Disclosure: Some links in this article are affiliate links. We may earn a commission on qualifying purchases at no extra cost to you.
Setting up voice control in Home Assistant used to mean stitching together a Raspberry Pi, a cheap USB microphone, and a pile of YAML. The Home Assistant Voice Preview Edition changes that: it's the first piece of hardware built by Nabu Casa specifically to give the platform a proper far-field microphone and a real speaker, wrapped in a case designed to sit on a shelf rather than behind a monitor.
It is not a smart speaker in the Alexa or Google Home sense. It has no cloud account of its own, no general-knowledge assistant baked in, and it does nothing at all without a running Home Assistant instance behind it. What it does well — local wake-word detection, hardware-level microphone muting, and per-room voice routing — depends entirely on how its voice pipeline is configured after unboxing.
This guide covers the full setup process, the pipeline decision that determines whether the device is fast or sluggish, what it can and can't do out of the box, the bugs that show up most often in the community, and how it stacks up against both mainstream voice assistants and other Wyoming-protocol satellites.
It's built for people who already run — or are about to run — Home Assistant and want a privacy-respecting way to talk to it, not for someone shopping for a general-purpose smart speaker.
What Makes the Voice Preview Edition Different
The Voice Preview Edition (Voice PE) is built around an ESP32-S3 microcontroller paired with an XMOS XU316 audio processor — the same class of chip used in commercial far-field voice products, handling hardware echo cancellation, stationary noise removal, and automatic gain control before audio ever reaches software. That's the piece most DIY satellites (a Raspberry Pi with a USB mic, or an M5Stack Atom Echo) skip entirely, and it's the reason Voice PE's wake-word detection holds up better across a room than earlier community builds.
The firmware is fully open source, built on ESPHome, and communicates with Home Assistant through the Wyoming protocol — the same protocol used by third-party satellites like the FutureProofHomes Satellite1. Nothing about the device is locked to Nabu Casa's cloud: it can run entirely on-device wake-word detection and hand audio to a fully local speech-to-text pipeline, or it can lean on Home Assistant Cloud for faster, more accurate transcription. That choice is the single biggest factor in how the device performs, and it's covered in detail below.
A physical mute switch cuts power to the microphones at the hardware level — verifiable with a multimeter, not just a software toggle — which is the kind of detail that matters to the privacy-first audience this device is aimed at. The multicolor LED ring doubles as a light entity inside Home Assistant, so it can reflect automation state, not just listening status.
The device launched in December 2024 and sold out within 23 minutes of its announcement. No hardware revision has followed; every improvement since has arrived through ESPHome firmware updates, including a reduction in peak CPU usage from 72% to 35% during simultaneous music and announcement playback.
Hardware & Specs
| Component | Specification |
|---|---|
| SoC | ESP32-S3, 16 MB flash, 8 MB octal PSRAM |
| Audio processor | XMOS XU316 (echo cancellation, noise removal, auto gain control) |
| Microphones | Dual MEMS array with XMOS-driven beamforming |
| Speaker | Internal mono speaker (voice feedback, not music-oriented) |
| Audio output | 3.5 mm stereo jack via TI AIC3204 DAC, 48 kHz |
| Physical controls | Multipurpose button (single/double/triple/long press), rotary dial, hardware mute switch |
| Feedback | Multicolor addressable LED ring, exposed as a light entity |
| Radios | 2.4 GHz Wi-Fi, Bluetooth 5.0 |
| Power | USB-C, 5V/2A (cable and power supply not included) |
| Expandability | Grove port, exposed PCB pads |
| Dimensions / weight | 84 × 84 × 21 mm, 96 g |
| Enclosure | Injection-molded polycarbonate |
| Operating range | 0–30°C (32–86°F), non-condensing |
| Firmware | ESPHome, open source, OTA-updated |
| Out-of-box wake words | "Okay Nabu," "Hey Jarvis," "Hey Mycroft" (on-device microWakeWord) |
Recommended for: Check Price on Amazon.
The device sits in entry-level pricing territory for purpose-built voice hardware, and it has a history of stocking out at official distributors — if the listing above shows unavailable, it's worth checking directly on Amazon rather than assuming the product has been discontinued.
