decode results between online and online2

[amr....@gmail.com]

Hi,

I am working on online decoding with kaldi and am having some issues concerning the decoding with the online-gmm-faster decoder in the first setup. I try to use the kaldi-gstreamer server (https://github.com/alumae/kaldi-gstreamer-server) with my own gmm models.

The problem is that the models I generated do not work very good with the online-gmm-faster although the results were correct with the  online2-wav-latgen decoder in online2. Is preparing the models for the old setup different from the new ?

I trained with my own data, using the training scripts in steps/ for offline decoding. I then used the script steps/online/prepare_online_decoding.sh to have the online models. 

Thanks for your help.

[Daniel Povey]

The script in steps/online/prepare_online_decoding.sh  is specifically to prepare models for the online2 decoding.  I think the other, 'online' decoding setup does not require specifically built models, but I don't recall exactly.  You'll have to search to see if there any example scripts that use that setup.

Dan

--
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[amr....@gmail.com]

Thanks for your answer.

When I tried to use the built models "as is" the results were also very bad compared to offline decoding so I thought it was meant to be prepared for online. Anyway, I will try to built models again. 

Ammar

[Daniel Povey]

The "old" online decoding setup does not do adaptation so the results will definitely be worse than the new one, that is expected.
Dan

[wupea...@gmail.com]

Hi, 

I'm confused, if I do SAT offline,  why do I need to retrain the model for online instead of directly using the old one trained? Is there any relation to the online CMVN stats? for using different online cmvn stats? waiting for you help, Thanks!

在 2015年9月2日星期三 UTC+8上午12:06:45，amr....@gmail.com写道：

[Daniel Povey]

Because the features behave a bit differently if you do online CMN.
Dan

[wupeachfan]

Thanks so much!

在 2016年7月14日星期四 UTC+8上午4:24:22，Dan Povey写道：