A Transformer Based Pitch Sequence Autoencoder with MIDI Augmentation

Despite recent achievements of deep learning automatic music generation algorithms, few approaches have been proposed to evaluate whether a single-track excerpt is composed by automatons or Homo sapiens. To tackle this problem, we apply masked language model based on ALBERT for composers classification. The aim obtain that can suggest the probability MIDI clip might be condition auto-generation hypothesis, and which trained with only AI-composed MIDI. In paper, amount parameters reduced, two methods data augmentation are as well refined loss function prevent overfitting. experiment results show our ranks \(3^{rd}\) in all 7 teams challenge CSMT (2020). Furthermore, inspiring method could spread other information retrieval tasks small dataset.