操作系统 ： Ubutu18.04_x64
gcc版本 ：7.4.0
数据准备及训练
数据地址： http://www.openslr.org/18/
在 egs/thchs30/s5 建立 thchs30-openslr 文件夹，然后把三个文件解压在了该文件夹下：
[mike@local thchs30-openslr]$ pwd /home/mike/src/kaldi/egs/thchs30/s5/thchs30-openslr [mike@local thchs30-openslr]$ tree -L 2 . ├── data_thchs30 │ ├── data │ ├── dev │ ├── lm_phone │ ├── lm_word │ ├── README.TXT │ ├── test │ └── train ├── resource │ ├── dict │ ├── noise │ └── README └── test-noise ├── 0db ├── noise ├── noise.scp ├── README ├── test.scp └── utils 14 directories, 5 files [mike@local thchs30-openslr]$
进入 s5 目录，修改脚本：
修改cmd.sh脚本，把原脚本注释掉，修改为本地运行：
export train_cmd=run.pl export decode_cmd="run.pl --mem 4G" export mkgraph_cmd="run.pl --mem 8G" export cuda_cmd="run.pl --gpu 1"
修改run.sh脚本，需要修改两个地方（cpu并行数及thchs30数据目录）：
n=8 #parallel jobs #corpus and trans directory #thchs=/nfs/public/materials/data/thchs30-openslr thchs=/home/mike/src/kaldi/egs/thchs30/s5/thchs30-openslr
执行 run.sh 即可开始训练。
查看训练效果(比如看tri1的效果) :
$ cat tri1/decode_test_word/scoring_kaldi/best_wer %WER 36.06 [ 29255 / 81139, 545 ins, 1059 del, 27651 sub ] exp/tri1/decode_test_word/wer_10_0.0
在线识别
1、安装 portaudio
cd tools ./install_portaudio.sh
2、编译相关工具
cd .. cd src make ext
3、建立目录结构并同步识别模型
从voxforge把online_demo拷贝到thchs30下，和s5同级，online_demo建online-data和work两个文件夹。 online-data下建audio和models，audio放要识别的wav，models建tri1，将s5下/exp/下的tri1下的 final.mdl 和 35.mdl 拷贝过去，把s5下的exp下的tri1下的graph_word里面的 words.txt 和 HCLG.fst 也拷过去。
模型同步脚本：
mike@local:online_demo$ cat 1_tri1.sh #! /bin/bash set -v srcDir="/home/mike/src/kaldi/egs/thchs30/s5/exp/tri1" dst="/home/mike/src/kaldi/egs/thchs30/online_demo/online-data/models/tri1/" echo $dst rm -rf $dst/*.* cp $srcDir/final.mdl $dst cp $srcDir/35.mdl $dst cp $srcDir/graph_word/words.txt $dst cp $srcDir/graph_word/HCLG.fst $dst echo "done" mike@local:online_demo$
4、修改脚本
注释如下代码：
#if [ ! -s ${data_file}.tar.bz2 ]; then # echo "Downloading test models and data ..." # wget -T 10 -t 3 $data_url; # if [ ! -s ${data_file}.tar.bz2 ]; then # echo "Download of $data_file has failed!" # exit 1 # fi #fi
修改模型：
ac_model_type=tri1
修改识别代码：
simulated) echo echo -e " SIMULATED ONLINE DECODING - pre-recorded audio is used " echo " The (bigram) language model used to build the decoding graph was" echo " estimated on an audio book's text. The text in question is" echo " "King Solomon's Mines" (http://www.gutenberg.org/ebooks/2166)." echo " The audio chunks to be decoded were taken from the audio book read" echo " by John Nicholson(http://librivox.org/king-solomons-mines-by-haggard/)" echo echo " NOTE: Using utterances from the book, on which the LM was estimated" echo " is considered to be "cheating" and we are doing this only for" echo " the purposes of the demo." echo echo " You can type "./run.sh --test-mode live" to try it using your" echo " own voice!" echo mkdir -p $decode_dir # make an input .scp file > $decode_dir/input.scp for f in $audio/*.wav; do bf=`basename $f` bf=${bf%.wav} echo $bf $f >> $decode_dir/input.scp