This directory contains some experiments on feature extraction using MFCC (Mel Frequency Cepstral Coefficient))and by HMM (based on HTK). Note that Data pepration is the same for all of the experiments. As the copying of data prepration is time consuming we are using one data prepration for all of the experiments. Additionally feature extraction is done under another experiment using MFCC. 1. The first step is data preperation which is implemented in data_prep directory. This step converts annotations (file with extension of .rec) to transcripts which can be read by htk (files with extension of .lab) and save them in the related directories. After running these instructions, a directory named annotations and two subdirectory under annotations will be created. you can run data_prep_annotated.sh by these instructions: cd ../../data/isip/exp/tuh_eeg/exp_0014/data_prep sh data_prep_annotated.sh 2. Then based on your experiments you can create a directory such as t_000_000. The second step should be feature extraction. As feature extraction files are generatted already, we just create a list of them by these instructions: cd ../t_000_000/feats sh feats_lists 3. Then third step is training. This script is used to to train HMM models using HTK toolbox. This should be submitted as a job on owlsnest. Read the instructions in http://www.hpc.temple.edu/owlsnest/OwlsnestUserGuide.html. Before that you should check for the available nodes by : showfree then replace the name of the nodes which are appropriate and available in the header of train.sh and run these: cd ../train qsub -k oe train.sh 4. The forth and fifth steps are scoring. This scripts is used to score (compute LL) of data against an HMM model using HTK toolbox and our own extension to it Note that every time you are submitting a job, you should check free nodes by showfree and replace the free and appropriate node in the header. cd ../scoring qsub -k oe score_train.sh qsub -k oe score_eval.sh 5. from now on there are two parallel steps. You can run decode and then eval or you can run swap and then det. First we are running decode: cd ../decode qsub -k oe decode.sh 6. after decode the next step is running eval: cd ../eval sh eval.sh 7. As it is mentioned in step4, the other path is running swap. For increasing the speed of calculation there are two jobs which should be submitted simultaneously. cd ../swap qsub -k oe swap_02_part1.sh qsub -k oe swap_02_part2.sh 8. The next step after swap is running det: cd ../det sh det_02.sh ----------------------------------------------------------------------------------------------------------------------------- Some notes: - proto defines a prototype for HMM which will be used by htk. In this file vector size is the length of feature vector. For example here it is 27 because we have 8 main feature and with d-energy it resulets 9 and with delta results 18 and with delta-delta it results 27. - Frame rate is 0.1. But window is not constant and we can change it. Here window is 0.2.