I think the future is probably going to resolve this one in time, and not that much time either. I don’t see why LLM technology shouldn’t eventually be able to perform adequate real-time translation, it’s just a matter of continuing to develop the process.
Some languages will be more difficult than others, and translation will always be imperfect. But we don’t need perfection, just better than our current fairly meh (but still impressively not bad) tools.