VoxForge
VoxForge
hi all!
my first post :-)
I was trying to run the hdecode tutorial and stumbled upon
"WARNING [-9999] no token survived to sent end! in ...
Segmentation fault"
I have trained cross-word triphones and am using configuration scripts according to the htk book 3_41 example.
The debugging brought me to HTKLVRec/HLVRec-traceback.c:93
with bestInst = dec->bestInst; (obviously the segfault because the bestInst member is still set to 0x0)
any suggestions as to what am I supposed to search for?
regards
--- (Edited on 10/18/2012 12:43 pm [GMT-0500] by mimik) ---
--- (Edited on 10/18/2012 12:45 pm [GMT-0500] by mimik) ---
Well, no software should segfault, but putting that aside the way to avoid the error message is to increase the beam using the -t option.
Tony
--
Dr Tony Robinson
CEO Cantab Research Ltd
Phone +44(0)1223 794497
--- (Edited on 18-October-2012 9:21 pm [GMT+0100] by TonyR) ---
Thank you for the the advice.
Regretfully, increasing the beam treshold does not seem to cut it for me.
I've done some debugging and found out, that when HDecode is reading mfc files, there is a conditional block swapping the bytes of the header. What happened is that the HCopy did not swap the byte order when the mfc data was being generated. For some reason both HDecode and HList attempt to activate the swap when reading the mfc file.
The header is no more of a threat as I've removed the swapping from the code.
I'm currently trying to examine [HDecode:DoRecognition][HLVRec-propagate:ProcessFrame] with the hope of finding some indications as to why no bestAlignInfo is being generated.
/mimik
--- (Edited on 10/31/2012 7:35 am [GMT-0500] by mimik) ---