heretikio/aipi-lite-selfhost

This project shows you how to take a small pocket AI gadget called the AIPI Lite, which is a rebadged version of a device called xiaozhi-esp32, and disconnect it completely from its maker's cloud servers. Instead of sending your voice data to a company, everything runs on hardware you own and control. The repo has two main parts. The first is new firmware for the device itself, rebuilt with an English interface and an English wake word, and pointed at your own server instead of the vendor's cloud. The second part is that server, a small Docker based stack you run yourself. It handles detecting when you are speaking, converting your speech to text, and converting text back to speech, all running on an ordinary CPU rather than a specialized chip. The only thing that leaves your home network is the actual request to a language model, and that language model is one you choose and control, using tools like vLLM, Ollama, llama.cpp, or LM Studio. The end result is a voice assistant with no subscription fees, no vendor tracking, and no dependency on a company staying in business or keeping its servers running. To use this, you need the physical AIPI Lite gadget or a similar device, a USB-C cable, a computer to flash the firmware, and a separate machine like a mini-PC or NAS to run the server software. The instructions walk you through backing up the original firmware first, so you can always restore the device to its factory state, then building and flashing the new firmware, then setting up the server and connecting the device to it. Each step includes checks so you know it worked before moving to the next one. The project is released under the MIT license, and the author is upfront that it builds on top of several other open source projects rather than being built entirely from scratch.