Every time an Alexa speaker hears its wake word, a recording leaves your house, crosses the internet, and gets processed on servers you do not control. That is a lot of trust to hand a small speaker sitting in your kitchen or bedroom. Home Assistant Voice Preview Edition is the Home Assistant team’s first hardware built to replace that arrangement, and it can run the entire voice chain inside your home.
The device is a compact puck with a speaker, dual microphones, a rotary dial ringed with LEDs, and a physical mute switch. It launched at around $59, and it handles wake word detection on the device itself. This guide covers pairing it, choosing local speech engines, and exposing the right devices so that recognition, command handling, and spoken replies all happen on hardware you own.

What You Need Before You Start
Voice PE is built around an ESP32-S3 chip paired with a dedicated XMOS audio processor that handles echo cancellation and noise suppression. It needs USB-C power and a 2.4 GHz Wi-Fi network, and it talks to your Home Assistant server over your local network. You also need a reasonably current Home Assistant release, since voice features have been changing quickly. If you have not built that foundation yet, setting up Home Assistant as a self-hosted smart home hub comes first.
The decision that shapes everything else is where speech-to-text runs. Whisper, through the faster-whisper add-on, is the accurate option, and it understands natural phrasing and handles accents well. It is also hungry for processor time, so on a Raspberry Pi the pause between finishing a sentence and hearing a response can stretch uncomfortably. Speech-to-Phrase takes the opposite approach. It listens only for phrases built from your actual device names, areas, and common commands, which makes it fast even on modest hardware. The cost is flexibility: ask it something outside its trained patterns and it will stumble. A small mini PC with an Intel N100-class chip generally has enough headroom for Whisper’s smaller models.
For spoken replies, Piper is the local text-to-speech engine, and it is also installed as an add-on. It offers a range of voices and quality levels, and the higher-quality voices sound noticeably more natural at the cost of slightly slower responses. Home Assistant Cloud offers a hosted alternative for both speech tasks, but it sends audio off your network, which defeats the purpose of this project.
Pairing the Device
Plug the puck in and wait for the LED ring to indicate it is ready. Home Assistant should discover it automatically through ESPHome, and a notification appears under Settings, then Devices and Services. Click Add, follow the prompts to connect it to your Wi-Fi, and accept any firmware update it offers before doing anything else. The default wake word is “Okay Nabu,” and you can switch to another option in the device’s settings. Detection runs on the puck itself, so nothing is streamed to the server until the wake word fires.
Test it by saying the wake word and asking it to turn on a light. If nothing happens, check the mute switch first.

Wiring Up a Fully Local Voice Pipeline
Open Settings, then Voice assistants, and add a new assistant. Install the Whisper or Speech-to-Phrase add-on and the Piper add-on from the add-on store, and they will register through the Wyoming protocol. Back in the assistant setup, set the conversation agent to Home Assistant, pick your speech-to-text engine, pick your Piper voice, and set the language. Then open the Voice PE device page and select that assistant in its pipeline dropdown. From that moment the chain runs entirely on your network.
The next step is where most people lose an evening. Assist can only control entities you have exposed to it, which you manage in the Expose tab under Voice assistants. Resist the urge to expose everything. A cleaner list means fewer misheard commands, and it keeps door locks and alarm panels out of reach of a stray phrase. Naming matters as much as exposure: “living room lamp” is easy to say, while a default entity name like “light.lr_lamp_2” is not. Adding aliases lets one device answer to several names, and assigning devices to areas means the puck knows which room it is in. That is why “turn off the lights” can affect only the room where you are standing.
If you want conversation beyond fixed commands, Home Assistant can hand requests to a local language model through the Ollama integration. The sensible configuration keeps the built-in intent handler first, so “turn on the porch light” is resolved instantly and only unrecognized requests fall through to the model. The trade-off is hardware. Running a model well generally calls for a capable GPU, and smaller models can be unreliable when asked to control devices, so treat this as an experiment rather than a requirement.

Be honest with yourself about what you are giving up. There are no Alexa skills, no built-in music catalog, and no shopping-by-voice. Timers work, to-do lists work, and music playback becomes possible through Music Assistant, but you will be assembling those pieces yourself. The speaker on the puck is also modest, which is why the 3.5 mm audio output matters: plug it into a decent powered speaker and the device’s weakest component stops being the one you hear.





