The issue of time-series segmentation refers to the division of continuous sensor data into discrete windows, each of which are processed individually and assigned a class label. Unlike offline data processing scenarios, segmentation for low-power wearable devices must balance the juxtaposed aims of accuracy and computational efficiency. In this paper, we propose a novel scheme for segmentation of sparse sensor data using an adaptive window size approach. Our results are benchmarked on an audio-based nutrition monitoring dataset, and show a reduction in processing overhead of 68% compared to the baseline with fixed window sizes.