Global Structure-Aware Drum Transcription Based on Self-Attention Mechanisms

This paper describes an automatic drum transcription (ADT) method that directly estimates a tatum-level score from music signal in contrast to most conventional ADT methods estimate the frame-level onset probabilities of drums. To score, we propose deep model consists encoder for extracting latent features and decoder estimating pooled at tatum level. capture global repetitive structure scores, which is difficult learn with recurrent neural network (RNN), introduce self-attention mechanism tatum-synchronous positional encoding into decoder. mitigate difficulty training self-attention-based insufficient amount paired data improve musical naturalness estimated regularized uses structure-aware masked language (score) pretrained extensive collection scores. The experimental results showed proposed outperformed RNN-based terms error rate F-measure, even when only limited was available so non-regularized underperformed model.