Open Corpus for Arabic Maqam Recognition (OCMR)

The Open Corpus for Arabic Maqam Recognition (OCMR) is a collection of melodic pitch features and expert maqam annotations for 1,525 recordings of Arabic maqam music spanning 33 maqamat. It is released as a companion to the ISMIR 2026 paper "An Open Corpus for Arabic Maqam Recognition." No audio is included. All features are non-invertible melodic representations derived from the recordings; the corpus is intended for computational research on maqam recognition and analysis. Primary features (one file per recording, organized by maqam): pcd_tomato/ 160-bin octave-folded pitch-class distributions (PCDs) from the tomato toolbox (predominant melody + Bozkurt pitch filter), Gaussian-smoothed and L1-normalized. These are the exact features used in the paper's experiments; peak-aligning them reproduces the reported 66.20% headline accuracy. Present for all 1,525 recordings. Supplementary features (no reported result depends on them): pitch_tomato/ tomato predominant-melody f0 tracks (one Hz value per line; 0.0 = unvoiced). Present for all 1,525 recordings; each tomato PCD is computed directly from its own stream. pitch_crepe/ CREPE (full model) f0 estimates with per-frame confidence, unfiltered. Present for all 1,525 recordings. CREPE is a monophonic pitch tracker, included for extractor comparison; it is not well suited to much of this predominantly polyphonic corpus. pcd_crepe/ 160-bin PCDs recomputed from the CREPE streams. Present for 1,460 recordings; 65 low-confidence recordings (dense pop, percussion, spoken/choral material) yield a stream but too few confidently-voiced frames for a PCD.  tdms_tomato/, tdms_crepe/ 48x48 Time-Delayed Melody Surfaces (Gulati et al. 2016; makam-tuned parameters from Ganguli et al. 2022), computed from the respective pitch streams. Metadata and documentation: metadata.csv per-recording metadata: stable id, maqam, family, artist/album/track, duration, region, secondary maqamat, modulation notes, per-feature availability flags, and archival provenance for the East African 78 rpm subset annotations.json expert maqam labels (single primary annotator, co-author Layth Sidiq). algorithms.json exact algorithms and parameters for every feature. README.md overview of the corpus, file layout, and instructions for reproducing the paper's aligned-PCD results QA_REPORT.md build report: recording and class counts, per-feature coverage, and provenance matches for the East African subset SHA256SUMS SHA-256 checksums for every feature and metadata file, for verifying download integrity Feature coverage: all 1,525 recordings have the tomato PCD, tomato predominant-melody stream, tomato TDMS, and CREPE stream; 1,460 also have a CREPE PCD. Every tomato feature derives from the same per-recording pitch stream, so the PCDs and TDMS are reproducible from the released streams. Per-recording feature availability is flagged in metadata.csv. The corpus includes a subset of digitized early East African / Hadrami 78 rpm shellac recordings (His Master's Voice, Gallotone, Mzuri labels, ca. 1938-1965), contributed by co-author Andrew J. Eisenberg with archival provenance metadata. Please cite the ISMIR 2026 paper when using this dataset. Code to reproduce the experiments: github.com