Saulo Verissimo has published whistle-esp32 on GitHub, a speech-to-text demo that runs entirely on an ESP32-S3. Hold a button, say something like “make it red” or “flash the light blue”, and the board’s onboard WS2812 LED does it. No network, no cloud, no phone. It transcribes free speech rather than matching a fixed list of phrases, and it handles commands in seven languages with automatic language detection.

The model is Whistle by Cactus Compute, Apache 2.0. Cactus ships no ESP32 build and no source for the engine, so this is an independent C99 implementation with no dependencies, checked tensor by tensor against the official library. The full model is larger than the board’s flash, so a script strips the laddered decoder down to depth 4.

Matrix multiplies run in int8 on the S3’s PIE vector unit, 16 multiply-adds per instruction. A 2-second English command is processed in a few seconds at 240 MHz.…