Your words stay yours.
Spoken, understood, forgotten — in a fraction of a second, on your own computer. Your script, your voice and your rehearsals never leave it.
Nothing leaves your PC.
Speech recognition and the speech coach run on your computer. There's no cloud for your content; the app's windows can't even reach the internet.
No recordings. Ever.
Your voice becomes words as you speak, and a moment later the sound is wiped from memory. It's not saved and not uploaded, so there is no audio file to leak.
Locked to you.
What's kept (your last script, your notes, the coach's marks) is encrypted with your Windows account. Or keep nothing at all, with one switch.
Invisible in screen shares.
The light, the prompter and your notes don't appear in Teams, Zoom, Meet or recordings.
Shared notes stay in the room. They travel only on your own Wi-Fi, end-to-end encrypted, with the key only in the link.
What happens to…
| Thing | Where it lives | How long | Leaves the PC? |
|---|---|---|---|
| Your voice | Memory, 10 ms at a time | Milliseconds, then overwritten | Never |
| The words you speak | Memory, the last few words | Replaced as you go | Never |
| Your script | Memory, plus an encrypted copy (or none) | Until you delete it | Never |
| The coach's analysis | Memory, plus an encrypted copy (or none) | Until you delete it | Never |
| Your licence | Windows Credential Manager | Until you deactivate | Only the key and an anonymous machine hash, at licence checks |
Don't take our word for it. Read the code.
This is the actual code that handles your data — not a summary.
fn wipe(samples: &mut [f32]) {
for s in samples.iter_mut() {
unsafe { std::ptr::write_volatile(s, 0.0) };
}
}
stream.accept_waveform(sr as i32, &chunk);
wipe(&mut chunk); // recognised: these samples are gone nowconnect-src ipc: http://ipc.localhostpub fn seal(plain: &[u8]) -> Option<Vec<u8>> {
CryptProtectData(&blob(plain), None, Some(&entropy), None, None,
CRYPTPROTECT_UI_FORBIDDEN, &mut out).ok()?;
…
}Everything that goes online
The licence check: your key, an anonymous hash of this PC, its name (your IT can withhold it), and the app version.
The one-time download of the speech models (open-source, verified byte for byte).
That's it. No analytics, no tracking pixels, no crash reports, no update pings, no AI training on your words.
For IT and security teams
One policy file, every user.
{ "keep_nothing": true, "downloads": false, "models_dir": "C:\\ProgramData\\Lumo\\models", "device_name": false }Firewall allowlist: the licence server, plus Hugging Face only if you don't provide the models.
Offline activation for restricted networks.
No telemetry, no logs of content.
We never receive scripts, transcripts or audio, so we're not a processor of your content.
- Request the security brief →
Verify it yourself
- 01
Block Lumo in Windows Firewall after setup — it keeps working.
- 02
Watch it with Process Monitor while you speak — no files written.
- 03
Open the saved script file — it starts with LPV1 and the rest is unreadable.
We publish everything that touches your data. Our prompter engine and the coach's know-how are our craft — they ship encrypted and run only on your machine.