Cloud voice assistants — Alexa, Google Assistant, Gemini for Home — send every “turn off the kitchen lights” through a data centre. Even with the new on-device chips, half the round-trip still leaves our network. The 2026 local-voice stack runs the whole pipeline on hardware we own, with response times under 200 milliseconds and zero data leaving the LAN. We tested seven desktop apps and self-hosted tools to find the best apps for local voice control of a smart home.
The test rig was a Home Assistant Yellow with a 16 GB Pi 5 compute module, plus a separate Intel NUC with 32 GB RAM running models. We measured latency, wake-word accuracy, and how each tool integrates with Home Assistant and the broader smart-home ecosystem. Whether we want a polished plug-and-play setup or a from-scratch DIY stack, one of these will fit.
What to look for in a local voice stack
- End-to-end local processing. Wake word, speech-to-text, intent matching, text-to-speech, action execution — all on our hardware.
- Latency. Sub-300ms feels instant; over 1 second feels like cloud-era Alexa.
- Hardware requirements. Some models need 8 GB+ of dedicated RAM. A Raspberry Pi 4 alone is usually not enough for the full stack.
- Wake-word accuracy. False positives wake the assistant accidentally; misses make us repeat ourselves. Trained per-user wake words work better than generic ones.
- Language support. English-first is the default; Spanish, French, German, Italian, Russian, Portuguese, and Japanese have decent coverage. Smaller languages lag.
- Integration with smart home platforms. Home Assistant integration is the gold standard. Tools that only work standalone are limited.
Quick comparison
| Tool | Best for | License | Local STT | Local TTS |
|---|---|---|---|---|
| Home Assistant | Full smart-home + voice in one | Apache 2.0 | Yes (via Whisper) | Yes (via Piper) |
| Whisper (faster-whisper) | Local speech-to-text engine | MIT | Yes | N/A |
| Piper | Local text-to-speech engine | MIT | N/A | Yes |
| Rhasspy | Older modular voice toolkit | MIT | Yes | Yes |
| Mycroft AI / OpenVoiceOS | Standalone voice assistant | Apache 2.0 | Yes | Yes |
| Willow | ESP32-based hardware + local stack | Apache 2.0 | Yes (server-side) | Yes |
| Ollama | Local LLM for intent matching | MIT | N/A | N/A |
The 7 best apps for local voice control
1. Home Assistant — best end-to-end stack
Home Assistant is the platform every other tool on this list orbits around. The 2026 Assist pipeline ties together a wake-word engine (microWakeWord on dedicated hardware), a speech-to-text engine (Whisper via the Wyoming Protocol), an intent matcher (Speech-to-Phrase or LLM-backed), and a text-to-speech engine (Piper) into one configurable pipeline.
The Home Assistant Voice Preview Edition hardware ($59) is the polished entry point — a small puck with a microphone array that runs the wake-word locally. The assist pipeline can stay 100% local even with this hardware.
Where it falls short: Setup requires planning — choosing the right Pi or NUC, deciding whether to run Whisper on CPU or GPU, configuring add-ons. Configuration UI helps but YAML still surfaces for advanced setups. The voice on Piper is good but not as natural as Alexa’s voice.
Pricing: Free server software; ~$59 for HA Voice PE; ~$130-$300 for compute hardware
Platforms: Home Assistant OS on a Raspberry Pi, NUC, x86 box, or in a VM
Download: home-assistant.io
Bottom line: Pick Home Assistant as the brain of the stack. Skip it only if we want a standalone voice assistant with no smart-home integration.
2. Whisper (faster-whisper) — best local speech-to-text
Whisper is OpenAI’s open-source speech-to-text model. The “faster-whisper” rewrite uses CTranslate2 to run 4-8x faster than the original, with transcription starting while we are still talking. It is the speech-to-text engine most Home Assistant Voice setups use under the hood, running as a Wyoming Protocol add-on.
For users who want the most accurate local speech-to-text in 2026, Whisper is the default. The “small.en” model fits in 1 GB RAM and handles common commands. “medium” or “large-v3” are more accurate but need more compute.
Where it falls short: Accuracy depends on model size — small models miss accents and unusual phrasing. GPU acceleration helps a lot but is not required. Latency on CPU-only setups can creep above 1 second for longer phrases.
Pricing: Free
Platforms: Docker on Windows, macOS, Linux; or HA add-on
Download: github.com/SYSTRAN/faster-whisper
Bottom line: Pick Whisper if we want best-in-class local speech-to-text. Skip it only if our hardware cannot run any model.
3. Piper — best local text-to-speech
Piper is the open-source text-to-speech engine built specifically for fast on-device synthesis. Voices are pre-trained per-language, and the inference time on a Pi 4 is well under 300ms for a typical response. The 2026 release added emotional inflection and 50+ language voices.
For users who want the response voice to feel natural without sending text to a cloud, Piper is the answer.
Where it falls short: Voice naturalness trails ElevenLabs and the proprietary cloud TTS providers. Some less-common languages have only one voice. Custom voice training is possible but requires GPU time.
Pricing: Free
Platforms: Docker on Windows, macOS, Linux; or HA add-on
Download: github.com/rhasspy/piper
Bottom line: Pick Piper as the voice of any local stack. Skip it only if we want cloud-quality naturalness.
4. Rhasspy — best modular older toolkit
Rhasspy is the original open-source voice assistant toolkit. It is modular by design: pick a wake-word engine, a speech-to-text engine, an intent recogniser, and a text-to-speech engine, and Rhasspy ties them together. It still works in 2026 but is no longer the recommended path for new setups.