done # online-wav-gmm-decode-faster --verbose=1 --rt-min=0.8 --rt-max=0.85 # --max-active=4000 --beam=12.0 --acoustic-scale=0.0769 # scp:$decode_dir/input.scp $ac_model/model $ac_model/HCLG.fst # $ac_model/words.txt '1:2:3:4:5' ark,t:$decode_dir/trans.txt # ark,t:$decode_dir/ali.txt $trans_matrix;; online-wav-gmm-decode-faster --verbose=1 --rt-min=0.8 --rt-max=0.85 --max-active=4000 --beam=12.0 --acoustic-scale=0.0769 --left-context=3 --right-context=3 scp:$decode_dir/input.scp $ac_model/final.mdl $ac_model/HCLG.fst $ac_model/words.txt '1:2:3:4:5' ark,t:$decode_dir/trans.txt ark,t:$decode_dir/ali.txt $trans_matrix;;
5、执行语音识别
将声音文件复制到 online-data/audio/ 目录，然后运行 run.sh 执行识别操作。
测试文本：
自然语言理解和生成是一个多方面问题，我们对它可能也只是部分理解。
识别效果如下：
mike@local:online_demo$ ls online-data/audio/ A1_00.wav mike@local:online_demo$ ./run.sh SIMULATED ONLINE DECODING - pre-recorded audio is used The (bigram) language model used to build the decoding graph was estimated on an audio book's text. The text in question is "King Solomon's Mines" (http://www.gutenberg.org/ebooks/2166). The audio chunks to be decoded were taken from the audio book read by John Nicholson(http://librivox.org/king-solomons-mines-by-haggard/) NOTE: Using utterances from the book, on which the LM was estimated is considered to be "cheating" and we are doing this only for the purposes of the demo. You can type "./run.sh --test-mode live" to try it using your own voice! online-wav-gmm-decode-faster --verbose=1 --rt-min=0.8 --rt-max=0.85 --max-active=4000 --beam=12.0 --acoustic-scale=0.0769 --left-context=3 --right-context=3 scp:./work/input.scp online-data/models/tri1/final.mdl online-data/models/tri1/HCLG.fst online-data/models/tri1/words.txt 1:2:3:4:5 ark,t:./work/trans.txt ark,t:./work/ali.txt File: A1_00 自然 语言 理解 和 生产 是 一个 多方面 挽 起 我们 对 它 可能 也 只是 部分 礼节 ./run.sh: 行 116: online-data/audio/trans.txt: 没有那个文件或目录 ./run.sh: 行 121: gawk: 未找到命令 compute-wer --mode=present ark,t:./work/ref.txt ark,t:./work/hyp.txt WARNING (compute-wer[5.5.421~1453-85d1a]:Open():util/kaldi-table-inl.h:513) Failed to open stream ./work/ref.txt ERROR (compute-wer[5.5.421~1453-85d1a]:SequentialTableReader():util/kaldi-table-inl.h:860) Error constructing TableReader: rspecifier is ark,t:./work/ref.txt [ Stack-Trace: ] /data/asr/kaldi_full/src/lib/libkaldi-base.so(kaldi::MessageLogger::LogMessage() const+0xb42) [0x7fce5ef61692] compute-wer(kaldi::MessageLogger::LogAndThrow::operator=(kaldi::MessageLogger const&)+0x21) [0x55bb3a782299] compute-wer(kaldi::SequentialTableReader<kaldi::TokenVectorHolder>::SequentialTableReader(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&)+0xc2) [0x55bb3a787c0c] compute-wer(main+0x226) [0x55bb3a780f60] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xe7) [0x7fce5e3cbb97] compute-wer(_start+0x2a) [0x55bb3a780c5a] kaldi::KaldiFatalError mike@local:online_demo$
可以看到，识别效果比较差。
本文中涉及训练数据及测试示例地址： https://pan.baidu.com/s/1OdLkcoDPl1Hkd06m2Xt7wA
可关注微信公众号后回复 19102201 获取提取码。
本文github地址：