Your voice stays with you.

The short version

• All inference is local. The detection model runs in your browser, on your device. Audio is never uploaded.
• Nothing you say is stored. Audio exists in memory for about a second, then it’s overwritten by the next second.
• Detections stay on your screen. The recent-cues list lives in the open tab and is gone when you close it.
• No accounts, no analytics, no ads. And no audio “samples collected to improve our service.”

How inference stays local

When you press start, your microphone streams into a short in-memory buffer. A compact acoustic model, running via WebAssembly inside your browser, scores that buffer for filler words many times a minute. The only thing the model produces is a label, a confidence, and a timestamp, and those stay in your tab’s memory. No audio, transcript, or detection ever appears in a network request.

What’s stored on your device

To start fast and work offline, Umbot keeps a copy of the app and the model files in your browser’s local caches, alongside a record of which model version was verified. That’s it. Clearing this site’s data in your browser removes everything.

What actually touches the network

Downloading the app, fetching the model, and checking for model updates are ordinary web requests to our servers. Like any website, our hosting provider sees standard request metadata such as IP address and browser version in its logs. Those requests carry no audio and no detection data, because that information never leaves your device.

Permissions

Umbot asks for microphone access only when you press start, and you can revoke it in your browser at any time. While listening, it may also hold a screen wake lock so your device doesn’t doze off mid-session. That’s the full list.

Changes and contact

Umbot is test software operated by Zaks.io, LLC, and this page will evolve with the product. When our practices change, the plain-language explanation here changes with them. Questions go to isaac [at] zaks.io.

Effective July 19, 2026.