There is a core tradeoff in real-time voice translation: waiting for more context improves accuracy but increases latency, while outputting early reduces latency but may result in translation errors.