timit实例训练

TIMIT S5实例：

首先，将TIMIT.ISO中的TIMIT复制到主文件夹。

1.进入对应的目录，进行如下操作：

zhangju@ubuntu :~$ cd kaldi-trunk/egs/timit/s5/

zhangju@ubuntu :~/kaldi-trunk/egs/timit/s5$

sudo local/timit_data_prep.sh /home/zhangju/TIMIT

会看到如下显示：

Creating coretest set.

MDAB0  MWBT0  FELC0  MTAS1  MWEW0  FPAS0  MJMP0  MLNT0  FPKT0  MLLL0  MTLS0  FJLM0  MBPM0  MKLT0  FNLP0  MCMJ0  MJDH0  FMGD0  MGRT0  MNJM0  FDHC0  MJLN0  MPAM0  FMLD0

# of utterances in coretest set = 192

Creating dev set.

FAKS0  FDAC1  FJEM0  MGWT0  MJAR0  MMDB1  MMDM2  MPDF0  FCMH0  FKMS0  MBDG0  MBWM0  MCSH0  FADG0  FDMS0  FEDW0  MGJF0  MGLB0  MRTK0  MTAA0  MTDT0  MTHC0  MWJG0  FNMR0  FREW0  FSEM0  MBNS0  MMJR0  MDLS0  MDLF0  MDVC0  MERS0  FMAH0  FDRW0  MRCS0  MRJM4  FCAL1  MMWH0  FJSJ0  MAJC0  MJSW0  MREB0  FGJD0  FJMG0  MROA0  MTEB0  MJFC0  MRJR0  FMML0  MRWS1

# of utterances in dev set = 400

Finalizing test

Finalizing dev

timit_data_prep succeeded.

于是在/home/zhangju/kaldi-trunk/egs/timit/s5文件夹下新生成data文件夹，其内包含local文件夹以及相关内容。

2 在终端输入：

local/timit_train_lms.sh data/local（下载、计算文本，用以建立语言模型）

local/timit_format_data.sh（处理与fst有关的东西）

3创建train的mfcc：

sudo steps/make_mfcc.sh data/train exp/make_mfcc/train mfccs 4

（要对train，dev，test创建）

会看到：

Succeeded creating MFCC features for train

sudo steps/make_mfcc.sh data/test exp/make_mfcc/test mfccs 4

会看到：

Succeeded creating MFCC features for test

sudo steps/make_mfcc.sh data/dev exp/make_mfcc/dev mfccs 4

会看到：

Succeeded creating MFCC features for dev

4训练单音素系统（monophone systom）

sudo steps/train_mono.sh data/train data/lang exp/mono

会显示：

Computing cepstral mean and variance statistics

Initializing monophone system.

Compiling training graphs

Pass 0

Pass 1

Aligning data

Pass 2

Aligning data

Pass 3

Aligning data

Pass 4

Aligning data

Pass 5

Aligning data

Pass 6

Aligning data

Pass 7

Aligning data

Pass 8

Aligning data

Pass 9

Aligning data

Pass 10

Aligning data

Pass 11

Pass 12

Aligning data

Pass 13

Pass 14

Pass 15

Aligning data

Pass 16

Pass 17

Pass 18

Pass 19

Pass 20

Aligning data

Pass 21

Pass 22

Pass 23

Pass 24

Pass 25

Aligning data

Pass 26

Pass 27

Pass 28

Pass 29

于是，新建了exp/mono文件夹

scripts/mkgraph.sh --mono data/lang exp/mono exp/mono/graph（制图）

会显示：

fsttablecompose data/lang/L.fst data/lang/G.fst

fstdeterminizestar --use-log=true

fstminimizeencoded

fstisstochastic data/lang/tmp/LG.fst

-0.000244359 -0.0912761

warning: LG not stochastic.

fstcomposecontext --context-size=1 --central-position=0 --read-disambig-syms=data/lang/tmp/disambig_phones.list --write-disambig-syms=data/lang/tmp/disambig_ilabels_1_0.list data/lang/tmp/ilabels_1_0

fstisstochastic data/lang/tmp/CLG_1_0.fst

-0.000244359 -0.0912761

warning: CLG not stochastic.

