How Tiny Can Analog Filterbank Features Be Made for Ultra-low-power On-device Keyword Spotting

Watch a 16-minute research symposium presentation exploring power optimization techniques for analog filterbank features in ultra-low-power keyword spotting systems. Learn how Columbia University PhD student Subhajit RAY investigates the architectural parameters of analog filterbanks to achieve significant power reductions while maintaining high accuracy. Discover the comparison between analog and digital approaches, understand the fundamentals of analog audio feature extraction, and explore how careful parameter selection led to a 33.6x power reduction with only a minimal 1.8% accuracy loss in a 10-word keyword spotting system. Follow along through key topics including introduction, motivation, analog vs digital comparison, analog audio feature extraction methods, power saving strategies, and a comprehensive summary of findings.