Abstract:
Choral music recordings are a particularly challenging target for source separation due to the choral blend and the inherent acoustical complexity of the ‘choral timbre’. Due to the scarcity of publicly available multi-track choir recordings, we create a dataset of synthesized Bach chorales. We apply data augmentation to alter the chorales so that they more faithfully represent music from a broader range of choral genres. For separation we employ Wave-U-Net, a time-domain convolutional neural network (CNN) originally proposed for vocals and accompaniment separation. We show that Wave-U-Net outperforms a baseline implemented using score-informed NMF (non-negative matrix factorization). We introduce score-informed Wave-U-Net to incorporate the musical score into the separation process. We experiment with different score conditioning methods and show that conditioning on the score leads to improved separation results. We propose a ‘score-guided’ model variant in which separation is guided by the score alone, bypassing the need to specify the identity of the extracted source. Finally, we evaluate our models (trained on synthetic data only) on real choir recordings and find that in the absence of a large training set of real recordings, NMF still performs better than Wave-U-Net in this setting. To our knowledge, this paper is the first to study source separation of choral music.
