Persian Musical Instruments Classification Using Polyphonic Data Augmentation

Repository for the paper “Persian Musical Instruments Classification Using Polyphonic Data Augmentation.”

This project introduces a culturally-informed polyphonic augmentation pipeline for classifying Persian musical instruments using a fine-tuned MERT backbone. The paper and this repository describe dataset creation, augmentation strategies, model design, and evaluation procedures.

---

🪕 Overview

We present a new approach to Persian instrument classification using synthetic polyphonic data generated by mixing monophonic recordings that share dastgāh (modal key) and tempo (BPM).

Key highlights:

• Introduces a novel Dastgāh + BPM polyphonic augmentation method.
• Uses MERT-v1-330M as the pretrained backbone for multi-label classification.
• Achieves:
  • ROC-AUC: 0.795 (Dastgāh + BPM)
  • Accuracy: 0.841 (Dastgāh-only)
• Evaluated on a manually annotated real-world test set of 491 polyphonic 5-second excerpts.

---

🎼 Dataset

Dataset on Huggingface

Summary

• Monophonic dataset: ~16,800 5-second clips across 10 instrument classes
  (Ney, Tar, Santur, Kamancheh, Daf, Tonbak, Piano, Violin, Sitar, Avaz)
• Polyphonic (synthetic) dataset: ~50,000 generated clips
• Test set: 491 manually labeled 5s real-world polyphonic excerpts

---

🎛️ Data Augmentation Modes

Mode	Description
RAMDOM	Randomly mixed clips
BPM	Same tempo, different dastgāh
DASTGAH	Same dastgāh, different tempos
DASTGAH_BPM	Same dastgāh and tempo (proposed method)

Finding: Tonal alignment (dastgāh) and tempo alignment both improve classification performance.

---

🧠 Model

• Backbone: MERT-v1-330M
• Classifier: Multi-label head (weighted layer aggregation + MLP + sigmoid)
• Loss: Binary Cross-Entropy with Logits (multi-label)
• Input: 5-second audio clips
• Output: 10 instrument probabilities

Architecture Summary

Audio (5s)
   ↓
MERT-v1-330M (pretrained)
   ↓
Weighted layer aggregation
   ↓
Two-layer MLP
   ↓
Sigmoid (multi-label output)

---

⚙️ Training Configuration

Parameter	Value
Backbone	mert-v1-330m
Optimizer	AdamW
Learning Rate	1e-4
Batch Size	16
Epochs	10
Loss	BCEWithLogitsLoss

Train Example

---

📊 Evaluation

Metrics

• Accuracy
• ROC-AUC
• F1-score (macro & micro)

Reported Results (from paper)

Data Augmentation	Accuracy	ROC-AUC	F1-score
Random	0.794	0.750	0.606
BPM	0.807	0.764	0.617
Dastgah	0.841	0.780	0.669
Dastgah + BPM	0.823	0.795	0.652

Evaluation conducted on 491 real-world 5s polyphonic samples.

---

📚 Citation

If you use this repository or dataset, please cite:

@article{persian_instruments_2025,
  title={Persian Musical Instruments Classification Using Polyphonic Data Augmentation},
  author={...},
  year={2025},
  journal={...}
}

---

🪄 Acknowledgements

• The monophonic dataset was curated from Persian traditional music recordings and annotated by expert musicians.
• Inspired by the rich modal system (dastgāh) of Persian music.
• The pretrained MERT model is provided by the Hugging Face team.

---

📜 License

This project is licensed under the MIT License — see the LICENSE file for details.