Homer speaker and brains

The Homer speaker is your home's voice assistant, and it runs on your computer. It hears “Homer, turn on the bedroom light”, works out what you said, does simple things itself and hands complex ones to a bigger model.

All setup is in the web version: Settings → Speaker. In the app, the same cards are on the Settings → Home tab. The cards are “Speaker names”, “Speaker speech and voice”, “Homer brains”, “Speaker microphones” and “Household voices”.

The speaker speaks the home's language

The speaker uses the language of the home: Russian or English. In an English home:

• you call it “Homer” or “Hey Homer”;
• it recognizes and answers in English;
• it tells the time in 12-hour format (“It's 3:05 PM”).

Alice and Yandex Station speak Russian only. In an English home they are not used: “Homer, turn on Alice” is off, and the brains don't hand anything to a Station.

The path of one phrase

Microphone→The name “Homer” (Vosk)→Recognition (speaches)→Quick light commands→Brains: Claude Haiku→Home tools→Voice (Piper)

1. The name. The speaker listens to the microphone all the time with Vosk, a light offline recognizer, and waits for its name. Without the name, nothing goes further.
2. Recognition. A phrase with the name goes to speaches, a speech recognition and synthesis server in Docker. Without speaches the speaker falls back to Vosk and the Windows voice.
3. Quick commands. Common commands like “turn on the living room light”, “brightness to half” or “turn everything off” are handled by the Homer server in a fraction of a second, with no neural network.
4. Brains. Everything else is answered by Claude Haiku. It sees a snapshot of the home: rooms, devices, states. It turns things on and off and calls tools, and hands multi-step jobs to Claude Opus.
5. Voice. The answer is spoken by the chosen Piper voice. In a Russian home, radio, fairy tales and weather can be handed to a Yandex Station.

If the main brain doesn't answer, the speaker asks a fallback brain: the same cloud model directly, then a local model on the computer.

If the computer with the speaker is off, a backup speaker on the home server can take over. It handles lights and quick commands itself, and complex things with a cloud model if the home uses a cloud with a key.

The living room has a third speaker, Smartbox: its microphone and loudspeaker connect to the computer over the network, or to the home server when the computer is off.

A satellite can also be your own board on ReSpeaker. And if the home has confirmed household voices, Homer knows who is speaking and keeps cameras, tasks, memory and plan editing away from guests.

What you need

• A Windows 10 or 11 computer for the speaker, and a microphone. A webcam with a microphone works well.
• Claude Code signed in to a Claude Pro or Max subscription. The brains run through it; no separate API key is needed.
• Docker Desktop for speaches. With an NVIDIA GPU, recognition is almost instant.
• Python 3.12 and packages:

  pip install numpy sounddevice vosk livekit comtypes mcp paho-mqtt yandex-music pystray

• For music: a Yandex Plus subscription and the Yandex Station integration connected to the home.

Where to start

1. Start the speaker on the computer.
2. Set up speaches and pick a voice.
3. Say “Homer, turn on the light” and look at the speaker log.
4. Optional: household voices, Homer via Alice, custom skills, Telegram log, connecting to Claude Code.