Edge AI / TinyML Keyword Spotting

An always-listening voice-command detector small enough for a microcontroller. A depthwise-separable CNN is trained on Google Speech Commands on a laptop CPU, quantised to full-integer int8 with the TensorFlow Lite converter (46.7 KB, 95.5 % on the official 12-class test set - the same as the float model) and deployed with TensorFlow Lite Micro on an ESP32 and with the LiteRT runtime on a Raspberry Pi. It recognises yes, no, up, down, left, right, on, off, stop and go, plus silence and unknown words.

Category: Edge AI / TinyML / Embedded Speech Recognition / Internet of Things

CSE Project 30 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.