Speech:Exps 0009

From Openitware
Jump to: navigation, search

Title: Tiny Train w/Test on Train #1


Author: Cedric Woodbury

Date: July 24, 2012

Purpose: A training w/decoding run for proof of concept of a working system


Once again I used the transcript from the train1 folder under SphinxTools-1.0 and attempted to run a train again. Then I would need to set up the Language Model and attempt a decode.


Once again it only produced the .semi files. However, I compared the sphinx_train.cfg file to the one from the old/train/train2 experiment. I found a difference in the configuration. That config file had the line #CFG_HMM_TYPE = '.semi.' commented out, and was using #CFG_HMM_TYPE = '.cont' instead. My config file had it the other way around. Apparently this is how it is generated by default when you run setup_Sphinx.pl. I made the changes and ran the train again. This time it produced the correct .cont files. I then created the Language Model and ran a decode. Both appeared to be successful.


I determined what the problem was with the train producing the wrong file types in the model_parameters folder and can now take steps to make sure future experiments work correctly.