CPJKU Submission to DCASE21: Cross-Device Audio Scene Classification with Wide Sparse Frequency-Damped CNNs

We describe the CP-JKU team’s submission for Task 1A Low-Complexity Acoustic Scene Classification with Multiple Devices of the DCASE2021 Challenge. We use Receptive Field (RF) regularized Convolutional Neural Network (CNN) with Frequency Damping as a baseline. We investigate widening the convolutional layers while keeping the number of parameters low by grouping and pruning. We apply iterative magnitude pruning to sparsify the weights of the models. Additionally, we investigate an adversarial domain adaptation approach.