Python module for evaluting ASR hypotheses (i.e. word error rate and word recognition rate).
This module depends on the editdistance project, for computing edit distances between arbitrary sequences.
The formatting of the output of this program is very loosely based around the same idea as the align.c program commonly used within the Sphinx ASR community. This may run a bit faster if neither instances nor confusions are printed.
Please let me know if you have any comments, questions, or problems.
You can clone this git repo and install using distutils:
git clone git@https://github.com/zweiein/asr-evaluation.git
cd asr-evaluation
python setup.py install
For command line usage, see:
wer --help
It should display something like this:
usage: wer [-h] [-i | -r] [--head-ids] [-id] [-c] [-p] [-m count] [-a] [-e]
ref hyp
Evaluate an ASR transcript against a reference transcript.
positional arguments:
ref Reference transcript filename
hyp ASR hypothesis filename
optional arguments:
-h, --help show this help message and exit
-i, --print-instances
Print all individual sentences and their errors.
-r, --print-errors Print all individual sentences that contain errors.
--head-ids Hypothesis and reference files have ids in the first
token? (Kaldi format)
-id, --tail-ids, --has-ids
Hypothesis and reference files have ids in the last
token? (Sphinx format)
-c, --confusions Print tables of which words were confused.
-p, --print-wer-vs-length
Print table of average WER grouped by reference
sentence length.
-m count, --min-word-count count
Minimum word count to show a word in confusions.
-a, --case-insensitive
Down-case the text before running the evaluation.
-e, --remove-empty-refs
Skip over any examples where the reference is empty.
For contributions, it's best to Github issues and pull requests. Proper testing and documentation suggested.
Code of conduct is expected to be reasonable, especially as specified by the Contributor Covenant