HTK下yes/no的识别

最新推荐文章于 2019-05-18 11:58:07 发布

雪糕蛋蛋

最新推荐文章于 2019-05-18 11:58:07 发布

阅读量1.3k

点赞数

分类专栏： HTK 语音识别文章标签：测试任务 function 工具网络 c

本文链接：https://blog.csdn.net/shuipi100/article/details/7801362

版权

HTK 同时被 2 个专栏收录

6 篇文章 0 订阅

订阅专栏

语音识别

4 篇文章 0 订阅

订阅专栏

孤立词识别的简单结构

data/: data/train data/test

analysis/:

training/:

model/:

def/:

test/:

第一步创建语音文件录音

命令：HSLab any_name.sig

yes/no识别

yes与no都录10次记为

yes01.sig yes02.sig…

no01.sig no02.sig…

第二步：声学分析

抽取yes和no的mfcc特征参数。保存在data/train/lab与data/train/sig下。

命令：HCopy -A -D -C analysis.conf -S targetlist.txt

其中 analysis.conf为抽取参数配置文件，

内容为：

# Example of an acoustical analysis configuration file

SOURCEFORMAT = HTK # Gives the format of the speech files

TARGETKIND = MFCC_0_D_A # Identifier of the coefficients to use

# Unit = 0.1 micro-second :

WINDOWSIZE = 250000.0 # = 25 ms = length of a time fram

TARGETRATE = 100000.0 # = 10 ms = frame perioficity

NUMCEPS = 12 # Number of MFCC coeffs (here from c1 to c2)

USEHAMMING = T # Use of Hamming function for windowing frames

PREEMCOEF = 0.97 # Pre-emphasis coefficient

NUMCHANS = 26 # Number of filterbank channels

CEPLIFTER = 22 # Length of cepstral liftering

# The End

targetlist.txt说明抽取源文件路径和目标文件的保存路径，

内容如下：

data/train/sig/yes01.sig data/train/mfcc/yes01.mfcc

data/train/sig/yes02.sig data/train/mfcc/yes02.mfcc

data/train/sig/yes03.sig data/train/mfcc/yes03.mfcc

data/train/sig/yes04.sig data/train/mfcc/yes04.mfcc

data/train/sig/yes05.sig data/train/mfcc/yes05.mfcc

data/train/sig/yes06.sig data/train/mfcc/yes06.mfcc

data/train/sig/yes07.sig data/train/mfcc/yes07.mfcc

data/train/sig/yes08.sig data/train/mfcc/yes08.mfcc

data/train/sig/yes09.sig data/train/mfcc/yes09.mfcc

data/train/sig/yes10.sig data/train/mfcc/yes10.mfcc

data/train/sig/no01.sig data/train/mfcc/no01.mfcc

data/train/sig/no02.sig data/train/mfcc/no02.mfcc

data/train/sig/no03.sig data/train/mfcc/no03.mfcc

data/train/sig/no04.sig data/train/mfcc/no04.mfcc

data/train/sig/no05.sig data/train/mfcc/no05.mfcc

data/train/sig/no06.sig data/train/mfcc/no06.mfcc

data/train/sig/no07.sig data/train/mfcc/no07.mfcc

data/train/sig/no08.sig data/train/mfcc/no08.mfcc

data/train/sig/no09.sig data/train/mfcc/no09.mfcc

data/train/sig/sil10.sig data/train/mfcc/sil10.mfcc

data/train/sig/sil01.sig data/train/mfcc/sil01.mfcc

data/train/sig/sil02.sig data/train/mfcc/sil02.mfcc

data/train/sig/sil03.sig data/train/mfcc/sil03.mfcc

data/train/sig/sil04.sig data/train/mfcc/sil04.mfcc

data/train/sig/sil05.sig data/train/mfcc/sil05.mfcc

data/train/sig/sil06.sig data/train/mfcc/sil06.mfcc

data/train/sig/sil07.sig data/train/mfcc/sil07.mfcc

data/train/sig/sil08.sig data/train/mfcc/sil08.mfcc

data/train/sig/sil09.sig data/train/mfcc/sil09.mfcc

data/train/sig/sil10.sig data/train/mfcc/sil10.mfcc

第三步：HMM原型定义

建立文件hmm_yes、hmm_no、hmm_sil保存在model/proto下。

hmm_yes内容如下：

~o <VecSize> 39 <MFCC_0_D_A>

~h "yes"

<NumStates> 6

<State> 2

<Mean> 39

0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

<Variance> 39

1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

<State> 3

<Mean> 39

0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

<Variance> 39

1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

<State> 4

<Mean> 39

0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

<Variance> 39

1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

<State> 5

<Mean> 39

0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

<Variance> 39

1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

<TransP> 6

0.0 0.5 0.5 0.0 0.0 0.0

0.0 0.4 0.3 0.3 0.0 0.0

0.0 0.0 0.4 0.3 0.3 0.0

0.0 0.0 0.0 0.4 0.3 0.3

0.0 0.0 0.0 0.0 0.5 0.5

0.0 0.0 0.0 0.0 0.0 0.0

hmm_no、hmm_sil有着相同的内容

第四步：HMM训练

1.初始化：

命令:

