Persian Traditional Music Retrieval

This repository contains code for analyzing and retrieving Persian traditional music using self-supervised pretrained models. The project evaluates three tasks: instrument classification, Dastgah recognition, and artist identification using the Nava dataset.

Features

• Apply and compare Music2vec, MusicHuBERT, and MERT.
• Fine-tune pretrained models for improved accuracy.
• Multi-layer feature fusion to enhance representations.
• Scripts for preprocessing, training, evaluation, and feature fusion.

Dataset

The Nava dataset includes:

• Instruments: Tar, Setar, Santur, Kamancheh, Ney, and more
• Dastgah modes: 7 primary Dastgah
• Artists: Solo performer annotations

⚠️ Dataset is not included due to licensing. Please obtain it through the following link: Nava Dataset

Results

• Instrument Classification: 99.64% accuracy
• Dastgah Recognition: 24.70% accuracy
• Artist Identification: 79.25% accuracy

Accuracy improved via fine-tuning and multi-layer feature fusion.

Citation

If you use this code, please cite:

@article{effectiveness2025BabaAli,
  title={On the effectiveness of self-supervised pre-trained models for Persian traditional music information retrieval},
  author={BabaAli, Bagher & Mohseni, Pouya},
  journal={-},
  year={2025}
}