make-h-transducer --disambig-syms-out=exp/mono/graph/disambig_tid.list --transition-scale=1.0 data/lang/tmp/ilabels_1_0 exp/mono/tree exp/mono/final.mdl

fstminimizeencoded

fstdeterminizestar --use-log=true

fsttablecompose exp/mono/graph/Ha.fst data/lang/tmp/CLG_1_0.fst

fstrmsymbols exp/mono/graph/disambig_tid.list

fstrmepslocal

fstisstochastic exp/mono/graph/HCLGa.fst

0.000331581 -0.091291

HCLGa is not stochastic

add-self-loops --self-loop-scale=0.1 --reorder=true exp/mono/final.mdl

5.for test in dev test ; do

  steps/decode_deltas.sh exp/mono data/$test data/lang exp/mono/decode_$test &

done（解码test数据集（test是*/s5/data中dev、test文件夹中的test文件夹））

终端输出结果是：[1] 2307

                      [2] 2308

6.scripts/average_wer.sh exp/mono/decode_*/wer > exp/mono/wer

会显示：

[1]-  完成                  steps/decode_deltas.sh exp/mono data/$test data/lang exp/mono/decode_$test

[2]+  完成                  steps/decode_deltas.sh exp/mono data/$test data/lang exp/mono/decode_$test

7从单音素系统中获得alignments：（分别从mono文件夹中的train,dev,test中获得）（用以训练其他系统）

steps/align_deltas.sh data/train data/lang exp/mono exp/mono_ali_train

会显示：

Computing cepstral mean and variance statistics

Aligning all training data

Done.

方法二：修改run.sh中的timit路径，但后直接运行run.sh

TIMIT S3实例

1 数据准备，输入：

local/timit_data_prep.sh  /home/zhangju/TIMIT

终端显示：

Creating coretest set.

MDAB0  MWBT0  FELC0  MTAS1  MWEW0  FPAS0  MJMP0  MLNT0  FPKT0  MLLL0  MTLS0  FJLM0  MBPM0  MKLT0  FNLP0  MCMJ0  MJDH0  FMGD0  MGRT0  MNJM0  FDHC0  MJLN0  MPAM0  FMLD0  （这是说话人的名字，前面加M，F分别表示男性和女性）

# of utterances in coretest set = 192 （核心测试集中有192句话）

Creating dev set.

FAKS0  FDAC1  FJEM0  MGWT0  MJAR0  MMDB1  MMDM2  MPDF0  FCMH0  FKMS0  MBDG0  MBWM0  MCSH0  FADG0  FDMS0  FEDW0  MGJF0  MGLB0  MRTK0  MTAA0  MTDT0  MTHC0  MWJG0  FNMR0  FREW0  FSEM0  MBNS0  MMJR0  MDLS0  MDLF0  MDVC0  MERS0  FMAH0  FDRW0  MRCS0  MRJM4  FCAL1  MMWH0  FJSJ0  MAJC0  MJSW0  MREB0  FGJD0  FJMG0  MROA0  MTEB0  MJFC0  MRJR0  FMML0  MRWS1

# of utterances in dev set = 400 （设备集中有400句话）

Finalizing test （完成test）

Finalizing dev （完成dev）

timit_data_prep succeeded.

输入：

local/timit_train_lms.sh data/local

终端显示为

Not installing the kaldi_lm toolkit since it is already there.

（kaldi_lm工具箱里有：

compute_perplexity计算复杂度（用于对语言模型作评估，复杂度越低越好）

discount_ngrams给n阶语法模型作平滑处理（留出频率给实际会出现的但ngram中没出现的词语组合）

get_raw_ngrams（得到原始n阶语法模型）

get_word_map.pl*（得到词语的映射表）

interpolate_ngrams（补充（修改）n阶语法模型）

finalize_arpa.pl（完成arpa（arpa是一种格式，协议），是interpolate_ngrams程序中调用的）

map_words_in_arpa.pl（得到arpa格式的词语）

merge_ngrams（合并、融合n阶语法模型）

merge_ngrams_online（在线合并、融合n阶语法模型）

optimize_alpha.pl（使alpha最优化）

prune_lm.sh（删去出现频率较低的数据）

prune_ngrams（删去出现频率较低的数据）

scale_configs.pl

train_lm.sh（训练语言模型）

uniq_to_ngrams

）

Creating phones file, and monophone lexicon (mapping phones to itself). （创建音子文件及单音素词典）

