New algorithm based on spectral distance maximization to deal with the overlapping partial problem in note–event detection

Harmonic matching pursuit (HMP) is an interesting signal processing tool for note–event detection in polyphonic music transcription. HMP decomposes an audio signal into harmonic atoms. Audio signals are well represented by harmonic atoms due to their strong harmonic content. However, HMP provides an inaccurate decomposition when musical notes with rational fundamental frequency relation are simultaneously played (the overlapping partial problem). In this paper, a signal processing algorithm dealing with this problem is proposed. The algorithm is based on maximizing a smoothness-based criterion of the spectral envelope for each harmonic atom resulting from the HMP decomposition. In this way, we obtain an improved harmonic decomposition, which achieves high accuracy rates in note–event detection when dealing with harmonically related simultaneous notes.