Live translation,
as it happens.
You run Earwyrm yourself, on hardware you own, and it translates your meeting while the meeting is still going. Everyone reads live captions in their own language, with the original alongside so you can check what was actually said. No outside cloud is in the loop unless you put one there.
What happens to a sentence.
A plate read left to right. Sound arrives at an ear on the left and draws a live waveform; underneath it, the words being heard assemble on a writing line, faint and unsteady at first and firming up as they travel. A vertical line partway across only lets a word past once it has stopped changing, so unfinished words bunch against it. Past that line each word drops into three lanes below, one per reader, each running a little behind the last and hurrying to catch up, and every word that crosses adds a mark to a running record along the bottom.
Example. Heard: The second batch failed the salt spray test, so we are moving the ship date to Friday. Shown to this viewer: El segundo lote no pasó la prueba de niebla salina, así que movemos la fecha de envío al viernes.
It listens through a browser tab.
So it doesn’t care whether you’re on Zoom, Teams, Meet, or sitting around a table.
It shows its work.
Faint words are a first guess and can still change. Once it has heard something the same way twice, the words go solid and stay put.
Everyone gets their own language.
One meeting, several languages at once. If a line is already in your language, you get it exactly as it was spoken.
Your meetings, your data.
The usual way to get live translation is to hand your meeting audio to somebody else’s cloud and trust their retention policy with what your people actually said. Earwyrm goes the other way. You run it on hardware you own, a Mac on your desk or a Linux server you own, and it can run entirely on local models, so speech recognition and translation both happen on your machine. Privacy and security are what owning that path buys you, and your meetings never have to leave the building.
We built it because translation demos fall apart in real meetings. People interrupt. They switch languages halfway through a sentence. They pause for half a second and keep going. Earwyrm has been through about a hundred of our own meetings, and captions land roughly a second behind the speaker.
Privacy.
This page runs no analytics and no outside scripts. It doesn’t ask our servers anything about you.
Earwyrm lives on a machine you own, and so do your recordings and transcripts. We never see them. Run it on local models and nothing leaves that machine at all. Point the translation step at an outside provider instead, and the text of what was said goes to them. That’s your call, not ours. What people said stays out of the logs unless you turn that on.
One thing Earwyrm doesn’t do yet is ask the people in the meeting for consent. Tell them it’s being captioned, and check the recording rules where you all are.
Before a pilot starts, we agree in writing what gets recorded, where it lives and when it’s deleted. Questions: security@phasewake.com
Try it on your own meetings.
We’re opening pilots for teams that work across languages. If that’s you, write to Thomas.