Creating biphone model（创建双音子模型）

Training biphone language model in folder data/local/lm （训练双音子语言模型）

Creating directory data/local/lm/biphone (创建目录data/local/lm/biphone )

Getting raw N-gram counts （）

Iteration 1/7 of optimizing discounting parameters

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.900000 phi=2.000000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.900000 phi=2.000000

discount_ngrams: for n-gram order 3, D=0.800000, tau=1.100000 phi=2.000000

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.675000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.675000 phi=2.000000

discount_ngrams: for n-gram order 3, D=0.800000, tau=0.825000 phi=2.000000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=1.215000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=1.215000 phi=2.000000

discount_ngrams: for n-gram order 3, D=0.800000, tau=1.485000 phi=2.000000

interpolate_ngrams: 60 words in wordslist

Perplexity over 11412.000000 words is 17.013357

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.460842

real  0m0.021s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.016472

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.464985

real  0m0.020s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.021475

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.471402

real  0m0.025s

user  0m0.012s

sys 0m0.000s

optimize_alpha.pl: alpha=-2.1628504673 is too negative, limiting it to -0.5

Projected perplexity change from setting alpha=-0.5 is 17.016472->17.0106241428571, reduction of 0.00584785714286085

Alpha value on iter 1 is -0.5

Iteration 2/7 of optimizing discounting parameters

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=0.600000, tau=0.550000 phi=2.000000

interpolate_ngrams: 60 words in wordslist

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=0.800000, tau=0.550000 phi=2.000000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.080000, tau=0.550000 phi=2.000000

Perplexity over 11412.000000 words is 17.011355

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.457880

real  0m0.018s

user  0m0.004s

sys 0m0.008s

Perplexity over 11412.000000 words is 17.011355

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.457880

real  0m0.022s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.011355

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.457880

real  0m0.019s

user  0m0.008s

sys 0m0.004s

optimize_alpha.pl: objective function is not convex; returning alpha=0.7

Projected perplexity change from setting alpha=0.7 is 17.011355->17.011355, reduction of 0

Alpha value on iter 2 is 0.7

Iteration 3/7 of optimizing discounting parameters

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.412500 phi=2.000000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.550000 phi=2.000000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.742500 phi=2.000000

interpolate_ngrams: 60 words in wordslist

Perplexity over 11412.000000 words is 17.011355

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.457880

real  0m0.020s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.011355

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.457880

real  0m0.019s

user  0m0.008s

sys 0m0.004s

Perplexity over 11412.000000 words is 17.011355

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.457880

real  0m0.021s

user  0m0.012s

sys 0m0.000s

optimize_alpha.pl: objective function is not convex; returning alpha=0.7

Projected perplexity change from setting alpha=0.7 is 17.011355->17.011355, reduction of 0

Alpha value on iter 3 is 0.7

Iteration 4/7 of optimizing discounting parameters

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=1.750000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.000000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.350000

interpolate_ngrams: 60 words in wordslist

Perplexity over 11412.000000 words is 17.011355

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.457880

real  0m0.018s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.011355

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.457880

real  0m0.018s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.011355

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.457880

real  0m0.023s

user  0m0.012s

sys 0m0.000s

optimize_alpha.pl: objective function is not convex; returning alpha=0.7

Projected perplexity change from setting alpha=0.7 is 17.011355->17.011355, reduction of 0

Alpha value on iter 4 is 0.7

Iteration 5/7 of optimizing discounting parameters

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.450000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.700000

interpolate_ngrams: 60 words in wordslist

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.600000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.700000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.810000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.700000

Perplexity over 11412.000000 words is 17.008195

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.454326

real  0m0.019s

user  0m0.008s

sys 0m0.004s

Perplexity over 11412.000000 words is 17.011355

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.457880

real  0m0.019s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.018212

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.465417

real  0m0.021s

user  0m0.012s

sys 0m0.000s

optimize_alpha.pl: alpha=-0.670499383475985 is too negative, limiting it to -0.5

Projected perplexity change from setting alpha=-0.5 is 17.011355->17.0064832142857, reduction of 0.00487178571427904

