Voices of the Mountains: Kurdish Bayati-Kurd Maqam Vocal Error Dataset

The first annotated corpus for time-localized vocal error detection in Kurdish maqam singing. It contains 50 songs in the Bayati-Kurd maqam performed by 13 vocalists (WAV, mono, 22,050 Hz; 1.81 hours in total), with 221 expert-annotated error spans in three classes: fine pitch error (150), rhythm error (46), and modal drift (25). Annotations are provided as JSON files with start and end times in seconds, error type, and a short free-text note, plus per-song metadata (anonymized singer code, tonic frequency, maqam). Singers are identified only by anonymized codes (S01-S13); all recordings were collected with informed consent for public release. The archive includes README.md and docs/ANNOTATION_PROTOCOL.md describing the data format and the three-level annotation validation process. The dataset accompanies the paper "Voices of the Mountains: Deep Learning-Based Vocal Error Detection System for Kurdish Maqams" (Khairaldeen and Hassani, University of Kurdistan Hewler). Training and evaluation code is available at github.com. Mirrors: data.krd and KurdishBLARK.