The original author now leads Home Assistant’s voice work, and most of Rhasspy’s strengths have been ported into the Wyoming Protocol ecosystem. Existing Rhasspy users can keep running it — new installs are better served by Home Assistant’s Assist pipeline.
Where it falls short: Setup is more involved than Home Assistant Voice. Documentation has not been updated as quickly as Wyoming docs. Some integrations have stalled.
Pricing: Free
Platforms: Docker on Windows, macOS, Linux
Download: rhasspy.readthedocs.io
Bottom line: Pick Rhasspy if we already run it. Skip it for new setups — Home Assistant’s Assist is the modern path.
5. Mycroft AI / OpenVoiceOS — best standalone voice assistant
OpenVoiceOS (the community fork of Mycroft AI after Mycroft’s company wound down) is the standalone open-source voice assistant. It is the closest open-source equivalent to a dedicated Alexa or Google Assistant device — a voice assistant first, smart-home integration second.
For users who want a voice assistant without committing to Home Assistant, OpenVoiceOS is the right pick. Skills cover music, timers, weather, smart-home control, and more.
Where it falls short: Smart-home integration is less polished than Home Assistant Voice. Community is smaller than Home Assistant’s. Hardware support for ready-made mic puck devices is limited.
Pricing: Free
Platforms: Docker on Linux primarily
Download: openvoiceos.org
Bottom line: Pick OpenVoiceOS if we want a standalone voice assistant. Skip it if Home Assistant integration is critical.
6. Willow — best ESP32 hardware + server stack
Willow is the open-source voice satellite project that runs on cheap ESP32-S3-BOX hardware ($50-80). The microphone array and wake-word detection run on the ESP32, while speech-to-text and intent matching run on a Willow Inference Server we host. The combination delivers under-300ms response times in the typical room.
For users who want multiple voice satellites around the house without paying Home Assistant Voice PE prices per room, Willow is the most economical option.
Where it falls short: ESP32-S3-BOX hardware needs flashing (familiar territory if we have done ESPHome). The inference server needs a moderately powerful machine (a Mac mini or NUC). Some integration steps still require terminal commands.
Pricing: ESP32-S3-BOX-3 around $80; inference server uses existing hardware
Platforms: Server on Linux, macOS; satellites on ESP32
Download: github.com/HeyWillow/willow
Bottom line: Pick Willow if we want cheap satellites in every room. Skip it if flashing ESP32 hardware feels intimidating.
7. Ollama — best for local LLM intent matching
Ollama is not a voice tool — it is the local LLM runtime that ties together the smart parts of a 2026 voice stack. By running a small Llama 3 or Mistral model on the local network, Home Assistant’s Assist pipeline can handle natural-language commands (“turn off everything in the bedroom except the fan”) that simple intent matchers miss.
For users who want their voice assistant to understand intent beyond rigid templates, Ollama is the upgrade.
Where it falls short: Requires a machine with 8 GB+ of RAM dedicated to the model. Response time depends on the model — Llama 3 8B is responsive on a Mac with M-series silicon; Mistral 7B is good on most x86 boxes. Pure CPU inference is slower.
Pricing: Free
Platforms: Native installer for Windows, macOS, Linux
Download: ollama.com
Bottom line: Pick Ollama to add LLM-backed intent to Home Assistant’s voice stack. Skip it if our hardware cannot host a model.
How to pick the right one
If we want the simplest local voice setup: Home Assistant + Home Assistant Voice PE hardware. Plug, follow the setup wizard, and we are running local in 30 minutes.
If we want the best speech-to-text accuracy: Whisper (faster-whisper) with the medium model on a GPU-equipped machine.
If we want the most natural local TTS voice: Piper with one of the high-quality voices.
If we want multiple voice satellites without breaking the bank: Willow with ESP32-S3-BOX hardware.
If we want a voice assistant without committing to Home Assistant: OpenVoiceOS.
If we want LLM-quality intent matching: Add Ollama with a 7B-8B model to a Home Assistant Assist pipeline.
If we want the gold-standard private setup: Home Assistant + Voice PE + Whisper + Piper + Ollama, all running on a NUC with 32 GB RAM on the LAN.
FAQ
Can I really run a local voice assistant without cloud at all? Yes. The 2026 stack (Home Assistant Assist + Whisper + Piper + optional Ollama) runs entirely on hardware we own. No audio, intent, or response leaves the LAN.
What hardware do I need? At minimum, a Raspberry Pi 4 with 4 GB RAM for Home Assistant itself, plus a separate small machine (a Pi 5 with 8 GB or any small Linux box) for Whisper. For LLM-backed intent matching, add another machine with 16 GB+ RAM for Ollama.
How does response time compare to Alexa? A well-tuned local stack responds in 150-300ms. Cloud Alexa typically takes 500-1500ms. Apple HomePod with HomeKit is between 200-500ms.
Will it work in Spanish, French, or other languages? Yes, with caveats. Whisper handles ~99 languages reasonably well. Piper has voices for 50+ languages. Home Assistant’s intent matcher works in many languages but smaller languages need more configuration.
Is Mycroft still alive? The Mycroft AI company wound down in 2023. The community fork, OpenVoiceOS (a.k.a. OVOS), keeps the project alive and actively releases updates. Existing Mycroft users can upgrade to OVOS.