Alpha value on iter 5 is -0.5

Iteration 6/7 of optimizing discounting parameters

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.300000, tau=0.337500 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.700000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 2, D=0.300000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.700000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.300000, tau=0.607500 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.700000

Perplexity over 11412.000000 words is 17.008198

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.454134

real  0m0.019s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.006972

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.452861

real  0m0.020s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.006526

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.452349

real  0m0.022s

user  0m0.012s

sys 0m0.000s

Projected perplexity change from setting alpha=0.280321158690507 is 17.006972->17.0064966287094, reduction of 0.000475371290633575

Alpha value on iter 6 is 0.280321158690507

Iteration 7/7 of optimizing discounting parameters

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.300000, tau=0.576145 phi=1.750000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.700000

interpolate_ngrams: 60 words in wordslist

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.300000, tau=0.576145 phi=2.350000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.700000

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.300000, tau=0.576145 phi=2.000000

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.700000

interpolate_ngrams: 60 words in wordslist

interpolate_ngrams: 60 words in wordslist

Perplexity over 11412.000000 words is 17.006845

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.452750

real  0m0.019s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.006575

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.452414

real  0m0.021s

user  0m0.012s

sys 0m0.000s

Perplexity over 11412.000000 words is 17.006336

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.452127

real  0m0.022s

user  0m0.012s

sys 0m0.000s

Projected perplexity change from setting alpha=0.690827338145686 is 17.006575->17.0062591109755, reduction of 0.000315889024498972

Alpha value on iter 7 is 0.690827338145686

Final config is:

D=0.4 tau=0.45 phi=2.0

D=0.3 tau=0.576144521410728 phi=2.69082733814569

D=1.36 tau=0.935 phi=2.7

Discounting N-grams.

discount_ngrams: for n-gram order 1, D=0.400000, tau=0.450000 phi=2.000000

discount_ngrams: for n-gram order 2, D=0.300000, tau=0.576145 phi=2.690827

discount_ngrams: for n-gram order 3, D=1.360000, tau=0.935000 phi=2.700000

Computing final perplexity

Building ARPA LM (perplexity computation is in background)

interpolate_ngrams: 60 words in wordslist

interpolate_ngrams: 60 words in wordslist

Perplexity over 11412.000000 words is 17.006029

Perplexity over 10833.000000 words (excluding 579.000000 OOVs) is 17.451754

17.006029

输入

local/timit_format_data.sh

终端显示：

Creating L.fst

Done creating L.fst

Creating L_disambig.fst

Done creating L_disambig.fst

Creating G.fst

arpa2fst -

\data\

Processing 1-grams

Processing 2-grams

Connected 0 states without outgoing arcs.

remove_oovs.pl: removed 0 lines.

G.fst created. How stochastic is it ?

fstisstochastic data/lang_test/G.fst

0 -0.0900995

fsttablecompose data/lang_test/L_disambig.fst data/lang_test/G.fst

How stochastic is LG.fst.

fstisstochastic data/lang_test/G.fst

0 -0.0900995

fstisstochastic

fsttablecompose data/lang/L.fst data/lang_test/G.fst

0 -0.0900994

How stochastic is LG_disambig.fst.

fsttablecompose data/lang_test/L_disambig.fst data/lang_test/G.fst

fstisstochastic

0 -0.0900994

First few lines of lexicon FST:

0  1      0.356674939

0  1  sil  1.20397282

1  2  aa  AA  1.20397282

1  1  aa  AA  0.356674939

1  1  ae  AE  0.356674939

1  2  ae  AE  1.20397282

1  1  ah  AH  0.356674939

1  2  ah  AH  1.20397282

1  1  ao  AO  0.356674939

1  2  ao  AO  1.20397282

timit_format_data succeeded.

输入：mfccdir=mfccs

for test in train test dev ; do

>  steps/make_mfcc.sh data/$test exp/make_mfcc/$test $mfccdir 4

> done

终端显示：

Succeeded creating MFCC features for train

Succeeded creating MFCC features for test

Succeeded creating MFCC features for dev

2 训练单音素系统，终端输入：

steps/train_mono.sh data/train data/lang exp/mono

终端显示：

Computing cepstral mean and variance statistics

Initializing monophone system.

