fstaddselfloops：

Adds self-loops to states of an FST to propagate disambiguation symbols through it

They are added on each final state and each state with non-epsilon output symbols

on at least one arc out of the state.  Useful in conjunction with predeterminize

Usage:  fstaddselfloops in-disambig-list out-disambig-list  [in.fst [out.fst] ]

E.g:  fstaddselfloops in.list out.list < in.fst > withloops.fst

in.list and out.list are lists of integers, one per line, of the

same length.

例如：fstaddselfloops 'echo 276 |' 'echo 306221 |'

prepare_lang.sh:

utils/make_lexicon_fst.pl --pron-probs $tmpdir/lexiconp_disambig.txt $sil_prob $silphone '#'$ndisambig | \

    fstcompile --isymbols=$dir/phones.txt --osymbols=$dir/words.txt \

    --keep_isymbols=false --keep_osymbols=false |  \

    fstaddselfloops  "echo $phone_disambig_symbol |" "echo $word_disambig_symbol |" | \

    fstarcsort --sort_type=olabel > $dir/L_disambig.fst || exit 1;

l.fst l_disambig.fst

---

fstarcsort

 fstarcsort --sort_type=ilabel|olabel [in.fst [out.fst]]

 fstarcsort --sort_type=olabel > $dir/L_disambig.fst || exit 1;

 fstarcsort --sort_type=olabel G.fst | 

---

fstrmepsilon

移除空转移（输入和输出都为空）

fstrmepsilon | fstarcsort --sort_type=ilabel > ./data/langzh/G.fst

---

先上脚本

if [[ ! -s $lang/tmp/LG.fst || $lang/tmp/LG.fst -ot $lang/G.fst || \

      $lang/tmp/LG.fst -ot $lang/L_disambig.fst ]]; then

  fsttablecompose $lang/L_disambig.fst $lang/G.fst | fstdeterminizestar --use-log=true | \

    fstminimizeencoded | fstpushspecial > $lang/tmp/LG.fst.$$ || exit 1;

  mv $lang/tmp/LG.fst.$$ $lang/tmp/LG.fst

  fstisstochastic $lang/tmp/LG.fst || echo "[info]: LG not stochastic."

fi

---

fsttablecompose

Usage:  fsttablecompose (fst1-rxfilename|fst1-rspecifier) (fst2-rxfilename|fst2-rspecifier) [(out-rxfilename|out-rspecifier)]

fsttablecompose $lang/L_disambig.fst $lang/G.fst $lang /LG.fst

---

fstdeterminizestar

确保每个状态相同的输入只有一条路径

---

fstminimizeencoded

Minimizes FST after encoding [similar to fstminimize, but no weight-pushing]

合并路径

---

fstpushspecial

权重集中在前面，便于搜索

---

fstisstochastic

Checks whether an FST is stochastic and exits with success if so.

检查是否包含<s> </s> ??

fstisstochastic ./data/langzh/G.fst

---

fstproject

fstproject --project_output=true ./data/langzh/G.fst > ./data/langzh/G_prj.fst

---

fstreverse

头尾互换

---

fstinvert

输入输出互换:

---

fstreplace

Usage: fstreplace.exe root.fst rootlabel [rule1.fst label1 ...] [out.fst]

PROGRAM FLAGS:

  --epsilon_on_replace: type = bool, default = false

  Create an espilon arc when recursing

  --modules: type = bool, default = false

  Use modules mode

fstreplace --epsilon_on_replace data\lang\tmp\main.fst.org -1 data\lang\tmp\test.fst.org 337 data\lang\tmp\pinyin.fst.org 259 data\lang\tmp\lm.fst

---

写在最后

HCLG=asl(min(rds(det(H' omin(det(C omin(det(LoG))))))))

fstarcsort --sort_type=ilabel data\lang\tmp\CLG_3_1.fst data\lang\tmp\CLG_3_1.fst

make-h-transducer --disambig-syms-out=data\lang\graph/disambig_tid.int --transition-scale=1.000000 --print-args=false data\lang\tmp/ilabels_3_1 data\model/tree data\model/final.mdl data\lang\tmp/output

fsttablecompose --print-args=false data\lang\graph\Ha.fst data\lang\tmp\CLG_3_1.fst data\lang\tmp/output

fstdeterminizestar --print-args=false --use-log=true data\lang\tmp/input data\lang\tmp/output

fstrmsymbols --print-args=false data\lang\graph\disambig_tid.int data\lang\tmp/input data\lang\tmp/output

fstrmepslocal --print-args=false data\lang\tmp/input data\lang\tmp/output

fstminimizeencoded --print-args=false data\lang\tmp/input data\lang\tmp/output

fstisstochastic --print-args=false data\lang\graph\HCLGa.fst

add-self-loops --print-args=false --self-loop-scale=0.100000 --reorder=true data\model\final.mdl data\lang\graph\HCLGa.fst data\lang\graph\HCLG.fst