Voice control in Home Assistant without sending audio to the cloud is not a workaround or an expert-only trick — it is a fully supported, officially documented feature that most beginners never discover because every mainstream article keeps pointing them back to Alexa or Google. If you came to HA specifically to get off those platforms, this post is for you.
This guide walks through how local voice processing works in Home Assistant, what you actually need to set it up, and what to expect once it is running. No cloud accounts. No subscription. No audio leaving your network.
Why Local Voice Control Matters to the HA Community
The central value split in the Home Assistant community is local control vs. cloud-dependent. People who run HAOS are, almost by definition, people who made a deliberate choice to stop relying on a third-party server to turn their own lights on. Handing your voice audio to a cloud service defeats most of that purpose.
The “it just works until it doesn’t” sentiment in this community applies double to cloud voice services — they change APIs, retire skills, or go offline during the exact moment you need them. A local voice pipeline fails for reasons you can actually diagnose, fix, and control.
The good news is that Home Assistant now ships with a complete, on-device speech recognition stack built in. You do not need a developer background to use it.
How the Local Voice Pipeline Works in Home Assistant
HA’s local voice pipeline chains together four components. Understanding what each one does makes troubleshooting far less frustrating when something breaks in the chain.
| Stage | What It Does | Default Tool |
|---|---|---|
| Wake Word | Listens for the trigger phrase without processing anything else | openWakeWord (on-device) |
| Speech-to-Text | Converts your spoken command to text | Whisper (local add-on) |
| Intent Processing | Parses the text and matches it to an entity or action | Conversation integration (built-in) |
| Text-to-Speech | Speaks the response back to you | Piper (local add-on) |
The glue that connects all of these stages is the Wyoming protocol — a lightweight, open protocol Home Assistant uses internally so each component can talk to the others. You will see “Wyoming” appear in the add-on store and in your integrations list. It is not something you configure deeply; it is just the pipe the audio travels through.
Whisper speech to text is the component most people spend the most time tuning. It is the same open-source model family released by OpenAI, but when you install it as a Home Assistant add-on it runs entirely on your own hardware — no API key, no external call, no data leaving your network.
Setting Up the Local Voice Pipeline: What You Actually Need
Before you install anything, check your hardware. Whisper is the most computationally heavy piece of the stack. On a Raspberry Pi 4, the smaller Whisper model runs acceptably; on a mini PC or a machine with even modest dedicated compute, it runs noticeably faster. If you are still deciding between hardware options, the Raspberry Pi vs mini PC comparison covers how that choice affects voice performance specifically.
You also need a microphone. The most common starting point in the community is a small USB microphone, though many people graduate to a dedicated device like a voice satellite — essentially a small ESP32-based board that handles the wake word locally and streams only the triggered audio to your HA instance. Building one of those is a separate project, but it is worth knowing that option exists.
The install sequence is straightforward once you know the order:
| Step | Where | What to Install or Enable |
|---|---|---|
| 1 | Add-on Store | Whisper (faster-whisper) |
| 2 | Add-on Store | Piper (text-to-speech) |
| 3 | Integrations | Wyoming Protocol (auto-discovers both add-ons) |
| 4 | Settings → Voice Assistants | Create a new pipeline, select Whisper + Piper |
| 5 | Settings → Voice Assistants | Enable wake word, choose openWakeWord model |
After the pipeline is saved, use the built-in Assist interface on your HA dashboard to test it before connecting any microphone hardware. Type a command first. If the intent is not recognized, the problem is in your entity names or your Conversation integration — not the audio chain. That distinction alone will save you an hour of troubleshooting.
What Works, What Does Not, and What to Expect
Local voice in HA is excellent at controlling entities: lights, switches, covers, climate, media players. Commands like “turn off the kitchen lights” or “set the living room to 22 degrees” work reliably once your entity names are clean and unambiguous.
It is not a general-purpose AI assistant. It will not answer trivia questions or set reminders the way a cloud assistant might. The Conversation integration is built to match commands to entities and trigger automations — that scope is intentional, and for a private smart home voice command system it is exactly the right trade-off.