Compiling training graphs

Pass 0

Pass 1

Aligning data

Pass 2

Aligning data

Pass 3

Aligning data

Pass 4

Aligning data

Pass 5

Aligning data

Pass 6

Aligning data

Pass 7

Aligning data

Pass 8

Aligning data

Pass 9

Aligning data

Pass 10

Aligning data

Pass 11

Pass 12

Aligning data

Pass 13

Pass 14

Pass 15

Aligning data

Pass 16

Pass 17

Pass 18

Pass 19

Pass 20

Aligning data

Pass 21

Pass 22

Pass 23

Pass 24

Pass 25

Aligning data

Pass 26

Pass 27

Pass 28

Pass 29

scripts/mkgraph.sh --mono data/lang_test exp/mono exp/mono/graph（制图）

终端显示：

fsttablecompose data/lang_test/L_disambig.fst data/lang_test/G.fst

fstminimizeencoded

fstdeterminizestar --use-log=true

fstisstochastic data/lang_test/tmp/LG.fst

0 -0.0901494

warning: LG not stochastic.

fstcomposecontext --context-size=1 --central-position=0 --read-disambig-syms=data/lang_test/tmp/disambig_phones.list --write-disambig-syms=data/lang_test/tmp/disambig_ilabels_1_0.list data/lang_test/tmp/ilabels_1_0

fstisstochastic data/lang_test/tmp/CLG_1_0.fst

0 -0.0901494

warning: CLG not stochastic.

make-h-transducer --disambig-syms-out=exp/mono/graph/disambig_tid.list --transition-scale=1.0 data/lang_test/tmp/ilabels_1_0 exp/mono/tree exp/mono/final.mdl

fsttablecompose exp/mono/graph/Ha.fst data/lang_test/tmp/CLG_1_0.fst

fstdeterminizestar --use-log=true

fstminimizeencoded

fstrmsymbols exp/mono/graph/disambig_tid.list

fstrmepslocal

fstisstochastic exp/mono/graph/HCLGa.fst

0 -0.0901494

HCLGa is not stochastic

add-self-loops --self-loop-scale=0.1 --reorder=true exp/mono/final.mdl

3 解码测试的数据集，输入

for test in dev test ; do

  steps/decode_deltas.sh exp/mono data/$test data/lang exp/mono/decode_$test &

done

终端显示：

[1] 16368

[2] 16369

3.1计算结果，输入：

scripts/average_wer.sh exp/mono/decode_*/wer > exp/mono/wer

终端显示：

[1]-  完成                  steps/decode_deltas.sh exp/mono data/$test data/lang exp/mono/decode_$test

[2]+  完成                  steps/decode_deltas.sh exp/mono data/$test data/lang exp/mono/decode_$test

4 从单音素系统中获得排列

创建排列用以训练其他系统，如ANN-HMM。

输入：

steps/align_deltas.sh data/train data/lang exp/mono exp/mono_ali_train

终端显示：

Computing cepstral mean and variance statistics

Aligning all training data

Done.

steps/align_deltas.sh data/dev data/lang exp/mono exp/mono_ali_dev

方法二：修改相应的TIMIT路径之后，直接运行run.sh

TIMIT S4实例此脚本是用于构建一个音位识别器

WORKDIR=/home/zhangju/ss4（自己找个有空间的路径作为WORKDIR）

mkdir -p $WORKDIR

cp -r conf local utils steps path.sh $WORKDIR

cd $WORKDIR

. path.sh（此文件中的环境变量KALDIROOT要自己修改路径，改到自己裝的kaldi文件中。KALDIROOT=/home/mayuan/kaldi-trunk（我用nano改的。））

local/timit_data_prep.sh --config-dir=$PWD/conf --corpus-dir=/home/zhangju/TIMIT --work-dir=$WORKDIR

