Learning Adversarial Transformer for Symbolic Music Generation.

Area of Science:

• Artificial Intelligence
• Music Information Retrieval
• Computational Musicology

Background:

• Symbolic music generation is challenging due to long, complex note sequences with recurring patterns.
• Transformer models show promise for long sequences but struggle with music's unique structure.
• Existing methods often adapt text generation strategies, leading to inconsistent music quality.

Purpose of the Study:

• To propose a novel adversarial transformer model for generating symbolic music with enhanced musicality.
• To address the limitations of existing transformer-based music generation methods.
• To improve the quality and consistency of generated music pieces.

Main Methods:

• Combined generative adversarial learning with self-attention networks.
• Employed adversarial objectives to guide long sequence generation, focusing on global and local structures.
• Introduced a novel Reward Each Generated Step (REGS) method, replacing time-consuming Monte Carlo search.
• Trained a discriminator with global and local loss objectives for reliable REGS.
• Guided the generator using a combination of adversarial and teacher forcing objectives.

Main Results:

• The proposed adversarial transformer generates music pieces with improved quality.
• The model effectively learns both global and local structures in music sequences.
• The REGS method provides an efficient alternative for reward computation in sequence generation.
• Experimental results show superior performance compared to original music transformers.

Conclusions:

• The novel adversarial transformer offers a promising approach for high-quality symbolic music generation.
• The integration of adversarial learning and self-attention effectively addresses challenges in long music sequence modeling.
• The proposed REGS method enhances the efficiency and convenience of training generative models for music.