arxiv wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations