Normalize WER computation

We should compute two values for the Word Error Rate:

• compute the WER without altering strings WER(label, pred)
• compute the WER after removing punctuation from label and prediction WER(label_no_punct, pred_no_punct)