# lattice/輸出

## dev/test
> [name=鴻欣]kaldi這馬攏用best case

## decoding
- kaldi模型訓練好，來辨識新的音檔內容
- [維特比演算法](https://zh.wikipedia.org/wiki/%E7%BB%B4%E7%89%B9%E6%AF%94%E7%AE%97%E6%B3%95)

#### 資料夾內容
- scoring/text.filt: 辨識的正確答案
- scoring/X.a.b.txt: 參數lm-scale=X word-insert=a.b的kaldi辨識結果
- wer_X_a.b: 參數lm-scale=X word-insert=a.b的錯誤率
- 檔案內的WER：word error rate詞錯誤率
- 檔案內的SER：setence error rate
- 咱討論會講的SER：syallable error rate
- CER：charactor error rate漢字錯誤計

0000000倪齊民-tong0000000-ku0000000 sian-sinn｜sian-sinn 詞
0000000倪齊民-tong0000000-ku0000000 sian-sinn｜sian-Kinn 詞 100%

0000000倪齊民-tong0000000-ku0000000 sian｜sian sinn｜sinn 音節
0000000倪齊民-tong0000000-ku0000000 sian｜sian sinn｜Kinn 音節 50 %

``
to｜to sia｜sia
tse｜tse sia｜sia
``
%WER 28.47 [ 24652 / 86587, 3375 ins, 3721 del, 17556 sub ]
#### 辨識指令
佇[kaldi/egs/taiwanese/s5c/走評估.sh](https://github.com/sih4sing5hong5/kaldi/blob/taiwanese/egs/taiwanese/s5c/%E8%B5%B0%E8%A9%95%E4%BC%B0.sh#L54-L59)

    graph_dir=exp/tri4/graph
    $train_cmd $graph_dir/mkgraph.log \
      utils/mkgraph.sh $lang exp/tri4 $graph_dir
    steps/decode_fmllr.sh --nj $nj --cmd "$decode_cmd" \
      --config conf/decode.config \
      $graph_dir data/dev exp/tri4/decode_train_dev

tri3,4的模型, decode結果
`exp/tri3/decode_train_dev/`
tri3,4的exp/tri3/final.mdl,exp/tri3/topo, ...

## prune, rescoring
##### [ Speech activity detection and endpointing in real time ASR ](https://groups.google.com/d/msg/kaldi-help/v9_b_r4r4Og/-g8n7fUDAQAJ)
- lm 斷句kah斷段，效果愛做實驗

##### [ prune language model or small text corpus ](https://groups.google.com/d/msg/kaldi-help/SyncOpoXpUQ/O3FJ9oLGBgAJ)
- prune了後,perplexity huān-sè 會變好
- GT 是 dan 經驗siōng hó ê discount

#### ctm
chain model出來ê[愛`*frame-subsampling-factor`，預設=3](http://kaldi-asr.org/doc/chain.html)，:

LOG (nnet3-latgen-faster[5.5.162~1420-ca32c]:CheckAndFixConfigs():nnet-am-decodable-simple.cc:294) Increasing [--frames-per-chunk](http://kaldi-asr.org/doc/dnn3_scripts_context.html) from 50 to 51 to make it a multiple of --frame-subsampling-factor=3

##### align
`steps/get_train_ctm.sh`

##### decoding

###### [ show-alignment at a frame level? ](https://groups.google.com/d/msg/kaldi-help/-H7a9IuhLLE/sUig5eT2AQAJ) show-alignments, ali-to-phones, copy-int-vector

#### 輸出
##### [Replacing unk in decode output by best matching phone combination ](https://groups.google.com/d/msg/kaldi-help/uGiExyOiyjc/VkDQfOB0AgAJ)
##### [ the difference between posteriors calculated by forward-backward and lattice-to-post ](https://groups.google.com/d/msg/kaldi-help/iBgeXN4diSY/27CD0v0wCwAJ)

