A small basic python implementation of WER (word error rate) and levenshtein
Project description
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
- Install with:
pip install simplepythonwer
- Import with:
from simplepythonwer import wer
- 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
- Providing an empty string or filled with whitespace ground-truth will intentionally raise a divide by zero.
- 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
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
simplepythonwer-1.0.3.tar.gz
(4.3 kB
view hashes)
Built Distribution
Close
Hashes for simplepythonwer-1.0.3-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 69d827d7c678636d2054d9768e8aa440c65b448fe6c8027bd255f640aedbba83 |
|
MD5 | 75211063c8588ae924e85dfd4c59f236 |
|
BLAKE2b-256 | d026fdc27ef32532432b84b612f6514688c7f9c5b1254f33228baa28978efdcc |