Multiple-feature fusion based onset detection for solo singing voice

Abstract

Onset detection is a challenging problem in automatic singing transcription. In this paper, we address singing onset detection with three main contributions. First, we outline the nature of a singing voice and present a new singing onset detection approach based on supervised machine learning. In this approach, two Gaussian Mixture Models (GMMs) are used to classify audio features of onset frames and non-onset frames. Second, existing audio features are thoroughly evaluated for this approach to singing onset detection. Third, feature-level and decision-level fusion are employed to fuse different features for a higher level of performance. Evaluated on a recorded singing database, the proposed approach outperforms state-of-the-art onset detection algorithms significantly.