VoxForge
Hi,
I have a large wav file (about 1 hour of speak), to train it, I have extract MFCC parameters then I used HCompV to the iinitialization. I would like to use HERest (10 times or iterations) to obtain the so called UBM. can I do it without phone.mlf file because I dont see how to use it in my case.
Thanks a lot
Hi Lyes,
>I would like to use HERest (10 times or iterations) to obtain the so
>called UBM. can I do it without phone.mlf file
Sorry, I am not familiar with UBM. You might try asking the HTK mailing list.
Ken
Hi ken,
In fact the problem of UBM is very simple. We have just to model all speakers with one set of GMM (one state of HMM). So my question is : I have one wav file and I'd like to train it. First, I had extract MFCC features then I used HCompV to initialize my model then I'm trying to obtain my HMM model by HERest as :
HERest -C config/config_train_mfcc -I label/label_EE4H_noSP.mlf -t 250.0 150.0 1000.0 -S list/listTrainFulPath.scp -H hmmsTrained/hmm0/macros -H hmmsTrained/hmm0/models -M hmmsTrained/hmm1 lib/wordList_noSP
where "listTrainFulPath.scp" contain "C:\htk-3.3\bd/train_f_1_1.mfcc"
my problem is what should I write in "label_EE4H_noSP.mlf" and "lib/wordList_noSP".
Can you help me in this case. Thanks
>my problem is what should I write in "label_EE4H_noSP.mlf" and "lib/wordList_noSP".
re: abel_EE4H_noSP.mlf - something analogous to: words.mlf?
re: lib/wordList_noSP - something analogous to: monophones0?
Ken