Abstract
Commonly music has an obvious hierarchical structure, especially for the singing parts which usually act as the main melody in pop songs. However, most current singing annotation datasets only record symbolic information of music notes, ignoring the structure of music. In this paper, we propose a hierarchical singing annotation dataset (HSD) consisting of 68 pop songs from YouTube. This dataset records the onset/offset time, pitch, duration, and lyric of each musical note in an enhanced LyRiCs format to present the hierarchical structure of music. We validate the labeling accuracy by comparing it with an automatic singing transcription (AST) dataset, indicating that HSD reaches high labeling accuracy while providing richer structural information.
Introduction: The Need for Hierarchy in Singing Voice Analysis
In pop music, the vocal melody is not merely a sequence of independent notes; it possesses a clear, nested structure where songs are composed of phrases, and phrases are composed of notes. Most existing datasets for singing voice analysis are "flat," focusing on individual note pitches and timings or lyric-to-audio alignment without capturing the overarching musical organization. This lack of structural context limits the development of advanced music information retrieval (MIR) tasks that require an understanding of musical phrasing and "breathing."
Methodology: Efficient Two-Stage Labeling
Creating high-quality hierarchical datasets is labor-intensive. To address this, we developed a two-stage "N&L" (Notation and Lyrics) labeling process to ensure both efficiency and accuracy:
- Initialization: We leverage existing musical notation and standard LRC (lyrics) files. By using a mathematical distribution formula, we generate initial note placements within known phrase timestamps, providing a high-quality starting point.
- Calibration: Human annotators perform a manual check. A key innovation here is efficiency via propagation: by primarily adjusting phrase boundaries, the notes within those boundaries are automatically scaled to fit, significantly reducing the need for manual note-by-note editing.
The resulting data is stored in an enhanced LRC format, extending the common lyrics format to include pitch and duration metadata in a human-readable and easily editable way.
Dataset Comparison and Features
The HSD dataset contains 68 pop songs (62 Chinese, 6 Japanese) sourced from YouTube. Compared to other industry-standard datasets like MIR-ST500, Jamendo, and RWC, HSD is unique in its comprehensive coverage of note-level lyrics and hierarchical structure.
Key Results and Validation
We validated the accuracy of HSD using the COnP (Correct Onset and Pitch) metric against the MIR-ST500 benchmark. At a 50ms tolerance, HSD achieved an accuracy of 97.07% (outperforming MIR-ST500's 96.49%), and at a 100ms tolerance, it reached 98.53%. These results demonstrate that our semi-automated labeling method produces data of equal or superior quality to specialized transcription datasets while offering significantly more structural depth.
Conclusion and Future Work
HSD provides a new standard for singing voice annotation by preserving the natural hierarchy of music. While currently focused on a specific set of pop songs, it serves as a robust foundation for future research in Automatic Lyric Transcription (ALT) and symbolic music generation. The dataset and labeling tools are open-sourced to encourage further community development.
Summary updated