Sing-Pass Retraining
1.A feature->B features MFCC_E_D->MFCC_E_D_Z
2.wide-band-->narrow-ban eg:16k->8k
命令行: HERest -r -C config -S trainList -I labs -H dir1/MMF -M dir2 hmmList
1. config
HFWDBKWD: PRUNEINIT = 250.0
HFWDBKWD: PRUNEINC = 250.0
HFWDBKWD: PRUNELIM = 1000.1
HFWDBKWD: MINFORPROB = 20.0
HFB: TRACE = 1
MAXTRYOPEN = 3
HPARM1: TARGETKIND = MFCC_E_D
HPARM2: TARGETKIND = MFCC_E_D_Z
HFWDBKWD: PRUNEINC = 250.0
HFWDBKWD: PRUNELIM = 1000.1
HFWDBKWD: MINFORPROB = 20.0
HFB: TRACE = 1
MAXTRYOPEN = 3
HPARM1: TARGETKIND = MFCC_E_D
HPARM2: TARGETKIND = MFCC_E_D_Z
2.config
HFWDBKWD: PRUNEINIT = 250.0
HFWDBKWD: PRUNEINC = 250.0
HFWDBKWD: PRUNELIM = 1000.1
HFWDBKWD: MINFORPROB = 20.0
HFB: TRACE = 1
MAXTRYOPEN = 3
HPARM1: TARGETKIND = MFCC_E_D
HPARM2: TARGETKIND = MFCC_E_D
HFWDBKWD: PRUNEINC = 250.0
HFWDBKWD: PRUNELIM = 1000.1
HFWDBKWD: MINFORPROB = 20.0
HFB: TRACE = 1
MAXTRYOPEN = 3
HPARM1: TARGETKIND = MFCC_E_D
HPARM2: TARGETKIND = MFCC_E_D
3. trainList:两列。第一列为原始模型的特征列表,第二列为需要更新模型的特征列表,一一对应,也就是说single-pass的两个模型的
训练数据相同.另外两个模型的
gauss数和state数也一致
4. 注意:HTEfile中末尾需添加set HESPR,这样命令行才会有-r(singal-pass 触发)
Two-Model Retraining
1. one model->another model
命令行: HERest -C config -S trainlist -I labs -H dir1/MMF -M dir3 hmmlist1
1. config
HPARM: TARGETKIND = MFCC_E_D_Z
HFWDBKWD: PRUNEINIT = 250.0
HFWDBKWD: PRUNEINC = 250.0
HFWDBKWD: PRUNELIM = 1000.1
HFWDBKWD: MINFORPROB = 20.0
HFB: TRACE = 1
MAXTRYOPEN = 3
ALIGNMODELMMF = model1/MMF
ALIGNHMMLIST = model1/tri.xwrd.cluster.list
HFWDBKWD: PRUNEINIT = 250.0
HFWDBKWD: PRUNEINC = 250.0
HFWDBKWD: PRUNELIM = 1000.1
HFWDBKWD: MINFORPROB = 20.0
HFB: TRACE = 1
MAXTRYOPEN = 3
ALIGNMODELMMF = model1/MMF
ALIGNHMMLIST = model1/tri.xwrd.cluster.list
其中:model1是已经训练好的较好的模型
dir1/MMF:需要更新的初始模型
dir3:更新后的模型
dir3/MMF与dir1/MMF gauss数和state保持一致
注意:做two-model之前需要去除dir1/MMF中的varFloor1对应项,更新后的模型在添上该项