VoxForge
VoxForge
I am working on sphinx for arabic language recognition, and I am using Buckwalter for all the files in sphinx (dictionary, phones,transcription,...etc)..the problem is that I always get sentence error =100% and WER=100% after the decoding process...I mostly tried everything to make this work but it did not...will really appreciate if someone can help
--- (Edited on 7/18/2012 2:50 am [GMT-0500] by hazemsafwat) ---
All you are telling us is that your system is broken - that's not enough information for anyone to help you.
Have you been through the standard build proceedure for English? Did that work?
Tony
--
Dr Tony Robinson
CEO Cantab Research Ltd
--- (Edited on 18-July-2012 9:10 am [GMT+0100] by TonyR) ---
Yes, I have tried before the an4 package which is in English
I also tried a simple experiment for arabic digits and it worked
Could you please tell me what information you want to know? Do u need to have a look at the files?
--- (Edited on 7/18/2012 3:15 am [GMT-0500] by hazemsafwat) ---
I went through all the steps in this tutorial and I have tried some experiments...
http://cmusphinx.sourceforge.net/wiki/tutorialam
they all work fine, but in this experiment it gives the error that I have previously stated after decoding. I opened the log file and I found many error messages related to
"
Failed to align audio to trancript: final state of the search is not reached
"
--- (Edited on 7/18/2012 3:44 am [GMT-0500] by hazemsafwat) ---
Listen to the audio for which you get the "Failed to align" message and check that it is the same as the text.
Also, what recognition rate are you getting on the Arabic digits? Is this the same acoustic models that give you 10)% WER?
Tony
--
Dr Tony Robinson
CEO Cantab Research Ltd
--- (Edited on 18-July-2012 9:49 am [GMT+0100] by TonyR) ---
Thanks a lot for your quick respose
Regarding your point
"
Listen to the audio for which you get the "Failed to align" message and check that it is the same as the text."
It was really the error at the beginning. I listened to all the audio files and I found that they were not matching the transcription files so I adjusted this already, So I am expecting this should be working now
"Also, what recognition rate are you getting on the Arabic digits? "
I got 0 WER when i have tested the arabic digits so this might be more than expexected, but this is because I have used same files for testing and training for the purpose of checking the system first how it goes with arabic letters and phonemes, taking into consideration for sure that sphinx 4 is supporting arabic phonems
After that I tested somehow more audio files for some egyptian females reading sentences about 55 files (training) and 10 files (testing) using arabic letters same as the arabic digits experiment..but it did not work and it was giving WER=100 so I went to another direction for using Buckwalter as an alternative since the arabic phonemes has "tashkeel" that I thought might causing a problem for sphinx...I stucked then in a problem related to sphinx not recognizing some symbols in BW like (*,>,<) as it creates a files on the pc with the name of the phones..so I came over that by trying a different alternative which is using Qalam table in the link below for substituting those symbols with other that are xml friendly...but again i got same error!!
--- (Edited on 7/18/2012 4:07 am [GMT-0500] by hazemsafwat) ---
I am doing my experiments on grapheme based ...knowing that I get my langauge model from the sphinx knowledge based tool in this link
http://www.speech.cs.cmu.edu/tools/lmtool-new.html
--- (Edited on 7/18/2012 4:18 am [GMT-0500] by hazemsafwat) ---