Balancing bias and performance in polyphonic piano transcription systems

Current state-of-the-art methods for polyphonic piano transcription tend to use high capacity neural networks. Most models are trained "end-to-end", and learn a mapping from audio input to pitch labels. They require large training corpora consisting of many audio recordings of different piano models and temporally aligned pitch labels. It has been shown in previous work that neural network-based systems struggle to generalize to unseen note combinations, as they tend to learn note combinations by heart. Semi-supervised linear matrix decomposition is a frequently used alternative approach to piano transcription-one that does not have this particular drawback. The disadvantages of linear methods start to show when they encounter recordings of pieces played on unseen pianos, a scenario where neural networks seem relatively untroubled. A recently proposed approach called "Differentiable Dictionary Search" (DDS) combines the modeling capacity of deep density models with the linear mixing model of matrix decomposition in order to balance the mutual advantages and disadvantages of the standalone approaches, making it better suited to model unseen sources, while generalization to unseen note combinations should be unaffected, because the mixing model is not learned, and thus cannot acquire a corpus bias. In its initially proposed form, however, DDS is too inefficient in utilizing computational resources to be applied to piano music transcription. To reduce computational demands and memory requirements, we propose a number of modifications. These adjustments finally enable a fair comparison of our modified DDS variant with a semi-supervised matrix decomposition baseline, as well as a state-of-the-art, deep neural network based system that is trained end-to-end. In systematic experiments with both musical and "unmusical" piano recordings (real musical pieces and unusual chords), we provide quantitative and qualitative analyses at the frame level, characterizing the behavior of the modified approach, along with a comparison to several related methods. The results will generally show the fundamental promise of the model, and in particular demonstrate improvement in situations where a corpus bias incurred by learning from musical material of a specific genre would be problematic.