERO) { /* found */ /* check if the arc already exist */ tflag = FALSE; if (to_node == from_node && wchmm->self_a[from_node] == a) { tflag = TRUE; } else if (to_node == from_node + 1 && wchmm->next_a[from_node] == a) { tflag = TRUE; } else { for (actmp = wchmm->ac[from_node]; actmp; actmp = actmp->next) { for(j=0;jn;j++) { if (actmp->arc[j] == to_node && actmp->a[j] == a) { tflag = TRUE; break; } } if (tflag == TRUE) break; } } if (tflag) break; /* add the arc to wchmm */ add_wacc(wchmm, from_node, a, to_node); return; /* exit function here */ } } j_internal_error("wchmm_link_hmm: No arc to endstate?\n"); } /** * * 木構造化辞書中の２単語中のある音素間を接続する. * * @param wchmm [i/o] 木構造化辞書 * @param from_word [in] 遷移元の単語のID * @param from_seq [in] 遷移元の単語中の接続する音素の位置 * @param to_word [in] 遷移先の単語のID * @param to_seq [in] 遷移先の単語中の接続する音素の位置 * * * Connect two phonemes in tree lexicon. * * @param wchmm [i/o] tree lexicon * @param from_word [in] source word ID * @param from_seq [in] index of source phoneme in @a from_word from which the other will be connected * @param to_word [in] destination word ID * @param to_seq [in] index of destination phoneme in @a to_word to which the other will connect * */ static void wchmm_link_subword(WCHMM_INFO *wchmm, int from_word, int from_seq, int to_word, int to_seq) { HMM_Logical *last; int lastp; last = wchmm->winfo->wseq[from_word][from_seq]; lastp = wchmm->offset[from_word][from_seq] + hmm_logical_state_num(last)-2 -1; wchmm_link_hmm(wchmm, lastp, wchmm->offset[to_word][to_seq], hmm_logical_trans(last)); } /**************************************************************/ /******** homophone processing: duplicating leaf nodes ********/ /**************************************************************/ /** * @note * * 同音語処理: * 木構造化辞書においてすべての単語は独立した最終状態を持つ必要があるため， * 同音語は注意深く扱う必要がある. このため，最初の木構造化辞書を構築した後, * 別の単語と完全に共有された単語(同音語), あるいは別の単語の一部として * 埋め込まれてしまっている単語を発見するとともに, その最終ノードを * コピーして新たな単語終端ノードを作る必要がある. * * * Homophones: * As all words need to have an uniq state as a final state in a lexicon tree, * homophones should be handled carefully. After primal tree has been made, * we look through the tree to find the fully shared or embedded words * (homophone or part of other word), and duplicate the last leaf node * to have uniq end state. * */ /** * * 単語終端状態の独立化：与えられた単語の終端ノードをコピーして， * 新たにある単語の最終状態として定義する. * * @param wchmm [i/o] 木構造化辞書 * @param node [in] 同音語の終端ノード番号 * @param word [in] 新たに登録する単語 * * * Isolation of word-end nodes for homophones: duplicate the word-end state, * link as the same as original, and make it the new word-end node of the * given new word. * * @param wchmm [i/o] tree lexicon * @param node [in] the word end node of the already existing homophone * @param word [in] word ID to be added to the tree * */ static void wchmm_duplicate_state(WCHMM_INFO *wchmm, int node, int word) /* source node, new word */ { int j, n; int n_src, n_prev; A_CELL2 *ac; HMM_Logical *lastphone; /* 1 state will newly created: expand tree if needed */ if (wchmm->n + 1 >= wchmm->maxwcn) { wchmm_extend(wchmm); } /* n: the target new node to which 'node' is copied */ n = wchmm->n; n_src = node; /* copy output probability info */ #ifdef PASS1_IWCD { RC_INFO *rcnew; LRC_INFO *lrcnew; wchmm->outstyle[n] = wchmm->outstyle[n_src]; if (wchmm->outstyle[n] == AS_RSET) { /* duplicate RC_INFO because it has its own cache */ rcnew = (RC_INFO *)mybmalloc2(sizeof(RC_INFO), &(wchmm->malloc_root)); memcpy(rcnew, wchmm->state[n_src].out.rset, sizeof(RC_INFO)); wchmm->state[n].out.rset = rcnew; } else if (wchmm->outstyle[n] == AS_LRSET) { /* duplicate LRC_INFO because it has its own cache */ lrcnew = (LRC_INFO *)mybmalloc2(sizeof(LRC_INFO), &(wchmm->malloc_root)); memcpy(lrcnew, wchmm->state[n_src].out.lrset, sizeof(LRC_INFO)); wchmm->state[n].out.lrset = lrcnew; } else { /* share same info, simply copy the pointer */ memcpy(&(wchmm->state[n].out), &(wchmm->state[n_src].out), sizeof(ACOUSTIC_SPEC)); } } #else /* ~PASS1_IWCD */ memcpy(&(wchmm->state[n].out), &(wchmm->state[n_src].out), sizeof(HTK_HMM_State *)); #endif lastphone = wchmm->winfo->wseq[word][wchmm->winfo->wlen[word]-1]; acc_init(wchmm, n); /* add self transition arc */ wchmm->self_a[n] = wchmm->self_a[n_src]; /* copy transition arcs whose destination is the source node to new node */ if (hmm_logical_state_num(lastphone) == 3) { /* = 1 state */ /* phone with only 1 state should be treated carefully */ if (wchmm->winfo->wlen[word] == 1) { /* word consists of only this phone */ /* no arcs need to be copied: this is also a start node of a word */ wchmm->offset[word][0] = n; /* index the new word-beginning node as startnode (old ststart) */ if (wchmm->lmtype != LM_PROB || word != wchmm->winfo->head_silwid) { wchmm->startnode[wchmm->startnum] = n; if (wchmm->category_tree) wchmm->start2wid[wchmm->startnum] = word; /* expand data area if necessary */ if (++wchmm->startnum >= wchmm->maxstartnum) wchmm_extend_startnode(wchmm); } } else { /* copy arcs from the last state of the previous phone */ n_prev = wchmm->offset[word][wchmm->winfo->wlen[word]-2] + hmm_logical_state_num(wchmm->winfo->wseq[word][wchmm->winfo->wlen[word]-2]) - 3; if(n_src == n_prev + 1) { add_wacc(wchmm, n_prev, wchmm->next_a[n_prev], n); } else { for(ac=wchmm->ac[n_prev];ac;ac=ac->next) { for(j=0;jn;j++) { if (ac->arc[j] == n_src) { add_wacc(wchmm, n_prev, ac->a[j], n); } } } } /* also update the last offset (== wordend in this case) */ wchmm->offset[word][wchmm->winfo->wlen[word]-1] = n; } } else { /* phone with more than 2 states */ /* copy arcs from/to the source node to new node */ for (n_prev = wchmm->offset[word][wchmm->winfo->wlen[word]-1]; n_prev < n_src; n_prev++) { if (n_src == n_prev + 1) { add_wacc(wchmm, n_prev, wchmm->next_a[n_prev], n); } else { for(ac=wchmm->ac[n_prev];ac;ac=ac->next) { for(j=0;jn;j++) { if (ac->arc[j] == n_src) { add_wacc(wchmm, n_prev, ac->a[j], n); } } } } if (n_prev == n_src + 1) { add_wacc(wchmm, n, wchmm->next_a[n_src], n_prev); } else { for(ac=wchmm->ac[n_src];ac;ac=ac->next) { for(j=0;jn;j++) { if (ac->arc[j] == n_prev) { add_wacc(wchmm, n, ac->a[j], n_prev); } } } } } } /* map word <-> node */ wchmm->stend[n] = word; /* 'n' is an end node of word 'word' */ wchmm->wordend[word] = n; /* the word end node of 'word' is 'n' */ /* new state has been created: increment the size */ wchmm->n++; } /** * * 木構造化辞書全体を走査して，すべての同音語について単語終端状態の独立化 * を行う. * * @param wchmm [i/o] 木構造化辞書 * * * Scan the whole lexicon tree to find already registered homophones, and * make word-end nodes of the found homophones isolated from others. * * @param wchmm [i/o] tree lexicon * */ static int wchmm_duplicate_leafnode(WCHMM_INFO *wchmm) { int w, nlast, n, narc, narc_model; boolean *dupw; /* node marker */ A_CELL2 *actmp; int dupcount; dupcount = 0; nlast = wchmm->n; dupw = (boolean *)mymalloc(sizeof(boolean) * nlast); for(n=0;nwinfo->num;w++) { n = wchmm->wordend[w]; if (dupw[n]) { /* if already marked (2nd time or later */ wchmm_duplicate_state(wchmm, n, w); dupcount++; /* duplicate */ } else { /* if not marked yet (1st time) */ /* try to find an arc outside the word */ { /* count number of model-internal arc from the last state */ HMM_Logical *lastphone; HTK_HMM_Trans *tinfo; int laststate, i; lastphone = wchmm->winfo->wseq[w][wchmm->winfo->wlen[w]-1]; laststate = hmm_logical_state_num(lastphone) - 2; tinfo = hmm_logical_trans(lastphone); narc_model=0; for(i=1;ia[laststate][i] != LOG_ZERO) narc_model++; } /* count number of actual arc from the last state in the tree */ narc = 0; if (wchmm->self_a[n] != LOG_ZERO) narc++; if (wchmm->next_a[n] != LOG_ZERO) narc++; for(actmp=wchmm->ac[n];actmp;actmp=actmp->next) narc += actmp->n; } /* if both number does not match, it means it is not a single word tail */ if (narc_model != narc) { /* word 'w' is embedded as part of other words at this node 'n' */ /* duplicate this node now */ wchmm_duplicate_state(wchmm, n, w); dupcount++; /* as new node has been assigned as word end node of word 'w', reset this source node as it is not the word end node */ wchmm->stend[n] = WORD_INVALID; } else { /* no arc to other node found, it means it is a single word tail */ /* as this is first time, only make sure that this node is word end of [w] */ wchmm->stend[n] = w; } /* mark node 'n' */ dupw[n] = TRUE; } } free(dupw); return(dupcount); } /**************************************************************/ /*************** add a word to wchmm lexicon tree *************/ /**************************************************************/ /** * * 木構造化辞書に新たに単語を追加する. 追加場所の情報として，現在の木構造化 * 辞書内で最もその単語と先頭から良くマッチする単語，およびそのマッチする長さ * を指定する. * * @param wchmm [i/o] 木構造化辞書 * @param word [in] 追加する辞書単語のID * @param matchlen [in] @a word と @a matchword の先頭からマッチする音素長 * @param matchword [in] 既存の木構造化辞書中で @a word と最もマッチする単語 * @param enable_iwsp [in] 単語間ショートポーズ機能使用時TRUEを指定 * * * Add a new word to the lexicon tree. The longest matched word in the current * lexicon tree and the length of the matched phoneme from the word head should * be specified to tell where to insert the new word to the tree. * * @param wchmm [i/o] tree lexicon * @param word [in] word id to be added to the lexicon * @param matchlen [in] phoneme match length between @a word and @a matchword. * @param matchword [in] the longest matched word with @a word in the current lexicon tree * @param enable_iwsp [in] should be TRUE when using inter-word short pause option * */ static boolean wchmm_add_word(WCHMM_INFO *wchmm, int word, int matchlen, int matchword, boolean enable_iwsp) { boolean ok_p; int j,k,n; int add_head, add_tail, add_to; int word_len, matchword_len; HMM_Logical *ltmp; int ato; LOGPROB prob; int ntmp; int ltmp_state_num; #ifdef PASS1_IWCD CD_Set *lcd = NULL; #endif int *out_from; int *out_from_next; LOGPROB *out_a; LOGPROB *out_a_next; /* for multipath handling */ int out_num_prev, out_num_next; int kkk; ok_p = TRUE; if (wchmm->hmminfo->multipath) { out_from = wchmm->wrk.out_from; out_from_next = wchmm->wrk.out_from_next; out_a = wchmm->wrk.out_a; out_a_next = wchmm->wrk.out_a_next; } /* * if (matchlen > 0) { * printf("--\n"); * put_voca(stdout, wchmm->winfo, word); * put_voca(stdout, wchmm->winfo, matchword); * printf("matchlen=%d\n", matchlen); * } */ /* variable abbreviations */ n = wchmm->n; word_len = wchmm->winfo->wlen[word]; matchword_len = wchmm->winfo->wlen[matchword]; /* malloc phone offset area */ wchmm->offset[word] = (int *)mybmalloc2(sizeof(int)*word_len, &(wchmm->malloc_root)); /* allocate unshared (new) part */ add_head = matchlen; add_tail = word_len - 1; add_to = matchlen - 1; if (wchmm->hmminfo->multipath) { /* make word-beginning node if needed */ if (matchlen == 0) { /* create word-beginning node */ wchmm->wordbegin[word] = n; wchmm->stend[n] = WORD_INVALID; acc_init(wchmm, n); wchmm->state[n].out.state = NULL; /* index the new word-beginning node as startnode (old ststart) */ wchmm->startnode[wchmm->startnum] = n; if (wchmm->category_tree) wchmm->start2wid[wchmm->startnum] = word; /* expand data area if necessary */ if (++wchmm->startnum >= wchmm->maxstartnum) wchmm_extend_startnode(wchmm); if (++n >= wchmm->maxwcn) wchmm_extend(wchmm); } else { wchmm->wordbegin[word] = wchmm->wordbegin[matchword]; } /* now n is at beginning of output state */ /* store the initial outgoing arcs to out_from[] and out_a[] */ out_num_prev = 0; if (matchlen == 0) { /* set the word-beginning node */ out_from[0] = wchmm->wordbegin[word]; out_a[0] = 0.0; out_num_prev = 1; } else { /*printf("%d(%s)\n", word, wchmm->winfo->woutput[word]);*/ /* on -iwsp, trailing sp is needed only when no phone will be created */ get_outtrans_list(wchmm, matchword, add_to, out_from, out_a, &out_num_prev, wchmm->winfo->maxwn, (enable_iwsp && add_tail - add_head + 1 <= 0) ? TRUE : FALSE); /*printf("NUM=%d\n", out_num_prev);*/ } } else { /* end of multipath block */ if (matchlen == 0) { if (wchmm->lmtype != LM_PROB || word != wchmm->winfo->head_silwid) { /* index the new word-beginning node as startnode (old ststart) */ wchmm->startnode[wchmm->startnum] = n; if (wchmm->category_tree) wchmm->start2wid[wchmm->startnum] = word; /* expand data area if necessary */ if (++wchmm->startnum >= wchmm->maxstartnum) wchmm_extend_startnode(wchmm); } } } if (add_tail - add_head + 1 > 0) { /* there are new phones to be created */ ntmp = n; for (j=add_head; j <= add_tail; j++) { /* for each new phones */ ltmp = wchmm->winfo->wseq[word][j]; ltmp_state_num = hmm_logical_state_num(ltmp); #ifdef PASS1_IWCD if (wchmm->ccd_flag) { /* in the triphone lexicon tree, the last phone of a word has left-context cdset */ if (wchmm->winfo->wlen[word] > 1 && j == wchmm->winfo->wlen[word] - 1) { if (wchmm->category_tree) { #ifdef USE_OLD_IWCD lcd = lcdset_lookup_by_hmmname(wchmm->hmminfo, ltmp->name); #else lcd = lcdset_lookup_with_category(wchmm, ltmp, wchmm->winfo->wton[word]); if (lcd == NULL) { /* no category-aware cdset found. This is case when no word can follow this word grammatically. so fallback to normal state */ jlog("WARNING: wchmm: no lcdset found for [%s::%04d], fallback to [%s]\n", ltmp->name, wchmm->winfo->wton[word], ltmp->name); lcd = lcdset_lookup_by_hmmname(wchmm->hmminfo, ltmp->name); } #endif } else { lcd = lcdset_lookup_by_hmmname(wchmm->hmminfo, ltmp->name); } if (lcd == NULL) { jlog("ERROR: wchmm: at word #%d: no lcdset found for [%s]\n", word, ltmp->name); ok_p = FALSE; } } } #endif /* PASS1_IWCD */ for (k = 1; k < ltmp_state_num - 1; k++) { /* for each state in the phone */ /* set state output prob info */ #ifdef PASS1_IWCD if (wchmm->ccd_flag) { /* output info of triphones needs special handling */ if (wchmm->winfo->wlen[word] == 1) { /* word with only 1 phone */ wchmm->outstyle[ntmp] = AS_LRSET; wchmm->state[ntmp].out.lrset = (LRC_INFO *)mybmalloc2(sizeof(LRC_INFO), &(wchmm->malloc_root)); (wchmm->state[ntmp].out.lrset)->hmm = ltmp; (wchmm->state[ntmp].out.lrset)->state_loc = k; if (wchmm->category_tree) { (wchmm->state[ntmp].out.lrset)->category = wchmm->winfo->wton[word]; } } else if (j == 0) { /* head phone of a word */ wchmm->outstyle[ntmp] = AS_RSET; wchmm->state[ntmp].out.rset = (RC_INFO *)mybmalloc2(sizeof(RC_INFO), &(wchmm->malloc_root)); (wchmm->state[ntmp].out.rset)->hmm = ltmp; (wchmm->state[ntmp].out.rset)->state_loc = k; } else if (j == wchmm->winfo->wlen[word] - 1) { /* last phone of a word */ wchmm->outstyle[ntmp] = AS_LSET; wchmm->state[ntmp].out.lset = &(lcd->stateset[k]); } else { wchmm->outstyle[ntmp] = AS_STATE; if (ltmp->is_pseudo) { jlog("WARNING: wchmm: word-internal phone should not be pseudo\n"); put_voca(stdout, wchmm->winfo, word); ok_p = FALSE; } wchmm->state[ntmp].out.state = ltmp->body.defined->s[k]; } } else { /* monophone */ if (ltmp->is_pseudo) { j_internal_error("wchmm_add_word: CDSET phoneme exist in monophone?