The Impact of Label Noise on a Music Tagger

TL;DR

This paper investigates the effect of label noise on training music taggers, demonstrating that even noisy labels contain valuable information and that curated labels yield the best results.

Contribution

It quantifies how noisy labels impact music tagging performance and introduces artificial corruption to analyze their informational value.

Findings

Curated labels lead to the highest performance metrics.

High levels of noise still enable successful learning.

Artificial corruption helps measure the contribution of noisy labels.

Abstract

We explore how much can be learned from noisy labels in audio music tagging. Our experiments show that carefully annotated labels result in highest figures of merit, but even high amounts of noisy labels contain enough information for successful learning. Artificial corruption of curated data allows us to quantize this contribution of noisy labels.