Multiple-f0 Estimation and Note Tracking for Piano Music for Mirex 2014 Using Temporal Evolution Information

In this submission for MIREX 2014 we utilize a novel piano music transcription algorithm based on the temporal evolution of piano notes. Most existing transcription algorithms, especially those based on Non-negative matrix factorization and Probabilistic latent component analysis, operate on a spectrogram on a frame-by-frame basis, i.e., they do not consider the temporal evolution of the notes of musical instruments, which in the case of percussive instruments is very characteristic. In our work, we propose a spectrogram factorization algorithm that uses the full spectrogram of sampled notes as template and a greedy algorithm to detect the templates to activate for a given chord. Combining this algorithm with a robust onset detection method will lead to a more accurate piano music transcription system. We also propose a new metric for spectrogram similarity.