Setting Up the Voice Preview Edition — Step by Step
Voice PE has no standalone mode. Before unboxing it, a running Home Assistant instance (Core, Supervised, or OS) updated to a current release is required, along with an admin account and the Home Assistant Companion app installed on a phone with Bluetooth permission granted — this is used for onboarding only and can be revoked afterward.
- Power it on. Connect USB-C to a 5V/2A supply (not included). A warm-white twinkle animation indicates the device is ready to pair. If the unit was previously linked to another Home Assistant hub, reset it first — it stores an encryption key tied to that original hub.
- Find it in the Companion app. The device boots into Improv-over-BLE mode and appears under Settings > Devices & Services > Discovered as "home-assistant-xx Improv via BLE." Tap Add, then Submit.
- Join Wi-Fi. Enter the SSID and password for a 2.4 GHz network — if the router auto-steers devices to 5 GHz, select the 2.4 GHz band explicitly.
- Authorize on the device. Press the center button when prompted, confirming physical possession of the hardware.
- Add the ESPHome node. Under Discovered, add "Home Assistant Voice XXX" and submit. The wizard checks for firmware updates at this stage, which accounts for most of the two-to-four-minute setup time.
- Choose the voice pipeline. This is the step that determines speed and accuracy — see the breakdown below.
- Expose entities and talk. Devices must be exposed to Assist under Settings > Voice assistants > Expose before they respond to voice commands, and area or entity names have to match what's spoken (or have an alias configured).
Anyone setting up their first Zigbee coordinator or installing HACS alongside Voice PE should expect a similar amount of one-time configuration — none of it is difficult, but skipping the entity-exposure step is the most common reason a fresh setup "doesn't respond to anything."
Choosing Your Voice Pipeline — Cloud, Speech-to-Phrase, or Local Whisper
The hardware is identical across every Voice PE unit; what changes performance is entirely software, and Home Assistant asks which pipeline to use during the "What hardware do you want to use?" step of setup. There are three practical paths.
Home Assistant Cloud. Speech-to-text and text-to-speech run on Nabu Casa's servers while intent matching stays local. This is the fastest and most accurate option for low-power hosts, and it covers the broadest set of languages. End-to-end latency runs roughly 500–900 ms. It requires a Nabu Casa subscription.
Speech-to-Phrase (local, low-power hardware). Runs sub-one-second on a Raspberry Pi or Home Assistant Green, entirely offline, but recognizes only a preset list of home-control phrases — no shopping lists, no named timers, no open-ended requests. It launched covering English, French, German, Dutch, Spanish, and Italian, which the Home Assistant team estimated covers close to 70% of its user base.
Local Whisper + Piper (mid-to-high-power hardware). Fully local speech-to-text and text-to-speech, recommended on hardware at least equivalent to an Intel N100. On that class of hardware, end-to-end latency runs approximately 600–1,200 ms. On weaker hardware — a Raspberry Pi 5 can technically run it — latency stretches to 3–10 seconds, which is the single most common source of "why is my Voice PE slow" complaints.
Layering a local or cloud LLM (via Ollama, OpenAI, or similar) on top of any of these adds natural-language flexibility — Voice PE's built-in intent system is literal and needs fairly exact phrasing without one — but pushes latency to roughly 1.5–3 seconds and means those specific queries are no longer local, even if the base STT/TTS pipeline is. Enabling "prefer handling commands locally" keeps simple commands fast and routes only unmatched phrases to the LLM.
| Pipeline | Hardware needed | Approx. latency | Offline? |
|---|---|---|---|
| Home Assistant Cloud | Any (Pi, Green) | ~500–900 ms | No |
| Speech-to-Phrase | Pi, Green | Sub-1 second | Yes, preset phrases only |
| Local Whisper + Piper | Intel N100 or better | ~600–1,200 ms | Yes |
| Local Whisper + Piper | Underpowered (Pi 5) | 3–10 seconds | Yes |
| + Local/cloud LLM | Varies | ~1.5–3 seconds | Depends on LLM |
Anyone weighing which host to run this pipeline on should compare it against the broader Home Assistant hardware options — a Home Assistant Green is enough for Speech-to-Phrase or Cloud, but local Whisper at usable speed calls for an N100-class mini PC.
