MulKINet: Multi-Stage Key-Invariant Convolutional Neural Networks for Accurate and Fast Cover Song Identification

Cover song identification (CSI) is a challenging task in the music information retrieval (MIR) community. The employment of convolutional neural networks (CNN) have significantly improved the performance of CSI systems, especially CNN designed to be invariant against key transpositions. In this paper, we propose MulKINet, a multi-stage CNN architecture that preserve the property of key invariance while its representational ability is substantially enhanced. Combined with three options for building blocks, channel and temporal attention mechanism, we present an accurate and fast CSI system.