Tempo vs. Pitch: understanding self-supervised tempo estimation

TL;DR

This paper investigates the robustness of self-supervised models for tempo estimation in music, analyzing how input representations and data distribution affect model performance through experiments with synthetic data.

Contribution

It provides new insights into the fragility of self-supervised tempo estimation models and explores mitigation strategies by dissecting the relationship between input representation and data distribution.

Findings

Self-supervised models are sensitive to data distribution variations.

Input representation significantly impacts model robustness.

Synthetic data experiments reveal potential mitigation pathways.

Abstract

Self-supervision methods learn representations by solving pretext tasks that do not require human-generated labels, alleviating the need for time-consuming annotations. These methods have been applied in computer vision, natural language processing, environmental sound analysis, and recently in music information retrieval, e.g. for pitch estimation. Particularly in the context of music, there are few insights about the fragility of these models regarding different distributions of data, and how they could be mitigated. In this paper, we explore these questions by dissecting a self-supervised model for pitch estimation adapted for tempo estimation via rigorous experimentation with synthetic data. Specifically, we study the relationship between the input representation and data distribution for self-supervised tempo estimation.