You will be evaluated based on the Levenshtein distance between your output sequences and the ground truth sequences for the test dataset of LibriSpeech.