\n"); put_voca(stdout, wchmm->winfo, word); ok_p = FALSE; } wchmm->outstyle[ntmp] = AS_STATE; wchmm->state[ntmp].out.state = ltmp->body.defined->s[k]; } #else /* ~PASS1_IWCD */ if (ltmp->is_pseudo) { j_internal_error("wchmm_add_word: CDSET phoneme exist in monophone?\n"); put_voca(stdout, wchmm->winfo, word); ok_p = FALSE; } wchmm->state[ntmp].out = ltmp->body.defined->s[k]; #endif /* PASS1_IWCD */ /* initialize other info */ acc_init(wchmm, ntmp); wchmm->stend[ntmp] = WORD_INVALID; if (! wchmm->hmminfo->multipath) { /* make transition arc from HMM transition info */ for (ato = 1; ato < ltmp_state_num; ato++) { prob = (hmm_logical_trans(ltmp))->a[k][ato]; if (prob != LOG_ZERO) { if (j == add_tail && k == ltmp_state_num - 2 && ato == ltmp_state_num - 1) { /* arc outside new part will be handled later */ } else { add_wacc(wchmm, ntmp, prob, ntmp + ato - k); } } } } ntmp++; /* expand wchmm if neccesary */ if (ntmp >= wchmm->maxwcn) wchmm_extend(wchmm); } /* end of state loop */ } /* end of phone loop */ if (wchmm->hmminfo->multipath) { /* On multipath version, the skip transition should be handled! */ /* make transition arc from HMM transition info */ ntmp = n; for (j = add_head; j <= add_tail; j++) { ltmp = wchmm->winfo->wseq[word][j]; ltmp_state_num = hmm_logical_state_num(ltmp); out_num_next = 0; /* arc from initial state ... need arc expansion from precious phone */ for (ato = 1; ato < ltmp_state_num; ato++) { prob = (hmm_logical_trans(ltmp))->a[0][ato]; if (prob != LOG_ZERO) { /* expand arc from previous HMM */ if (ato == ltmp_state_num - 1) { /* to final state ... just register states for next expansion */ for(kkk=0; kkka[k][ato]; if (prob != LOG_ZERO) { if (ato == ltmp_state_num - 1) { /* to final state ... register states for next expansion */ out_from_next[out_num_next] = ntmp; out_a_next[out_num_next] = prob; out_num_next++; } else { add_wacc(wchmm, ntmp, prob, ntmp + ato - k); } } } ntmp++; } /* end of state loop */ /* swap out list for next phone */ for(kkk=0;kkkhmminfo->multipath && enable_iwsp && add_tail - add_head + 1 > 0) { /* there are new phones to be created */ int ntmp_bak; /* set short pause state info */ ntmp_bak = ntmp; if (wchmm->hmminfo->sp->is_pseudo) { for(k = 1;k < hmm_logical_state_num(wchmm->hmminfo->sp) - 1; k++) { wchmm->outstyle[ntmp] = AS_LSET; wchmm->state[ntmp].out.lset = &(wchmm->hmminfo->sp->body.pseudo->stateset[k]); acc_init(wchmm, ntmp); wchmm->stend[ntmp] = WORD_INVALID; ntmp++; if (ntmp >= wchmm->maxwcn) wchmm_extend(wchmm); } } else { for(k = 1;k < hmm_logical_state_num(wchmm->hmminfo->sp) - 1; k++) { wchmm->outstyle[ntmp] = AS_STATE; wchmm->state[ntmp].out.state = wchmm->hmminfo->sp->body.defined->s[k]; acc_init(wchmm, ntmp); wchmm->stend[ntmp] = WORD_INVALID; ntmp++; if (ntmp >= wchmm->maxwcn) wchmm_extend(wchmm); } } ntmp = ntmp_bak; /* connect incoming arcs from previous phone */ out_num_next = 0; for (ato = 1; ato < hmm_logical_state_num(wchmm->hmminfo->sp); ato++) { prob = hmm_logical_trans(wchmm->hmminfo->sp)->a[0][ato]; if (prob != LOG_ZERO) { /* to control short pause insertion, transition probability toward the word-end short pause will be given a penalty */ prob += wchmm->hmminfo->iwsp_penalty; if (ato == hmm_logical_state_num(wchmm->hmminfo->sp) - 1) { /* model has a model skip transition, just inherit them to next */ for(kkk=0; kkkhmminfo->sp)->a[0][hmm_logical_state_num(wchmm->hmminfo->sp)-1] == LOG_ZERO) { /* to make insertion sp model to have no effect on the original path, the skip transition probability should be 0.0 (=100%) */ prob = 0.0; for(kkk=0; kkkhmminfo->sp) - 1; k++) { for (ato = 1; ato < hmm_logical_state_num(wchmm->hmminfo->sp); ato++) { prob = hmm_logical_trans(wchmm->hmminfo->sp)->a[k][ato]; if (prob != LOG_ZERO) { if (ato == hmm_logical_state_num(wchmm->hmminfo->sp) - 1) { out_from_next[out_num_next] = ntmp; out_a_next[out_num_next] = prob; out_num_next++; } else { add_wacc(wchmm, ntmp, prob, ntmp + ato - k); } } } ntmp++; } /* swap work area for next */ for(kkk=0;kkk node on wchmm */ for (j=0;joffset[word][j] = wchmm->offset[matchword][j]; } else if (add_tail < j) { /* shared tail part (should not happen..) */ wchmm->offset[word][j] = wchmm->offset[matchword][j+(matchword_len-word_len)]; } else { /* newly created part */ wchmm->offset[word][j] = n; n += hmm_logical_state_num(wchmm->winfo->wseq[word][j]) - 2; } } if (wchmm->hmminfo->multipath) { /* create word-end node */ /* paranoia check if the short-pause addition has been done well */ if (enable_iwsp && add_tail - add_head + 1 > 0) { n += hmm_logical_state_num(wchmm->hmminfo->sp) - 2; if (n != ntmp) j_internal_error("wchmm_add_word: cannot match\n"); } /* create word-end node */ wchmm->wordend[word] = n; /* tail node of 'word' is 'n' */ wchmm->stend[n] = word; /* node 'k' is a tail node of 'word' */ acc_init(wchmm, n); wchmm->state[n].out.state = NULL; /* connect the final outgoing arcs in out_from[] to the word end node */ for(k = 0; k < out_num_prev; k++) { add_wacc(wchmm, out_from[k], out_a[k], n); } n++; if (n >= wchmm->maxwcn) wchmm_extend(wchmm); if (matchlen == 0) { /* check if the new word has whole word-skipping transition */ /* (use out_from and out_num_prev temporary) */ out_num_prev = 0; get_outtrans_list(wchmm, word, word_len-1, out_from, out_a, &out_num_prev, wchmm->winfo->maxwn, enable_iwsp); for(k=0;kwordbegin[word]) { jlog("ERROR: *** ERROR: WORD SKIPPING TRANSITION NOT ALLOWED ***\n"); jlog("ERROR: Word id=%d (%s[%s]) has \"word skipping transition\".