Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM
Voice, Speech & Multimodal

Home Assistant Assist Review: Local Voice Control and LLM Agents

·6 min readBy Hans Kuepper · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI

Assist (part of Home Assistant Core 2026.10.3, released on 23 October 2026) is the voice assistant inside the Home Assistant smart-home platform, and it can run fully on your own hardware. It is not a standalone app: the full local pipeline needs Home Assistant OS, and an LLM agent is an optional, partly experimental add-on. Every statement here comes from the project documentation, not from testing.

Key Takeaways

  • Maintenance: the repository was last pushed to on 24 October 2026, so no stale-project warning applies.
  • License: Apache-2.0 open source, free to run; only Home Assistant Cloud and some voice hardware cost money.
  • History: Assist was introduced in Home Assistant 2023.2; the Ollama integration followed in 2024.4.
  • Scope: a voice layer inside a larger platform, so you run (or already run) a Home Assistant server.
  • LLM control of the home is experimental and works only with models that support tools.

📍 In One Sentence

Assist is the open-source voice assistant inside Home Assistant that turns spoken or typed commands into smart-home actions, locally or with cloud and LLM help.

💬 In Plain Terms

You speak to a microphone device, a speech engine turns your words into text, Home Assistant matches the text to a device action or asks an LLM, and a voice reads the answer back.

📌Note: This review is based on the Home Assistant documentation, the GitHub release page and the Nabu Casa pages, checked on 24 October 2026. PromptQuorum has not installed or timed the pipeline, so recognition quality, latency and LLM reliability are not assessed.

Read in:
🇺🇸en🇩🇪de🇫🇷fr🇯🇵ja🇨🇳zh🇪🇸es🇧🇷pt🇸🇦ar🇰🇷ko

Assist is the voice assistant built into Home Assistant, the open-source smart-home software. It can run entirely on your own hardware, with Whisper or Speech-to-Phrase for speech recognition and Piper for speech output connected through the Wyoming protocol, and it can hand requests to an LLM such as one served by Ollama. This review covers only that voice and LLM pipeline, not Home Assistant as a whole, and it is based on the project documentation rather than hands-on testing.

Home Assistant Assist Review: Local Voice Control and LLM Agents

What Is Assist?

Assist is the voice and text assistant built into Home Assistant that maps commands to smart-home actions. The documentation says it was introduced in Home Assistant 2023.2 and works in the companion app, on the Home Assistant Voice Preview Edition and on devices built with ESPHome.

This review covers the Assist pipeline only: wake word, speech-to-text, conversation agent and text-to-speech. Home Assistant itself, with its device integrations and automations, is out of scope. The Wyoming Protocol integration, introduced in 2023.5, lets Home Assistant use local speech and wake word services.

Get It

Assist is not a separate download: you get it by installing Home Assistant, then adding speech services as apps.

Option
How to get it
Home Assistant OSInstallation guide: Green, Raspberry Pi, Odroid, x86-64, VMs
ContainerDocker or similar on Linux; no access to apps, so no Whisper or Piper apps
Companion appGoogle Play, F-Droid, App Store
Voice hardwareVoice Preview Edition: $69 excl. tax or €59 incl. tax (MSRP)
Home Assistant CloudNabu Casa: €7.50/month or €75/year excl. tax; 31-day trial
Websitehome-assistant.io/voice_control
Privacy policyhome-assistant.io/privacy
Source codehome-assistant/core, Apache-2.0

Version reviewed: Home Assistant Core 2026.10.3, released on 23 October 2026 per GitHub. Prices are as listed on the Voice Preview Edition and Nabu Casa pages on 24 October 2026; the Nabu Casa page also offers other currencies that were not checked.

Requirements

The documentation gives timings for only a few reference devices, so sizing beyond them is guesswork.

Operating system

What the sources say:
Home Assistant OS for the full pipeline; Container installs have no apps

Speech-to-text

What the sources say:
Speech-to-Phrase: under 1 second on a Green or Pi 4. Whisper: about 8 seconds on a Pi 4, under 1 second on an Intel NUC

Text-to-speech

What the sources say:
Piper, optimized for the Pi 4: 1.6 seconds of speech per second with medium-quality models

LLM agent

What the sources say:
An Ollama server reachable over the network; only models that support tools can control the home

Exposed entities

What the sources say:
Fewer than 25 recommended for Ollama; context window defaults to 8k in Home Assistant, 2k in Ollama

Wake word

What the sources say:
Voice Preview Edition: on-device "Okay Nabu", "Hey Jarvis" or "Hey Mycroft"; openWakeWord is listed for Wyoming

Quality varies by language, so check the language table for each option before you buy hardware.

How to Get Started

These steps follow the documentation; PromptQuorum has not run them.

