SimplePythonWER

The purpose of this repo is to provide a well tested basic python implementation of levenshein / WER so it can be shared across projects. It's based on this with a couple of minor changes.

Getting Started

1. Install with: pip install simplepythonwer
2. Import with: from simplepythonwer import wer
3. Use with:

>>> wer("the cat sat on the mat", "the mat sat on the cat")
0.3333333333333333

Features

• Simple, minimal and only in python with 0 external dependencies
• It is versioned and can be pip installed
• Provide examples with tests to ensure it's working correctly

Caveats and Gotchas

• It's possible to have greater than 100% WER if the ASR result is many times larger than the ground-truth, this is normal. It's sometimes a good idea to cap the results at a 100% with min function e.g. min(wer(ground_truth, new_asr_string), 1.0), otherwise you could be exposed to unlimited error rate that could skew your averages.

Change Log

• v1.0.0 - First release - Minor ~15% speed improvements compared to original
• v1.0.1 - Fixed pip packaging and added install steps. Exclude tests from pip
• v1.0.2 - Fixed pip packaging issue
• v1.0.3 - Fixed divide by zero error when the ground truth is zero length (including evaluates to zero length since it's just whitespace)

Tests

Run with: PYTHONPATH=$(pwd) python3 -m unittest discover . Results:

rob@rob-T480s:~/projects/SimplePythonWER (master)$ PYTHONPATH=$(pwd) python3 -m unittest discover .
..
----------------------------------------------------------------------
Ran 9 tests in 0.001s

OK

Speed Improvements

from simplepythonwer.simplepythonwer import *
import timeit
sentence = "the cat sat on the mat"*5
print(timeit.timeit('levenshtein(sentence, sentence[::-1])', number=10000, globals=globals()))
print(timeit.timeit('levenshtein_original(sentence, sentence[::-1])', number=10000, globals=globals()))
38.16882774699479
44.751817572047