VoxForge
I've got julius 4.1.5, the latest nightly 8khz acoustic model from this site, and the grammar from here: http://www.keithv.com/software/giga/lm_giga_64k_nvp_3gram.zip
I've then tried to run an 8khz wave file through julius using that model and grammar, but I must be using the wrong grammar dictionary file. I get lots of errors like:
Error: voca_load_htkdict: line 3: triphone "*-1+k" or biphone "1+k" not found
Error: voca_load_htkdict: line 3: triphone "1-k+ah" not found
Error: voca_load_htkdict: the line content was: \'cause 1 k ah z
Error: voca_load_htkdict: line 4: triphone "*-1+ah" or biphone "1+ah" not found
The grammar came with 3 .dic files (hdecode, hvite, and sphinx). I've tried the first two (because I'm not using sphinx here) but both result in a ton of errors similar to those above. I had found a thread on the forums here where someone said they used one of the .dic files from the same grammar package I downloaded. What am I doing wrong?
--- (Edited on 8/31/2010 7:52 pm [GMT-0500] by Visitor) ---
>grammar from here: http://www.keithv.com/software/giga/lm_giga_64k_nvp_3gram.zip
I think that the statistical language model from Keith's site has words with triphones/biphones that are not in the VoxForge acoustic model.
You might try re-running step 10 of the acoustic model creation process on the hmmdef files in the hmm10 folder of a dev version of the VoxForge acoustic model:
XXX_HTK_AcousticModel-2010-09-09_8kHz_16bit_MFCC_O_D_devel.tgz 09-Sep-2010 05:53 40.6M
using one of the HTK pronunciation dictionary (HDecode or HVite) files from Keith's site.
Ken
--- (Edited on 9/10/2010 12:01 am [GMT-0400] by kmaclean) ---