A reader assumes whisper.cpp on a microcontroller is a hack, but 1.7.5’s q4_0 quantization plus ggml’s mmap means the 24 MB model never has to fit in RAM.