Description of the experiment:
Input of a command like this:
./gibbs -m 3 -example_index example/phn_files/3_10_1.norm.phn -ex_data example/raw_files/3_10_1.jnas.raw -query_list example/query.2.list -out example/output_files/3_10_1.jnas.out -config example/config_sparse_timit -b ./ -snapshot example/snapshot
-example_index ==> land mark files for the given keyword A. The easiest thing to do here is to create a landmark for each frame. -ex_data ==> headerless mfcc feature files for the given Keyword A. -query_list ==> paths to the samples that A is going to be compared to, i.e., B1 ~ B20. -out ==> path to the output file that saves the result -config ==> configuration file that used to train the model -b ./ ==> (not used) -snapshot ==> model file
In query.*.list:
1st col ==> land mark files for B1B20
2nd col ==> raw mfcc feature data for B1B20
3rd col ==> used to extract the name, e.g., 3_12_1, for each example.
The output file: 1st line ==> path of the output file 2nd line ==> decoded unit sequence of A every line after ==> the file name of compared example and the score (the larger the better)