Hi Marion,

I have started to do my own experiments with Julius but my current results are still way below those of HDecode (To be more specific my percentage of correct words is 93% for HDecode and 82% for Julius). I am curious about your experiments, namely:

• What is your testing set?
• Do you have any results that compare the accuracies achieved with WSJ and VoxForge acoustic models?
• How did you deal with unseen triphones? I have downloaded the WSJ acoustic model but did not find the file with decision trees so I cannot make new triphones that are not included in tiedlist.

Tomas

--- (Edited on 4/18/2009 2:29 am [GMT-0500] by tpavelka) ---