simple-wer

A simple command line tool to calculate WER for ASR.

Installation

pip install simple-wer

Python API Usage

from simple_wer import simple_wer

# results is a list of (utterance_id, reference_words, hypothesis_words)
results = [
    ("utt1", ["I", "LOVE", "YOU"], ["i", "LOVE", "YOU"]),
    ("utt2", ["I", "HATE", "YOU"], ["I", "hate", "YOU"]),
]

# Use with custom logdir and test name
simple_wer(results, logdir="logs", test_name="my_test")

# Or let logdir default to a temp directory and test_name to current timestamp
simple_wer(results)

The results parameter is a list of tuples, each containing:

• utterance_id (str): unique identifier for the utterance
• reference (list[str]): reference transcript as a list of words
• hypothesis (list[str]): hypothesis transcript as a list of words

When logdir is None (default), a temporary directory is created automatically. When test_name is None (default), the current timestamp (microsecond precision) is used.

Command Line Usage

$ simple-wer --help
Usage: simple-wer [OPTIONS] REFS HYPS

  The shell entry point to simple-wer, a simple tool to calculate WERs for
  ASR.

  The ``refs`` and ``hyps`` are the paths for references and hypothesis, their
  format should be ``utt_id \t refenence/hypothesis``.

Options:
  --test-name TEXT  The name for current test, it is used to distinguish the
                    file names for different tests.
  --logdir TEXT     The directory to write the WER results.
  --help            Show this message and exit.

For example:

The refs.txt contains:

1       I LOVE YOU
2       I HATE YOU

The hyps.txt contains:

1       i LOVE YOU
2       I hate YOU

Run the following command:

simple-wer --test-name example refs.txt hyps.txt

The logs look like:

2023-05-16 11:16:05,253 INFO [simple_wer.py:21] The transcripts are stored in logs/recogs-example.txt
2023-05-16 11:16:05,255 INFO [simple_wer.py:123] [example] %WER 33.33% [2 / 6, 0 ins, 0 del, 2 sub ]
2023-05-16 11:16:05,257 INFO [simple_wer.py:29] Wrote detailed error stats to logs/errs-example.txt

There will be two files (recogs-example.txt and errs-example.txt) in logs,

recogs-example.txt looks like:

1:      ref=['I', 'LOVE', 'YOU']
1:      hyp=['i', 'LOVE', 'YOU']

2:      ref=['I', 'HATE', 'YOU']
2:      hyp=['I', 'hate', 'YOU']

errs-example.txt looks like:

%WER = 33.33
Errors: 0 insertions, 0 deletions, 2 substitutions, over 6 reference words (4 correct)
Search below for sections starting with PER-UTT DETAILS:, SUBSTITUTIONS:, DELETIONS:, INSERTIONS:, PER-WORD STATS:

PER-UTT DETAILS: corr or (ref->hyp)  
1:      (I->i) LOVE YOU
2:      I (HATE->hate) YOU

SUBSTITUTIONS: count ref -> hyp
1   I -> i
1   HATE -> hate

DELETIONS: count ref

INSERTIONS: count hyp

PER-WORD STATS: word  corr tot_errs count_in_ref count_in_hyp
i   0 1 0 1
hate   0 1 0 1
I   1 1 2 1
HATE   0 1 1 0
YOU   2 0 2 2
LOVE   1 0 1 1