Cloud voice assistants are still convenient, but every microphone in a home is a privacy decision that gets renegotiated every year. The home lab crowd has moved a step further: quantized LLMs running on a single Proxmox node, wake words handled on a $20 satellite, and Alexa unplugged. The apps below are the ones that actually make a fully local voice stack work in 2026, ranked by how close they get to “say the thing, the thing happens” without an internet round-trip.
The eight picks below all run on a desktop, NAS, or home lab box, and most pair with cheap dedicated hardware for the room microphones. Each section lists what the app does well, what it does badly, and what hardware footprint to expect.
What to look for in a local voice stack
A local assistant is several apps glued together. The pieces that matter:
- Wake word. Runs constantly on a low-power device. Look for small models with good rejection of false triggers.
- Speech-to-text. Whisper-family models are the current default. Pick by latency and language coverage.
- Intent handling. The “what does this sentence mean” layer. Either explicit grammars or an LLM doing the reasoning.
- Text-to-speech. Local TTS quality varies widely. Modern neural voices are unrecognizable from cloud output.
- LLM backend. For open-ended questions, a local Mistral, Llama, or Qwen variant via Ollama or LocalAI.
- Smart-home integration. Connection to Home Assistant, openHAB, Zigbee2MQTT, or similar.
- Hardware footprint. CPU-only is possible but slow. A consumer GPU with 8GB+ VRAM transforms the experience.
Quick comparison
| App | Best for | Free plan | Cost | Standout |
|---|---|---|---|---|
| Home Assistant + Voice PE | Full stack | Yes | Free (hardware sold separately) | First-party voice satellite, deepest integration |
| Rhasspy | DIY voice pipeline | Yes | Free | Profiles per language, swappable components |
| Open Voice OS | Mycroft successor | Yes | Free | Skills marketplace and active community |
| Willow | Low-latency wake | Yes | Free | ESP32-S3 firmware for cheap satellites |
| Mycroft AI | Classic open-source | Yes | Free | Battle-tested codebase, smaller community now |
| Ollama | Local LLM backend | Yes | Free | One-line model install, REST API |
| LocalAI | OpenAI-compatible local LLM | Yes | Free | Drop-in for tools that expect the OpenAI API |
| Wyoming Protocol | Audio pipeline glue | Yes | Free | Standard transport between mic, STT, intent, TTS |
The apps
1. Home Assistant + Voice Preview Edition, the integrated default
Home Assistant Voice is now the default recommendation for a self-hosted voice stack. The Voice Preview Edition satellite handles wake word and microphone array in a $59 puck, and the server side runs on whatever Home Assistant already lives on. The Year of the Voice work is done, and the result is a stack that controls devices, reads sensors, and reports calendar events without ever touching a cloud API.
Where it falls short: open-ended conversational use is still better with a separate LLM in the loop. The Voice PE hardware is the smoothest path, but it does sell out.
Pricing: Free software, Voice PE hardware around $59.
Platforms: Home Assistant runs on Linux, Windows (via container), macOS (via container), or Home Assistant OS on a Pi or NUC.
Download: Home Assistant
Bottom line: Start here. The first-party satellite plus Home Assistant Assist covers most of what a household needs without any other components.
2. Rhasspy, the modular DIY stack
Rhasspy is the toolkit version: each piece (wake word, STT, intent, TTS) is swappable, with profiles per language, and a configuration UI that does not hide the pipeline. It pairs neatly with Home Assistant via the Wyoming protocol and has the broadest language coverage in this list.
Where it falls short: the build-it-yourself feel is intentional. Expect to spend an evening picking models, recording audio samples, and tuning the wake word.
Pricing: Free, open source under MIT.
Platforms: Linux, container on Windows or macOS.
Download: Rhasspy 3 docs
Bottom line: The right pick if you want full control of each pipeline stage. Skip if you want a black box that works.
3. Open Voice OS, the Mycroft successor
Open Voice OS picked up the Mycroft codebase after the project’s commercial arm wound down, and the community has carried it forward. The skills marketplace covers everything from timers to weather to home control, and the framework runs on Pi, x86, and Mark II hardware. The 2026 cycle added better LLM-backed intent handling.
Where it falls short: documentation is split between old Mycroft pages and the new Open Voice OS site. Some legacy skills still need porting.
Pricing: Free, open source.
Platforms: Linux on Raspberry Pi, x86 desktop, container on macOS.
Download: Open Voice OS
Bottom line: A good middle ground between Home Assistant’s integrated approach and Rhasspy’s bare-bones toolkit. Worth a look for households that want the skill ecosystem.
4. Willow, the latency-focused satellite
Willow’s stack runs on a $20 ESP32-S3 board with an array of microphones, talks to a server you self-host, and gets a wake-to-response time in the few-hundred-millisecond range. The codebase has matured into one of the better cheap satellite options for households with several rooms to cover.
