Can Home Assistant use ChatGPT or a local LLM for voice?

Yes. Home Assistant's voice assistant, Assist, can hand requests to a language model. You can connect a local model through an Ollama server on your own network, or a cloud model through the OpenAI integration, which uses OpenAI's API. The OpenAI integration adds a conversation agent powered by OpenAI, and it is a paid service ¹. Commands still go to the built-in handler first, and only requests it cannot understand reach the model ².

This answer uses the project's own integration documentation and its 2025 and 2026 posts, checked on 17 October 2026. Our full guide to Home Assistant voice covers the rest of the pipeline.

The model handles what the built-in commands miss

Assist's built-in sentences cover lights, locks, timers, lists and scenes. A question, a request for advice, or a command phrased in a way the built-in list does not match goes to the model. The project's example is that turning on the kitchen light stays with Assist, while a request for help in a dark kitchen goes to the AI ².

The default exists for speed and cost. Before the project added the prefer-local setting, it noticed that many easy-to-run commands were being sent to a language model, which slowed things down and wasted tokens ³. Keep the default on, so that a model only handles what the built-in list cannot.

Speech-to-Phrase cannot pass requests to a model

The fast local speech engine does not work with a model as a fallback. The project says you cannot use an LLM as a fallback with Speech-to-Phrase, because unexpected commands cannot be converted into text for the model to process ⁴.

That decides your hardware. If you want a model to answer questions, you need an open speech engine such as Whisper. Our question on the best speech to text for Home Assistant explains the speed and hardware trade-off, and it recommends at least an Intel N100 for Whisper.

Ollama keeps requests on your own network

Ollama runs the model on a server you control. Home Assistant has three limits to know about before you start.

• Context window. Home Assistant defaults to an 8k context window, which is larger than the 2k default on the Ollama server itself ⁵.
• Entities. The project recommends exposing fewer than 25 entities when you experiment with local LLMs, and it warns that smaller models make more mistakes ⁶.
• Tools. Only Ollama models that support tools can control Home Assistant ⁷. A model without tool support can answer a question, but it cannot act on a device.

The project's advice on fewer than 25 entities is for experiments, not a fixed limit, and the warning about smaller models is the reason to start small and test each command.

OpenAI sends requests to a paid cloud service

The OpenAI integration requires an API key. The project says it is a paid service, and it advises you to watch your costs in the OpenAI portal ¹. Each request is billed to your own account, so the cost grows with use.

The integration has a setting for storing requests. The project says that when it is enabled, OpenAI stores requests and responses, which you can view in your OpenAI dashboard logs. The setting is disabled by default ⁸. That default applies to the Home Assistant setting only. Your OpenAI account settings and OpenAI's own policies also apply.

Local or cloud depends on privacy, cost and control

What to do

1. Set up the built-in pipeline first. Confirm your commands work with Assist before you add a model.
2. Choose the speech engine. Use Whisper if you want a model to handle open requests, because Speech-to-Phrase cannot pass requests to a model.
3. Choose Ollama for privacy, and pick a model that supports tools. Expose fewer than 25 entities to start.
4. Choose OpenAI for capability, and watch your costs in the OpenAI portal before you test.
5. Leave storage off in the OpenAI integration unless you want requests and responses in your dashboard logs.
6. Check the answers. Smaller models can make mistakes controlling devices, so test each command before you rely on it.

For the device that carries the voice, read what the Voice Preview Edition is. For how language models change speech recognition in general, read our guide to speech to text and LLMs. A large language model is the model that interprets and answers, and an intent match is the simpler built-in path that handles most commands first.