Cross-version Singing Voice Detection in Opera Recordings: Challenges for Supervised Learning

Abstract

In this paper, we approach the problem of detecting segments of singing voice activity in opera recordings. We consider three state-of-the-art methods for singing voice detection based on supervised deep learning. We train and test these models on a novel dataset comprising three annotated performances (versions) of Richard Wagner’s opera “Die Walküre.” The results of our cross-version experiments indicate that the models do not sufficiently generalize across versions even in the case that another version of the same musical work is available for training. By further analyzing the systems’ predictions, we highlight certain correlations between prediction errors and the presence of specific singers, instrument families, and dynamic aspects of the performance. With these findings, our case study provides a first step towards tackling singing voice detection with deep learning in challenging scenarios such as Wagner’s operas.