People speaking different languages need translations while a speaker is still talking, not after the whole sentence. Waiting too long makes conversation awkward, but translating too early causes mistakes, and spoken errors cannot be taken back.
The system learns from full sentence pairs when a partial phrase is safe to translate. It speaks in short chunks on a fixed time grid. A tunable setting controls how cautious the model is, so the same model can be tuned for speed or accuracy.
In tests it matched or beat current streaming translation systems in quality, and adapting the setting on recorded conversation improved accuracy with almost no added delay. That gives real products a practical speed-versus-quality dial from the same model.