Tools
How much of a meeting survives in a second language?
When a meeting runs in everyone’s second language, meaning leaks out twice — once when the speaker translates a thought into English, and again when a non-native listener parses it. Those two losses don’t average; they multiply. If each step keeps about 60%, only 0.6 × 0.6 ≈ 36% of the original idea lands. Drag the sliders below to see it for your own team — then see what changes when only one conversion has to happen.
Expressing a thought in a second language, you reach for the word you have, not the word you wanted.
On the other end, another non-native speaker parses your accent, your speed, a half-known idiom.
Two losses stack: 60% × 60% — they multiply, not average.
One machine conversion in the middle replaces two human ones — a single loss instead of two stacked.
These figures are an illustration to make the shape of the problem visible — that two losses compound instead of cancelling out. They are not a benchmark, and the second bar is not a claim about machine-translation accuracy; no translation is perfect. The point is simply that one conversion beats two stacked.
The fix isn’t “better English”
You can spend years closing the fluency gap and still have native speakers setting the pace. The reframe is to stop making people translate themselves in real time. Let the speaker talk in their own language, at full expression. Let the listener read in theirs, at full comprehension. The translation happens once, in the middle, by a machine — one conversion step, and a smaller loss, replacing two big human ones. You don’t make the loss disappear; you stop it from compounding, and you take the social tax off speaking up, because nobody is performing in a second language anymore.
Questions
- Why do the two losses multiply instead of average?
- A meeting in a shared second language leaks meaning in two separate places: the speaker only gets part of the idea into the sentence, and the listener only catches part of that. Because the listener works on what already survived the speaker, the fractions multiply. If each step keeps about 60%, what lands is 0.6 × 0.6 ≈ 36% — well under half — not the 60% an average would suggest.
- Are these percentages a real measurement?
- No. They are an illustration to make the shape of the problem visible — that two losses compound rather than cancel out. They are not a benchmark, and the second bar is not a claim about machine-translation accuracy. The exact figure is not the point; the compounding is.
- How does per-person translation change the math?
- Instead of two people each straining through a second language, the speaker talks in their own language at full expression and the listener reads in theirs at full comprehension. The translation happens once, in the middle, by a machine. Machine translation is not perfect either, but it is one conversion step replacing two human ones — a single smaller loss instead of two stacked.
- What does Sageio actually do?
- Add bot@sageio.net to your Google Meet calendar invite and it joins on its own — nothing to install. Each participant picks the language they want captions in, so the same meeting serves everyone at once, with captions appearing in about two seconds. Sageio translates into 20+ languages, with Asian languages treated as first-class. Within about five minutes of the call ending, a searchable transcript and summary arrive. (Today this runs on Google Meet; Zoom and Microsoft Teams are coming soon.)