commoncorrections 1.0.12

A small python implementation of common ASR corrections

CC - CommonCorrections

A simple repo that is used to correct common ASR outputs. The aim is not on mistakes but different ways of transcribing the same thing with a focus on how something may sound as opposed to the shortened form. The primary use case is to align the ground-truth and output from ASRs just before the WER is calculated.

Static Examples

there's -> there is
google.com -> google dot com

Dynamic Examples

1 2 3 -> one two three
53.4 -> fifty three point four
23:59 -> twenty three fifty nine

Features

1. Designed to be used and fast (ish) with Pandas dataframes
2. Lots of built in corrections for free
3. Ability to easily extend with private corrections

Getting Started

1. Install with: pip install commoncorrections
2. Import with: from commoncorrections import CommonCorrections

Usage Examples

Turn numbers into words:

>>> cc = CommonCorrections()
>>> print(cc.correct_str("1 2 3"))
one two three

Turn times into words:

>>> cc = CommonCorrections()
>>> print(cc.correct_str("23:59"))
twenty three fifty nine

Correct a pandas dataframe:

df = pd.DataFrame(data={"transcript": ['5 4 3', "123 the time is 1:23"],
                             "asr_1": ["five four three", "one two three the time is one twenty three"],
                             "filename": ["./my_local_file.wav", "file2.wav"]})
cc = CommonCorrections()

# to correct only specific columns 
new_df = cc.correct_df(df, column_list=['transcript', 'asr_1'])
# to apply to whole dataframe
new_whole_df = cc.correct_df(df)

mypy Type Checks

I tested installing mypy to check that types are compatible

(py) rob@rob-T480s:~/projects/CommonCorrections/commoncorrections (master)$ mypy commoncorrections.py
Success: no issues found in 1 source file

Change Log

• v1.0.0 - First release
• v1.0.1 - Fixed packaging issue
• v1.0.3 - Fixed pip packaging issue
• v1.0.4 - Fixed pip packaging issue
• v1.0.5 - Fixed issue single digits
• v1.0.6 - Fixed case where dataframe contains a non-str type (e.g. int)
• v1.0.7 - Fixed adding additional dict works and added print(cc) object
• v1.0.8 - Fixed print bug with repl
• v1.0.9 - Added some words with space in default corrections csv
• v1.0.10 - Typo in some corrections
• v1.0.11 - Added test case and fixed mistake
• v1.0.12 - Fixed pinning requirements