Voice Mode gives LumaBrowser a spoken conversation loop where both halves run on your own machine. Whisper listens, your local model thinks, and a neural voice answers out loud. There is no cloud speech service in the path, no GPU requirement, and no recording of your microphone anywhere off the computer.
Click the microphone button in the chat composer and the loop starts. From there it is hands-free: the browser decides when you have finished speaking, transcribes, answers, and listens again.
A local voice detector watches the microphone and adapts to the room's noise floor, so it triggers on speech rather than on a fan or a keyboard.
A rolling one second buffer sits in front of the detector, so the beginning of your sentence is never clipped off while it decides you were talking.
After a short pause the utterance is closed and sent to transcription. Very short blips and known silence artefacts are discarded rather than sent to the model.
A live transcript bubble above the composer updates while you talk, so you can tell it is hearing you correctly before you finish.
The reply is spoken sentence by sentence as it is generated, so audio starts before the model has finished writing the whole answer.
Start talking while it is speaking and it stops mid-sentence, drops the queued audio, abandons the turn, and listens to you instead.
The thing that makes a voice assistant feel broken is having to wait politely for it to finish. Voice Mode is built around cutting in.
Speaking while it talks stops playback immediately, cancels the speech still being synthesized, and aborts the model's turn.
Cutting in also works during the quiet stretch while the model reasons or runs a tool, not only once audio has started.
The words that triggered the interruption are captured from the pre-roll buffer, so what you said becomes the next message rather than being thrown away.
While speaking, the bar to interrupt is deliberately raised and the moment after each clip starts is ignored, so the assistant's own voice does not cut itself off.
In voice mode the model is steered toward short spoken sentences, and tables, bullets, headings, and code blocks are kept out of the audio.
Chain-of-thought is disabled automatically for spoken turns, because silent reasoning is dead air when you are waiting for a voice to answer.
Most voice assistants stream your microphone to a data centre. This one does not have a network call in the path at all. Here is every hop your voice takes.
The audio never leaves the machine. The transcribed words follow your model choice. Once your speech becomes text, it is sent to whichever chat model the conversation uses, exactly like a message you typed. Pair Voice Mode with a local model and the entire conversation stays on your computer end to end. Point it at a hosted provider and the text goes there, as any typed message would. The recording itself is never uploaded in either case.
The only network traffic this feature generates is the one-time download of the engines and voices when you first set it up.
Both halves are open source and downloaded on first use from the setup popover. You pick a voice quality tier; everything else is chosen for you.
| Piece | What runs | Notes |
|---|---|---|
| Listening | whisper.cpp, with Whisper Small as the default model | Multilingual out of the box. A smaller English-only model and a larger, more accurate model are also available. |
| Speaking, Low | Piper LibriTTS-R, English | The fastest option and the smallest download. A good choice on modest hardware. |
| Speaking, Medium | Kokoro 82M, the default | Good quality across nine languages, balanced against speed. |
| Speaking, High | Kokoro 82M English, full precision | The best sounding option, English only. |
Speech synthesis runs on the CPU by design, so it never competes with your language model for video memory. Speech recognition also defaults to the CPU, with an optional NVIDIA build on Windows for faster transcription. Both engines unload themselves after a few minutes idle.
Voice Mode is deliberately cheap to run. It asks for no video memory at all, so it sits alongside a large local model without taking anything from it.
One-click setup for both engines. An optional NVIDIA build speeds up transcription if you have a CUDA card, though a plain CPU install works fine.
One-click setup for both engines on x64. Transcription runs on the CPU.
The speech engine has no prebuilt macOS binary upstream, so Voice Mode cannot be set up on a Mac today. Every other local AI feature in LumaBrowser works normally there.
Neither engine needs a GPU and neither reserves video memory. Speech synthesis is CPU-only by design so your language model keeps the whole card.
Voice Mode is new, and the feel of it is still being tuned. Echo handling and the pause length that ends your turn are the two things most sensitive to your room and microphone. Headphones or a headset give the most reliable experience, since they remove the assistant's own voice from the microphone entirely.
It lives in the AI tab. The microphone button is in the chat composer there. The browser side panel and the shared web client do not have it.
Voice picking is by quality tier. You choose Low, Medium, or High, which selects the underlying voice model. Choosing an individual speaker or adjusting the speaking rate is not exposed yet.
Voice Mode ships with LumaBrowser, free to download. Pair it with a local model and you get a spoken assistant that works on a plane, behind a firewall, or anywhere you would rather not send your voice to somebody else's server.