Predicting Missing Music Components with Bidirectional Long Short-Term Memory Neural Networks

Successfully predicting missing components (entire parts or voices) from complex multipart musical textures has attracted researchers of music information retrieval and music theory. However, these applications were limited to either two-part melody and accompaniment (MA) textures or four-part Soprano-Alto-Tenor-Bass (SATB) textures. This paper proposes a robust framework applicable to both textures using a Bidirectional Long-Short Term Memory (BLSTM) recurrent neural network. The BLSTM system was evaluated using frame-wise accuracies on the Nottingham Folk Song dataset and J. S. Bach Chorales. Experimental results demonstrated that adding bidirectional links to the neural network improves prediction accuracy by 3% on average. Specifically, BLSTM outperforms other neural-network based methods by 4.6% on average for four-part SATB and two-part MA textures (employing a transition matrix). The high accuracies obtained with BLSTM on both two-part and four-part textures indicated that BLSTM is the most robust and applicable structure for predicting missing components from multi-part musical textures.