Multi-Stage contextual attention aggregation network for generating Tibetan folk song melodies

The application of deep learning in music generation has become a prominent research focus in artificial intelligence. Tibetan folk song melodies, recognized as a national intangible cultural heritage, are characterized by their wide range, rhythmic freedom, distinct melodic contours, and strong ethnic identity. However, research on Tibetan folk song melody generation remains largely unexplored. Existing music generation models struggle to capture the global structural logic and frequency-domain characteristics specific to Tibetan folk melodies, often resulting in distorted outputs. To address this challenge, we propose a Tibetan Folk Song Melody Generation Network based on multi-stage context-aware attention aggregation. Our approach integrates a multi-stage context-aware attention mechanism (MCOA-Attention), leveraging multi-scale convolution to extract local contextual features of the music sequence while correlating key musical elements with the original sequence to capture dynamic sub-structural information. Additionally, we design a Music Reasoning Capsule Neural Network (MRCN-Module) based on the Transformer architecture, encoding music sequences into low-level capsules and employing dynamic routing to map key information from low-level to high-level capsules. This process models the logical relationships between sub-sequences, enabling the generation of melodies that maintain both structural coherence and musical logic. Furthermore, we introduce a Spatiotemporal Wavelet Convolution Module (STW-CNN), which applies multi-level wavelet decomposition to progressively refine frequency-domain features of Tibetan folk melodies. By utilizing LSTMs to capture long-term dependencies across different frequency components, the model effectively learns the unique, melodious characteristics of traditional Tibetan music. In addition, we construct a dedicated Tibetan folk song dataset to evaluate the proposed model. Experimental results demonstrate that our approach achieves notable advantages in both subjective and objective evaluation metrics, generating Tibetan folk song melodies with realistic timbre and pitch fidelity.