Exploring the limits of decoder-only models trained on public speech recognition corpora