Learning a Representation for Cover Song Identification Using Convolutional Neural Network

Abstract

Cover song identification represents a challenging task in the field of Music Information Retrieval (MIR) due to complex musical variations between query tracks and cover versions. Previous works typically utilize hand-crafted features and alignment algorithms for the task. More recently, further breakthroughs are achieved employing neural network approaches. In this paper, we propose a novel Convolutional Neural Network (CNN) architecture based on the characteristics of the cover song task. We first train the network through classification strategies; the network is then used to extract music representation for cover song identification. A scheme is designed to train robust models against tempo changes. Experimental results show that our approach outperforms state-of-the-art methods on all public datasets, improving the performance especially on the large dataset.

Detecting Cover Songs with Pitch Class Key-Invariant Networks

“…Our proposed network employs a feature extraction module to convert an input CQT spectrogram of a track to a fixed length embedding, similar to [14]. PiCKINet, outlined graphically in Fig.…”

“…Deep Learning (DL) methods have recently become the focus of CSI research, with improved performance recorded. While some DL-CSI methods have employed CCMs [8,9], embedding based approaches are more popular [10][11][12][13][14][15][16]. Pro loss functions [12,13,15], or classification based learning with the embedding extracted from the penultimate network layer at test time [10,11,14,16].…”

“…Semitone Constant-Q Transform (CQT) features [23], from which chroma is often extracted, are also used in DL-CSI. The earlier such networks [11,14] consider some pitch invariance by using filters of octave-based dimensions in early layers, leading to receptive fields and latent representations of several octaves range. Alternatively, the state-of-the-art ByteCover [16] uses a structure-agnostic ResNet-50, resulting in large frequency range receptive fields.…”

“…In this paper we propose a Pitch Class Key-Invariant Network (PiCKINet) for CSI. PiCKINet affords use of the better-performing CQT input features [14,16], which are transformed to a pitch class dimension activation map by a key-invariant multi-octave filter. As PiCKINet does not employ octave-based pooling as in [10,13], there is no need to employ filters with large temporal dimensions; key-invariant convolutions are applied to pitch class dimension activation maps throughout the network.…”

Bytecover: Cover Song Identification Via Multi-Loss Training

“…Roughly speaking, existing deep learning methods to CSI can be classified into two categories. The first category of methods, e.g., [8][9][10], treats CSI as a multi-class classification problem, where each version group is considered as an unique class. Convolutional neural networks (CNNs) are trained to classify music tracks in the training set and during retrieval, the network's penultimate layer is used to generate feature for audio matching.…”