Firefly algorithm-based LSTM model for Guzheng tunes switching with big data analysis

Area of Science:

• Computational Musicology
• Artificial Intelligence in Music
• Algorithmic Composition

Background:

• Guzheng tune progression relies on complex melodic motif transitions, posing challenges for automated generation.
• Existing methods struggle to navigate the vast creative space for musically consistent elaborations.

Purpose of the Study:

• To develop a specialized model for generating musically consistent Guzheng tune transitions.
• To enhance the exploration of creative possibilities in Guzheng music generation.

Main Methods:

• Enhanced Firefly Algorithm (FA) with adaptive diversity preservation and swim parameters.
• Specialized stacked LSTM architecture with residual connections and conditioned embedding vectors.
• Unsupervised learning of Guzheng-specific melody embeddings using a variational autoencoder.
• Adversarial generative models for realistic Guzheng tune synthesis and evaluation.

Main Results:

• Achieved a 63% reduction in reconstruction error compared to standard FA optimization.
• Outperformed baselines in motif capture, modality coherence ( <2% dissonant pitch errors), and rhythmic cadence retention.
• Generated tune transitions rated highly for naturalness, novelty, and stylistic faithfulness by users.

Conclusions:

• The proposed FA-enhanced LSTM model significantly improves the fidelity and musicality of generated Guzheng tune transitions.
• The model effectively captures unique harmonic signatures and long-range temporal dependencies in Guzheng music.
• This work offers a robust framework for AI-driven algorithmic composition in traditional Chinese music.