Advancing deep learning for expressive music composition and performance modeling.

Area of Science:

• Artificial Intelligence
• Music Information Retrieval
• Machine Learning

Background:

• AI music generation faces challenges in long-term structure and emotional nuance.
• Deep learning models have advanced AI music composition and transcription.

Purpose of the Study:

• To comparatively analyze Long Short-Term Memory (LSTM) networks, Transformer models, and Generative Adversarial Networks (GANs) for AI music.
• To evaluate AI music generation and transcription using the MAESTRO dataset.

Main Methods:

• Comparative analysis of LSTM, Transformer, and GAN architectures.
• Dual evaluation framework: objective metrics (perplexity, harmonic consistency, rhythmic entropy) and subjective Mean Opinion Score (MOS) human evaluations.
• Utilized the MAESTRO dataset for training and evaluation.

Main Results:

• Transformer models achieved the best performance on objective metrics and MOS (4.3).
• Objective metrics for Transformers: perplexity 2.87, harmonic consistency 79.4%.
• Human compositions achieved the highest perceptual quality (MOS: 4.8).

Conclusions:

• Transformer models demonstrate superior capabilities for expressive AI music generation.
• Future AI music systems require emotion-aware modeling and human-AI collaboration.
• Reinforcement learning is crucial for bridging the gap between AI and human music.