Why it’s interesting
Smart Mouth Billy Bass turns an old novelty singing fish into a voice character that lives on a local network. An ESP32-S3 inside the fish listens for the creator's custom “Hey Billy” wake word, then passes audio to a GPU host. whisper.cpp transcribes it, Qwen3.5-9B through llama.cpp creates a response, and Kokoro speaks while the mouth, tail, and body move. No cloud AI provider is required. The creator reports roughly two seconds to first audio and demonstrates seven languages. The appeal is both the absurd talking fish and the unusually transparent path from plastic toy to reproducible local-AI device.
How to try it
Anyone can watch the final demo, installed-board photos, and build notes on the creator's page without an account. Reproducing it is real hardware work: you need a Big Mouth Billy Bass, an ESP32-S3, the custom PCB, soldering and disassembly tools, plus a GPU machine running the speech and language stack on the LAN. Firmware, backend code, wake-word model, schematic, and PCB files are in the repository. Revision one was built and installed; revision two's schematic fixes a known nFAULT grounding issue but has not been manufactured, so builders should read that warning first.
Requirements and sources
Ryan Drake published the completed build and code on September 17, 2026. The repository is Apache-2.0. The creator's photos and video document the work but have no stated open-media licence, so the official finished-build photo is used only for identification. This is marked paid and device-required because it needs a physical toy, electronics work, and a local GPU. It requires no cloud-model account or third-party login.
[Full build log](https://existens.org/billy/) · [Code and board files](https://github.com/ryandrake08/billy)