# 入門

這規包文件有高明達老師、呂仁園老師、呂道誠學長佮薛丞宏的討論紀錄(2017)

---

[github](https://github.com/kaldi-asr/kaldi)佮[官方文件](http://kaldi-asr.org/doc/)。
目前包[匯出Kaldi格式資料](https://github.com/sih4sing5hong5/tai5-uan5_gian5-gi2_hok8-bu7/blob/master/%E8%87%BA%E7%81%A3%E8%A8%80%E8%AA%9E%E6%9C%8D%E5%8B%99/management/commands/%E5%8C%AF%E5%87%BAKaldi%E6%A0%BC%E5%BC%8F%E8%B3%87%E6%96%99.py)佮[taiwanese](https://github.com/sih4sing5hong5/kaldi/tree/taiwanese/egs/taiwanese/s5c)訓練script。

### 入手
1. 需Unix，建議使用[Mint Linux](https://www.linuxmint.com/download.php)作業系統。若是Mac需要先走`brew install automake autoconf wget libtool sox`
2. [裝kaldi](http://kaldi-asr.org/doc/install.html)
1. 入去`tools` ，而且照`INSTALL`做
2. 入去`src` ，而且照`INSTALL`做
3. 走範例
1. 入去`egs/yesno/s5`，照伊的`README`指示
- `bash run.sh`，幾分鐘內程式會走了，對硬體較無要求
2. 入去`egs/vystadial_en/s5`，照伊的`README`指示
- 照家己的需要改`cmd.sh`
- `bash run.sh`
- CPU時間需要5000~6000分鐘
3. `egs/voxforge/s5`，走`bash getdata.sh`佮`time bash -x run.sh`
- 純CPU時間
- `real 10081m51.386s`、`user 32927m9.961s`、`sys 294m11.191s`
4. 看[導覽教學](http://kaldi-asr.org/doc/tutorial.html)
- 伊的例，語料愛錢，所以提第三步的`yesno`佮`vystadial_en`比較
- `vystadial_en`的`data`資料夾，叫做`lang_prep`
5. 照[資料準備](http://kaldi-asr.org/doc/data_prep.html)
- 一步一步做
6. 修改範例
- [taiwanese](https://github.com/sih4sing5hong5/kaldi/tree/taiwanese/egs/taiwanese/s5c)
- 自`swbd`改的
- 參考`timit`佮`swbd`的`sgmm2`
- `egs/voxforge/s5/`的`run.sh`上清氣，較好看
- `iban` `sgmm2`的`rescore`佮`big lm`
- 用細lm來`decode`才閣用大lm來`rescore`(15.84%)，不如直接用大lm來`decode`(15.32%)
- `voxforge`有`mpe`
- `nnet3`：`wsj`、`swbd`
- `wsj/s5/local/run_sgmm2.sh` `quinphone`
- 簡單瞭解各個做法：[2013_interspeech_dnn](http://www.danielpovey.com/files/2013_interspeech_dnn.pdf)

### 其他參考文章
#### [new in KALDI](https://groups.google.com/d/msg/kaldi-help/tzyCwt7zgMQ/8pW58rYwAAAJ)
有一份110頁的pdf介紹HMM, GMM
#### [三本冊冊單](https://groups.google.com/forum/#!msg/kaldi-help/3LBSzmploC0/xDPK2vuXDgAJ)
#### [WFST bible and Dan's thesis](https://groups.google.com/d/msg/kaldi-help/3LBSzmploC0/ff35yEUqEgAJ)

### 其他相關軟體
- sox
- [srilm](http://www.speech.sri.com/projects/srilm/download.html)

##### 安裝
- [sudo apt-get install zlib1g-dev](https://groups.google.com/d/msg/kaldi-help/lke3CPycg6U/M4aciiXIAgAJ)
- [Problem in Kaldi Installation , with openfst-1.6.5 under Fedora 27 Linux
](https://groups.google.com/d/msg/kaldi-help/5smFF2RW9do/P3Ml5iITAgAJ)

### 回報問題
##### [ Regarding Cholesky decomposition failed in Kaldi ](https://groups.google.com/d/msg/kaldi-help/lEly8S-PqwA/grnum5jIAgAJ)
- 教按怎用`gdb`
##### [Re: Some problems when training CNN in nnet3](https://groups.google.com/d/msg/kaldi-help/wg6S1_qKgYw/hi2gqpm7AwAJ)
- 教按怎用`gdb`