DLR : Toward a deep learned rhythmic representation for music content analysis

TL;DR

This paper introduces DLR, a deep learning-based rhythmic representation learned directly from raw audio, which enhances music tagging tasks by capturing compact rhythmic features.

Contribution

It proposes a novel method to learn rhythmic representations directly from raw audio using a 1D convolutional network, improving music informatics tasks.

Findings

DLR captures compact rhythmic information

DLR improves multi-tagging performance

Visualizations show meaningful rhythmic features

Abstract

In the use of deep neural networks, it is crucial to provide appropriate input representations for the network to learn from. In this paper, we propose an approach to learn a representation that focus on rhythmic representation which is named as DLR (Deep Learning Rhythmic representation). The proposed approach aims to learn DLR from the raw audio signal and use it for other music informatics tasks. A 1-dimensional convolutional network is utilised in the learning of DLR. In the experiment, we present the results from the source task and the target task as well as visualisations of DLRs. The results reveals that DLR provides compact rhythmic information which can be used on multi-tagging task.