Split for the big mixed dataset

Corresponding to the experiment in Section IV.E, Fig. 7 in the paper. This big mixed dataset uses all the datasets listed in Tab. I of the paper as sources and splits them to create a joint training and test dataset. For MusicNet, we use the test set MuN-10a. For Schubert_Winterreise, we employ the strict neither split. For Bach10, we use numbers 7--10 for testing and the remaining numbers for training and validation. PHENICX-Anechoic, we take the Bruckner and Mozart examples as test set. Since the ChoralSingingDataset is our only choir dataset, we enrich it by creating for each piece also a three-part mix (once leaving out each of the four parts, named nosop, noalt, noten, nobas) and use the resulting five mixes from Bruckner's Locus iste for testing, the other two pieces for training. For Bach10, PHENICX-Anechoic, and ChoralSingingDataset, this corresponds to a work or song split since all recordings in these datasets stem from the same performers and were recorded in the same session. Finally, we use the full TRIOS dataset for testing since it comprises chamber music recordings very similar to MusicNet (but with very high annotation quality). For one work, an excerpt from Brahms' horn trio op. 40, this corresponds to a version split since this piece is part of MusicNet as well; for the other four pieces, to a neither split. The combined test set comprises 42 pieces and, thus, is comparable in size to MuN-10full used before. The training set is much larger, but we increase the strides so that the number of training and validation examples is comparable to our remaining experiments on MusicNet.

Training files

Validation files

Test files