wav2letter, hybrid non-end-to-end ASR models, and other models trained on the same benchmark datasets.