DBTMPE: Deep Bidirectional Transformers-Based Masked Predictive Encoder Approach for Music Genre Classification

Music is a type of time-series data. As the size data increases, it challenge to build robust music genre classification systems from massive amounts Robust require large labeled data, which necessitates time- and labor-intensive data-labeling efforts expert knowledge. This paper proposes musical instrument digital interface (MIDI) preprocessing method, Pitch Vector (Pitch2vec), deep bidirectional transformers-based masked predictive encoder (MPE) method for classification. The MIDI files are considered as input. converted vector sequence by Pitch2vec before being input into MPE. By unsupervised learning, MPE based on transformers designed extract representations automatically, musicological insight. In contrast other deep-learning models, such recurrent neural network (RNN)-based enables parallelization over time-steps, leading faster training. To evaluate performance proposed experiments were conducted Lakh dataset. During training, approximately 400,000 segments utilized MPE, recovery accuracy rate reached 97%. task, indicators more than 94%. experimental results indicate that improves compared with state-of-the-art models.