VoxForge
VoxForge
Hi Rauf,
the choce of TARGETKIND can have a pretty big impact on resulting accuracy, see this post (you can find my configs in the package I posted yesterday).
Unfortunatelly my results do not agree with what is reported in the literature, see this. I guess further investigation is needed.
Regarding your offer to post your training scripts, I have some ideas, but first I need to finish a couple of other things. I will write about it later.
--- (Edited on 23.04.2009 08:54 [GMT+0200] by tpavelka) ---
I have started a new thread to discuss what should be done about these training scripts.
--- (Edited on 23.04.2009 11:12 [GMT+0200] by tpavelka) ---
Hi all,
Finally I found out the reason of poor results with adapted models. As I said, when using adapted models with HDecode, it showed warnings "no token survived to the sent end"
I was performing adaptation steps in HTK Book. I changed the HADAPT:SPLITTHRESH parameter in config.rc file, from 1000 to 5000, and it started working.
Results of HDecode with speaker-independent models was 65% and it increased to 73% with adapted models.
Thank you for your help, also thank you Tomas for your adaptaptation script
--- (Edited on 4/23/2009 10:10 am [GMT-0500] by Rauf) ---
--- (Edited on 4/23/2009 10:12 am [GMT-0500] by Rauf) ---
--- (Edited on 4/23/2009 10:33 am [GMT-0500] by Rauf) ---
Hi
>What is the reason for reviewing of the HVite log in alignment? Is it that
>HTK may drop some sentences because of overpruning and not report
>any errors (unless you enable tracing)?
I think it was just a way to catch mismatches between transcription and speech audio, but which were not bad enough to stop HVite's alignment process.
However, I think I remember you and Nickolay (nsh) talking about this recently (can't find the post though...), and both of you coming to the conclusion that this is not the optimal approach to catch these types of errors...
Ken
--- (Edited on 4/22/2009 11:46 am [GMT-0400] by kmaclean) ---