OnDA: On-device Channel Pruning for Efficient Personalized Keyword Spotting

Researchers Matteo Risso, Alessio Burrello, and Daniele Jahier Pagliari from Politecnico di Torino have introduced OnDA, a breakthrough system for efficient personalized keyword spotting that couples weight adaptation with architectural optimization. Unlike traditional approaches that only fine-tune model weights on-device, OnDA implements online structured channel pruning—dynamically removing unnecessary neural network channels based on actual user data patterns. This dual adaptation approach addresses the critical challenge of always-on voice assistants needing to adapt to individual users and environments while operating under severe latency and energy constraints on edge devices.

On the technical front, OnDA achieves remarkable efficiency gains: up to 9.63x model compression compared to unpruned baselines while maintaining equivalent task performance measured as accuracy at 0.5 false alarms per hour. When deployed on NVIDIA's Jetson Orin Nano embedded GPU, the system demonstrates 1.52x/1.57x improvements in latency and 1.64x/1.77x improvements in energy consumption during online training and inference respectively, compared to weights-only adaptation. The research, submitted to Interspeech 2026, represents a significant advancement toward truly personalized, battery-friendly voice interfaces that can learn continuously without compromising responsiveness or privacy.