I trained models (16 GMM tied-state)using TIMIT training data and a back-off bigram languague model (HLStats) with its corresponding wordnetwork (HBuild), this network was of 1MB size. When i tested on the TEST data the result obtained was 80.29%, it has taken a time of 4 1/2 hours. But when i tested using the bigram model (not back-off bigram) and its corresponding wordnetwork(this was of 1GB) that was built using same HBuild, the result was 95.39%. The issue in here is the latter test took 2 1/2 days. So is this time normal or is there any way to reduce this testing time. Please suggest if there's a solution for this. My professor told that Sphinx doesnt take this much time and it takes to the max an hour for testing (the same TIMIT database). I searched on web to know the testing time but could not find.