jxlarrea/voice-satellite-card-integration

What it solves

Voice Satellite turns any web browser—on a tablet, phone, or computer—into a hands-free, private voice assistant for Home Assistant. It eliminates the need for dedicated physical hardware like Alexa or Google Home by allowing any device with a browser to act as a voice-activated satellite for controlling devices, asking questions, and receiving announcements.

How it works

The project operates as a global engine that loads across all Home Assistant pages. Once a satellite entity is assigned to a browser, it begins listening for wake words. It supports three on-device wake word detection engines:

• vsWakeWord: A WebGPU-accelerated engine specifically trained for wall-mounted tablets and far-field capture.
• microWakeWord: A CPU-based engine that works on all devices with low latency.
• openWakeWord: A WebGPU-accelerated engine with a broad pre-trained keyword library.

When a wake word is detected, the browser streams audio to Home Assistant's Assist pipeline (STT, conversation agent, and TTS). The system also exposes the browser as a media_player entity, allowing it to play audio, video, and camera streams full-screen.

Who it’s for

Home Assistant users who want to deploy voice-activated control across their home using existing tablets or phones, particularly those using wall-mounted kiosks.

Highlights

• On-device wake word detection: Three different engines to balance performance and compatibility across hardware.
• Dual wake words: Ability to load two different wake words simultaneously, routing each to separate Assist pipelines.
• Media integration: Full-screen playback of local video, HLS/MJPEG camera streams, and TTS audio.
• Customizable visuals: 10 built-in skins and CSS override support for the voice overlay and activity bar.
• Kiosk Satellite companion app: An Android app that allows wake word detection to run natively, enabling the microphone to work even when the screen is off.
• LLM Tool support: Experimental ability to display rich visual results (weather, stocks, Wikipedia) inline when paired with a tool-capable LLM.