کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Architecture Design and Branching Strategies for Sound Event Detection Using Cross Separation Index

نویسندگان :

Parmis Vaghef Davari¹ Hamed Hosseini² AliReza Beigy³ Mehdi Tale Masouleh⁴ Ahmad Kalhor⁵

1- University of Tehran 2- University of Tehran 3- University of Tehran 4- University of Tehran 5- University of Tehran

کلمات کلیدی :

Sound Event Detection،Cross Separation Index،Network Design،Branching،Pruning،NAO Robot

چکیده :

Sound Event Detection(SED) is of prime importance in environmental monitoring, speech recognition, and autonomous robotic systems. In this work, the authors present a Convolutional Neural Network(CNN)-based framework that classifies and analyzes audio signals while optimizing the architecture of the network on the basis of interpretability-driven design. Specifically, they introduced a metric called Cross Separation Index(CSI), which quantifies class-wise separability across the network layers as a guide through two key architectural refinements: branching and pruning. The proposed approach assesses the evolution of CSI along the network to identify redundant or stalled layers so that they are subjected to pruning with the aim of reducing computational complexity without degrading performance. Experimental results show that up to 60.2% reduction in inference time for ResNet and 46.4% for DenseNet obtained by the CSI-guided pruning results in more than 89% fewer trainable parameters with less than 1.5% accuracy drop from the original models. Second, a CSI-driven branching allows separate processing for classes that become stagnant versus those that keep improving and is able to reduce the inference time by up to 43% in ResNet and about 36% in DenseNet without sacrificing classification quality. Overall, the proposed framework improves the efficiency and interpretability of SED while providing a principled basis for lightweight and high-performing models in robotic auditory perception and human-centered environments. The ResNet-Branched model was further validated in a realworld deployment on a NAO robot that successfully identified and verbally repeated approximately 94.2% of the test instances across seven key audio classes: blender, water, chopping, door knock, light switch, toaster and dishes.