Privacy by design
Zero network requests. Your voice never leaves your machine.
Not a slogan. A verifiable fact — captured in Wireshark.
Proof, captured live
1,683 packets · 0 external connections
We ran Wireshark while using OFFTYPE for voice input. The filter `ip.addr != 127.0.0.1 && ip.addr != ::1` shows all external network traffic. The result speaks for itself.
Filter: `ip.addr != 127.0.0.1 && ip.addr != ::1` · 1,683 packets captured · 0 external connections during speech recognition
How your data flows
Four stages. All happen in-process on your device.
Microphone captures audio
Audio is captured by the system microphone and sent directly to the local speech server. It never passes through any network interface.
Local ASR model converts speech to text
The ASR model (SenseVoice-Small or Qwen3-ASR) runs entirely in-process. The audio never leaves RAM.
Local AI adds punctuation and corrects
Optionally, a local 1.5B-parameter model adds punctuation, translates, or rewrites the text — all processed locally.
Text is injected at your cursor
The final text is injected directly into the target application via the clipboard. It is never sent anywhere else.
Privacy isn't a setting. It's the architecture.
There is no opt-in. There is no tracking. There is no cloud.
Download for Windows