$ conda create --name cscores --file requirements.txt
$ source activate cscores
$ pip install editdistance
Edit the config
file, you should set the kaldi alignement phone file for the
language you are evaluating
In the test directory I include three files:
- mandarin.phn: the gold phoneme alignments
- mandarin.txt: the baseline of ZRTools used in the Challenge 2017 in the tde evaluation format.
- mandarin.out: the raw output of ZRTools discorery for the Challenge 2017, in the ZRTools format.
the script use to compute the ned for a file with the class tde format (mandarin.txt) is
compute_ned.py
, for example for the test mandarin.txt file you can do:
$ . config
$ ./compute_ned.py test/mandarin.txt | tee mandarin.log
to decode the log file I use:
$ sed -rn '/^.*within: NED=(.*) std=.*pairs=(.*)$/\1 \2/p' mandarin.log
to compute ned from the raw output of ZRTools I do:
$ ./zrtools2eval.py test/mandarin.out > mandarin.class
$ ./compute_ned.py mandarin.class | tee mandarin.log