Chakavak: A Multi-aspect Dataset for Automatic Tagging of Microtonal Music

Overview

Chakavak is a comprehensive dataset for automatic tagging and analysis of Persian classical music (also known as Iranian traditional music), which features microtonal characteristics. The dataset provides multi-aspect tags for each audio segment, covering dastgah (modal system), instruments, rhythmic properties, and performance type classifications.

Dataset Description

The Chakavak dataset contains 7,000+ labeled audio segments from Persian classical music recordings. Each segment is approximately 30 seconds long and is tagged with multiple aspects:

Tag Categories:

1. Dastgah (Modal Systems):

   • Chahargah
   • Homayoun
   • Mahur
   • Nava
   • Rast-Panjgah
   • Segah
   • Shur
   • Abuata
   • Afshari
   • Bayat-Esfahan
   • Bayat-Tork
   • Dashti

2. Musical Instruments:

   • Tar
   • Setar
   • Santur
   • Ney
   • Kamanche
   • Violin
   • Piano
   • Tonbak
   • Oud
   • Qanun
   • Daf

3. Performance Attributes:

   • Rhythm (with/without rhythm)
   • Avaz (singing style)
   • Man_voice (male voice presence)
   • Solo/Ensemble performance

Repository Structure

chakavak/
├── LICENSE                        # CC BY-NC-SA 4.0 License file
├── environment.yml                # Conda environment specification
├── data/
│   ├── audio/                     # Directory for MP3 files (available upon request)
│   ├── metadata/                  # Directory for CSV and JSON metadata
│   │   ├── chakavak_tags.csv      # Original dataset labels
│   │   ├── metadata.json          # Processed metadata in JSON format
│   │   ├── mfcc/                  # MFCC features for each audio file
│   │   ├── spectrograms/          # Spectrogram images for visualization
│   │   ├── chromagrams/           # Chroma features extraction
│   │   └── pitch/                 # Pitch tracking features
│   └── splits/                    # Directory for dataset splits
│       ├── standard/              # Standard train-validation-test split (70-15-15%)
│       ├── k_fold_5/              # 5-fold stratified cross-validation
│       ├── k_fold_10/             # 10-fold stratified cross-validation
│       ├── grouped_k_fold_5/      # 5-fold grouped stratified cross-validation
│       └── grouped_k_fold_10/     # 10-fold grouped stratified cross-validation
├── notebooks/
│   ├── 01_data_exploration.ipynb  # Dataset exploration and statistics
│   ├── 02_feature_extraction.ipynb # Generate metadata features
│   ├── 03_data_splitting.ipynb    # Create and analyze dataset splits
│   └── 04_example_usage.ipynb     # Example usage of dataset features
└── scripts/
    ├── extract_features.py        # Scripts for feature extraction
    ├── create_splits.py           # Script to create dataset splits
    ├── visualize_features.py      # Scripts for visualization
    └── utils.py                   # Utility functions

Installation and Setup

1. Clone the repository:

   git clone https://github.com/mehdikiani/chakavak.git
   cd chakavak

2. Create the conda environment:

   conda env create -f environment.yml
   conda activate chakavak

3. Request audio files: The audio files are not included in the repository due to size limitations. Please contact the authors to request access to the audio files.

Using the Dataset

Loading the Dataset

import pandas as pd

# Load the dataset labels
tags_df = pd.read_csv('data/metadata/chakavak_tags.csv')

# Preview the data
print(tags_df.head())

# Check the distribution of dastgah (modal systems)
dastgah_columns = ['chahargah', 'homayoun', 'mahur', 'nava', 
                   'rast_panjgah', 'segah', 'shur', 'abuata', 
                   'afshari', 'bayat_esfahan', 'bayat_tork', 'dashti']
dastgah_distribution = tags_df[dastgah_columns].sum().sort_values(ascending=False)
print(dastgah_distribution)

Using Extracted Features

import numpy as np
import json

# Load metadata with file paths
with open('data/metadata/metadata.json', 'r') as f:
    metadata = json.load(f)

# Load MFCC features for a specific file
file_id = "abuata_042_720_750"
mfcc_features = np.load(f"data/metadata/mfcc/{file_id}.npy")

print(f"MFCC shape: {mfcc_features.shape}")

Using Data Splits

# Load a standard train/val/test split
train_df = pd.read_csv('data/splits/standard/train.csv')
val_df = pd.read_csv('data/splits/standard/validation.csv')
test_df = pd.read_csv('data/splits/standard/test.csv')

print(f"Train set: {len(train_df)} samples")
print(f"Validation set: {len(val_df)} samples")
print(f"Test set: {len(test_df)} samples")

# Or use a specific k-fold split
fold_0 = pd.read_csv('data/splits/k_fold_5/fold_0.csv')
print(f"Fold 0 size: {len(fold_0)} samples")

Feature Extraction

The repository includes various pre-extracted features for music analysis:

1. MFCC (Mel-Frequency Cepstral Coefficients): Captures timbral characteristics

   • Shape: (n_frames, 20) for each audio segment
   • Found in data/metadata/mfcc/

2. Chroma Features: Represents pitch class distribution, useful for modal analysis

   • Shape: (n_frames, 12) for each audio segment
   • Found in data/metadata/chromagrams/

3. Pitch Tracking: Fundamental frequency estimation using CREPE

   • Found in data/metadata/pitch/

4. Spectrograms: Visual frequency representation

   • Found in data/metadata/spectrograms/

To extract features for new audio files, use the provided scripts:

from scripts.extract_features import extract_all_features

# Extract all features for a given audio file
extract_all_features('path/to/audio.mp3', 'output_id')

Citation

If you use this dataset in your research, please cite:

Kiani, M., Ramezani, R., & Ayanbod, M. H. (2025). Chakavak: A Multi-aspect Dataset for Automatic Tagging of Microtonal Music. 
Computer Engineering Faculty, University of Isfahan, & Music Faculty, Art University, Isfahan, Iran.
https://github.com/mehdikiani/chakavak

License

This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0). See the LICENSE file for details.

Acknowledgments

We thank all the musicians and artists whose works have been included in this dataset.

Contact

For any questions or to request access to the audio files, please contact:

• Mehdi Kiani: mkiani3000@gmail.com