Fusing Memory and Attention for Symbolic Music Generation

This repository accompanies the paper:

Fusing Memory and Attention: A Study on LSTM, Transformer, and Hybrid Architectures for Symbolic Music Generation

The project provides a systematic empirical analysis of LSTM, Transformer, and a Hybrid Transformer–LSTM architecture for symbolic melody generation, with an emphasis on the trade-off between local melodic continuity and global structural coherence.

---

Repository Structure

This repository is intentionally organized by research objective, with each branch corresponding to a distinct experimental track reported in the paper.

main/
├── dataset/
│   ├── processed/
│   │   ├── train_dataset
│   │   ├── train_mapping
│   │   ├── val_dataset
│   │   ├── val_mapping
│   │   ├── test_dataset
│   │   └── test_mapping
│   └── seeds.txt
└── preprocess.py

lstm/                # LSTM baseline experiments
transformer/         # Transformer baseline experiments
hybrid/              # Proposed Transformer–LSTM hybrid model
hybrid-ablation/     # Controlled ablation studies on the hybrid model

Branch Overview

• main (this branch)

  • Contains the fully preprocessed dataset used across all experiments
  • Includes the reproducible preprocessing pipeline (preprocess.py)
  • Serves as the canonical entry point for the project

• lstm

  • Training and evaluation scripts for the LSTM baseline
  • Generated melodies, metric summaries, and logs

• transformer

  • Decoder-only Transformer model implementation
  • Generation outputs and evaluation results

• hybrid

  • Proposed Transformer Encoder + LSTM Decoder architecture
  • Full training, inference, and evaluation pipeline

• hybrid-ablation

  • Ablation experiments isolating architectural components
  • Encoder depth, attention width, decoder capacity, embeddings, positional encoding, and regularization studies

---

Dataset

All experiments are conducted on the Deutschl subset of the Essen Folk Song Collection (available via the Humdrum KernScores repository: Deutschl collection), consisting of monophonic symbolic melodies in Kern (.krn) format.

Preprocessing Highlights

• Key normalization:

  • All major keys → C major
  • All minor keys → A minor

• Duration filtering to a fixed rhythmic vocabulary

• Symbolic encoding using:

  • MIDI pitch numbers
  • Explicit rest tokens
  • Duration continuation markers (_)

• Fixed sequence length: 64 tokens

• Train/Validation/Test split: 80 / 6 / 14 (fixed random seed)

The processed datasets in this branch are identical across all models, ensuring fair architectural comparison.

---

Experimental Summary

Across all experiments, 1,000 melodies per architecture were generated from a shared set of fixed seed sequences and evaluated using 17 musically grounded metrics, spanning both local and global musical properties.

Key Findings

• LSTM

  • Strong local continuity and stylistic consistency
  • Weak long-range structure retention

• Transformer

  • Improved global dependency modeling
  • Occasional irregular phrasing and local instability

• Hybrid Transformer–LSTM (Proposed)

  • Best overall balance between:

    • Local melodic smoothness
    • Global structural coherence

  • Highest scores in pitch variance, motif diversity, entropy, and perceptual originality

A structured human listening study further confirmed that the hybrid model produces melodies rated higher in musicality, coherence, and overall quality.

---

Reproducibility

• All preprocessing steps are deterministic
• Identical seed sequences are used across models
• Metric computation is shared across experimental branches

To regenerate the dataset from raw .krn files:

python preprocess.py

Note: Raw .krn files are not redistributed due to dataset licensing. Please obtain the Essen Folk Song Collection separately.

---

Citation

If you use this code or dataset in your research, please cite the paper:

BibTeX

@article{GHOSHAL2026131173,
  title = {Fusing memory and attention: A study on LSTM, transformer and hybrid architectures for symbolic music generation},
  journal = {Expert Systems with Applications},
  volume = {308},
  pages = {131173},
  year = {2026},
  issn = {0957-4174},
  doi = {https://doi.org/10.1016/j.eswa.2026.131173},
  url = {https://www.sciencedirect.com/science/article/pii/S0957417426000874},
  author = {Soudeep Ghoshal and Sandipan Chakraborty and Pradipto Chowdhury and Himanshu Buckchash},
  keywords = {Deep learning LSTM, Musical structure analysis, Sequence modeling, Symbolic music generation, Transformer}
}

---

License

This repository is released for academic and research use only.

Please refer to the individual dataset and library licenses for additional restrictions.

---

Contact

For questions, clarifications, or collaboration inquiries:

Soudeep Ghoshal Email: soudeepghoshal.sg@gmail.com

---

This repository is structured to emphasize clarity, reproducibility, and architectural insight rather than leaderboard optimization.