HInit -A -D -T 1 -S trainlist_yes.txt -M model/hmm0 -H model/proto/hmm_yes yes

trainlist_yes.txt内容为：

data/train/mfcc/yes01.mfcc

data/train/mfcc/yes02.mfcc

data/train/mfcc/yes03.mfcc

data/train/mfcc/yes04.mfcc

data/train/mfcc/yes05.mfcc

data/train/mfcc/yes06.mfcc

data/train/mfcc/yes07.mfcc

data/train/mfcc/yes08.mfcc

data/train/mfcc/yes09.mfcc

data/train/mfcc/yes10.mfcc

no、sil类似地初始化。

HInit -A -D -T 1 -S trainlist_no.txt -M model/hmm0 -H model/proto/hmm_no no

HInit -A -D -T 1 -S trainlist_sil.txt -M model/hmm0 -H model/proto/hmm_sil sil

2．训练

命令： HRest -A -D -T 1 -S trainlist_yes.txt -M model/proto -H model/hmm0/hmm_yes yes

no、sil类似地训练。

HRest -A -D -T 1 -S trainlist_no.txt -M model/proto -H model/hmm0/hmm_no no

HRest -A -D -T 1 -S trainlist_sil.txt -M model/proto -H model/hmm0/hmm_sil sil

第五步：任务定义

1．建立语法规则和字典

建立语法规则文件gram.txt（在YesNoSystem文件夹下），内容为：

* Task grammar

$WORD = YES | NO;

( { SIL } [ $WORD ] { SIL } )

建立字典文件dict.txt（在YesNoSystem文件夹下），内容为：

YES [yes] yes

NO [no] no

SIL [sil] sil

不要忘记在文件的末尾添加新行(如果不添加，那么最后一个入口将被忽略)。

2．建立任务网络

命令：HParse -A -D -T 1 gram.txt net.slf

HSGen -A -D -s net.slf dict.txt

第六步：识别

1．建立文件hmmlist.txt（在YesNoSystem文件夹下）文件，内容为：

yes

sil

注意每一列最后留一个空格。每个元素通过新行字符隔开，不要忘记在最后一个元素后面插入新行。

2．待识别信号并抽取mfcc参数（文件test.mfcc）

HCopy -A -D -C analysis.conf -S targetlist_test.txt

targetlist_test.txt说明抽取源文件路径和目标文件的保存路径，

内容如下：

data/test/sig/yestest01.sig data/test/mfcc/yestest01.mfcc

data/test/sig/yestest02.sig data/test/mfcc/yestest02.mfcc

data/test/sig/yestest03.sig data/test/mfcc/yestest03.mfcc

data/test/sig/notest01.sig data/test/mfcc/notest01.mfcc

data/test/sig/notest02.sig data/test/mfcc/notest02.mfcc

data/test/sig/notest03.sig data/test/mfcc/notest03.mfcc

data/test/sig/siltest01.sig data/test/mfcc/siltest01.mfcc

data/test/sig/siltest02.sig data/test/mfcc/siltest02.mfcc

data/test/sig/siltest03.sig data/test/mfcc/siltest03.mfcc

3．命令：

Hvite -H model/proto/hmm_yes -H model/proto/hmm_no -H model/proto/hmm_sil -i result/reco_yestest01.mlf -w net.slf dict.txt hmmlist.txt data/test/mfcc/yestest01.mfcc

Hvite -H model/proto/hmm_yes -H model/proto/hmm_no -H model/proto/hmm_sil -i result/reco_yestest02.mlf -w net.slf dict.txt hmmlist.txt data/test/mfcc/yestest02.mfcc

Hvite -H model/proto/hmm_yes -H model/proto/hmm_no -H model/proto/hmm_sil -i result/reco_yestest03.mlf -w net.slf dict.txt hmmlist.txt data/test/mfcc/yestest03.mfcc

Hvite -H model/proto/hmm_yes -H model/proto/hmm_no -H model/proto/hmm_sil -i result/reco_notest01.mlf -w net.slf dict.txt hmmlist.txt data/test/mfcc/notest01.mfcc

Hvite -H model/proto/hmm_yes -H model/proto/hmm_no -H model/proto/hmm_sil -i result/reco_notest02.mlf -w net.slf dict.txt hmmlist.txt data/test/mfcc/notest02.mfcc

Hvite -H model/proto/hmm_yes -H model/proto/hmm_no -H model/proto/hmm_sil -i result/reco_notest03.mlf -w net.slf dict.txt hmmlist.txt data/test/mfcc/notest03.mfcc

Hvite -H model/proto/hmm_yes -H model/proto/hmm_no -H model/proto/hmm_sil -i result/reco_siltest01.mlf -w net.slf dict.txt hmmlist.txt data/test/mfcc/siltest01.mfcc

Hvite -H model/proto/hmm_yes -H model/proto/hmm_no -H model/proto/hmm_sil -i result/reco_siltest02.mlf -w net.slf dict.txt hmmlist.txt data/test/mfcc/siltest02.mfcc

Hvite -H model/proto/hmm_yes -H model/proto/hmm_no -H model/proto/hmm_sil -i result/reco_siltest03.mlf -w net.slf dict.txt hmmlist.txt data/test/mfcc/siltest03.mfcc

识别结果自动存放在文件reco.mlf

#!MLF!#

"yes01.rec"

0 34300000 yes -17771.425781

性能测试

创建主标签文件：

手动创建ref.mlf，记作包含整个测试集的正确副本，即通过手工标注的副本。内容如下

#!MLF!#

"*/yestest01.lab"

1605625 3667500 sil

3667500 9288125 yes

9288125 11770000 sil