An Attack/Decay Model for Piano Transcription

We demonstrate that piano transcription performance for a known piano can be improved by explicitly modelling piano acoustical features. The proposed method is based on non-negative matrix factorisation, with the following three refinements: (1) introduction of attack and harmonic decay components; (2) use of a spike-shaped note activation that is shared by these components; (3) modelling the harmonic decay with an exponential function. Transcription is performed in a supervised way, with the training and test datasets produced by the same piano. First we train parameters for the attack and decay components on isolated notes, then update only the note activations for transcription. Experiments show that the proposed model achieves 82% on note-wise and 79% on frame-wise F-measures on the ‘ENSTDkCl’ subset of the MAPS database, outperforming the current published state of the art.