Kazakos Evangelos ; Nagrani Arsha ; Zisserman Andrew ; Damen Dima

Abstract

We propose a two-stream convolutional network for audio recognition, thatoperates on time-frequency spectrogram inputs. Following similar success invisual recognition, we learn Slow-Fast auditory streams with separableconvolutions and multi-level lateral connections. The Slow pathway has highchannel capacity while the Fast pathway operates at a fine-grained temporalresolution. We showcase the importance of our two-stream proposal on twodiverse datasets: VGG-Sound and EPIC-KITCHENS-100, and achieve state-of-the-artresults on both.