Temporal Modeling of Classical Piano Music using GRU-based Recurrent Neural Networks

This Paper introduces a recurrent neural network framework for music generation focusing on modeling the timing of notes in classical piano music. The system utilizes stacked recurrent neural networks with gated recurrent units (GRU), trained on Maestro dataset, which includes over 200 hours of aligned MIDI and audio recordings-from global piano compositions. Distinct from traditional methods involving generative adversarial networks or transformer models, our approach highlights how deep temporal modeling of note-onset sequences enables the generation of coherent, long-form compositions while maintaining stylistic fidelity. The proposed model attains a note prediction accuracy of 89.6 percent on the validation set by employing curriculum learning strategies that incrementally address polyphonic complexity. Quantitative assessment based on the Music Theory Similarity Index (MTSI) demonstrates an 82.4 percent alignment with classical harmonic progressions, outperforming baseline LSTM models by 14.2 percent. The primary contribution lies in the introduction of a dynamic temperature sampling mechanism, which effectively balances exploration and exploitation during generation, thereby mitigating the over-smoothing phenomenon commonly observed in recurrent neural network-based systems. This study advances automated composition systems by illustrating that targeted temporal modeling of single-instrument repertoires can achieve higher stylistic fidelity than generalized multi-instrument frame-works. The findings indicate strong potential for applications in computer-assisted composition and music education, particularly in preserving historical performance practices through AI-based analysis and generation. This work addresses these challenges by leveraging Gated Recurrent Units (GRUs), a variant of recurrent neural networks known for their efficiency and effectiveness in modeling sequential data with long-term dependencies.