# Kaldi Alignent 實作

參考文章
---
- [Jonathan Hui](https://medium.com/@jonathan_hui/speech-recognition-kaldi-35fec0320496)
- [Josh Meyer](http://jrmeyer.github.io/asr/2019/08/17/Kaldi-troubleshooting.html)

腳本
---
1. 從lattice 中生出phone alignment

#!/bin/bash
stage=0
experiment=exp_DAAIKIDAug_Decept_AddADOSNoise

fold=0
. ./path.sh
. ./utils/parse_options.sh

lattice_path=${experiment}/chain/ADOS_tdnn_fold${fold}_transfer/decode_test_kid${fold}_groundtruth/

if [ $stage -le 0 ]; then
  rm -f ${lattice_path}/phone_readable_fold${fold}.ctm
  for lat in $(find $lattice_path -iname "lat.*.gz");do     
    filename=${lat##*/}; IFS="." digit=( $filename );
    lattice-align-phones --replace-output-symbols=true ${experiment}/chain/ADOS_tdnn_fold${fold}_transfer/final.mdl  ark:"gunzip -c ${lat}|" ark:${lattice_path}/Phone_lat.${digit}.gz
    lattice-to-ctm-conf ark:${lattice_path}/Phone_lat.${digit}.gz -| utils/int2sym.pl -f 5 data/lang/phones.txt - >> ${lattice_path}/phone_readable_fold${fold}.ctm 
  done
  mkdir -p ${lattice_path}/Phone_level_decoding; mv ${lattice_path}/phone_readable_fold${fold}.ctm ${lattice_path}/Phone_level_decoding/
fi

1. 從lattice 中生出word level alignment
這個部份kaldi已經有實踐了，可以用下列指令

Usage: steps/get_ctm.sh [options] <data-dir> <lang-dir|graph-dir> <decode-dir>

e.g. steps/get_ctm.sh --frame-shift 0.03  data/test_kid0 exp_Deceipt_ADOSfold0/chain/ADOS_tdnn_fold0_transfer/graph/ exp_Deceipt_ADOSfold0/chain/ADOS_tdnn_fold0_transfer/decode_test_kid0/

其中當你的model是採用chain的setting的時候frame-shift要設為0.03

2. 從alignment 中生出word level alignment 採用

steps/get_train_ctm.sh --cmd utils/parallel/run.pl --frame-shift 0.03 data/ADOS_afraidangrysad_hires data/lang exp_ADOShappyDAAIKidallDeceiptformosaCSRC/chain/ADOS_tdnn_fold_transfer_ali/ exp_ADOShappyDAAIKidallDeceiptformosaCSRC/chain/ADOS_tdnn_fold_transfer_ali/
e.g.

簡單快速看alignment方法：

e.g.
show-alignments exp_ADOShappyDAAIKidallDeceiptformosaCSRC/chain/ADOS_tdnn_fold_transfer_ali/phones.txt exp_ADOShappyDAAIKidallDeceiptformosaCSRC/chain/ADOS_tdnn_fold_transfer_ali/final.mdl ark,t:"gunzip -c exp_ADOShappyDAAIKidallDeceiptformosaCSRC/chain/ADOS_tdnn_fold_transfer_ali/ali.*.gz|" > Alignments.audacity.ali

###### tags: `kaldi`