VA5: a $20 ESP32-S3 voice assistant with no personality until you write one

VA5 (Voice Assistant 5), by Udi Tirosh, is a push-to-talk voice assistant built from an ESP32-S3 dev board, an INMP441 microphone, a MAX98357A amplifier and speaker, a 1.28in GC9A01 round display and one big button, in a round 3D-printed case. The author puts the parts at under $20. Hold the button, talk, let go, and it answers out loud through the OpenAI Realtime API over a secure WebSocket, using your own OpenAI API key (so you pay OpenAI for usage directly). The point is that it has no character until you give it one: a web page served by the device lets you write a persona prompt of up to 12 KB, pick one of ten OpenAI voices and a reply language, and swap the animated emoji face for your own GIF, PNG, JPG or MP4, with no reflashing. Personas export as a .zip, and a Browse gallery pulls ready-made ones from the repo (Ada Lovelace, Marie Curie, Isaac Newton, Abraham Lincoln, Winston Churchill, Nietzsche, Yoda, Daffy Duck and more). It came out of classrooms, where students research a figure, write the persona and then have to hold a conversation with it. The model also calls a few built-in tools: it picks the face's mood, changes its own volume when you ask it to speak up, and can show its IP address on screen. The board's built-in status LED shows listening, thinking and speaking states, and the device remembers up to six Wi-Fi networks, set up through a captive portal on first boot. You can breadboard it with jumper wires, or order the 2-layer KiCad PCB from the gerbers in the repo; an optional PCM5102A DAC adds a 3.5 mm headphone jack. Firmware is Arduino (ESP32 core 3.3.8), built in the Arduino IDE or PlatformIO, and the emoji frames go onto a separate LittleFS partition. The board needs to be the N16R8 version (16 MB flash, 8 MB PSRAM). GPL-3.0 licensed.