Translated audio can't start before there's something to translate, so a live interpreter — human or AI — always trails the original speech by some margin. That's true of every simultaneous-interpretation system that has ever existed, not a limitation specific to any one product. What actually differs between implementations is how that trailing margin is managed: whether the delay stays bounded and predictable, or whether a long sentence is allowed to pile up a growing backlog behind the words already on screen.
Why the delay can never be zero
Simultaneous interpretation — the kind professional human interpreters do in a booth at the UN or a conference — requires hearing enough of a sentence to understand its meaning before translating it. A word-for-word translation the instant each word is spoken produces nonsense in any language pair with different word order, so some amount of "wait and listen" is unavoidable before translation can begin. This is a structural property of the task itself, not an engineering shortfall: professional human simultaneous interpreters are trained specifically to manage a perceptible, unavoidable lag behind the speaker, and it's a well-studied part of how the profession works. Real-time AI interpretation runs into the identical constraint — it has to hear enough to translate correctly, and that takes a moment.
Why the delay isn't something you tune down
Older speech-translation pipelines stack three sequential steps: transcribe the speech to text, machine-translate that text, then synthesize new audio from the translation — three separate delays added end to end. A native simultaneous model collapses that into one continuous process that transcribes, translates, and speaks in a single pass, which removes two of those three sequential delays. What it doesn't remove is the model's own internal pacing: a native simultaneous model manages its own timing internally, deciding moment to moment how much context to wait for before committing to an interpretation. There's no manual "faster" dial to turn, because the model — not a client-side setting — owns that decision.
What happens when a long sentence pulls audio behind the captions
Live captions and live translated audio are produced by the same underlying process, but they don't always finish at exactly the same moment — a long or complex sentence can leave the translated audio trailing further behind the caption text than usual. Left unmanaged, that gap would only grow over a long session. Instead, adaptive playback pacing watches the size of that gap and, when it grows past a threshold, very slightly speeds up the queued audio — using time-compression that preserves pitch, so it doesn't sound sped-up — until the audio catches back up to the live caption. The alternative approaches (skipping words entirely, or letting the gap grow indefinitely) are both worse: one loses content, the other makes a long session feel increasingly disconnected from what's actually being said.
What voice-activity detection does before any of this
Before any of the above even applies, on-device speech detection filters out silence and non-speech sound so it's never sent for translation in the first place. This isn't primarily a latency optimization — its main job is making sure the pipeline isn't wasting effort "translating" background noise or dead air — but it does mean the system isn't burning its own pacing budget on audio that was never going to produce a translation anyway.
FAQ
Can the delay be turned down manually? There's no direct latency dial — the translation model manages its own pacing. Scenario presets (meeting, film, conference) change how aggressively silence gets skipped before audio is sent for translation, which affects what reaches the model, not how fast the model itself responds.
Does a longer sentence always mean a longer wait? Generally, a little — longer utterances give the model more context and tend to translate more accurately as a result, which is why very short fragments are sometimes batched rather than translated word by word.
Is this the same as watching a professionally dubbed film? No. A professional dub is translated and re-recorded well before anyone watches it, with no time pressure. This is happening live, in the moment — a small, unavoidable lag is simply part of what "live" means here, the same way it's part of what live human interpretation means.
Davut Akca, Founder, VoxisLive. More about VoxisLive.