Introduction
The Ethiopian Orthodox Tewahedo Church (EOTC) preserves one of the world's oldest liturgical music traditions, known as Yaredawi Zema, dating back to the 6th century. Traditionally an oral tradition requiring decades of study, this music remains digitally underrepresented. The Zema Dataset is the first comprehensive computational resource designed to bridge this gap, providing over 10 hours of annotated audio to facilitate Music Information Retrieval (MIR) tasks.
Central to Yaredawi Zema is a sophisticated notation system comprising "neumes" (symbols) and "sireyu" (melodic roots). These notations, as shown in Figure 1, dictate the melodic realization of the text. However, the performance is highly contextual, varying by religious season and "reading tone," presenting significant challenges for automated transcription and analysis.
Methodology
The creation of the Zema Dataset followed a multi-stage pipeline to ensure high-quality, machine-learning-ready data. Audio was sourced from a single expert scholar to maintain stylistic consistency and segmented into manageable clips. Lyrics were extracted via OCR from liturgical texts and manually corrected.
A critical phase involved multi-layer annotation. Beyond temporal word alignment, experts annotated "reading tones" and "chanting options"—instances where a single phrase allows for multiple valid melodic interpretations. Figure 2 illustrates the complexity of these annotations, highlighting how specific textual elements are linked to musical features and notation layers.
Dataset Characteristics
The resulting Zema Dataset focuses on Se’atat Zema (Horologium Chants) and includes:
- Total Duration: 10.21 hours of audio across 369 audio-text pairs.
- Annotations: Word-level boundaries, chanting modes (Ge’ez, Ezil, Araray), and reading tones.
- Visual Cues: The dataset accounts for functional color-coding in manuscripts, where red notations signify the Ezil mode, while black notations are used for Araray and Ge’ez.
Experimental Results
To validate the dataset's utility, the authors performed a Chanting Mode Classification task using a Support Vector Machine (SVM) with RBF kernels. The model achieved an overall accuracy of 69.11%. Notably, the Ezil mode was classified with 85% accuracy, likely due to its distinct, gentle, and melodic character. Conversely, the Ge’ez mode proved more challenging (43% accuracy) due to its simpler style and lower pitch range, which often overlaps with other modes in the feature space.
Conclusion and Future Work
The Zema Dataset provides a foundational resource for the computational study of Ethiopian liturgical music. While currently limited to a single source and specific chant book, it opens doors for advanced tasks like automated Ge'ez speech recognition and complex lyric-to-audio alignment. Future efforts will focus on expanding the dataset to include multiple performers and broader liturgical collections like the Digua to account for greater stylistic diversity.
Summary updated