VoxForge
VoxForge
Hi Satish,
My understanding is that the Sphinx HUB4 and WSJ acoustic models are in a Sphinx binary format. Julius uses HTK ascii format files.
I did run across a site by Keith Vertanen where he create binary HTK Acoustic Models
using the WSJ corpus. I've emailed him to see if we can get ASCII
versions of his Acoustic Models - so they can be used with Julius.
Ken
--- (Edited on 1/ 1/2007 9:52 pm [GMT-0500] by kmaclean) ---
Hi Satish,
I emailed LEE Akinobu (Julius maintainer), and he provided me with the following method to convert binary HTK hmmdefs file to ASCII:
% touch null.hed
% HHEd -H binary_hmmdefs -w ascii_hmmdefs null.hed tiedlist
The tiedlist file is included in the AM archives.
Also please look at
http://kt-lab.ics.nitech.ac.jp/hts-users/spool/2006/msg00280.html.
I got Julian to work with Keith Vertanen's htk_wsj_all_10000_32.zip file, though recognition (with my voice) was not that good. It may be that the Julian configuration file needs to be tweaked, or that the Acoustic Model needs to be adapted to my voice - which can be done since he provides the 'stats 'file with the distribution.
VoxForge provides some instruction for adapting Acoustic Models to your own voice (see this link), but using HTK version 3.2.1. I could not get adaptation to work using HTK version 3.3. I have not had a chance to try it with HTK rel 3.4.
Ken
--- (Edited on 1/ 2/2007 1:40 pm [GMT-0500] by kmaclean) ---
--- (Edited on 1/ 2/2007 1:42 pm [GMT-0500] by kmaclean) ---