Hello, I am going to run DNN in Kaldi.
In the script,
egs/rm/s5/local/nnet/run_cnn.sh
at line 33:
//Split the training set
utils/subset_data_dir_tr_cv.sh $dir ${dir}_tr90 ${dir}_cv10
By default the script will use --cv-utt-percent option(for 10%)
But I couldn't split the file utt2spk to tr90 and cv10, and I try it for a long time....
In the file utt2spk, I set:
utt1 global
utt2 global
.
.
etc.
Is anything wrong?
Thank you very much.
Last edit: spy 2015-07-21
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Possibly it is trying to do a split where validation-set speakers are
distinct from training-set speakers.
In any case, you should never set up your utt2spk that way, and in the
data-prep docs I warn against doing so. Instead just make each speaker-id
identical to its utterance-id.
Hello, I am going to run DNN in Kaldi.
In the script,
egs/rm/s5/local/nnet/run_cnn.sh
at line 33:
//Split the training set
utils/subset_data_dir_tr_cv.sh --cv-spk-percent 10 $dir ${dir}_tr90
${dir}_cv10
By default the script will use --cv-utt-percent option(for 10%)
But I couldn't split the file utt2spk to tr90 and cv10, and I try it for a
long time....
In the file utt2spk, I set:
utt1 global
utt2 global
.
.
etc.
Hello, I am going to run DNN in Kaldi.
In the script,
egs/rm/s5/local/nnet/run_cnn.sh
at line 33:
//Split the training set
utils/subset_data_dir_tr_cv.sh $dir ${dir}_tr90 ${dir}_cv10
By default the script will use --cv-utt-percent option(for 10%)
But I couldn't split the file utt2spk to tr90 and cv10, and I try it for a long time....
In the file utt2spk, I set:
utt1 global
utt2 global
.
.
etc.
Is anything wrong?
Thank you very much.
Last edit: spy 2015-07-21
Possibly it is trying to do a split where validation-set speakers are
distinct from training-set speakers.
In any case, you should never set up your utt2spk that way, and in the
data-prep docs I warn against doing so. Instead just make each speaker-id
identical to its utterance-id.
Next time please ask at one of the lists here:
http://kaldi-asr.org/forums.html
Dan
On Mon, Jul 20, 2015 at 6:13 PM, spy h7558h@users.sf.net wrote:
Dan, thanks for your help.