I used the perfect script local/online/multisplice.sh to do
DNN traninig. The DNN of this script is and the result is rests
on tri3b setp.
My qst is that the result of this step (that take very long time) not improved the WER and
was even a little less good for the one of tri3b.
I can say that as I progressed through the stages tri1 until tri5
i improved the WER. but in this stage (DNN) i didn't get better WER.
Can someone pls help me?
Thanks!
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
You'll have to provide more details. I don't see a script with that exact
name in any of the example directories.
If you're applying this to your own data, it's possible that the number of
parameters is not right.
Dan
I used the perfect script local/online/multisplice.sh to do
DNN traninig. The DNN of this script is and the result is rests
on tri3b setp.
My qst is that the result of this step (that take very long time) not
improved the WER and
was even a little less good for the one of tri3b.
I can say that as I progressed through the stages tri1 until tri5
i improved the WER. but in this stage (DNN) i didn't get better WER.
I'm not sure how long your training utterances are, but 30k doesn't sound
like a lot to me.
You have to remember that Fisher is a huge database, there is about 1900
hours of speech, IIRC. So the parameter settings (pnorm-input-dim,
pnorm-output-dim, number of layers) that are suitable for Fisher will not
be suitable for your setup. Maybe the settings in the WSJ example (if one
exists) would be more suitable. Try to find an example script that has
about the same number of hours of speech.
Dan
Hello
I used the perfect script local/online/multisplice.sh to do
DNN traninig. The DNN of this script is and the result is rests
on tri3b setp.
My qst is that the result of this step (that take very long time) not improved the WER and
was even a little less good for the one of tri3b.
I can say that as I progressed through the stages tri1 until tri5
i improved the WER. but in this stage (DNN) i didn't get better WER.
Can someone pls help me?
Thanks!
You'll have to provide more details. I don't see a script with that exact
name in any of the example directories.
If you're applying this to your own data, it's possible that the number of
parameters is not right.
Dan
On Sat, Mar 28, 2015 at 2:50 PM, atuk atuk123@users.sf.net wrote:
Thanks for your answer.
I used fisher_english/s5/local/online/run_nnet2_multisplice.sh
My data is about 30K utt , the corpus about 30M and the lexicon
about 60K words. Can you pls help me about the parameters?
Thank you very much!
I'm not sure how long your training utterances are, but 30k doesn't sound
like a lot to me.
You have to remember that Fisher is a huge database, there is about 1900
hours of speech, IIRC. So the parameter settings (pnorm-input-dim,
pnorm-output-dim, number of layers) that are suitable for Fisher will not
be suitable for your setup. Maybe the settings in the WSJ example (if one
exists) would be more suitable. Try to find an example script that has
about the same number of hours of speech.
Dan
On Sun, Mar 29, 2015 at 2:31 AM, atuk atuk123@users.sf.net wrote: