edit_distance.py 文件源码

def compute_per(ref, hyp, normalize=True):
    """Compute Phone Error Rate.
    Args:
        ref (list): phones in the reference transcript
        hyp (list): phones in the predicted transcript
        normalize (bool, optional): if True, divide by the length of str_true
    Returns:
        per (float): Phone Error Rate between str_true and str_pred
    """
    # Build mapping of phone to index
    phone_set = set(ref + hyp)
    phone2char = dict(zip(phone_set, range(len(phone_set))))

    # Map phones to a single char array
    # NOTE: Levenshtein packages only accepts strings
    phones_ref = [chr(phone2char[p]) for p in ref]
    phones_hyp = [chr(phone2char[p]) for p in hyp]

    per = lev.distance(''.join(phones_ref), ''.join(phones_hyp))
    if normalize:
        per /= len(ref)
    return per