Wake word accuracy is the most common pain point. openWakeWord’s default models are decent but not flawless. False positives (the wake word triggering on TV audio) and false negatives (not triggering when you speak) are both normal. The community workaround is to train a custom wake word or position the microphone away from speakers. If you are running a presence detection setup alongside voice control, keep in mind that two audio-aware systems on the same device can sometimes compete for resources.
Whisper model size is a tunable knob. The “tiny” model is fastest and least accurate; “small” is the practical sweet spot for most hardware; “medium” and above require significantly more RAM and CPU. Start with “small” and only move up if accuracy is genuinely causing you problems in daily use.
For remote access without a subscription — so you can use voice control outside your home network — see the remote access without Nabu Casa guide. The local voice pipeline runs on your LAN, so remote use requires a separate secure tunnel solution regardless of which voice setup you choose.
If you want a structured path through the full HA setup — hardware, integrations, automations, and voice — the Home Assistant Smart Home for Beginners guide from The Skill Mill covers each stage in order with honest notes on where things commonly break. It is designed for people who are new to HAOS and want a map, not just a list of features.
For deeper background on open-source voice technology and privacy standards, the Electronic Frontier Foundation and Mozilla both publish accessible writing on why on-device processing matters for personal data — worth a read if you want the broader context behind the technical choices HA has made.
Frequently Asked Questions
Does this work completely offline, even if my internet goes down?
Yes. Once Whisper and Piper are installed and the Wyoming protocol integration is configured, the entire pipeline runs on your local network. An internet outage does not affect voice control at all.
Do I need a Nabu Casa subscription for local voice control?
No. Nabu Casa is an optional paid service for remote access and some cloud-based integrations. The local voice pipeline described here has no dependency on it.
Which Whisper model size should a beginner start with?
Start with “small.” It balances accuracy and speed well on most hardware people run HAOS on. You can switch models in the add-on configuration and restart without losing your pipeline settings.
Can I use an existing Alexa or Google Home device as a microphone for the local pipeline?
Not directly. Those devices are designed to send audio to their own cloud services and do not expose a raw audio stream to HA. A USB microphone, an ESPHome-based voice satellite, or a supported device running the HA companion app are the practical options.
What languages does Whisper support for local speech recognition in Home Assistant?
Whisper supports a wide range of languages because the underlying model was trained on multilingual data. In the add-on configuration you set the language code, and Piper has separate voice models for different languages. Check the current add-on documentation for the specific language list, as it expands with updates.
My voice command is understood but the wrong entity is controlled — what is happening?
This is almost always an entity naming issue. HA’s Conversation integration matches spoken words to entity friendly names. If two entities have similar names (“bedroom light” and “bedroom lamp”), the match can be ambiguous. Rename entities to be distinct, then test with the Assist text interface before testing with the microphone again.
Have you set up a local voice pipeline in HA, and what was the piece that gave you the most trouble — hardware, wake word tuning, or something in the YAML side of things? Leave a comment below and share what worked.
{“@context”:”https://schema.org”,”@type”:”BlogPosting”,”headline”:”Voice Control in Home Assistant Without Sending Audio to the Cloud”,”description”:”Run voice control in Home Assistant without sending audio to the cloud. Set up Whisper, the Wyoming protocol, and a fully local voice pipeline step by step.”,”keywords”:[“local voice assistant home automation”,”offline voice control smart home”,”privacy-focused voice commands home assistant”,”Wyoming protocol Home Assistant voice”,”on-device speech recognition home assistant”,”Home Assistant voice assistant no internet”,”Whisper speech to text home assistant local”,”self-hosted voice control smart home”,”Home Assistant voice pipeline local processing”,”private smart home voice commands without cloud”],”datePublished”:”2026-09-03T12:17:41.199Z”,”dateModified”:”2026-09-03T12:17:41.199Z”,”author”:{“@type”:”Organization”,”name”:”The Skill Mill”},”mainEntityOfPage”:”https://blog.theskillmillbooks.com/voice-control-in-home-assistant-without-sending-audio-to-the/?utm_source=skillmillblog&utm_medium=blog&utm_campaign=home-assistant-smart-home&utm_content=voice-control-in-home-assistant-without-sending-audio-to-the”}
Tools that help: our Vault & Press spreadsheet · MineStock Pro.
Leave a Reply