Voice
Two modes, one conversation loop.
Modes
| Mode | Behaviour |
|---|---|
| Manual (default) | Click Speak and talk. Recording ends after 5 seconds of silence, Jarvis answers, then the mic is released. Hold Space for push-to-talk. |
| Active listening | The mic stays open for an on-device wake-word detector only. "Hey Jarvis" wakes it at once; a bare "Jarvis" is confirmed locally. After a reply the conversation stays open for follow-ups without the wake word. It ends when you say goodbye or after 10 seconds of silence. |
In Active listening, no audio leaves your Mac for the wake word: the check runs with a local speech model.
The wake-word models are installed separately with apps/desktop/scripts/install-wake-models.sh into the app's user data folder. Manual mode needs nothing extra.
Languages
Settings, then Language. Choose a preferred language and the allowed ones (default Portuguese, English, Spanish). Each turn Jarvis decides the language from your words and the recogniser, and replies in it.
Speech engines
| Stage | Default | Fallback |
|---|---|---|
| Speech to text | Deepgram, when you have added a key | Local whisper.cpp (also used offline) |
| Reply voice | ElevenLabs, when you have a key | Kokoro (English), then the macOS voice |
The reply voice speed is adjustable in Settings (0.7 to 1.2).
Interrupting
Speak over Jarvis and it stops, keeping the part you heard. Esc stops it too.
Decision engine (optional)
Settings, then Decision engine (JEV), can classify a finished utterance (stop, pause, continue, ignore) before the model answers. It is off by default and needs its own key.