Local voice for Home Assistant: Whisper + Piper in Docker
A voice assistant that works when the internet is down, and that never sends your conversations to anyone's cloud, is three small containers away. Home Assistant's Assist pipeline splits voice into jobs — wake word, speech-to-text, the conversation itself, and text-to-speech — and the Wyoming protocol lets each job run in its own Docker container on your NAS. Here is the speech half: Whisper listening, Piper talking.
Published Nov 17, 2026
How the pipeline fits together
Say the wake word near a satellite (a Voice Preview Edition, an ESPHome device, or a phone) and the audio travels a fixed path: the wake-word service says 'that was for me,' Whisper transcribes the sentence, Home Assistant's conversation agent works out what to do, and Piper speaks the reply. Every hop can live on your hardware — this guide installs the three Wyoming services as containers; pairing them with a satellite is covered in the voice & music chapter.
The Compose stack
Three small services, three ports, no database, nothing to configure beyond a voice and a model choice:
services:
wyoming-whisper:
image: rhasspy/wyoming-whisper:latest
container_name: wyoming-whisper
restart: unless-stopped
ports:
- "10300:10300"
volumes:
- ./whisper/data:/data
command: --model base-int8 --language en
wyoming-piper:
image: rhasspy/wyoming-piper:latest
container_name: wyoming-piper
restart: unless-stopped
ports:
- "10200:10200"
volumes:
- ./piper/data:/data
command: --voice en_US-lessac-medium
wyoming-openwakeword:
image: rhasspy/wyoming-openwakeword:latest
container_name: wyoming-openwakeword
restart: unless-stopped
ports:
- "10400:10400"
command: --preload-model 'ok_nabu'
- Whisper model choice is a speed dial.
tiny-int8is fastest and fine for clear commands near a mic;base-int8is the sweet spot on a small NAS CPU;smalland up transcribe noticeably better but can take several seconds per sentence on modest hardware. Start at base-int8. - Piper voices download on first use into the data volume — the first reply takes a while, every one after is quick.
en_US-lessac-mediumis a natural default; the voice catalogue is large if you want a different character. - openWakeWord's
ok_nabuis the friendly default wake phrase; 'alexa' and 'hey_jarvis' models can be preloaded the same way.
Connect them to Home Assistant
- In Home Assistant: Settings → Devices & Services → Add Integration → Wyoming Protocol. Add it three times — host = your NAS IP, ports 10300 (Whisper), 10200 (Piper), 10400 (openWakeWord).
- Settings → Voice Assistants → Add Assistant. Name it something honest like 'Local'. Set conversation agent to Home Assistant, speech-to-text to Whisper, text-to-speech to Piper, wake word to openWakeWord.
- Point a satellite at this pipeline (each voice satellite picks which assistant it uses in its own settings) and say the wake phrase, then a command with an obvious result — 'turn on the kitchen light' beats 'what's the weather' for a first test.
What to expect (honestly)
On a small NAS CPU, Whisper base-int8 answers in about one to three seconds — a beat slower than a cloud assistant, fast enough that you stop noticing. Piper is effectively instant. What you give up versus the big cloud assistants is open-ended chit-chat and web knowledge; what you gain is that 'turn off the bedroom lights' works during an internet outage, at 2 a.m., with zero audio leaving your house. For controlling a home — the actual job — the local pipeline is not the compromise option anymore. It is the dependable one.
Performance tip: if transcription feels sluggish, drop to tiny-int8 before blaming the hardware — for short, clear smart-home commands its accuracy is genuinely hard to distinguish from the bigger models.
Troubleshooting
- Satellite hears the wake word but nothing happens: the satellite is pointed at a different pipeline — check which assistant is selected in the satellite's settings.
- Replies are silent: Piper's first-run voice download may still be going; check the wyoming-piper log, then confirm the satellite's volume isn't the problem with a test announcement.
- Garbled transcriptions: almost always microphone placement (too far, behind a TV, near a fan) before it is model size. Move the satellite, then consider a bigger model.
- Everything works except when the NAS reboots: the containers need
restart: unless-stopped(already in the Compose above) and the satellite needs the NAS back before it can listen — give it a minute.
Choosing something to talk to
The containers are the brainstem; you still need ears and a mouth in the room. The easiest satellite is the Home Assistant Voice Preview Edition (below) — purpose-built, with the wake word handled on-device. ESPHome satellites are the DIY route: a small board plus a microphone array, covered in our ESPHome guide, and the route to multi-room coverage on a budget. The Home Assistant phone app can act as a satellite too — a fine way to test this whole stack for free before buying hardware. Whichever you choose, placement beats hardware: ear-height, a metre or more from walls and TVs, out of the kitchen's fan noise. A well-placed cheap satellite outperforms a premium one shoved behind a soundbar, every time.
Gear for this build
Prices are editorial and change often; buttons show live prices. Some links may earn a commission at no cost to you — see the disclosure page.
Official documentation and references
- Home Assistant — voice control and Assist documentation
- Home Assistant — the Wyoming integration for local voice services
- Wyoming Faster Whisper on GitHub — the local speech-to-text server
- Wyoming Piper on GitHub — the local text-to-speech server
Frequently asked questions
Does local voice work without internet?
Yes — that is the point of this stack. Wake word, transcription, the conversation agent and speech all run on your hardware. The only things that fail offline are the ones that inherently need the internet, like asking for a weather forecast from an online provider.
What hardware does Whisper need?
Less than people fear. On a modern low-power NAS CPU, base-int8 transcribes short commands in one to three seconds. Whisper is the one service here that appreciates a stronger CPU; if your NAS is very old, tiny-int8 keeps it responsive.
Can I mix local and cloud in one pipeline?
Yes. Pipelines are assembled per piece — a common setup is local wake word and Piper speech with a cloud conversation agent for smarter answers, or the reverse. Each assistant you define in Home Assistant picks its own combination.

