Private Voice Assistants for Home

Wyoming, Voice PE, Rhasspy, OVOS — local STT / TTS / wakeword without the cloud.

The 2026 voice story is "local, private, and Home Assistant native." Cloud assistants (Alexa / Google) still answer faster, but local has caught up enough for lights / scenes / timers. For the broader speech stack see Transcription (Whisper) and AI / LLMs. For the LLM brain behind a voice assistant see Home Assistant LLM Integrations. For the hub that ties it together see Smart Home Hubs.

The default — Home Assistant Wyoming stack

★ Wyoming Protocol — Home Assistant's open protocol for wakeword + STT + TTS + intent over a network. Each piece is a separate service; mix and match.
- https://github.com/rhasspy/wyoming
★ HA Voice PE (Voice Preview Edition) — official $59 hardware satellite. ESP32-S3 + microphone array + speaker. Pairs to HA in two clicks; runs Wyoming under the hood. The "buy this and stop fiddling" recommendation in 2026.
★ Whisper (faster-whisper Wyoming wrapper) — STT add-on; runs on CPU for tiny.en, GPU recommended for medium / large. Local. Multilingual.
★ Piper — local TTS; tiny ONNX models; runs fast on a Pi. Surprisingly good voices. The default HA TTS.
- https://github.com/rhasspy/piper
★ openWakeWord — local wakeword detection; trainable per-user; "Hey Jarvis," "OK Nabu," etc.
- https://github.com/dscripka/openWakeWord
microWakeWord — TFLite-Micro wakeword that runs on-device on the ESP32-S3, so the satellite can wake itself before streaming audio.

Voice satellites (the hardware that listens)

★ HA Voice PE — see above. Official, supported, gets firmware updates from the HA team.
★ ESP32-S3-BOX-3 (Espressif) — well-built dev kit; ESPHome firmware turns it into a Wyoming satellite.
Atom Echo (M5Stack) — tiny, cheap, single-mic; fine for one-room.
WaveShare ESP32-S3 mic boards — bare-board option; bring your own enclosure.
Apollo Automation MS24 / MSR-2 — combo mmWave presence + voice satellite; ESPHome based; popular 2026 pick.
Old phones / tablets — run Home Assistant Assist in the browser as a "voice browser tab."

Older / legacy / sunsetting

Rhasspy — the original local-first voice stack. In transition — the maintainer (Mike Hansen) joined Nabu Casa and is rebuilding the components as Wyoming services. Rhasspy v3 work is effectively where the modern HA voice stack came from. New installs should go straight to Wyoming + HA Assist.
Mycroft AI — community classic; company shut down 2023. The OSS lives on as OVOS (below). Don't deploy Mycroft itself in 2026.
Picovoice (Porcupine, Cheetah, Leopard) — commercial; closed-source models with a free tier. High accuracy. Reach for if openWakeWord can't make your wakeword reliable enough.
- https://picovoice.ai

OVOS / Open Voice OS

OVOS — community fork of Mycroft; full open-source assistant; skill ecosystem. More ambitious than HA Assist — it's a complete assistant, not just home-control intents.
- https://github.com/OpenVoiceOS
Neon AI — sister project; OVOS with extra features.

If you want a "Mycroft-style" full assistant outside HA, OVOS is the choice. If you just want "lights off in the bedroom," HA Assist + Wyoming is simpler.

STT / TTS engines you can plug into Wyoming

★ Whisper / faster-whisper — STT; local; multilingual.
★ Piper — TTS; local; default.
Vosk — older STT engine; tiny models; runs on a Pi Zero. Lower accuracy than Whisper.
Coqui TTS — was the modern OSS TTS; company wound down 2024 but the models live on; some HA users still run it.
OpenAI Whisper API / ElevenLabs — cloud STT/TTS; HA supports them as fallback. Pay-per-use.

Voice + LLM (the 2026 thing)

See Home Assistant LLM Integrations for the full setup. The pattern:

Wakeword detected on device → audio to HA.
Whisper transcribes locally.
Transcription → Ollama (local) or OpenAI / Claude / Gemini (cloud) with HA's exposed entities.
LLM picks an HA service call OR generates a chat response.
Piper speaks the result back.

Local end-to-end on a Pi 5 + Coral USB takes ~3-6 seconds; on a real GPU (3060+) sub-second is achievable.

Pick this if…

Default 2026 voice setup: HA Voice PE + Whisper + Piper + openWakeWord.
Build-your-own satellite: ESP32-S3-BOX-3 + ESPHome.
Combo presence + voice: Apollo MS24.
Full Mycroft-style assistant outside HA: OVOS.
Wakeword that just won't trigger: Picovoice Porcupine.
Cloud fallback for hard accents / multilingual: Nabu Casa Cloud Voice or OpenAI Whisper API.