Accuracy gets all the questions in the demo. Delay decides whether anyone is still reading in week three.
Where the seconds go
Three things happen between a word leaving the speaker's mouth and a word appearing on a phone. The sound has to be recognised as speech, the sentence has to be complete enough to translate, and the result has to travel to the device.
The middle step is the expensive one, and it is expensive for a reason that is not technical.
A sentence cannot be translated until it has enough of itself
Word order differs between languages. In German the verb often arrives at the end; in Persian and Turkish the structure resolves late; in English the subject and verb come early. Translate before the sentence lands and you produce a confident sentence that turns out to be wrong when the speaker finishes.
Any system that feels instant is doing one of two things: guessing and correcting itself on screen, which is far more distracting than a short delay, or working with a language pair close enough to get away with it. Between English and Spanish you can cheat. Between English and Japanese you cannot.
This is also why we translate only the final version of each utterance and never the interim guesses. An interim guess that reaches the audience and then changes is worse than a sentence that arrives whole a second later.
What one to two seconds feels like
Slightly behind, and then invisible. Within about a minute, readers stop noticing the offset entirely — the same way film subtitles stop feeling out of step. The gap is conspicuous at exactly two moments:
- The room laughs and a reader has not got there yet.
- The speaker asks a question and expects hands up.
Both are worth mentioning to whoever is at the front. One extra beat after a punchline or a question closes the gap at no cost.
A good human interpreter is also behind
People find this surprising. A skilled simultaneous interpreter also runs several seconds behind the speaker, for exactly the same reason — they are waiting for the sentence to make sense. If you have ever watched interpretation at a conference, you already accepted that delay without noticing it.
The difference is that the interpreter's delay varies with how hard the sentence is, and a system's delay is more consistent. Consistency is easier to read.
Three things that make it worse
Long, unpunctuated speech
A speaker who runs three thoughts together with no landing point forces the system to wait. This is the largest controllable factor and it belongs entirely to the person holding the microphone.
A weak audio signal
When the audio is quiet, more of the work goes into hearing at all. We measured two broadcasts of the same service on the same engine: one with speech arriving at 1500–4700 produced 23 correct windows out of 26; one at 260–710 produced garbage. Long before it produces garbage, weak audio produces slow — because the system holds on, waiting for something confident enough to use.
A congested network
Not the translation's fault, but it lands the same way. If forty people join one access point at once, the bottleneck is the number of simultaneous connections rather than the amount of data.
What does not make it worse
Adding languages. Each listener's language is produced independently, so the tenth language does not slow down the first. This surprises people who assume translation is sequential.
The one number worth asking a provider for
Ask for a figure, not an adjective — then ask to watch it on your audio rather than their demo file. A demo uses a clear speaker, a good microphone and a friendly language pair. Your Tuesday evening will not be that, and the difference between those two tests is the entire answer.
Spoken translation is further behind than the text
Worth knowing if you offer read-aloud as well as captions, because listeners will notice and readers will not.
A sentence has to be written before it can be spoken, so the audio necessarily starts after the text appears — and then it takes as long to say as it takes to say. A listener is therefore a little further behind the room than a reader is, and the gap grows slightly during a fast passage before recovering in the next pause.
This is not a defect to be tuned away; it is what speech is. The practical consequence is that read-aloud suits teaching better than it suits anything interactive. For a talk it is fine. For a question-and-answer session, readers will keep up and listeners will not.
When the video stream is the real problem
If you also stream the service, check which arrives first before you blame the translation.
Video is usually delayed more than captions are — often by ten or twenty seconds, depending on how it is delivered. If your captions run two seconds behind reality and your video runs fifteen, an online viewer sees the sentence before they hear it. That reads as broken even though both halves are working correctly.
Open your own stream on a phone next to the translation and see. Most streaming setups let you add a delay to bring the two into line, and five minutes of that is worth more than any amount of optimisation elsewhere.
What to tell the person at the front
Three habits, and they are the whole briefing:
- Finish your sentences. A sentence with a clear end can be translated at once; three thoughts run together cannot.
- Pause a beat after a question or a joke. One extra beat closes the gap entirely and costs nothing.
- Repeat questions from the floor before answering them, or the readers get an answer to nothing.
None of these are translation requirements as such. They are what makes a talk easier to follow for everyone in the room, and the translation simply makes the cost of ignoring them visible.
Do not optimise past the point it matters
Once you are at a second or two, stop. Effort spent shaving off half a second is effort not spent on the microphone, and the microphone is where the actual quality lives.
A reader will forgive a two-second lag indefinitely and will not forgive a sentence the speaker never said. If you have a choice between faster and more accurate, the answer is not close.
Judge it from a seat
InMyTongue delivers subtitles and spoken translation within a second or two of the speaker, and the only fair test is to sit in the room and read along. New accounts include free credit. Try it on a real talk.