Singing voice detection in opera recordings: A case study on robustness and generalization

Abstract

Automatically detecting the presence of singing in music audio recordings is a central task within music information retrieval. While modern machine-learning systems produce high-quality results on this task, the reported experiments are usually limited to popular music and the trained systems often overfit to confounding factors. In this paper, we aim to gain a deeper understanding of such machine-learning methods and investigate their robustness in a challenging opera scenario. To this end, we compare two state-of-the-art methods for singing voice detection based on supervised learning: A traditional approach relying on hand-crafted features with a random forest classifier, as well as a deep-learning approach relying on convolutional neural networks. To evaluate these algorithms, we make use of a cross-version dataset comprising 16 recorded performances (versions) of Richard Wagner’s four-opera cycle Der Ring des Nibelungen. This scenario allows us to systematically investigate generalization to unseen versions, musical works, or both. In particular, we study the trained systems’ robustness depending on the acoustic and musical variety, as well as the overall size of the training dataset. Our experiments show that both systems can robustly detect singing voice in opera recordings even when trained on relatively small datasets with little variety.