Omniscio documentation
Browse all documentation
  1. Getting Started13
  2. Sessions & Agents115
  3. Inbox & Notifications64
  4. Projects & Tasks95
  5. Automation & Scheduling81
  6. Knowledge & Memory26
  7. AI Features61
  8. Integrations100
  9. Plugins & Marketplace33
  10. Cloud & Teams56
  11. Settings & Customization59
  12. Account & Billing28
  13. Troubleshooting85
  14. CLI & API Reference22
  15. Legal & Policies4
  16. Uncategorised22

Wake word detection

Wake word detection arms the voice system hands-free: say a hotword and Omniscio starts listening. It is off by default, needs the master Voice toggle, and its bundled keyword heads are a known accuracy gap on real speech.

What it is

Wake word detection is an always-listening hotword that arms Omniscio's voice system without a keypress. Say a phrase and the app begins listening, so the keyboard never has to be in reach. It is off by default, so it is strictly opt-in, and it needs the master Voice toggle on to do anything at all.

Where to find it

Settings -> Voice Control -> Voice Input tab -> the Wake Words card, where the engine picker, keyword selector, sensitivity, action mode, focus-only switch and a Test button live. Settings search answers to "wake word". It requires both wakeWordEnabled and voiceEnabled; turning it on while Voice is off shows a master-gate banner instead of silently doing nothing.

How it behaves

  • Two engines, one live. The bundled free engine (openWakeWord, no key and no account) is the default; the paid Picovoice engine is retired and hidden from the picker, and a stale Picovoice selection silently falls back to the free engine.
  • The bundled keywords are computer and hey-omniscio, with hey-omniscio ticked by default; the retired jarvis keyword is dropped on load.
  • Action mode is a picker: command (the default — the transcript is parsed and dispatched through the same intent parser Alt+V uses) or dictation (typed into the active session).
  • Focus-only listening is on by default; sensitivity defaults to 0.7; audio is fed in 16 kHz mono batches with a rate limit of 8 batches per second.
  • Known limitation, stated plainly: neither surviving keyword head scores near the 0.60 fire line on real speech, so hands-free wake is a known gap until a head is trained from real voice clips — synthetic text-to-speech training does not transfer.

For agents

  • Service: src/main/services/wake-word-service.ts.
  • Free engine: src/main/services/wake-word/openwakeword-engine.ts.
  • IPC: src/main/ipc/wake-word-handlers.ts.

Related

Last verified 2026-10-06