The abstract from the paper is the following:
We show for the first time that learning powerful representations from speech audio alone followed by fine-tuning on
transcribed speech can outperform the best semi-supervised methods while being conceptually simpler.