Facebook AI cuts by more than half the error rate of unsupervised speech recognition

The use of unlabeled data has big implications for work with languages that, in an AI context, are low-resource, such as Amharic, Swahili, Kyrgyz, and Tatar.