Cactus Compute's Whistle runs offline speech recognition on a $10 ESP32S3

Cactus Compute has released Whistle, a 16.9 MB open-source speech-to-text model that runs entirely offline on a $10 ESP32S3 microcontroller. It supports seven languages and uses 8-bit integer math and dual-core processing for real-time recognition on constrained hardware.

• The model is 16.9 MB and supports seven languages, including English and German
• It runs offline on an ESP32S3 with 16 MB flash and 8 MB RAM
• Average response time is about 10 seconds, with 78% accuracy on color commands
• Released under Apache 2.0, it uses an INMP441 digital microphone