wav2letter, conventional hybrid ASR models, and other models trained on the same benchmark datasets.