TARREAN: A Novel Transformer with a Gate Recurrent Unit for Stylized Music Generation.

Area of Science:

• Artificial Intelligence
• Music Information Retrieval
• Deep Learning

Background:

• AI-driven music generation is a growing research area.
• Existing Transformer-based models face challenges with coherence and computational expense.
• Temporal dependencies and causal structures are crucial for realistic music sequences.

Purpose of the Study:

• To propose a novel Transformer-based model, TARREAN, for improved AI music generation.
• To address issues of coherence and high computational costs in current music generation methods.
• To enhance the quality and human-likeness of AI-generated music.

Main Methods:

• Developed TARREAN, a Transformer model integrating a gate recurrent unit (GRU) and root mean square norm restriction (RMS Norm).
• Employed masked multi-head attention to maintain causal structure during training.
• Utilized a compound word method for encoding music sequences into discrete symbol-event combinations.

Main Results:

• TARREAN demonstrated improved temporal coherence in generated music sequences.
• The model achieved higher subjective satisfaction scores (4.34) compared to Transformer-XL + REMI (3.79).
• Objective evaluations showed a 15% improvement in temporal coherence over traditional methods, with reduced computational costs.

Conclusions:

• TARREAN effectively mitigates discontinuity issues and enhances music generation quality.
• The proposed model offers a significant improvement in coherence and efficiency for AI music generation.
• TARREAN represents a promising advancement in creating more human-aligned AI-generated music.