Speech-to-text AI running on an ESP32-S3 (real run, no cloud)
I ran a speech-to-text AI model on an ESP32-S3 microcontroller. No cloud. No laptop doing the inference. 🎙️➡️📟 This is the real, unedited run: • The browser records my voice and streams 16 kHz audio to the board over USB • The ESP32-S3 runs Whistle (55M parameters, by Cactus Compute) entirely on the chip • 6.4 s of speech → text back in 20 s How it fits on a 16 MB board: ✂️ The full model is 16.9 MB, so I cut the decoder from 8 layers to 4 → 14.5 MB, about 11.5% word error rate (9.7% for the full model) ⚡ Inference engine written from scratch in C, using the S3 vector unit, both cores and 2-bit weights read straight from flash 🚀 From 88.9 s down to 9.9 s for a 2.25 s clip It's slow and it's a small model, so it makes mistakes. But a microcontroller turning speech into text on its own is real. Full write-up on my blog: https://amal.vs.binarybit.in/blog/6adf1f677027b216982a9889 Code: https://github.com/alchemist123/STT-Model-ESP32 Model: Whistle by Cactus Compute (Apache-2.0) #ESP32 #EdgeAI #TinyML #SpeechToText #Embedded #AI