\n", word, wchmm->winfo->wname[word], wchmm->winfo->woutput[word]); jlog("ERROR: All HMMs in the word:\n "); for(kkk=0;kkkwinfo->wlen[word];kkk++) { jlog("%s ", wchmm->winfo->wseq[word][kkk]->name); } jlog("\n"); jlog("ERROR: has transitions from initial state to final state.\n"); jlog("ERROR: This type of word skipping is not supported.\n"); ok_p = FALSE; } } } wchmm->n = n; } else { wchmm->n = n; k = wchmm->offset[word][word_len-1] + hmm_logical_state_num(wchmm->winfo->wseq[word][word_len-1])-2 -1; wchmm->wordend[word] = k; /* tail node of 'word' is 'k' */ wchmm->stend[k] = word; /* node 'k' is a tail node of 'word' */ if (matchlen != 0 && add_tail - add_head + 1 > 0) { /* new part has been created in the above procedure: */ /* now make link from shared part to the new part */ wchmm_link_subword(wchmm, matchword,add_to,word,add_head); } } return(ok_p); } /*************************************************************/ /**** parse whole structure (after wchmm has been built) *****/ /*************************************************************/ /** * * 木構造化辞書を走査し，単語の終端状態から外への次遷移確率のリストを作成する. * (non multipath) * * @param wchmm [i/o] 木構造化辞書 * * * Scan the lexicon tree to make list of emission probability from the word end * state. (non multipath) * * @param wchmm [i/o] tree lexicon * */ static void wchmm_calc_wordend_arc(WCHMM_INFO *wchmm) { WORD_ID w; HTK_HMM_Trans *tr; LOGPROB a; for (w=0;wwinfo->num;w++) { tr = hmm_logical_trans(wchmm->winfo->wseq[w][wchmm->winfo->wlen[w]-1]); a = tr->a[tr->statenum-2][tr->statenum-1]; wchmm->wordend_a[w] = a; } } #ifdef SEPARATE_BY_UNIGRAM /********************************************************************/ /****** for separation (linearization) of high-frequent words *******/ /********************************************************************/ /** * * unigram確率でソートするための qsort コールバック関数. * * @param a [in] 要素1 * @param b [in] 要素2 * * @return 演算の結果の符合を返す. * * * qsort callback function to sort unigram values. * * @param a [in] element #1 * @param b [in] element #2 * * @return the result of comparison. * */ static int compare_prob(LOGPROB *a, LOGPROB *b) { if (*a < *b) return (1); if (*a > *b) return (-1); return(0); } /** * * 1-gramスコアの上位 N 番目の値を求める. * * @param winfo [in] 単語辞書 * @param n [in] 求める順位 * * @return 上位 N 番目の uni-gram 確率の値を返す. * * * Get the Nth-best unigram probability from all words. * * @param winfo [in] word dictionary * @param n [in] required rank * * @return the Nth-best unigram probability. * */ static LOGPROB get_nbest_uniprob(WCHMM_INFO *wchmm, int n) { LOGPROB *u_p; WORD_ID w; LOGPROB x; WORD_INFO *winfo; NGRAM_INFO *ngram; winfo = wchmm->winfo; ngram = wchmm->ngram; if (n < 1) n = 1; if (n > winfo->num) n = winfo->num; /* store all unigram probability to u_p[] */ u_p = (LOGPROB *)mymalloc(sizeof(LOGPROB) * winfo->num); for(w=0;wnum;w++) { if (ngram) { x = uni_prob(ngram, winfo->wton[w]) #ifdef CLASS_NGRAM + winfo->cprob[w] #endif ; } else { x = LOG_ZERO; } if (wchmm->lmvar == LM_NGRAM_USER) { x = (*(wchmm->uni_prob_user))(wchmm->winfo, w, x); } u_p[w] = x; } /* sort them downward */ qsort(u_p, winfo->num, sizeof(LOGPROB), (int (*)(const void *,const void *))compare_prob); /* return the Nth value */ x = u_p[n-1]; free(u_p); return(x); } #endif /**********************************************************/ /****** MAKE WCHMM (LEXICON TREE) --- main function *******/ /**********************************************************/ #define COUNT_STEP 500 ///< Word count step for debug progress output /** * * 与えられた単語辞書と言語モデルから木構造化辞書を構築する. この関数は * 処理が遅く，Julianで"-oldtree"オプション指定時のみ使用されます. オプション * 非指定時およびJuliusでは代わりに build_wchmm2() が用いられます. * * @param wchmm [i/o] 木構造化辞書 * @param lmconf [in] 言語モデル(LM)設定パラメータ * * * Build a tree lexicon from given word dictionary and language model. * This function is slow and only used when "-oldtree" option is specified * in Julian. Julian without that option and Julius uses build_wchmm2() * instead of this. * * @param wchmm [i/o] lexicon tree * @param lmconf [in] language model (LM) configuration parameters * * @callgraph * @callergraph */ boolean build_wchmm(WCHMM_INFO *wchmm, JCONF_LM *lmconf) { int i,j; int matchword=0, sharelen=0, maxsharelen=0; int num_duplicated; #ifdef SEPARATE_BY_UNIGRAM LOGPROB separate_thres; LOGPROB p; #endif boolean ok_p; /* lingustic infos must be set before build_wchmm() is called */ /* check if necessary lingustic info is already assigned (for debug) */ if (wchmm->winfo == NULL || (wchmm->lmvar == LM_NGRAM && wchmm->ngram == NULL) || (wchmm->lmvar == LM_DFA_GRAMMAR && wchmm->dfa == NULL) ) { jlog("ERROR: wchmm: linguistic info not available!!