Kaldi命令词识别

$ tree -d
├── speaker001
├── speaker002
├── speaker003
├── speaker004
├── speaker005
├── speaker006
├── speaker007
├── speaker008
├── speaker009
├── speaker010
└── Levoice.list

└┤ tree speaker001
speaker001
├── 00001.wav
├── 00002.wav
├── 00003.wav
...
└── 00200.wav

└┤ head -n 2 Levoice.list 
speaker001/00001.wav	三六零通讯录	5.6
speaker001/00002.wav	三六五日历	2.8

 !/bin/bash

#if need cv or not
needcv=true

# rename wav files by add prefix by "speaker"
start_path=`pwd`
for dirname in $(ls | grep "speaker")
do
    #get first filename
    filename=$(ls $dirname | head -n 1)
    if [[ $filename =~ "speaker" ]]; then
        echo "files in $dirname have already renamed, passing..."
    else
        echo "now rename flies with prefix speakers"
        echo $dirname
        cd $dirname
        #in centos rename
        rename "00" $dirname"_00" "00"*
        # ubuntu using follows
        #rename "s/00/$dirname""_00/" 00*
        cd ..
    fi
done

# devide file to train, cv and test
cd $start_path
rm -rf  test train cv  && mkdir test train cv

i=1
for dirname in $(ls | grep "speaker")
do
    if [ $i -lt 9 ];then
        cp $dirname/* train
    else
        cp $dirname/* test
    fi
    let i=$i+1

done

function rand(){
    min=$1
    max=$(($2-$min+1))
    num=$(($RANDOM+1000000000))
    echo $(($num%$max+$min))  
}

count=0
array=("0" "0" "0" "0")
#ls -al train
if [ needcv ]; then
    for file in $(ls train | grep "speak")
    do
        array[$count]=$file
        let count=$count+1
        if [ $count -eq 4 ];then
            rnd=$(rand 0 3)
            mv train/${array[$rnd]} cv
            #echo ${array[$rnd]}
            let count=0
        fi
    done
    echo "cv files prepared over, examples number is $(ls cv | wc -l)"
fi
echo "train files number is $(ls train | wc -l)"
echo "test files number is $(ls test | wc -l)"

speaker001/00001.wav	三六零通讯录	5.6
---->
speaker001_00001.wav	三六零通讯录	5.6

$ tree -L 1
|-- cmd.sh // 运行配置目录
|-- conf  // 配置文件目录
|-- local //存放run.sh 中调用的脚本工具，需要自己编写
|-- path.sh //Kaldi 工具和库目录添加到PATH
|-- run.sh // top层脚本，运行该脚本训练数据和测试， 需要自己编写
|-- steps // kaldi 脚本工具， 复制到工程目录下
|-- tools // kaldi 脚本工具， 复制到工程目录下
`-- utils // kaldi 脚本工具， 复制到工程目录下

export train_cmd=run.pl
export decode_cmd="run.pl --mem 4G"
export mkgraph_cmd="run.pl --mem 8G"

$ ls
decode_dnn.config  fbank.conf  mfcc.conf
$ more mfcc.conf 
--use-energy=false   # only non-default option.
--sample-frequency=8000
$ more decode_dnn.config 
beam=18.0 # beam for decoding.  Was 13.0 in the scripts.
lattice_beam=10.0 # this has most effect on size of the lattices.
$ more fbank.conf 
--sample-frequency=8000
--num-mel-bins=40

#!/bin/bash
# 2017-3-23 by zqh 

# This file prepares files needed in kaldi
# including text, wav.scp, utt2spk, spk2utt
# output: 
#   data/train dir include infomation of train data
#   data/test dir include infomation of test data
#   data/cv dir include infomation of cross validation data

run_dir=$1
dataset_dir=$2

cd $run_dir
echo "prepare data in data/{train, test, cv}"
mkdir -p data/{train,test,cv}

#create text, wav.scp, utt2spk, spk2utt
(
i=0
for dir in train cv test; do
    echo "clean dir data/$dir"
    cd $run_dir/data/$dir
    rm -rf wav.scp utt2spk spk2utt word.txt text  
    #phone.txt
    for data in $(find $dataset_dir/$dir/*.wav | sort -u | xargs -i basename {} .wav);do
        let i=$i+1
        spkid=$(echo $data | awk -F"_" '{print "" $1}')
        uttid=$data
        echo $uttid $dataset_dir/$dir/$data.wav >> wav.scp
        echo $uttid $spkid >> utt2spk
        # gen word.txt
        echo $uttid $(cat $dataset_dir/utt2word.txt | grep $uttid | awk '{print "" $2}') >> word.txt
        # gen phone.txt TODO
    done
    cp word.txt text
    sort wav.scp -o wav.scp
    sort utt2spk -o utt2spk
    sort text -o text
    # sort phone.txt -o phone.txt
done
echo "all file number is $i"
) || exit 1

utils/utt2spk_to_spk2utt.pl data/train/utt2spk > data/train/spk2utt
utils/utt2spk_to_spk2utt.pl data/cv/utt2spk > data/cv/spk2utt
utils/utt2spk_to_spk2utt.pl data/test/utt2spk > data/test/spk2utt

[username@hostname dict]$ pwd
/home/username/dataset_wakeup/resource/dict
[username@hostname dict]$ ls
extra_questions.txt  lexiconp.txt  lexicon.txt  nonsilence_phones.txt  optional_silence.txt  silence_phones.txt

$ more lexicon.txt 
SIL sil
<SPOKEN_NOISE> sil
三六零通讯录 s an1 l iu4 l ing2 t ong1 x vn4 l u4
三六五日历 s an1 l iu4 uu u3 r iz4 l i4
三D图库	s an1 d i4 t u2 k u4

ngram-count -order 1 -text words.txt -lm word.arpa

-order  指定n-gram的n是多少,默认是3
-text   提供输入的语料文件,统计该语料中的n-gram
-lm     指定输出的lm文件
-vocab  用来指定对哪些词进行n-gram统计
-wbdiscount1 表示1gram Witten-Bell discounting 
Note:参数顺序无所谓