If it has to be live (i.e. works in real-time), doing the translation in a server elsewhere is probably a bad idea due to the latency.

It makes more sense to run an optimised translation model locally - ideally in the AirPods directly but if not, on the connected iPhone.