\n"); return FALSE; } ok_p = TRUE; #ifdef SEPARATE_BY_UNIGRAM /* 上位[separate_wnum]番目の1-gramスコアを求める */ /* 1-gramスコアがこの値以上のものは木から分ける */ separate_thres = get_nbest_uniprob(wchmm, lmconf->separate_wnum); #endif #ifdef PASS1_IWCD #ifndef USE_OLD_IWCD if (wchmm->category_tree) { if (wchmm->ccd_flag) { /* 全てのカテゴリID付き lcd_set を作成 */ lcdset_register_with_category_all(wchmm); } } #endif #endif /* PASS1_IWCD */ /* wchmmを初期化 */ wchmm_init(wchmm); /* カウンタリセット */ wchmm->separated_word_count=0; jlog("STAT: wchmm: Building HMM lexicon tree (left-to-right)\n"); for (i=0;iwinfo->num;i++) { if (wchmm->lmtype == LM_PROB) { if (i == wchmm->winfo->head_silwid || i == wchmm->winfo->tail_silwid) { /* 先頭/末尾の無音モデルは木構造化せず， * 先頭の無音単語の先頭への遷移，末尾単語の末尾からの遷移は作らない*/ /* sharelen=0でそのまま */ if (wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp) == FALSE) { jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n"); ok_p = FALSE; } continue; } #ifndef NO_SEPARATE_SHORT_WORD if (wchmm->winfo->wlen[i] <= SHORT_WORD_LEN) { /* 長さの短い単語を木構造化しない(ここでは1音節) */ /* sharelen=0でそのまま */ if (wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp) == FALSE) { jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n"); ok_p = FALSE; } wchmm->separated_word_count++; continue; } #endif #ifdef SEPARATE_BY_UNIGRAM if (wchmm->ngram) { p = uni_prob(wchmm->ngram, wchmm->winfo->wton[i]) #ifdef CLASS_NGRAM + wchmm->winfo->cprob[i] #endif ; } else { p = LOG_ZERO; } if (wchmm->lmvar == LM_NGRAM_USER) { p = (*(wchmm->uni_prob_user))(wchmm->winfo, i, p); } if (p >= separate_thres && wchmm->separated_word_count < lmconf->separate_wnum) { /* 頻度の高い単語を木構造化しない */ /* separate_thres は上位separate_wnum番目のスコア */ if (wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp) == FALSE) { jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n"); ok_p = FALSE; } wchmm->separated_word_count++; continue; } #endif } /* 最も長く音素を共有出来る単語を探す */ maxsharelen=0; for (j=0;jcategory_tree && wchmm->lmtype == LM_DFA) { if (wchmm->winfo->wton[i] != wchmm->winfo->wton[j]) continue; } sharelen = wchmm_check_match(wchmm->winfo, i, j); if (sharelen == wchmm->winfo->wlen[i] && sharelen == wchmm->winfo->wlen[j]) { /* word に同音語が存在する */ /* 必ず最大の長さであり，重複カウントを避けるためここで抜ける */ maxsharelen = sharelen; matchword = j; break; } if (sharelen > maxsharelen) { matchword = j; maxsharelen = sharelen; } } if (wchmm_add_word(wchmm, i, maxsharelen, matchword, lmconf->enable_iwsp) == FALSE) { jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n"); ok_p = FALSE; } } #if 0 /* 木構造を作らない */ for (i=0;iwinfo->num;i++) { if (wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp) == FALSE) { jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n"); ok_p = FALSE; } } #endif jlog("STAT: %5d words ended (%6d nodes)\n",i,wchmm->n); if (! wchmm->hmminfo->multipath) { /* 同一音素系列を持つ単語同士の leaf node を2重化して区別する */ num_duplicated = wchmm_duplicate_leafnode(wchmm); jlog("STAT: %d leaf nodes are made unshared\n", num_duplicated); /* 単語の終端から外への遷移確率を求めておく */ wchmm_calc_wordend_arc(wchmm); } /* wchmmの整合性をチェックする */ check_wchmm(wchmm); /* factoring用に各状態に後続単語のリストを付加する */ if (!wchmm->category_tree) { #ifdef UNIGRAM_FACTORING if (wchmm->lmtype == LM_PROB) { /* 同時に前もってfactoring値を計算 */ make_successor_list_unigram_factoring(wchmm); jlog("STAT: 1-gram factoring values has been pre-computed\n"); } else { make_successor_list(wchmm); } #else make_successor_list(wchmm); #endif /* UNIGRAM_FACTORING */ if (wchmm->hmminfo->multipath) { /* 構築された factoring 情報をスキップ遷移および文頭文法ノードにコピー */ adjust_sc_index(wchmm); } #ifdef UNIGRAM_FACTORING if (wchmm->lmtype == LM_PROB) { /* 単語間LMキャッシュが必要なノードのリストを作る */ make_iwcache_index(wchmm); } #endif /* UNIGRAM_FACTORING */ } jlog("STAT: done\n"); return ok_p; } /** * * 与えられた単語辞書と言語モデルから木構造化辞書を構築する. * この関数は bulid_wchmm() と同じ処理を行いますが， * 最初に単語を音素列でソートして音素列の似た順に単語を並べるため， * より高速に木構造化を行うことができる. とくにオプション指定をしない * 限り，Julius/Julianではこちらが用いられる. * * @param wchmm [i/o] 木構造化辞書 * @param lmconf [in] 言語モデル(LM)設定パラメータ * * * Build a tree lexicon from given word dictionary and language model. * This function does the same job as build_wchmm(), but it is much * faster because finding of the longest matched word to an adding word * is done by first sorting all the words in the dictoinary by their phoneme * sequence order. This function will be used instead of build_wchmm() * by default. * * @param wchmm [i/o] lexicon tree * @param lmconf [in] language model (LM) configuration parameters * * @callgraph * @callergraph */ boolean build_wchmm2(WCHMM_INFO *wchmm, JCONF_LM *lmconf) { int i,j, last_i; int num_duplicated; WORD_ID *windex; #ifdef SEPARATE_BY_UNIGRAM LOGPROB separate_thres; LOGPROB p; #endif boolean ok_p; boolean ret; /* lingustic infos must be set before build_wchmm() is called */ /* check if necessary lingustic info is already assigned (for debug) */ if (wchmm->winfo == NULL || (wchmm->lmvar == LM_NGRAM && wchmm->ngram == NULL) || (wchmm->lmvar == LM_DFA_GRAMMAR && wchmm->dfa == NULL) ) { jlog("ERROR: wchmm: linguistic info not available!!