- It waits for the Turkish verb
- Turkish puts the verb last, and the verb carries the negation — interpret half a sentence and "I will not come" becomes "I will come". For verb-final languages we hold until the sentence closes; for the others interpretation starts flowing before you finish.
- You still hear them, quietly
- Most translation tools replace the other person with a robot voice, and the silence in between reads as a dropped call. Here their real voice stays underneath the interpretation — you keep their tone, their hesitation, and whose turn it is.
- No echo, on a real phone line
- Full acoustic echo cancellation runs per listener, so what we play into a handset never loops back through its microphone. Without it, one side hears themselves being interpreted and the conversation collapses.
- Interrupt like a human
- Start talking and no new interpretation is played into your ear — but nothing stops listening. Your interjection is transcribed and interpreted at the same time, the way a human interpreter handles being cut off.
- The delay is on screen
- A pause in an interpreted call is ambiguous: is the other person thinking, or is the line struggling? We measure the actual wait from silence to speech and show it live, rated Excellent, Good or Slow.
- See what was relayed
- When you stop speaking, a caption shows you exactly what the other side was told and when. On a phone call you only ever hear your own silence — this gives you the window back.
- 880 ms median
- Measured against real utterances from live call logs, not a synthetic benchmark.
- Warmed before you speak
- The first two translations of a cold call took 6+ seconds — so the connection is warmed while the other side is still ringing.
- Wrong-language guard
- Two layers stop a mistaken language from being turned into confident nonsense and spoken to the other side.
- One voice, both languages
- The interpreter sounds like a single person throughout, not two different robots.
- Optional video
- Camera off by default. Simulcast, adaptive streaming and packet-loss protection keep audio first when the network drops.
- Optional one-time passcode
- Stored hashed, constant-time compared, spent on first use — for links that get forwarded.