Onset detection via separation of harmonic content from musical notes

Abstract

A novel method for onset detection in single-channel audio recordings is presented and evaluated. Here, source separation techniques are used as a preprocessing stage for extracting the harmonic content of musical notes. The residual channel is then used to estimate an onset detection function whose peaks align with note transitions. Several tests are conducted on a selected dataset in order to evaluate its performance and compare with alternative algorithms. The results provide evidence that the proposed residual-based method can achieve comparable levels of accuracy without the need of previous training stages.