Where it falls short: this is the wake-word and audio transport layer, not a complete assistant. Pair with Home Assistant or a local LLM for the actual answers.
Pricing: Free, with hardware boards available from several vendors for around $20-50.
Platforms: Server side runs on Linux. Client is ESP32-S3 firmware.
Download: Willow project
Bottom line: The right answer when you want many cheap, always-on microphones around the house without a Raspberry Pi in each room.
5. Mycroft AI, the original
The original Mycroft project is in maintenance, but the codebase still installs and runs, and many households built on it years ago are still going. The hardware Mark II ships of the original Kickstarter still work, and the skills framework is well documented.
Where it falls short: smaller active community, slower bug fixes, and most new development has moved to Open Voice OS.
Pricing: Free, open source.
Platforms: Linux on Pi or x86.
Download: Mycroft GitHub
Bottom line: Worth knowing about for historical reasons or if a Mark II already sits on a kitchen counter. New deployments should pick Open Voice OS instead.
6. Ollama, the LLM backend most stacks reach for
Ollama is now the default way to run a quantized local LLM on a Mac, Linux box, or Windows machine. One command pulls a model, another exposes it on an HTTP endpoint, and Home Assistant Assist, Rhasspy, and Open Voice OS can all talk to it through a simple integration. Llama, Mistral, Qwen, and Phi variants all work out of the box.
Where it falls short: not a voice tool on its own. It is the brain you plug into one of the assistants above.
Pricing: Free, open source.
Platforms: Linux, macOS, Windows.
Download: Ollama
Bottom line: The path of least resistance to “local LLM in the loop”. Pair with a 7B-13B model for fast responses on a single 8GB+ GPU.
7. LocalAI, the OpenAI-compatible drop-in
LocalAI exposes the same HTTP shape as the OpenAI API, so any tool that already speaks OpenAI can be pointed at a local model instead. Useful when you have a Home Assistant blueprint, Node-RED flow, or third-party automation that expects an OpenAI-shaped endpoint and you want to remove the cloud dependency.
Where it falls short: setup is more involved than Ollama. The performance varies more with model and backend choices.
Pricing: Free, open source.
Platforms: Linux (docker), macOS (docker), Windows (docker or WSL2).
Download: LocalAI
Bottom line: Pick this when an existing automation expects the OpenAI API and you want a local backend with no code changes.
8. Wyoming Protocol, the audio pipeline glue
Wyoming is the small but important protocol that lets Home Assistant, Rhasspy, Open Voice OS, and the satellite hardware speak to each other. Adopting it as the common transport between wake word, STT, intent, and TTS makes it easy to swap any component without rewriting the rest.
Where it falls short: it is plumbing, not a user-facing app. Most people use it without realizing they are using it.
Pricing: Free, open source.
Platforms: Wherever the components run (Linux, container on others).
Download: Wyoming docs
Bottom line: Worth knowing about because it is the reason a Home Assistant + Rhasspy + Ollama stack works without custom code in the middle.
How to pick the right one
- If you want the smoothest path, pick Home Assistant + Voice PE and stop reading.
- If you want every pipeline stage swappable, pick Rhasspy.
- If you want a skill marketplace, pick Open Voice OS.
- If you want $20 satellites in every room, pick Willow.
- If you want local LLM answers in the loop, pair any of the above with Ollama or LocalAI.
- Stack two or three at most. The Wyoming protocol means they slot together cleanly; do not feel pressure to pick just one.
FAQ
Can a local voice assistant really match Alexa or Google Assistant?
For “turn on the kitchen lights” and “set a timer”, yes. For open-ended questions, modern local LLMs are close enough on factual answers that most households do not notice the difference. The latency on a midrange GPU is now within the same envelope as a cloud call.
Do I need a GPU to run a local voice assistant?
For the voice pipeline alone (wake, STT, intent, TTS), no. A Raspberry Pi 5 or NUC handles it. For LLM-backed conversational use, a discrete GPU with 8GB+ VRAM transforms the experience. CPU-only LLM inference is possible but feels slow.
Is Home Assistant Assist private?
Yes, in its default Whisper-plus-local-intent configuration nothing leaves the network. If you enable an OpenAI or other cloud LLM as the conversation backend, queries to that backend leave the network. The wake word and STT can stay local even then.
What does the cheapest fully local setup cost?
Around $130 total: a Raspberry Pi 4 or 5 ($55-80), a Voice PE puck or Willow board ($20-59), and a USB microphone if you go DIY. LLM features add the cost of a discrete GPU; a used RTX 3060 12GB in the $200-250 range is the common pick.