What the Voice PE Can Do Out of the Box
Without any LLM attached, Voice PE's built-in intents cover the core of daily smart-home control: turning devices on and off, adjusting brightness and color, setting or nudging a thermostat, starting and querying timers (named timers require Cloud or Whisper, not Speech-to-Phrase), delayed commands like "turn off the bedroom lights in ten minutes," adding items to a shopping or to-do list, reading back a sensor value, checking the time or date, controlling media players, and triggering any automation or scene through the button or the LED-ring light entity. Broadcasts across multiple satellites and proactive "ask a question" flows were both added through firmware updates after launch, and a 2025 update let a single satellite run two wake words and pipelines at once for multilingual households.
What it doesn't do without an LLM: answer general-knowledge questions, take casually-phrased requests the way Alexa tolerates, or natively control music services like Spotify or Netflix. It also isn't a good standalone music speaker — the internal speaker is built for voice feedback, and anyone wanting real audio should route it through the 3.5 mm jack into Music Assistant or an external amp.
The feature most competitors underplay is Area-based routing: because each Voice PE satellite is assigned to a room in Home Assistant, saying "turn off the lights" affects the room the speaker is in, not the whole house. That's a genuine structural advantage over a single centralized Echo, and it scales cleanly with additional satellites — each one costs the same regardless of how many rooms already have coverage.
Language coverage is broader than most buyers expect: Home Assistant's voice pipeline processes 51 languages for commands and recognizes wake words in 22 languages, well past Alexa's roughly 8 or Google Assistant's roughly 20. The catch is that "supported locally" requires local speech-to-text, sentence support, and local text-to-speech all at once, and local STT is consistently the bottleneck — Home Assistant Cloud fills the gap for languages the local pipeline doesn't yet cover.
Common Issues and Fixes
| Issue | Cause | Fix |
|---|---|---|
| Onboarding fails on IoT VLAN / "Unable to connect" despite Wi-Fi | mDNS not bridging across subnets, or SSID band conflict | Check mDNS config, use a flat 2.4 GHz SSID during setup, try switching Wi-Fi channel (e.g., 6 to 11) |
| Disconnects after router reboot | Known ESPHome Wi-Fi stack/beacon-timeout issue | Update to current firmware |
| Wi-Fi broke after auto-updating to early 2025 firmware | Firmware corruption during OTA on early units | Reflash via the ESPHome web installer |
| Several-second STT lag after a Home Assistant Core update | Core regression affecting Voice PE/Wyoming STT | Update to a later point release |
| "Sorry, I am not aware of any device named X" | Entity not exposed to Assist, or name mismatch | Expose the entity under Settings > Voice assistants > Expose; match area/entity names exactly or add an alias |
| Wake word missed in a noisy room | Far-field pickup degraded by ambient noise | Speak closer/louder; stay on the on-device "Okay Nabu" wake word rather than server-side alternatives, which add extra latency |
| Auto-updates stop working after using ESPHome Builder | "Taking control" intentionally disables OTA updates | Only take control when writing custom YAML (e.g., adding a Grove sensor), and rebuild/flash manually afterward |
None of these are dealbreakers on their own, but they're common enough in the community that budgeting time for one of them during initial setup is realistic, especially on non-standard network setups like a segmented IoT VLAN.
Home Assistant Voice PE vs Alexa and Google Home
The comparison mainstream buyers actually care about isn't hardware — it's whether Voice PE can replace an Echo or Nest speaker outright. It can't, and it isn't trying to.
Alexa and Google Home win on casual, forgiving phrasing, general-knowledge answers, mature music-service integration, and wake-word robustness tuned over years of cloud-scale data. Voice PE wins on local processing (when configured that way), a hardware mute switch that physically disconnects the microphones, Area-based multi-room routing that doesn't require buying into a specific ecosystem, open and hackable firmware, no mandatory subscription, and language breadth that outpaces both mainstream assistants on paper.
Anyone who's already gone through linking Alexa or Google Home to Home Assistant for voice control on existing hardware should think of Voice PE as a different tool, not an upgrade path — it's for people who want the voice interface itself to be local and open, not just the automations behind it.
Home Assistant Voice PE vs Other Wyoming Satellites
Voice PE isn't the only hardware option for this stack, and it isn't automatically the best fit for every situation.
