Real-time conversation

Conversation mode turns the chat into a call: you talk, the agent listens while you speak, thinks and answers out loud. You can cut it off mid-sentence and keep talking, just like with a person. It is not dictation (record → transcribe → send): audio travels both ways at the same time.

Under the hood it uses WebRTC with a LiveKit server running on your own machine, so the audio never leaves your computer beyond whatever the model you picked already sends.

Turning it on

1. Flip the switch.

   Go to Settings → Voice and reading and enable “Real-time conversation (LiveKit)” (it ships turned off). It’s full-duplex voice over local WebRTC: interruptions, turn detection and real streaming.

2. Pick the microphone.

   In the chat composer, the Microphone selector lists the detected devices (System default if you’d rather not choose). If you just plugged in a headset, press Detect again.

3. Pick the voice it will answer with.

   It is the same read-aloud voice (Voice): the engine and voice you have configured there are the ones that speak in the conversation — Edge TTS, Groq, Kokoro, Pocket or Chatterbox. You can switch it live, without hanging up: the next sentence already comes out with the new voice.

4. Press the conversation button in the chat (“Real-time voice conversation (Live Mode)”).

   Since 3.6.0 there is nothing to download: the voice engine ships inside the installer. It starts in seconds.

5. Talk.

   The agent greets you with “I’m listening.” and the conversation begins.

Conversation mode: status, waves and live captions of what you say and what it answers.

Conversation mode: status, waves and live captions of what you say and what it answers.

What you see while you talk

The conversation window keeps telling you where it is:

Status	What it means
Connecting to the voice agent…	Starting the local server and the agent. A few seconds.
Listening to you…	The microphone is open and it’s your turn.
Processing…	You finished speaking and the model is thinking.
Speaking…	The agent is answering out loud.

On top of that:

• Live captions: what you say and what the agent answers show up in writing as they happen, without waiting for the end of the sentence.
• Voice interruption: if you start talking while the agent is speaking, it goes quiet instantly and switches to listening (“I’m listening…”). Nothing to press.
• “Are you still there?”: after about 20 seconds of silence, the agent asks whether you’re still there instead of staying mute.
• Retry: if the connection fails, the panel doesn’t hang — a Retry button appears.
• End conversation closes the call and returns the chat to its normal mode. What was said stays in the chat like any other message.

Note

The agent knows it is talking by voice: it answers in sentences meant to be listened to, and when the answer includes code it doesn’t dictate it — it says “I’ll leave the code on screen.” and writes it in the chat.

What it needs

Nothing to install or download. Since 3.6.0 the Windows, Linux and macOS installer carries the whole voice engine inside:

• The Node interpreter and the voice agent, packaged for your platform.
• The LiveKit server (version 1.13.7), listening on port 7880 of your own machine.
• The models: the Silero voice detector and the end-of-turn detector travel inside the package. Recognition uses the same local Whisper as dictation.

That’s why Conversation mode works with no internet connection. The only things that need the network are the AI model that answers (unless you use a local one) and, if you pick Edge TTS or Groq as the voice, the synthesis of that voice.

What you do need:

• A connected AI provider: the “brain” of the conversation is the same model as the chat. If there’s no provider, voice mode warns you: “Dictation needs a connected AI model. Connect a provider and try again.”
• A configured TTS engine for the spoken answer — see Voice.

Note

If you’re coming from an installation older than 3.6.0, the engine you had already downloaded to ~/.context/livekit/ keeps working: the app prefers the one shipped in the installer and uses that one as a fallback.

It works in the 17 languages of the interface: recognition and the greeting follow the language you have set in the app.

Tip

Also from your phone: with the mobile app on the same WiFi network you join the same conversation, with captions, pause and voice commands that act on the phone itself — see Mobile app.

Tip

Prefer a conversation without the main window in front of you? The Context Voice widget leaves a floating orb always on top, and it also understands commands for the app (“create a new chat”, “switch to model X”).

If something goes wrong

“the LiveKit server didn’t start” : Press Retry. If it keeps happening, something is occupying port 7880 or an antivirus is blocking the server binary; restart the application.

“the voice agent closed before registering” / “the voice agent didn’t register within 60 s” : This happened when starting twice in a row or when an orphan process from a previous attempt was left behind. Close the conversation, wait a few seconds and press Retry; if it persists, restart the application.

“this platform has no official livekit-server build” : Conversation mode needs an official LiveKit binary for your system and architecture. On that machine you still have the usual dictation and read-aloud.

It triggers on its own with keyboard noise : The voice detector is calibrated to ignore the clicks of a mechanical keyboard (a real word lasts longer than 250 ms, a click doesn’t). If it still fires, move the microphone away from the keyboard or use a headset with a mic.

The microphone shows up but no audio comes in : “Virtual” microphones (the ones from a capture card, some software ones) publish silence. Pick another device in the Microphone selector. On Windows, if access is blocked, the app says so: “Windows blocked access to the microphone. Go to Settings → Privacy & security → Microphone…”.

I picked a Groq voice and it answers with a different one : Some Groq voices require accepting their terms in the Groq console before they can be used. If the configured engine fails, the app doesn’t hang up the call: it synthesizes that sentence with Edge TTS so you can keep talking. Accept the terms in your Groq account and the chosen voice will sound again.

The microphone is in use by another application : Close it (another video call is usually the culprit) and try again.