New 'next iRDT' extractor runs 10x faster with zero multiplications

Keyword spotting on tiny devices has long relied on feature extractors like MFCC or CNN-based learned features — both computationally heavy for ultra-low-power microcontrollers. In a new arXiv paper, Radu Dogaru and Ioana Dogaru introduce 'next iRDT,' a feature extractor that eliminates multiplication entirely, using only additions, comparisons, and simple bitwise operations. This drastically cuts both computational complexity and energy consumption, making it a perfect fit for TinyML platforms where every operation costs battery life.

Tested on Google's 12-class keyword spotting dataset, next iRDT achieves accuracy comparable to MFCC and CNN baselines when paired with simple classifiers, and hits 94.7% validation accuracy with a more sophisticated classifier. Crucially, processing time is at least an order of magnitude (10x+) shorter than MFCC on CPU, and the hardware footprint is minimal. This suggests on-device voice control for wearables, smart sensors, and battery-powered IoT could become dramatically more efficient, enabling always-on listening without the usual power penalty.