Error rate calculation for handwriting recognition. pip install editdistance==0.4

ErrorRateMetric.py

import editdistance
  
  
def cer(y_true: str, y_pred: str) -> float:
    assert isinstance(y_true, str) and isinstance(y_pred, str)
    n_err = editdistance.eval(y_true.rstrip(), y_pred.rstrip())
    return n_err / len(y_true)

  
def wer(transcript: str, reference: str) -> float:
    assert isinstance(transcript, str) and isinstance(reference, str)
    transcript = transcript.split()
    reference = reference.split()
    n_err = editdistance.eval(transcript, reference)
    return n_err / len(reference)

  
if __name__ == "__main__":
    # exact case
    cer_score = cer("hello world", "hello world")
    wer_score = wer("hello world", "hello world")
    assert cer_score == 0.
    assert wer_score == 0.
    
    # one deletion case
    cer_score = cer("hello world", "hello world!")
    wer_score = wer("hello world", "hello world!")
    assert cer_score == 1. / 11.
    assert wer_score == 1. / 2.
    
    # one substitution case
    cer_score = cer("hello world", "hallo world")
    wer_score = wer("hello world", "hallo world")
    assert cer_score == 1. / 11.
    assert wer_score == 1. / 2.

    # one insertion case
    cer_score = cer("hello world", "ello world")
    wer_score = wer("hello world", "ello world")
    assert cer_score == 1. / 11.
    assert wer_score == 1. / 2.