Real-time and energy efficient signal feature extraction has become increasingly important for machine-learning-enabled smart sensor systems in mobile and Edge applications. As considerable scientific and technological efforts have been devoted to developing tactile sensing with prospective applications in many fields, such as smart prosthetics, remote palpation, and robotic surgery with the sense of touch; in this paper, we develop a parallel hardware-software signal feature extraction method and apply it to a dataset of tactile texture classification. Being easily parallelizable, a set of passband-power feature extraction blocks compute signal power in various passbands and can be clock gated for accuracy-energy trade-offs controlled by a proposed feature summarization algorithm. Our experimental results on the tactile dataset have shown that the proposed method works at high levels of parallelization and realtimeness, performs with lower computational complexity, and achieves accuracy levels comparable to those of convolutional neural networks.