Building smallest neural Text to speech system that runs on microcontroller cheaper than cup of coff

Most TTS is a cloud API call or a multi-gigabyte model. I wanted the opposite — the smallest, most honest neural TTS I could make, small enough to run somewhere with no server at all: a microcontroller, a browser tab, whatever you've already got. What it turned into: sanoTTS, 30 trained voices across 16 languages, ranging from 294,642 parameters (337 KB) up to 2.27M. The smallest one runs on a bare ESP32-S3 dev board — a chip that costs about $3 — with no internet connection, no