Using just ten minutes of labeled data and
pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER.