MaqamNet — Musicologically Interpretable Arabic Maqam Identification

Code and derived data for the paper:

J. Bakarji, K. Allami, M. El Asal, and D. Abou Jaoude, "Musicologically Interpretable Deep Learning Framework for Arabic Maqam Identification and Analysis", in Proc. of the 27th Int. Society for Music Information Retrieval Conf. (ISMIR), Abu Dhabi, UAE, 2026.

MaqamNet bypasses explicit tonic estimation by encoding the tonic as a latent shift variable: raw octave-wrapped pitch histograms are circularly cross-correlated with maqam and jins templates from the DiArMaqAr archive, and a ResNet with circular padding classifies the resulting response tensor end-to-end.

Setup

python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

Python 3.10+ is sufficient. All experiments run on CPU (no GPU required); approximate runtimes below are for a recent laptop CPU.

Data

The data/ directory contains everything needed to reproduce the paper — derived pitch tracks only, no audio:

Dataset	Contents here	Source
Arabic Oud Taqasim (104)	recording list + per-piece pitch distributions (whole-piece and windowed histograms); no audio or pitch contours	commercial recordings
Cairo Congress 1932 (94 used)	pitch tracks (pickles) + metadata with annotated tonics	ORD-CC32 corpus (Bozkurt et al.)
Maqam-478 (421 used)	pYIN pitch caches (.npy)	Quranic recitation dataset (Shahriar et al.)
DiArMaqAr	tuning systems, maqamat, ajnas (JSON)	Music-Intelligence-Lab/DiArMaqAr

See data/README.md for details and licensing notes. The unified corpus after restricting to the 7 maqam families shared across datasets is 619 recordings (104 oud + 94 Cairo + 421 Maqam-478). Cairo Congress and Maqam-478 experiments reproduce fully from this repository alone; oud-dependent rows additionally require regenerating pYIN pitch tracks from the recordings listed in data/arabic_oud/metadata.csv (the provided oud histograms are the distribution-level inputs the features are built from). All scripts use identical stratified 5-fold splits (fixed seed).

Reproducing the paper

Paper result	Command	Runtime
Table 1 (tonic-noise sensitivity) + Table 2 "Raw xcorr argmax" row	python scripts/tonic_noise_and_xcorr.py	~15 min
Table 2 oracle rows (RF, 2D CNN)	python scripts/unified_arabic_analysis.py	~1 h
Table 2 "MaqamNet (8 tpls)" + Table 4 ablation rows	python scripts/template_conv_ablation.py	~1.5 h
Table 2 "MaqamNet (36 tpls)" (ch64 sweep)	python scripts/template_conv_deep_tune.py	~3 h
Table 2 unwrapped-pitch baseline (per-subset)	python scripts/unwrapped_pitch.py	~1 h
Table 4 seed means (±std)	python scripts/seed_variance_ablation.py	~2 h
Sec. 5.3 validation-based protocol + Maqam-478-only + separability stats	python scripts/protocol_validation.py	~1 h
Discussion: absolute-pitch standardization analysis	python scripts/m478_standardization_check.py	~2 min
Figure 2 (t-SNE of four representations)	python figure-code/fig_tsne_representations.py	~45 min
Figure 3 (Grad-CAM / saliency grid)	python figure-code/fig_gradcam_combined.py	~1 h

Outputs are written to results/ and figures/. Neural results vary by ±1 point (overall) across seeds; Table 4 quantifies this (seed_variance_ablation.py).

Model and training configuration

MaqamNet (best configuration, "MaqamNet (36 tpls)" in Table 2):

• Input: (N, 36, 20, 30) — 36 template channels (8 maqam + 28 jins DiArMaqAr templates, Ronzevalle 1904 tuning), 20 proportional temporal windows, 30 pitch-class bins (40 cents/bin). Each channel is the circular cross-correlation of the window's raw octave-wrapped pitch histogram with one template.
• Architecture (~716K parameters): circular-padded conv stem (5×5, 36→64) with BatchNorm/ReLU; 2 residual blocks (64→128) with circular padding on the pitch axis, BatchNorm, dropout, MaxPool2d(2); AdaptiveAvgPool2d((2,4)); FC 128 → 7 classes.
• Optimization: Adam, lr 1e-3, weight decay 1e-4, batch size 32, up to 250 epochs, ReduceLROnPlateau(patience=10, factor=0.5) on the training loss, early-stopping patience 40, dropout 0.4.
• Evaluation protocols: (a) headline tables: stratified 5-fold CV (composite maqam-dataset stratification, seed 42), best test-fold epoch; (b) validation-based (Sec. 5.3): identical folds, 15% of each training fold held out for validation, early stopping on validation only, test fold evaluated once at the best-validation epoch.

The unwrapped-pitch baseline CNN operates on (N, 40, 240) pitch-time matrices (0–4800 cents above 55 Hz, no octave wrapping, no tonic information), channels (32, 64, 128), dropout 0.3, same optimizer.

Citation

@inproceedings{bakarji2026maqamnet,
  author    = {Bakarji, Joseph and Allami, Khyam and El Asal, Mohammad and Abou Jaoude, Dany},
  title     = {Musicologically Interpretable Deep Learning Framework for {A}rabic Maqam Identification and Analysis},
  booktitle = {Proc. of the Int. Society for Music Information Retrieval Conf.},
  year      = {2026}
}

For the DiArMaqAr archive, please also cite:

@inproceedings{allami2026diarmaqar,
  author    = {Allami, Khyam and Elkhansa, Ibrahim and El Asal, Mohammad and Bakarji, Joseph},
  title     = {Culture-Specific Computational Musicology: A Case Study of {DiArMaqAr} (Digital {A}rabic Maqam Archive)},
  booktitle = {Proc. of the Int. Society for Music Information Retrieval Conf.},
  year      = {2026}
}

License

Code is released under the MIT License (see LICENSE). The derived pitch data and DiArMaqAr JSON files carry the terms described in data/README.md.