Kalamكلام

Kalam collects nothing. Your voice is transcribed on your own computer, and neither audio, transcripts, nor any usage data ever leave it.

What Kalam does with your voice

When you hold the dictation key, Kalam records from your microphone and transcribes the audio using speech-recognition models that run locally in your browser (via WebGPU or CPU). The finished text is inserted into the page you are typing in. The audio is processed in memory and discarded; it is never written to disk, never uploaded, and never shared with the websites you use Kalam on.

What we collect

Nothing. Kalam has no servers, no accounts, no analytics, no telemetry, no crash reporting, and no advertising identifiers. We cannot see your audio, your transcripts, your settings, or even how often you use the extension.

The only network traffic

The first time you use a speech model, Kalam downloads its files from huggingface.co (Hugging Face), the public repository that hosts them, and caches them in your browser. This is an ordinary web download: Hugging Face sees the request the way any website sees a visit (e.g. your IP address), governed by the Hugging Face privacy policy. No audio, text, or personal data is ever included. After models are downloaded, Kalam performs no network requests at all — dictation works with your internet disconnected.

What stays on your device

Speech models — cached in your browser's storage; delete them any time from the popup's Models tab.
Settings — your engine, language, and hotkey choices, stored in Chrome's local extension storage.
Performance stats — the duration of your last transcription, shown in the popup.

Removing the extension removes all of it.

Why Kalam asks for its permissions

Microphone — captured only while you hold the dictation key, in an extension-private page; released the instant you let go (Chrome's recording indicator reflects this).
Access to websites — needed to detect the hotkey and insert your transcript into the page you're typing in. Kalam does not read, collect, or transmit page content.
Unlimited storage — speech models are large (80 MB–3 GB) and are cached locally so they download only once.

Verify it yourself

Kalam is open source (MIT). The code that makes these promises is public at github.com/madforstrength/kalam. You can also verify behavior directly: open your browser's network panel while dictating (no requests), or disconnect from the internet entirely (dictation keeps working).

Children

Kalam collects no data from anyone, including children.

Changes

If this policy ever changes, the new version will be published at this address with an updated effective date, and the change will be visible in the repository's public history.

Contact

Questions or concerns: open an issue at github.com/madforstrength/kalam/issues.