Voice Mode
Voice mode turns a chat into a hands-free, turn-based conversation: Libre listens, transcribes what you said, generates a reply, and speaks it back — then listens again. Interrupt a spoken reply just by talking over it. Everything runs through the same speech-to-text and text-to-speech contracts as the composer's dictation and read-aloud features; voice mode adds only orchestration, never a new audio path.
Starting a conversation
Open a chat and press the voice-mode button next to the microphone in the composer. The button appears whenever speech input is available: either a provider speech-to-text model (the same catalog as composer dictation, documented in the speech-to-text page) or the browser's own speech recognition. Spoken replies use your configured text-to-speech model and voice — including a saved voice with active consent — and are primed against browser autoplay policy by the opening gesture itself.
The conversation loop
One turn is active at a time, moving through four visible phases:
| Phase | What is happening |
|---|---|
| Listening | The microphone records; a pause of about 1.5 s ends your turn |
| Transcribing | The recording is transcribed by the selected provider or the browser |
| Thinking | The transcript is sent as a normal chat message and the reply generates |
| Speaking | The reply plays through your text-to-speech settings |
Controls while the overlay is open:
- Done speaking ends your turn immediately — useful in noisy rooms or when no audio analysis is available for automatic endpointing.
- Mute pauses capture without leaving the conversation; unmuting starts a fresh turn.
- Skip reply stops playback and returns to listening.
- Barge-in: start talking while the reply plays and playback stops, returning to listening for your next turn.
- Close tears the session down: the microphone stream stops, any in-flight transcription aborts, and playback cancels.
Failures never end the conversation. A microphone error, a rejected recording, a transcription failure, or a reply timeout surfaces inline and voice mode returns to listening for the next turn.
Governance and privacy
Voice mode is governed by its own access mode (voice-mode) in User
Management's Voice access card, separate from speech-to-text,
text-to-speech, and voice cloning; the VOICE_MODE_ACCESS_MODE
environment variable pins it. Transcription and speech synthesis run
under their own feature gates too, so restricting either disables that
part of the loop. Recordings follow the speech-to-text contract:
validated, bounded, and ephemeral — a turn's audio exists only long
enough to transcribe it, and only the transcript enters the chat as a
normal message.
Boundaries
- Voice mode is turn-based, not full-duplex: it does not stream microphone audio and model speech simultaneously over one realtime connection.
- Automatic endpointing and barge-in need audio analysis; when the browser cannot provide it, the manual Done speaking control ends the turn.
- Turn-based replies run one at a time; words spoken during a barge-in are not carried into the next turn's transcript.