Speech Recognition Without the Internet Now Fits on a $10 Chip (Whistle)

We ported Cactus Compute’s Whistle, a 16.9 MB speech-to-text model that understands seven languages, to a $10 ESP32-S3 microcontroller, where it runs completely offline with no Wi-Fi and no cloud. To make it fit, we trimmed the model’s decoder down to four layers and sped up the matrix math with 8-bit integers and both CPU cores, so you can hold a button, speak a command, and watch the onboard LED respond. We also break down how Whistle turns your voice into text, and the full project is on GitHub if you want to build one yourself. 🔗 Relevant Links Whistle: https://cactuscompute.com/blog/whistle Whistle on ESP32 S3: https://github.com/andrisgauracs/whistle-esp32 ❤️ More about us Radically better observability stack: https://betterstack.com/ Written tutorials: https://betterstack.com/community/ Example projects: https://github.com/BetterStackHQ 📱 Socials Twitter: https://twitter.com/betterstackhq Instagram: https://www.instagram.com/betterstackhq/ TikTok: https://www.tiktok.com/@betterstack LinkedIn: https://www.linkedin.com/company/betterstack 📌 Chapters: 0:00 Speech-to-text on a $10 chip 0:50 Meet Whistle by Cactus Compute 2:03 Whistle vs OpenAI Whisper 3:04 The model is bigger than the chip 3:45 Shrinking the model to fit 5:06 How speech becomes text 6:46 The hardware 7:59 Live demo 9:42 From 24 s to ~10 s 10:32 Accuracy 11:28 Recap and credits