Guitar note onset detection based on a spectral sparsity measure

摘要

The detection of note onsets is gaining a growing interest in audio signal processing research due to its wide range of applications in music information retrieval. We propose a new note onset detection algorithm NINOS2 exploiting the spectral sparsity difference between different parts of a musical note. When compared to the popular state-of-the-art LogFiltSpecFlux algorithm, the proposed algorithm shows up to 61% better performance for automatically annotated guitar melodies as well as chord progressions. We also propose an additional performance measure to assess the relative position of detected onsets w.r.t. each other.