Settings that follow you from device to device, a realtime voice you can choose for a wake word, and a camera that can tell a person from the cat.
- Share settings between devices. Pick a setting on the setup page (bass, the clock's look, the night hours, the slideshow and more, 61 in all) and tick the other devices that should keep it in step. Each device has to agree before it links, and a change on one reaches the rest within seconds. Nothing is shared until you turn it on, and nothing that could quiet an alarm or the microphones can be shared.
- Realtime voice (experimental). A wake word can be answered by a speech-to-speech model at OpenAI, xAI or Google Gemini, in place of its usual assistant. It sends what you say, as audio, to that company, and they bill for it. Off until you choose a wake word for it on the setup page; a daily limit in minutes; it can use your Home Assistant's own tools, and it is locked out of changes after reading outside text such as a calendar or the web. Not yet tried against the providers' own servers.
- Person detection (experimental, Show and Spot). With presence detection on, the camera can tell a person from anything else that moves, and Home Assistant gets a Person sensor. It runs on the device; no picture leaves it. Off by default, and not yet tried on a device.
- The on-device model code is hardened against malformed model files on 32-bit processors.
Untried on hardware: the three features above. Please open an issue if something doesn't work.