The voice loop
A conversation with ORBIS is a loop: it hears you, understands what you said, thinks about it, and speaks back — then waits for your next turn. This page explains the shape of that loop and the ideas that make it feel like talking to someone rather than typing at a prompt.
For the knobs, see the Voice reference; for fixes, see Voice isn't working.
Hearing
Capture. The native audio engine reads your microphone with low latency. ORBIS is Apple-Silicon-only partly because it uses native audio instead of the browser — it's what makes the round-trip fast and reliable.
Echo guard. The orb is also playing audio (its own voice). Echo cancellation removes that from the input so ORBIS doesn't hear — and reply to — itself.
Turn-taking (VAD). Voice-activity detection watches for you to start and stop speaking. ORBIS doesn't transcribe a fixed window; it waits for a natural pause, then treats what you said as one turn. That's why a brief pause mid-sentence is fine, but a long silence ends your turn.
Transcription (STT). The captured turn becomes text. With the local Whisper/Parakeet backends this is segmented — it happens after you stop talking. (Streaming backends, which transcribe as you speak, are a future option.) See STT backends.
Understanding & thinking
The text goes to the language model, which acts as a router first. For each turn it decides among:
- Just answer — a normal spoken reply.
- Use a tool — e.g. set a reminder, list reminders.
- Delegate — hand a heavier task to one of your configured agents and narrate the result when it returns.
While it's working it may emit short acknowledgements ("on it…") so the orb isn't silent during a slow tool call or delegation. These are deliberately brief and don't become part of the conversation.
Speaking
The reply text is synthesized by the TTS backend into the orb's voice and played back through the native engine. The orb's visual state tracks the loop — idle → listening → thinking → speaking — so you can see where it is.
Barge-in
You don't have to wait for ORBIS to finish. If you start talking while the orb is speaking, the barge-in gate stops playback and starts listening — the same way a person stops mid-sentence when you cut in. (Some barge-in features are conservative by default to avoid the orb interrupting itself on its own audio tail.)
Why it's local
Everything above — capture, echo cancellation, VAD, and the default STT/TTS — runs on your Mac. Audio doesn't leave the machine unless you point STT or TTS at a hosted endpoint. That's the single-owner, on-your-hardware posture ORBIS is built around.