A script to perform significance tests on the outputs of ASR models.
Only the reference transcription and the models' textual outputs are needed. The texts need to be in trn format, more info here
run_sign_tests.sh ref-text hyp-text-1 hyp-text-2 ... hyp-text-N
the first parameter is the filename of the transcript. After that, a list of files can be given, which will be compared.
Important: the script uses sclite and sc_stat, these files must be accesible, if you use Kaldi run path.sh before this script.
The script will produce several output files, which start with the prefix significance_report. The most important one is the significance_report.stats.unified that summarizes all the tests.