The M5Stack Atom Echo remains the cheapest way to add another Wyoming satellite to a room — it lacks the XMOS DSP and has a much weaker microphone and speaker, but it's a reasonable choice for expanding voice coverage into a space where far-field pickup doesn't matter much (a small bathroom, a hallway).
The FutureProofHomes Satellite1 sits above Voice PE: it uses the same ESP32-S3/XMOS XU316 combination but adds amplified audio output capable of driving an external speaker, built-in temperature/humidity/lux sensing, an optional mmWave presence sensor, and more physical buttons. It's aimed at buyers who want the satellite to double as a real speaker and a presence sensor, not just a microphone — some users who've tried both prefer it specifically for that reason.
Voice PE's position is the "reference implementation" — the first far-field satellite with real microphone performance sold as a finished, supported product rather than a DIY kit, at a price closer to the Atom Echo than the Satellite1.
Who Should Buy the Home Assistant Voice Preview Edition?
Buy it if: Home Assistant is already running (or about to be), local/private voice control matters more than casual phrasing, and the goal is per-room voice commands without a subscription requirement.
Skip it if: the goal is a single all-purpose speaker for music, general questions, and smart-home control without first setting up Home Assistant — a standard Echo or Nest speaker will do that with far less configuration.
Consider the Satellite1 instead if: the satellite also needs to double as a real speaker or a presence sensor.
Consider an Atom Echo instead if: the only goal is cheap voice coverage in a secondary room where far-field accuracy isn't critical.
Frequently Asked Questions
Q: Does the Home Assistant Voice Preview Edition work without internet?
Yes, if the pipeline is configured for local processing — Speech-to-Phrase or local Whisper plus Piper — the device functions entirely offline for supported commands. Home Assistant Cloud, by contrast, requires an internet connection for speech-to-text and text-to-speech.
Q: Is it as fast as Alexa or Google Home?
It depends entirely on the pipeline. Home Assistant Cloud runs close to Alexa's response time at roughly 500–900 ms. Local Whisper on an Intel N100-class host lands around 600–1,200 ms. On underpowered hardware, or with an LLM in the loop, response times stretch to several seconds.
Q: How do I set up the voice pipeline?
During onboarding, Home Assistant asks which hardware to use for voice processing: Home Assistant Cloud, or a "do it yourself" option that installs the Whisper and Piper add-ons for a fully local setup. The choice can be changed later under Settings > Voice assistants.
Q: Can I use a custom wake word?
Only "Okay Nabu," "Hey Jarvis," and "Hey Mycroft" ship as on-device wake words. Custom wake-word training exists in the community but isn't an officially supported, built-in feature.
Q: Why is my Voice PE responding slowly?
The most common cause is running local Whisper on hardware below the recommended Intel N100 baseline, or routing every command through an LLM instead of letting simple commands resolve locally. Switching to Home Assistant Cloud or Speech-to-Phrase, or enabling "prefer handling commands locally," usually resolves it.
Q: Can the Voice Preview Edition replace Alexa or Google Home entirely?
Not for most households. It has no native answer for general-knowledge questions or mainstream music-service voice control without adding an LLM, and its intent matching is more literal than Alexa's without one. It's built for Home Assistant users who prioritize local and private voice control over an assistant that answers anything.
Conclusion
The Voice Preview Edition is the first voice hardware built specifically for Home Assistant that behaves like a finished product rather than a DIY project, and the far-field microphone array backed by the XMOS XU316 processor is a real, measurable step up from earlier community satellites. But the device is only as good as the pipeline sitting behind it — the difference between a snappy voice assistant and a frustrating multi-second delay comes down entirely to whether Speech-to-Phrase, Home Assistant Cloud, or local Whisper on adequate hardware is doing the work.
It isn't a replacement for Alexa or Google Home in households that want general-knowledge answers or effortless natural language, and it isn't the right pick for anyone who wants a speaker before they want a smart-home platform. For an existing Home Assistant setup, though, it fills a gap nothing else on the market does cleanly: local, per-room voice control with a physical mute switch and no subscription requirement.
Anyone building out the rest of the local-control stack around it should also look at setting up Zigbee2MQTT for device connectivity, Matter integration in Home Assistant for newer hardware, and the broader Home Assistant hardware comparison to pick a host capable of running local Whisper if that's the end goal.
