Home Assistant Voice: Speech to Text and Voice Commands

Assist, the voice assistant built into Home Assistant, can run entirely on your own hardware. A spoken request passes through four stages: a wake word, speech to text, a matcher for built-in commands or an optional language model, and a spoken reply. For the fixed commands most homes use, such as turning on lights or setting a timer, Speech-to-Phrase is the fastest local choice. It transcribes in under one second on a Raspberry Pi 4 ¹. For open speech, Whisper is the local option, and it is slower: about 8 seconds per command on a Raspberry Pi 4 ².

This guide follows Home Assistant's own documentation and blog posts, and it notes where those sources disagree. The figures are as published, and we checked them on 17 October 2026. If you are new to voice control, start with our overview of voice assistants compared with dictation. Our guides on voice and AI cover the wider field.

Assist turns speech into Home Assistant actions in four stages

Each request goes through four jobs, and Home Assistant lets you choose the engine for each one. A wake word starts a request, speech to text turns your words into text, an intent matcher or a large language model works out the action, and text to speech speaks the answer.

The wake word is the stage most people notice first. Home Assistant detects it inside the system rather than on each device, so a cheap microphone can still act as a voice satellite. The project says that 5 voice satellites can stream audio at once without overwhelming a Raspberry Pi 4 ³. The project's own hardware for this is the Voice Preview Edition, which our question on what it is explains.

Piper is the text-to-speech engine that the project recommends for local use. On a Raspberry Pi, using medium quality models, it can generate 1.6 seconds of voice in one second ⁴. That is a speaking speed, not a command speed, so it does not affect how quickly Assist understands you.

Setting up the local pipeline takes five steps, according to the project's guide:

1. Install the speech-to-text app you chose, either Speech-to-Phrase or Whisper, and install Piper for text to speech.
2. Start the apps. Under Settings > Devices & services, the Wyoming integration should discover both services. Select Add for each.
3. Go to Settings > Voice assistants and select Add assistant.
4. Under Conversation agent, select Home Assistant. Under Speech-to-text, select the engine you chose, and under Text-to-speech, select Piper. Choose the language for each.
5. Expose your devices to Assist, or the assistant cannot control them by voice ⁵.

Be careful with old search results. If you search for Piper on GitHub, you may find a repository that is archived. The original Piper repository was archived by its owner on October 6, 2025 and is read-only, and development has moved to a different repository ⁶. When a guide points to the old repository, use the setup steps above instead.

Speech-to-Phrase is fast because it only listens for your own commands

Speech-to-Phrase is fast because it does not try to transcribe open speech. Home Assistant calls it a close-ended speech model: it transcribes what it knows ⁷. Its speed comes from that narrow choice, and its limits come from the same choice.

The project built it for low-powered hardware, and the reason is who runs Home Assistant. Its opt-in analytics show that over 50 percent of Home Assistant OS users run their homes on affordable, low-powered machines such as the Home Assistant Green or a Raspberry Pi ⁸. A tool that runs on those machines fits the hardware that most Home Assistant OS users have.

On the Raspberry Pi 5, the project reports 150 milliseconds per command, seven times faster than the Pi 4 ⁹. The chart below puts the published figures side by side. Whisper appears twice because the project gave two figures for the same Pi 4 test, and the newer one is the current documentation.

The speed has a price. Speech-to-Phrase recognizes only the phrases it was built from. The project's own documentation lists what falls outside that set: shopping lists, naming a timer and broadcasts are not usable out of the box ¹⁰. Its vocabulary is also tied to your home. The README says Speech-to-Phrase combines pre-defined sentence templates with the names of things from your Home Assistant ¹¹. That is why it recognizes your own room names, and why a command that uses a name it was not built from will not match.

Language support is the other limit. When Speech-to-Phrase launched, it supported English, French, German, Dutch, Spanish and Italian, and the project said that covered nearly 70 percent of its users ¹². Check your language before you choose this engine, because the project's guide says quality can change quite a bit from one language to another.

Whisper understands open speech but needs a faster computer

Whisper is the open-ended option. Home Assistant describes it as a model that tries to transcribe everything, at the cost of slower processing ¹³. It can handle a shopping list, a named timer or a broadcast, and it can pass the text to a language model. The price is hardware.

The current documentation says Whisper takes around 8 seconds per command on a Raspberry Pi 4, and under a second on an Intel NUC ². The February 2025 blog post gave a lower bound of at least 5 seconds for the same Pi 4 setup ¹⁴. The two figures differ, and the project has not reconciled them in the sources we checked. Either way, on the same Pi 4, Whisper takes longer than Speech-to-Phrase, which transcribes in under one second.

For that reason the project recommends a faster machine for fully local voice. If you want to use Whisper, its advice is at least an Intel N100 or an equivalent processor, which runs the Whisper Base model. The project says that model runs reasonably fast for English and Spanish, which have large public datasets to train on ¹⁵.

Hardware limits also decide which languages Whisper can serve well. The project's example is French: it may work well enough with the largest Whisper model, which needs a GPU, but it is unusable on a Raspberry Pi or even an N100-class PC ¹⁶.

