It started with the Onju Voice. I liked the concept — repurposing a Google Home Mini into something open and local — but the more I read about it, the more the same complaint kept coming up: a single I2S line shared between the microphones and the speaker. Half-duplex by design — it can’t listen and play audio at the same time. That’s not a firmware problem, it’s a fundamental hardware limitation.

At the time, I was already building my own DIY voice assistants using INMP441 microphones and MAX98357 DAC around an ESP32-S3. I liked them, but they had one problem I couldn’t fix: noise. They worked fine in a quiet room, close up. Anywhere else, they were unreliable. When the Home Assistant Voice Preview Edition was announced — ESP32-S3 paired with an XMOS XU316 for proper audio processing — I started thinking about combining that with the Onju concept. A proper drop-in replacement, not just a dev board stuffed into a case.

I’d designed PCBs before, but simple ones: room controllers, temperature sensors, roller blind motors. This was a different scale of complexity. I sat on the idea for a few months before finally opening KiCad in May 2025.

The first test batch arrived in July. It had two fundamental design errors and was essentially unusable. Back to the drawing board.

The second batch arrived in August. Visually, it looked right, but the XMOS wouldn’t respond on I2C — stuck at boot. And I hadn’t added an XTAG debug interface to the board, so I had no way to dig into why. Another lesson learned the hard way.

The third batch arrived in November. First fully working units. Not perfect, but stable and usable.

The fourth batch came in January 2026 — the final refinements, the beta tester units, and the board that’s been running in my own home ever since.

After four batches, eight months, a lot of mistakes, and tons of improvements, it’s finally here. And it works! You can back our campaign on Crowd Supply if you’d like one of your own.