Monolingual speech recognition models beat multilingual models ~30x bigger