Voice & music, without the cloud assistant
You can replace an Alexa-style assistant with one that runs in your own house: wake word on the device, commands processed locally, and only genuine questions ever needing a cloud AI — your choice, not a default.
How a voice room is built
Each room gets a small satellite (microphone + wake word) and a speaker. They can be one box or separate — separate is usually better value, because the satellite can be cheap while the speaker is chosen for sound. Replies and timer chimes play on the room speaker; music is handled by Music Assistant, below.
Satellites, by budget
- Best value: Seeed ReSpeaker Lite kit, about $38 complete — the same XMOS audio processor as the official satellite, with light assembly.
- Default pick: Home Assistant Voice Preview Edition, $58.95 — ready-made, first-party supported, line-out to a real speaker.
- Best ears for noisy rooms: premium satellites around $69.99 with stronger far-field microphone arrays.
- One-box option: FutureProofHomes Satellite1.1, $132.99 — satellite and speaker in one unit; its Sendspin multi-room support has shipped in firmware v0.2.x.
The speech pipeline that feels fast
Split the work: everyday commands ('turn off the kitchen lights') go to Speech-to-Phrase with Piper for replies — near-instant and fully local. Open questions go to Whisper (small, int8) for transcription, the AI you choose for the answer, and a natural local voice such as Kokoro via Speaches to speak it. That split is the difference between an assistant that feels snappy and one that pauses for the cloud on every sentence.
Music Assistant
Music Assistant is the music brain: it connects services like Apple Music and internet radio streams, groups speakers from different brands, and plays through affordable streamers (from about $34.99) or speakers like the microphone-free Sonos Era 100 SL (about $149 on holiday promotion). Speakers are the most price-sensitive part of any build — a $30 smart speaker you already own is a perfectly good starting point while you decide.
Does your voice leave the house?
With the pipeline above: commands, no. Wake word and command processing are local. Only questions you route to a cloud AI send text (not raw audio, unless you choose a cloud speech service) — and you can run a local model instead, even without a gaming PC, as the guide's AI chapter explains.
Next: switches & sensors · Build list & costs