\n"); return FALSE; } ok_p = TRUE; wchmm->separated_word_count = 0; jlog("STAT: Building HMM lexicon tree\n"); if (wchmm->lmtype == LM_PROB) { #ifdef SEPARATE_BY_UNIGRAM /* compute score threshold beforehand to separate words from tree */ /* here we will separate best [separate_wnum] words from tree */ separate_thres = get_nbest_uniprob(wchmm, lmconf->separate_wnum); #endif } #ifdef PASS1_IWCD #ifndef USE_OLD_IWCD if (wchmm->category_tree) { if (wchmm->ccd_flag) { /* when Julian mode (category-tree) and triphone is used, make all category-indexed context-dependent phone set (cdset) here */ /* these will be assigned on the last phone of each word on tree */ lcdset_register_with_category_all(wchmm); } } #endif #endif /* PASS1_IWCD */ /* initialize wchmm */ wchmm_init(wchmm); /* make sorted word index ordered by phone sequence */ windex = (WORD_ID *)mymalloc(sizeof(WORD_ID) * wchmm->winfo->num); for(i=0;iwinfo->num;i++) windex[i] = i; if (wchmm->category_tree && wchmm->lmtype == LM_DFA) { /* sort by category -> sort by word ID in each category */ wchmm_sort_idx_by_category(wchmm->winfo, windex, wchmm->winfo->num); { int last_cate; last_i = 0; last_cate = wchmm->winfo->wton[windex[0]]; for(i = 1;iwinfo->num;i++) { if (wchmm->winfo->wton[windex[i]] != last_cate) { wchmm_sort_idx_by_wseq(wchmm->winfo, windex, last_i, i - last_i); last_cate = wchmm->winfo->wton[windex[i]]; last_i = i; } } wchmm_sort_idx_by_wseq(wchmm->winfo, windex, last_i, wchmm->winfo->num - last_i); } } else { /* sort by word ID for whole vocabulary */ wchmm_sort_idx_by_wseq(wchmm->winfo, windex, 0, wchmm->winfo->num); } /* * { * int i,w; * for(i=0;iwinfo->num;i++) { * w = windex[i]; * printf("%d: cate=%4d wid=%4d %s\n",i, wchmm->winfo->wton[w], w, wchmm->winfo->woutput[w]); * } * } */ /* incrementaly add words to lexicon tree */ /* now for each word, the previous word (last_i) is always the most matched one */ last_i = WORD_INVALID; for (j=0;jwinfo->num;j++) { i = windex[j]; if (wchmm->lmtype == LM_PROB) { /* start/end silence word should not be shared */ if (i == wchmm->winfo->head_silwid || i == wchmm->winfo->tail_silwid) { /* add whole word as new (sharelen=0) */ if (wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp) == FALSE) { jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n"); ok_p = FALSE; } continue; } #ifndef NO_SEPARATE_SHORT_WORD /* separate short words from tree */ if (wchmm->winfo->wlen[i] <= SHORT_WORD_LEN) { if (wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp) == FALSE) { jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n"); ok_p = FALSE; } wchmm->separated_word_count++; continue; } #endif #ifdef SEPARATE_BY_UNIGRAM if (wchmm->ngram) { p = uni_prob(wchmm->ngram, wchmm->winfo->wton[i]) #ifdef CLASS_NGRAM + wchmm->winfo->cprob[i] #endif ; } else { p = LOG_ZERO; } if (wchmm->lmvar == LM_NGRAM_USER) { p = (*(wchmm->uni_prob_user))(wchmm->winfo, i, p); } /* separate high-frequent words from tree (threshold = separate_thres) */ if (p >= separate_thres && wchmm->separated_word_count < lmconf->separate_wnum) { if (wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp) == FALSE) { jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n"); ok_p = FALSE; } wchmm->separated_word_count++; continue; } #endif } if (last_i == WORD_INVALID) { /* first word */ ret = wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp); } else { /* the previous word (last_i) is always the most matched one */ if (wchmm->category_tree && wchmm->lmtype == LM_DFA) { if (wchmm->winfo->wton[i] != wchmm->winfo->wton[last_i]) { ret = wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp); } else { ret = wchmm_add_word(wchmm, i, wchmm_check_match(wchmm->winfo, i, last_i), last_i, lmconf->enable_iwsp); } } else { ret = wchmm_add_word(wchmm, i, wchmm_check_match(wchmm->winfo, i, last_i), last_i, lmconf->enable_iwsp); } } if (ret == FALSE) { jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n"); ok_p = FALSE; } last_i = i; } /* end of add word loop */ /*j_printerr("\r %5d words ended (%6d nodes)\n",j,wchmm->n);*/ /* free work area */ free(windex); if (wchmm->hmminfo->multipath) { jlog("STAT: lexicon size: %d nodes\n", wchmm->n); } else { /* duplicate leaf nodes of homophone/embedded words */ jlog("STAT: lexicon size: %d", wchmm->n); num_duplicated = wchmm_duplicate_leafnode(wchmm); jlog("+%d=%d\n", num_duplicated, wchmm->n); } if (! wchmm->hmminfo->multipath) { /* calculate transition probability of word end node to outside */ wchmm_calc_wordend_arc(wchmm); } /* check wchmm coherence (internal debug) */ check_wchmm(wchmm); /* make successor list for all branch nodes for N-gram factoring */ if (!wchmm->category_tree) { #ifdef UNIGRAM_FACTORING if (wchmm->lmtype == LM_PROB) { /* for 1-gram factoring, we can compute the values before search */ make_successor_list_unigram_factoring(wchmm); jlog("STAT: 1-gram factoring values has been pre-computed\n"); } else { make_successor_list(wchmm); } #else make_successor_list(wchmm); #endif /* UNIGRAM_FACTORING */ if (wchmm->hmminfo->multipath) { /* Copy the factoring data according to the skip transitions and startword nodes */ adjust_sc_index(wchmm); } #ifdef UNIGRAM_FACTORING if (wchmm->lmtype == LM_PROB) { /* make list of start nodes that needs inter-word LM cache */ make_iwcache_index(wchmm); } #endif /* UNIGRAM_FACTORING */ } //jlog("STAT: done\n"); #ifdef WCHMM_SIZE_CHECK if (debug2_flag) { /* detailed check of lexicon tree size (inaccurate!) */ jlog("STAT: --- memory size of word lexicon ---\n"); jlog("STAT: wchmm: %d words, %d nodes\n", wchmm->winfo->num, wchmm->n); jlog("STAT: %9d bytes: wchmm->state[node] (exclude ac, sc)\n", sizeof(WCHMM_STATE) * wchmm->n); { int count1 = 0; int count2 = 0; int count3 = 0; for(i=0;in;i++) { if (wchmm->self_a[i] != LOG_ZERO) count1++; if (wchmm->next_a[i] != LOG_ZERO) count2++; if (wchmm->ac[i] != NULL) count3++; } jlog("STAT: %9d bytes: wchmm->self_a[node] (%4.1f%% filled)\n", sizeof(LOGPROB) * wchmm->n, 100.0 * count1 / (float)wchmm->n); jlog("STAT: %9d bytes: wchmm->next_a[node] (%4.1f%% filled)\n", sizeof(LOGPROB) * wchmm->n, 100.0 * count2 / (float)wchmm->n); jlog("STAT: %9d bytes: wchmm->ac[node] (%4.1f%% used)\n", sizeof(A_CELL2 *) * wchmm->n, 100.0 * count3 / (float)wchmm->n); } jlog("STAT: %9d bytes: wchmm->stend[node]\n", sizeof(WORD_ID) * wchmm->n); { int w,count; count = 0; for(w=0;wwinfo->num;w++) { count += wchmm->winfo->wlen[w] * sizeof(int) + sizeof(int *); } jlog("STAT: %9d bytes: wchmm->offset[w][]\n", count); } if (wchmm->hmminfo->multipath) { jlog("STAT: %9d bytes: wchmm->wordbegin[w]\n", wchmm->winfo->num * sizeof(int)); } jlog("STAT: %9d bytes: wchmm->wordend[w]\n", wchmm->winfo->num * sizeof(int)); jlog("STAT: %9d bytes: wchmm->startnode[]\n", wchmm->startnum * sizeof(int)); if (wchmm->category_tree) { jlog("STAT: %9d bytes: wchmm->start2wid[]\n", wchmm->startnum * sizeof(WORD_ID)); } #ifdef UNIGRAM_FACTORING if (wchmm->lmtype == LM_PROB) { jlog("STAT: %9d bytes: wchmm->start2isolate[]\n", wchmm->isolatenum * sizeof(int)); } #endif if (!wchmm->hmminfo->multipath) { jlog("STAT: %9d bytes: wchmm->wordend_a[]\n", wchmm->winfo->num * sizeof(LOGPROB)); } #ifdef PASS1_IWCD jlog("STAT: %9d bytes: wchmm->outstyle[]\n", wchmm->n * sizeof(unsigned char)); { int c; c = 0; for(i=0;in;i++) { switch(wchmm->outstyle[i]) { case AS_RSET: c += sizeof(RC_INFO); break; case AS_LRSET: c += sizeof(LRC_INFO); break; } } if (c > 0) jlog("STAT: %9d bytes: wchmm->out (RC_INFO / LRC_INFO)\n", c); } #endif if (!wchmm->category_tree) { int c = 0; #ifdef UNIGRAM_FACTORING jlog("STAT: %9d bytes: wchmm->scword[]\n", sizeof(WORD_ID) * wchmm->scnum); #else for(i=1;iscnum;i++) { c += wchmm->sclen[i]; } jlog("STAT: %9d bytes: wchmm->sclist[]\n", c * sizeof(WORD_ID) + wchmm->scnum * sizeof(WORD_ID *)); jlog("STAT: %9d bytes: wchmm->sclen[]\n", wchmm->scnum * sizeof(WORD_ID)); #endif #ifdef UNIGRAM_FACTORING if (wchmm->lmtype == LM_PROB) { jlog("STAT: %9d bytes: wchmm->fscore[]\n", wchmm->fsnum * sizeof(LOGPROB)); } #endif } { int count, n; A_CELL2 *ac; count = 0; for(n=0;nn;n++) { for(ac=wchmm->ac[n];ac;ac=ac->next) { count += sizeof(A_CELL2); } } jlog("STAT: %9d bytes: A_CELL2\n", count); } } #endif /* WCHMM_SIZE_CHECK */ return ok_p; } /** * * 木構造化辞書のサイズなどの情報を標準出力に出力する. * * @param wchmm [in] 木構造化辞書 * * * Output some specifications of the tree lexicon (size etc.) to stdout. * * @param wchmm [in] tree lexicon already built * * @callgraph * @callergraph */ void print_wchmm_info(WCHMM_INFO *wchmm) { int n,i, rootnum; if (wchmm->hmminfo->multipath) { rootnum = wchmm->startnum; } else { if (wchmm->lmtype == LM_PROB) { rootnum = wchmm->startnum + 1; /* including winfo->head_silwid */ } else if (wchmm->lmtype == LM_DFA) { rootnum = wchmm->startnum; } } jlog(" Lexicon tree:\n"); jlog("\t total node num = %6d\n", wchmm->n); if (wchmm->lmtype == LM_PROB) { jlog("\t root node num = %6d\n", rootnum); #ifdef NO_SEPARATE_SHORT_WORD #ifdef SEPARATE_BY_UNIGRAM jlog("\t(%d hi-freq. words are separated from tree lexicon)\n", wchmm->separated_word_count); #else jlog(" (no words are separated from tree)\n"); #endif /* SEPARATE_BY_UNIGRAM */ #else jlog(" (%d short words (<= %d phonemes) are separated from tree)\n", wchmm->separated_word_count, SHORT_WORD_LEN); #endif /* NO_SEPARATE_SHORT_WORD */ } if (wchmm->lmtype == LM_DFA) { jlog("\t root node num = %6d\n", rootnum); } for(n=0,i=0;in;i++) { if (wchmm->stend[i] != WORD_INVALID) n++; } jlog("\t leaf node num = %6d\n", n); if (!wchmm->category_tree) { jlog("\t fact. node num = %6d\n", wchmm->scnum - 1); } } /* end of file */