Motif Transformer: Generating Music With Motifs

Music is composed of a set regular sound waves, which are usually ordered and have large number repetitive structures. Important notes, chords, music fragments often appear repeatedly. Such repeated (referred to as motifs) the soul song. However, most generated by existing generation methods can not distinct motifs like real music. This study proposes novel multi- encoders model called Motif Transformer generate containing more motifs. The constructed using an encoder-decoder framework that includes original encoder, bidirectional long short term memory-attention encoder (abbreviated bilstm-attention encoder), gated decoder. Where taken from transformer’s short-term memory network (BILSTM) attention mechanism; Both encode input encoded information representations decoder; decoder decodes entire passed enhances model’s ability capture in manner with significantly fragments. In addition, order better measure generating motifs, this evaluation metric used Experiments on multiple field metrics show proposed smoother beautiful music, contains