目录
kaldi中脚本东西比较多,一层嵌一层,不易阅读。
本文以yesno为例,直接使用kaldi编译的工具,书写简易训练步骤,方便学习kaldi工具的使用。 注意:转载请注明出处。
yesno训练
- 准备数据
- 在yesno/s5下新建文件夹:
mkdir easy
,后续的操作将在easy文件夹中执行。 - 拷贝s5下
./path
到easy
文件夹中,./path的作用是能直接调用工具,不用添加工具所在路径,类似于设置环境变量。 - 本脚主要便于理解kaldi工具的使用,一些批处理和数据下载并没有做,需要运行一遍
yesno/s5/./run.sh
生成训练所需输入。 - 到
s5/data/train
下拷贝wav.scp
到easy目录下作为训练输入,因为wav.scp
是相对路径也需要拷贝waves_yesno/
到easy下。 - 拷贝词典到目录下:拷贝
s5/input
到easy目录下。 准备数据结束,可以写自己的脚本了。
- 在yesno/s5下新建文件夹:
先给出整体脚本如下:
#!/bin/bash. ./path# feature extraction:# a series of light command [ compute-mfcc + copy-feats -> compute-cmvn-stats -> apply-cmvn -> add-deltas ] # the data flow transition [ wav -> mfcc.ark,scp -> cmvn.ark,scp -> delta.ark ]mkdir mfcccompute-mfcc-feats --verbose=2 --config="../conf/mfcc.conf" scp,p:wav.scp ark:- | copy-feats --compress=true ark:- ark,scp:mfcc/mfcc.ark,mfcc/mfcc.scpcompute-cmvn-stats scp:mfcc/mfcc.scp ark:mfcc/cmvn.arkapply-cmvn ark:mfcc/cmvn.ark scp:mfcc/mfcc.scp ark:- | add-deltas ark:- ark:mfcc/delta.ark# prepare dict for lang:# input data [ lexicon_nosil.txt lexicon.txt phones.txt ]# output data [ lexicon.txt lexicon_words.txt nonsilence_phones.txt optional_silence.txt silence_phones.txt ]mkdir -p lang/dictcp input/lexicon_nosil.txt lang/dict/lexicon_words.txtcp input/lexicon.txt lang/dict/lexicon.txtcat input/phones.txt | grep -v SIL > lang/dict/nonsilence_phones.txtecho "SIL" > lang/dict/silence_phones.txtecho "SIL" > lang/dict/optional_silence.txtecho "Dictionary preparation succeeded"# generate [ topo ] for acoustic modelutils/gen_topo.pl 3 5 2:3 1 > lang/lang/topo# from [lexicoin phone word] -> [L.fst word.txt] for [G.fst train.fst HCLG.fst]utils/prepare_lang.sh --position-dependent-phones false lang/dict "" lang/local lang/lang# train monophic acoustic model# 1.from [topo 39] -> 0.mdl treegmm-init-mono --train-feats=ark:mfcc/delta.ark lang/lang/topo 39 mono/0.mdl mono/tree# 2.from [L.fst 0.mdl tree word.txt text] -> train.fst# compile-train-graphs [options] compile-train-graphs mono/tree mono/0.mdl lang/lang/L.fst 'ark:sym2int.pl -f 2- lang/lang/words.txt text|' ark:lang/lang/graphs.fsts# 3.from [graphs.fst] equally align the train data -> [ euqal.ali ]# align-equal-compiled align-equal-compiled ark:lang/lang/graphs.fsts ark:mfcc/delta.ark ark:mono/equal.ali# 4.from [equal.ali delta.ark mdl] -> [ 0.acc ]gmm-acc-stats-ali mono/0.mdl ark:mfcc/delta.ark ark:mono/equal.ali mono/0.acc# 5.from [0.mdl 0.acc] -> [ 1.mdl ] # parameter est: gmm-est mono/0.mdl mono/0.acc mono/1.mdlx=1numliter=40numgauss=11while [ $x -lt $numliter ]; do # 6.from [1.mdl graphs.fst] align the data by new model -> [ 1.ali ] gmm-align-compiled --beam=6 --retry-beam=20 mono/$x.mdl ark:lang/lang/graphs.fsts ark:mfcc/delta.ark ark:mono/$x.ali # 4.from [equal.ali delta.ark mdl] -> [ 0.acc ] gmm-acc-stats-ali mono/$x.mdl ark:mfcc/delta.ark ark:mono/equal.ali mono/$x.acc # 5.from [x.mdl x.acc] -> [ x+1.mdl ] gmm-est --mix-up=$numgauss --power=0.25 mono/$x.mdl mono/$x.acc mono/$[$x+1].mdl numgauss=$[$numgauss+25] x=$[$x+1]donecp mono/$x.mdl mono/final.mdl# Graph compilation # from [input/task.arpabo word.txt] -> G.fstarpa2fst --disambig-symbol=#0 --read-symbol-table=lang/lang/words.txt input/task.arpabo lang/lang/G.fstfstisstochastic lang/lang/G.fst# from [final.mdl G.fst L.fst tree] -> HLCG.fstutils/mkgraph.sh lang/lang mono mono/graph
分块详解
首先进行特征提取:
