Tap the mic once and have a normal spoken conversation. You see your words appear while you speak, it notices when you have finished, answers, and reads the answer aloud. Everything below runs on non-US models, and your voice recording never leaves your device.
The first time, the browser downloads the Moonshine speech model (about 155 MB; smaller and Whisper options in Settings). It is cached afterwards, so later visits start instantly. The badge at the top says whether it runs on your GPU (fast) or CPU.
As soon as you start talking, a pulsing placeholder bubble appears. Your words fill it only once the speech model has finished, so you never see half-heard guesses. The reply gets the same placeholder while Mistral is thinking.
A voice-activity detector learns your room's background noise and watches the mic level. Once you have been quiet for the "End of turn" time (700 ms by default), your turn is over. Short pauses mid-sentence do not cut you off; raise the slider if you pause a lot.
Each word lights up as it is spoken, like karaoke, so you can read along. Mistral streams the answer word by word. As soon as a full sentence arrives it goes to Fish Audio, so the voice starts before the whole answer is written. If one Mistral model is busy, the next one takes over automatically.
Talk over the reply and it drops in volume at once; as soon as your words are recognised it stops and answers you instead. A cough or a door slam only dips it for a moment. While it is speaking, the mic needs a louder voice to trigger, so its own voice through your speakers does not interrupt it. Headphones work best. You can switch interrupting off in Settings.
| Data | Goes to | Kept? |
|---|---|---|
| Your voice recording | Nowhere | Transcribed in your browser, then discarded. |
| What you said (text) | Mistral, France | No. Sent only to zero-data-retention endpoints served by Mistral itself. |
| The reply text | Fish Audio | No. Zero-data-retention endpoint, used only to make the voice. |
| Reply audio | This tab, briefly | Each spoken sentence is deleted from memory the moment it finishes playing. |
| Conversation | Nowhere | Private mode (on by default) forgets each exchange once it is answered: no earlier context is sent, and the previous exchange disappears from the screen when the next one starts. Nothing is written to disk, a server or a database. |
The OpenRouter key stays on the server, behind the login, and the browser never sees it. The speech model itself is downloaded once from Hugging Face; nothing you say is sent there.