Open speech is also harder to transcribe. In the Whisper paper, the Large V2 model had a 17.6 percent word error rate on CallHome, a test set of conversations between friends and family ¹⁷. That figure is for a large model on casual speech, not for the Base model that Home Assistant recommends, so treat it as a sign of how hard open speech is. Our guide to word error rate explains how these percentages are calculated, and the glossary entry for word error rate gives the definition.

Pick the engine by hardware and by what you say

The right engine depends on two things: the computer you run Home Assistant on, and whether you will speak only commands or open requests. The table sets out the choice for each option, using the project's own figures.

If your home needs only lights, locks, timers and scenes, Speech-to-Phrase on a Raspberry Pi is enough, and it needs the least hardware of the three. If you want to speak freely or pair Assist with a language model, choose Whisper on an N100-class machine or better. If your language has no good local model, Home Assistant says that for some languages it believes Home Assistant Cloud provides the best experience ¹⁸. Our question on the best speech to text for Home Assistant gives the short answer.

Home Assistant Cloud processes speech for you, off your network

Home Assistant Cloud is the option that does not run speech on your own hardware. It is an optional subscription, and the project says it supplies the text-to-speech and speech-to-text engines used by Assist, while Home Assistant itself works fully without it ¹⁹. This guide does not quote its price.

The cloud route sends your audio to a service outside your home. The project says its speech processing uses Microsoft Azure, which it describes as bound by commercial terms and conditions that do not retain or store your data. The project also says that Home Assistant Cloud itself does not keep any record of your voice, data or commands ²⁰. These are the vendor's statements about its own service and its supplier. We did not read Microsoft's terms for this guide, so if you need to rely on them, read the terms directly before you decide.

The name is changing. The project says the service is planned to be renamed Home Assistant Link on December 2, 2026 ²¹. Pages and search results written before the change may still say Cloud, which is the name used throughout this guide.

Built-in commands cover devices, areas, timers and lists

The built-in sentences are the commands Assist understands without a language model. The project's starter list covers six groups, each with its own examples in the documentation. Our question on the voice commands Assist understands walks through the most useful ones.

Every device or area you name must first be exposed to Assist. The project's setup steps say that, if you have not done so, you will not be able to control the device by voice ⁵. Names must also match the ones in Home Assistant exactly, or you need an alias for them.

Timers have their own rules. A timer you start by voice is a voice timer, and it is separate from the timer helper entities you can use in dashboards, scripts and automations ²².

The list is long but not complete. The project says the supported sentences are updated for each language and cannot all be listed, and it points to the test sentences in its GitHub repository for your language. The page also explains that the second part of each test file name is the intent. If Assist does not understand you, the project's advice is to rephrase the sentence or to check that the entity or area name matches what is in Home Assistant.

Built-in commands do not handle everything. Requests such as asking for advice, or a command phrased in an unusual way, need a language model. The next section covers that.

A language model can handle requests the built-in commands do not

Assist and a language model are not competitors in the default setup. Home Assistant says Assist handles commands first, and only questions or commands it does not understand go to the AI you have set up. Its example is that turning on the kitchen light is handled by Assist, while a request for help in a dark kitchen goes to the AI ²³. The prefer handling commands locally setting, which can be chosen for each assistant, makes the same rule explicit.

That order is also the reason Speech-to-Phrase cannot be paired with a language model in the usual way. The project's June 2025 post says you cannot use an LLM as a fallback with it, because unexpected commands cannot be converted into text for the model to process ²⁴. If you want a model to handle the requests that the built-in commands miss, you need an open speech engine such as Whisper.

Home Assistant supports two ways to connect a model, and they differ in where the request goes and who pays for it.

The Ollama settings matter most for a local setup. The project says that smaller models are more likely to make mistakes, and that they may not reliably keep a conversation going while they control Home Assistant. The advice to expose fewer than 25 entities is a practical ceiling for a small model, and it is the project's general guidance rather than a test of one model.

The OpenAI option keeps the cost in the open. Each request is billed to your own OpenAI account, and the project advises you to watch your costs and set usage limits. Its storage setting is off by default, which means requests and responses are not kept in your OpenAI dashboard logs unless you turn it on. That default describes the Home Assistant setting only. Your account and OpenAI's own policies apply as well.

If you want only commands, Speech-to-Phrase on a Raspberry Pi is the fastest local setup, and you need no language model. If you want Assist to answer questions, follow a conversation, and act on devices as well, pair Whisper on an N100-class machine with a local or cloud language model. For the wider question of how language models change speech recognition, read our guide to speech to text and LLMs. For the difference between an assistant that acts and a dictation tool that types, see voice assistants compared with dictation, and for agents that hold a conversation on their own, see AI voice agents. Our question on whether Home Assistant can use ChatGPT or a local LLM for voice gives the short answer, and the Google Assistant voice commands guide compares a different assistant.

For the speech engine underneath, the guide to on-device speech recognition explains how local recognition differs from cloud recognition, and the guide to running Whisper locally covers the open-source model in more depth. The real-time speech recognition guide explains how streaming recognition trades some accuracy for speed. The glossary entry for voice assistant defines the wider category.