### 其他參考文章

* decode
* https://groups.google.com/forum/m/#!msg/kaldi-help/tAGblb-8Hy0/oMWnpgvhAAAJ
* lm、fst
http://vpanayotov.blogspot.tw/2012/06/kaldi-decoding-graph-construction.html
* acwt參數
* https://groups.google.com/forum/#!topic/kaldi-developers/IBhn9ndmjQI
* lattice
* https://groups.google.com/forum/#!topic/kaldi-help/cXX7y3Hvf3w
* https://groups.google.com/forum/#!topic/kaldi-help/QqUQoX816GE
* https://groups.google.com/forum/#!topic/kaldi-help/UiVD5WPA8fI
* 檢查聲學模型
* https://groups.google.com/forum/m/#!msg/kaldi-help/HU5FT32EguU/cIMjnWRUAQAJ
real-time decoding
* https://groups.google.com/forum/#!searchin/kaldi-help/recording|sort:relevance/kaldi-help/xogaf6nAF3E/xSfh0gFuLAAJ

### lattices
##### [Query on converting phoneme lattice to word lattice](https://groups.google.com/d/msg/kaldi-help/APrZQNaF-S4/6KYXyGo_AwAJ)
- 用`utils/show-lattice.sh`來顯示lattice圖
- cost愛規條path看才有意義，因為圖無frame的資訊
- word lattice變phoneme lattice有資訊無去

##### [ how can we compose with syllable lattices to get word lattices](https://groups.google.com/d/msg/kaldi-help/Dx3gqHQd1Hw/kXpPuIPaAwAJ)
有介紹lattice=>fst，佇fst操作

### decoding
##### [reduce the beams來加速](https://groups.google.com/d/msg/kaldi-help/Jj1yVqr-rbQ/SiiOIm-tBwAJ)

##### [ Re: missing words while do-endpointing is true](https://groups.google.com/d/msg/kaldi-help/XIcyX3weV9w/EHQIC4VPEQAJ)
The --do-endpointing is not really segmentation. It is for the scenario where you want to just stop the recognition at a certain point (e.g. to demonstrate how you might do that for an interactive application). So it will just discard the rest of the file.

##### [Decoding longer utterances of AMI dev](https://groups.google.com/d/msg/kaldi-help/Qraqjz83Tw4/5u9JBSaVBQAJ)
`Determinization finished earlier than the beam for utterance AMI_ES2011b_SDM_FEE041_0107148_0107863
LOG (latgen-faster:DecodeUtteranceLatticeFaster():decoder-wrappers.cc:111) Log-like per frame for utterance AMI_ES2011b_SDM_FEE041_0107148_0107863 is 0.713222 over 713 frames.
`代表` the lattice is not as deep as it might have been.`，無啥要緊

##### [ Is it possible to score in kaldi by ignoring the compound word splitting in particular language (Telugu). ](https://groups.google.com/d/msg/kaldi-help/vQetEk1h4m4/5zTYJYGdBQAJ)
用`local/wer_output_filter`調整輸出結果，方便比較

##### [ Recovering "OOV" phones after SAU conversion ](https://groups.google.com/d/msg/kaldi-help/8UV7ngdp72I/yYG4qnBZBwAJ)
辨識結果莫出現`unk`的改法：`fstprint`共fst轉做txt，共unk相關的提掉，才用`fstcompile`轉轉`txt`

##### [ The right way of adding new words to existing ngram LM. ](https://groups.google.com/d/msg/kaldi-help/d1KIOgREd84/qBBg98JCBAAJ)
[add_unigrams_arpa.pl](https://github.com/kaldi-asr/kaldi/blob/master/egs/wsj/s5/utils/lang/add_unigrams_arpa.pl)會當直接佇arpa加unigram，而且加的機率攏會當調整

##### [ Replacing <unk> in decode output by best matching phone combination ](https://groups.google.com/d/msg/kaldi-help/uGiExyOiyjc/tXlfTKVKCAAJ)
- 提掉oov的兩種可能做法

##### [Some Kaldi Notes](http://jrmeyer.github.io/asr/2016/02/01/Kaldi-notes.html)
`L.fst`,`L_disambig.fst`,`G.fst`,...的介紹