# Voice **Record a whole meeting with a live transcript, hand off tasks by voice, or keep talking to an agent hands-free.** ![A meeting recorder with a live transcript, an Ask tab that answers from what was said and sends requests to Another Me, and the recording, transcript and follow-up draft landing in the Another Me chat](../assets/feature-voice.jpg) ## Turning voice on Voice runs wherever the deployment can run it: TeamHub's `TEAMHUB_VOICE_MEDIA` is `on` (the default), `LINGER_SECRET_SEED` is set, and `ffmpeg` and `ffprobe` are on `PATH`. `TEAMHUB_VOICE_MEDIA=off`, a missing seed, or missing ffmpeg keep voice off; TeamHub still starts and logs `voice.media_unavailable` for the last two. Each feature then needs its provider key: [Soniox](https://soniox.com) for recording and dictation, [xAI](https://x.ai) for Grok Live voice mode; either works without the other. Set a key in `SONIOX_API_KEY` or `XAI_API_KEY`, or save it in the app when voice asks for it. A saved key applies at once, without a restart; an environment key wins over a saved one and cannot be replaced from the app. Node Services starts the matching voice workers within seconds of a save and stops them when a key goes away. Saved keys are shared by everyone on the server, and any signed-in person can save or replace one. Live transcription and archive recovery send the audio to the Soniox account whose key is in effect, so on a server several people share, set `SONIOX_API_KEY` and `XAI_API_KEY` in the environment: that locks the keys against replacement from the app. - The Linger app always runs the LiveKit server and ffmpeg it bundles; save the keys in the app. Keys written by hand into `user.env` in the instance folder (`~/.simplexai/instances/desktop/`) act as environment keys: they win over the app, and changing or removing one means editing that file and reopening the app. A computer the app has no bundled voice tools for keeps voice off. - Docker Compose from this checkout always runs LiveKit; save the keys in the app, or set them in `.env` and run `docker compose up --detach --wait` again. - `pnpm dev` always starts a LiveKit server on this computer, with the ports and keys `.env` names. Install it first (`brew install livekit` on macOS, `curl -sSL https://get.livekit.io | bash` on Linux); recording also needs `ffmpeg` and `ffprobe` on `PATH`. Without `livekit-server` on `PATH`, `pnpm dev` says so and voice cannot connect. A Native installation's LiveKit and ffmpeg listen on `127.0.0.1` only (one TCP port for signalling, one UDP port for media), so no media leaves the computer except to the speech services (Soniox, xAI) and the voice model Node Services downloads when its voice role starts. Browsers allow the microphone only on `localhost` or HTTPS, so voice from another computer needs an HTTPS proxy in front of port 3000. ## When a recording ends As soon as a recording stops, Another Me posts the recording and its transcript in its chat. A meeting summary follows under them a few seconds later: a title and a few topics with short bullets, written from the transcript by your Another Me's model. Grok Live reads the summaries of the recordings among the chat's recent messages, so it can answer about a meeting that just ended without opening its transcript. The full meeting notes (summary, decisions, action items) take a few minutes more and appear in Library › Meetings; the card says so and links to them once they are written. If the summary cannot be written, the recording and transcript stay as they are.