#!/bin/bash. ./path# 特征提取: compute-mfcc-feats, copy-feats# 输入为:wav.scp 输出为:mfcc.ark,mfcc.scpcompute-mfcc-feats --verbose=2 --config="../conf/mfcc.conf" scp,p:wav.scp ark:- | copy-feats --compress=true ark:- ark,scp:mfcc/mfcc.ark,mfcc/mfcc.scp# 计算均方归一化矩阵:# 输入为:mfcc.ark,mfcc.scp 输出为:mfcc/cmvn.ark,mfcc/cmvn.scpcompute-cmvn-stats scp:mfcc/mfcc.scp ark,scp:mfcc/cmvn.ark,mfcc/cmvn.scp# 计算一阶二阶差分:# 输入为:mfcc/cmvn.ark,mfcc/cmvn.scp 输出为:delta.arkapply-cmvn scp:mfcc/cmvn.scp scp:mfcc/mfcc.scp ark:- | add-deltas ark:- ark:mfcc/delta.ark
然后,准备训练所需的词典,音素文件,词文件等。
yesno里准备好了,直接拷贝即可。
# prepare dict for lang:# input data [ lexicon_nosil.txt lexicon.txt phones.txt ]# output data [ lexicon.txt lexicon_words.txt nonsilence_phones.txt optional_silence.txt silence_phones.txt ]mkdir -p lang/dictcp input/lexicon_nosil.txt lang/dict/lexicon_words.txtcp input/lexicon.txt lang/dict/lexicon.txtcat input/phones.txt | grep -v SIL > lang/dict/nonsilence_phones.txtecho "SIL" > lang/dict/silence_phones.txtecho "SIL" > lang/dict/optional_silence.txtecho "Dictionary preparation succeeded"
生成声学拓扑结构。
生成 L.fst
word.txt
用来生成G.fst
train.fst
HCLG.fst
。其中utils/prepare_lang.sh
所需全部输入为上一步生成的dict
文件。
# generate [ topo ] for acoustic modelutils/gen_topo.pl 3 5 2:3 1 > lang/lang/topo# from [lexicoin phone word] -> [L.fst word.txt] for [G.fst train.fst HCLG.fst]utils/prepare_lang.sh --position-dependent-phones false lang/dict "" lang/local lang/lang
训练单音素模型
- 流程如下 :
- 利用生成的声学拓扑初始化模型
- 生成训练图
- 初始化对齐
- 生成统计量
- 模型参数估计
- {重新对齐生,成统计量,模型参数估计}x10
- 生成并导出最终模型:
# train monophic acoustic model# 1.from [topo 39] -> 0.mdl treegmm-init-mono --train-feats=ark:mfcc/delta.ark lang/lang/topo 39 mono/0.mdl mono/tree# 2.from [L.fst 0.mdl tree word.txt text] -> train.fst# compile-train-graphs [options]compile-train-graphs mono/tree mono/0.mdl lang/lang/L.fst 'ark:sym2int.pl -f 2- lang/lang/words.txt text|' ark:lang/lang/graphs.fsts# 3.from [graphs.fst] equally align the train data -> [ euqal.ali ]# align-equal-compiled align-equal-compiled ark:lang/lang/graphs.fsts ark:mfcc/delta.ark ark:mono/equal.ali# 4.from [equal.ali delta.ark mdl] -> [ 0.acc ]gmm-acc-stats-ali mono/0.mdl ark:mfcc/delta.ark ark:mono/equal.ali mono/0.acc# 5.from [0.mdl 0.acc] -> [ 1.mdl ] # parameter est: gmm-est mono/0.mdl mono/0.acc mono/1.mdlx=1numliter=40numgauss=11while [ $x -lt $numliter ]; do # 6.from [1.mdl graphs.fst] align the data by new model -> [ 1.ali ] gmm-align-compiled --beam=6 --retry-beam=20 mono/$x.mdl ark:lang/lang/graphs.fsts ark:mfcc/delta.ark ark:mono/$x.ali # 4.from [equal.ali delta.ark mdl] -> [ 0.acc ] gmm-acc-stats-ali mono/$x.mdl ark:mfcc/delta.ark ark:mono/equal.ali mono/$x.acc # 5.from [x.mdl x.acc] -> [ x+1.mdl ] gmm-est --mix-up=$numgauss --power=0.25 mono/$x.mdl mono/$x.acc mono/$[$x+1].mdl numgauss=$[$numgauss+25] x=$[$x+1]donecp mono/$x.mdl mono/final.mdl
最后合成语言模型:
# Graph compilation # from [input/task.arpabo word.txt] -> G.fstarpa2fst --disambig-symbol=#0 --read-symbol-table=lang/lang/words.txt input/task.arpabo lang/lang/G.fstfstisstochastic lang/lang/G.fst# from [final.mdl G.fst L.fst tree] -> HLCG.fstutils/mkgraph.sh lang/lang mono mono/graph
运行结果:
建立解码脚本
解码指令较简单一个指令即可:
#Usage: gmm-latgen-faster [options] model-in (fst-in|fsts-rspecifier) features-rspecifier lattice-wspecifier [ words-wspecifier [alignments-wspecifier] ]gmm-latgen-faster --max-active=7000 --beam=13 --lattice-beam=6 --acoustic-scale=0.083333 \--allow-partial=true --word-symbol-table=lang/lang/words.txt mono/final.mdl \mono/graph/HCLG.fst ark:mfcc/delta.ark "ark:|gzip -c > result/lat.gz"
可以得到识别结果不是很好,没关系,主要用这个例子来理解kaldi是怎么样使用工具的。
转载请注明出处: