On the Importance of Time and Pitch Relativity for Transformer-Based Symbolic Music Generation

This paper describes experimental investigation of music representations to draw the full potential of the Transformer with the self-attention mechanism for symbolic music generation. To use sequence-to-sequence model like the Transformer originally proposed for natural language processing, one typically serializes a musical score into a sequence of event- or note-based tokens without concern for the impact on the quality of generated music. The semantic invariance of music with respect to the time and pitch shifts is attributed to the positional relativity of musical notes over the time-pitch plane in which beats and pitch classes are repeated at intervals of bars and octaves, respectively. We here hypothesize that the capability of the self-attention mechanism to learn the musically meaningful rhythm, melody, and harmony is limited because the relativity and cyclicity of time and pitch information are not explicitly represented in the token sequence. To solve this problem, we propose a cyclicity-aware relative time and pitch encoding unique to music for the attention mechanism. Comprehensive evaluation using the POP909 dataset demonstrated that the proposed Transformer works better with event- or note-based score tokenization¹.