  1. 1
    Install Home Assistant OS
    Why it matters: Use Home Assistant OS on a Green, Raspberry Pi, x86-64 machine or VM if you want the apps for local speech.
  2. 2
    Add the speech services
    Why it matters: Start the Whisper, Piper or Speech-to-Phrase apps; the Wyoming integration discovers them and you add them from there.
  3. 3
    Create a voice assistant
    Why it matters: Under Settings, Voice assistants, choose the speech engines and the conversation agent: the built-in one or an LLM such as Ollama.
  4. 4
    Add a microphone
    Why it matters: Use the companion app, a Voice Preview Edition or an ESPHome device, and expose only the entities you want controlled.

Features Confirmed by the Sources

Everything below comes from the Home Assistant documentation; none of it has been independently tested.

  • Built-in and custom sentences. Community-contributed sentences handle commands such as turning on a light; you can add your own under custom_sentences and handle them with intent_script, or start automations with sentence triggers.
  • Wyoming services. The Wyoming Protocol integration connects Speech-to-Phrase, Whisper, Piper and openWakeWord, and can point to a service on another device on your network, which helps for heavy engines such as Whisper.
  • LLM conversation agents. Ollama and OpenAI can answer through the Assist API and, with control enabled, act on the entities you expose.
  • Voice hardware. Besides the Voice Preview Edition, the voice page lists ESPHome ESP32 satellites (a $13 build), an experimental Linux-Voice-Assistant for x64 or ARM64 systems, and an analog phone setup.
  • Tuning. Fine-tuning options for Whisper and Piper are documented on each app's page.

Data Flow and Privacy

With Home Assistant Cloud, only the audio is sent to Nabu Casa for speech-to-text and text-to-speech, while the actions themselves still run on your own Home Assistant, per the documentation. An Ollama agent talks to a server you host, and the Ollama page adds no privacy statement. The OpenAI integration is a cloud service that works only with the official OpenAI endpoint, and the entities you expose are what the model can control or report on.

The Home Assistant privacy policy, last updated on 27 August 2026, says nothing is sent to Analytics unless you opt in, and it does not mention voice data. It states that Home Assistant Cloud is a separate commercial service under its provider's own privacy policy, which PromptQuorum did not review.

📌Note: These are the project's declarations, not audit results. PromptQuorum has not inspected network traffic or the source code of the voice pipeline.

Who Should Use It

  • Existing Home Assistant users who want voice control without sending audio to a cloud service.
  • Tinkerers with a spare server or Raspberry Pi who accept choosing and tuning speech engines per language.
  • People already running Ollama who want to try an LLM agent on their home entities and accept its experimental status.

What We Could Not Verify

  • Recognition quality and speed. The timings above are the documentation's; PromptQuorum measured nothing.
  • LLM reliability. The documentation warns that smaller models make more mistakes with control enabled; how often is untested here.
  • Hardware sizing. No minimum RAM or CPU is stated for Whisper or for Ollama models, and no model is recommended.
  • Cloud data handling. The Nabu Casa privacy policy and retention terms were not reviewed.

Competitors and Alternatives

Voxa

Platforms:
Windows, macOS, Linux
Price / license:
Free / MIT
Key difference:
Local-first desktop voice assistant with optional cloud voice
Articles about Voxa (4)

Also mentioned in:

Dograh

Platforms:
Web (self-hosted)
Price / license:
Free self-hosted / BSD-2-Clause
Key difference:
Voice-agent platform with telephony and bring-your-own model

Assist is a pipeline inside a home-automation server, so these are the nearest voice assistants in the directory rather than direct substitutes; confirm each app's current price, license and platforms on its own page.

Compare in directory

Frequently Asked Questions

Do I need an LLM to use Assist?

No. The built-in sentence matching handles device commands without one; an LLM agent is optional, and the Ollama page calls its control of Home Assistant experimental.

What can Speech-to-Phrase not do?

It is a close-ended model that understands only a subset of Assist commands, so shopping lists, timers and broadcasts are not usable out of the box; Whisper tries to transcribe everything but is slower.

Does Ollama work with sentence triggers?

No, the Ollama page says the integration does not work with sentence triggers. For small models it suggests two Ollama integrations on the same model, one without control for conversation and one with control for commands.

Will Assist work in my language?

Language support is rated separately for local processing and for Home Assistant Cloud on a four-step scale from Not supported to Fully supported, so check the table for your language and engine first.

Can a phone listen for a wake word?

The voice page says Android phones can use wake word detection, with "Hey Jarvis" and "Hey Nabu" as examples, and it does not describe the feature for iPhone.

Is the OpenAI integration free?

No. The documentation calls it a paid service that needs an API key and advises monitoring costs and setting usage limits.

Verdict

Assist is an actively developed way to add voice control to a Home Assistant system, with an offline route described in the documentation plus optional cloud and LLM routes. Against that, it needs a Home Assistant installation, the full local pipeline needs Home Assistant OS, Whisper is slower on small boards per the documentation, and LLM control is experimental. It suits existing Home Assistant users who want voice that stays in the house and are willing to tune it; anyone who wants a standalone voice assistant should look at the alternatives above.

Sources

← Back to Power Local LLM

MCP