最后编辑于 ：
©著作权归作者所有,转载或内容合作请联系作者
  • 人面猴
    序言：七十年代末，一起剥皮案震惊了整个滨河市，随后出现的几起案子，更是在滨河造成了极大的恐慌，老刑警刘岩，带你破解...
    沈念sama阅读 199,175评论 5赞 466
  • 死咒
    序言：滨河连续发生了三起死亡事件，死亡现场离奇诡异，居然都是意外死亡，警方通过查阅死者的电脑和手机，发现死者居然都...
    沈念sama阅读 83,674评论 2赞 376
  • 救了他两次的神仙让他今天三更去死
    文/潘晓璐 我一进店门，熙熙楼的掌柜王于贵愁眉苦脸地迎上来，“玉大人，你说我怎么就摊上这事。” “怎么了？”我有些...
    开封第一讲书人阅读 146,151评论 0赞 328
  • 道士缉凶录：失踪的卖姜人
    文/不坏的土叔 我叫张陵，是天一观的道长。 经常有香客问我，道长，这世上最难降的妖魔是什么？ 我笑而不...
    开封第一讲书人阅读 53,597评论 1赞 269
  • 港岛之恋（遗憾婚礼）
    正文 为了忘掉前任，我火速办了婚礼，结果婚礼上，老公的妹妹穿的比我还像新娘。我一直安慰自己，他们只是感情好，可当我...
    茶点故事阅读 62,505评论 5赞 359
  • 恶毒庶女顶嫁案：这布局不是一般人想出来的
    文/花漫 我一把揭开白布。 她就那样静静地躺着，像睡着了一般。 火红的嫁衣衬着肌肤如雪。 梳的纹丝不乱的头发上，一...
    开封第一讲书人阅读 47,969评论 1赞 275
  • 城市分裂传说
    那天，我揣着相机与录音，去河边找鬼。 笑死，一个胖子当着我的面吹牛，可吹牛的内容都是我干的。 我是一名探鬼主播，决...
    沈念sama阅读 37,455评论 3赞 390
  • 双鸳鸯连环套：你想象不到人心有多黑
    文/苍兰香墨 我猛地睁开眼，长吁一口气：“原来是场噩梦啊……” “哼！你这毒妇竟也来了？” 一声冷哼从身侧响起，我...
    开封第一讲书人阅读 36,118评论 0赞 254
  • 万荣杀人案实录
    序言：老挝万荣一对情侣失踪，失踪者是张志新（化名）和其女友刘颖，没想到半个月后，有当地人在树林里发现了一具尸体，经...
    沈念sama阅读 40,227评论 1赞 294
  • 护林员之死
    正文 独居荒郊野岭守林人离奇死亡，尸身上长有42处带血的脓包…… 初始之章·张勋 以下内容为张勋视角 年9月15日...
    茶点故事阅读 35,213评论 2赞 317
  • 白月光启示录
    正文 我和宋清朗相恋三年，在试婚纱的时候发现自己被绿了。 大学时的朋友给我发了我未婚夫和他白月光在一起吃饭的照片。...
    茶点故事阅读 37,214评论 1赞 328
  • 活死人
    序言：一个原本活蹦乱跳的男人离奇死亡，死状恐怖，灵堂内的尸体忽然破棺而出，到底是诈尸还是另有隐情，我是刑警宁泽，带...
    沈念sama阅读 32,928评论 3赞 316
  • 日本核电站爆炸内幕
    正文 年R本政府宣布，位于F岛的核电站，受9级特大地震影响，放射性物质发生泄漏。R本人自食恶果不足惜，却给世界环境...
    茶点故事阅读 38,512评论 3赞 302
  • 男人毒药：我在死后第九天来索命
    文/蒙蒙 一、第九天 我趴在偏房一处隐蔽的房顶上张望。 院中可真热闹，春花似锦、人声如沸。这庄子的主人今日做“春日...
    开封第一讲书人阅读 29,616评论 0赞 19
  • 一桩弑父案，背后竟有这般阴谋
    文/苍兰香墨 我抬头看了看天上的太阳。三九已至，却和暖如春，着一层夹袄步出监牢的瞬间，已是汗流浃背。 一阵脚步声响...
    开封第一讲书人阅读 30,848评论 1赞 255
  • 情欲美人皮
    我被黑心中介骗来泰国打工， 没想到刚下飞机就差点儿被人妖公主榨干…… 1. 我叫王不留，地道东北人。 一个月前我还...
    沈念sama阅读 42,228评论 2赞 344
  • 代替公主和亲
    正文 我出身青楼，却偏偏与公主长得像，于是被迫代替她去往敌国和亲。 传闻我的和亲对象是个残疾皇子，可洞房花烛夜当晚...
    茶点故事阅读 41,772评论 2赞 339