Multi-Template Shift-Variant Non-Negative Matrix Deconvolution for Semi-Automatic Music Transcription

For the task of semi-automatic music transcription, we extended our framework for shift-variant non-negative matrix deconvolution (svNMD) to work with multiple templates per instrument and pitch. A k-means clustering based learning algorithm is proposed that infers the templates from the data based on the provided user information. We experimentally explored the maximum achievable transcription accuracy of the algorithm and evaluated the prospective performance in a realistic setting. The results showed a clear superiority of the Itakura-Saito divergence over the Kullback-Leibler divergence and a consistent improvement of the maximum achievable accuracy when each pitch is represented by more than one spectral template.