VoxForge
Hello,
As I tried to launch Julius with the acoustic model and lexicon from the following file found in the Nightly Builds downloads section :
http://www.repository.voxforge1.org/downloads/Nightly_Builds/AcousticModel-2010-10-24/HTK_AcousticModel-2010-10-24_16kHz_16bit_MFCC_O_D.zip
...thanks to this command :
   ./julius/julius -nostrip -walign -demo -b 1500 -output 1 -n 1 -lmp 9 -8 -lmp2 10 -10 -input adinnetmfcc -d ./AM_LM/EN/f3b3bingramEnglish_alice_text -h ./AM_LM/EN/2010-10-25_16kHz_16bit_MFCC/hmmdefs -v ./AM_LM/EN/2010-10-25_16kHz_16bit_MFCC/dict
...I obtained such kind of error messages in my console :
Error: voca_load_htkdict: line 1027: triphone "b-ey+dh" not found
Error: voca_load_htkdict: the line content was: BATHINGÂ Â Â Â Â Â Â Â [BATHING]Â Â Â Â Â Â b ey dh ix ng
Error: voca_load_htkdict: line 1031: triphone "ax-t+ae" not found
Error: voca_load_htkdict: line 1031: triphone "y-ax+n" not found
Error: voca_load_htkdict: the line content was: BATTALIONÂ Â Â Â Â Â [BATTALION]Â Â Â Â b ax t ae l y ax n
Error: voca_load_htkdict: line 1035: triphone "t-uw+z" not found
Error: voca_load_htkdict: the line content was: BATTUESÂ Â Â Â Â Â Â Â [BATTUES]Â Â Â Â Â Â b ae t uw z
Error: voca_load_htkdict: line 1036: triphone "iy-ax+n" not found
Error: voca_load_htkdict: the line content was: BAVIAANÂ Â Â Â Â Â Â Â [BAVIAAN]
My language model is trained (3 forward 3 backward bi N gram) on a text I found on internet, thanks to the srilm 1.5.9 tool. NB : I use Julius 4.1.3.
So as far as I understand, the definition of the triphones is not found in the hmmdefs file given in argument. Am I right ?
If I am right, could someone redirect me to a description of this directory or tell me where I could find some compatible hmmdefs, hmmlist and lexicon ?
But I feel that I can be completely wrong, as I did not give any hmmlist as an argument to Julius (it should work without it anyway), so there should be no way for Julius to find discrepencies between those both files, as it reads just one of them. Furthermore, the files all came from the same archive, so even if by chance Julius was considering the hmmlist file, there is very few chances to find discrepencies. I suppose I did something wrong.
I really would be grateful if somebody helped me with this issue.
Thanks in advance.
Pierre.
--- (Edited on 10/25/2010 11:52 am [GMT-0500] by pierresendorek) ---
--- (Edited on 10/25/2010 11:58 am [GMT-0500] by pierresendorek) ---
>So as far as I understand, the definition of the triphones is
>not found in the hmmdefs file given in argument. Am I right ?
Your language model contains words with triphones that are not in the pronunciation dictionary used to train the VoxForge acoustic model.
If you want, figure which words are missing and create pronunciations for them, and we can add the to the VoxForge dictionary.
--- (Edited on 10/29/2010 9:36 pm [GMT-0400] by kmaclean) ---