Toward Part-Aware Choral Transcription with singing voice assignment

One-line takeaway

PawCT enables end-to-end part-aware choral transcription with improved accuracy over existing methods.

提出了一种端到端的合唱音频转谱方法PawCT，能识别SATB各声部并生成独立乐谱，显著提升转谱准确率。

Key points

PawCT introduces the first end-to-end neural framework for part-aware choral transcription, jointly modeling vocal part assignment and note-level transcription. It leverages pitch range priors and ordered continuity to improve accuracy and structural consistency. Experiments on YouChorale show significant gains over baselines, with robust performance across datasets.

• PawCT enables end-to-end part-aware choral transcription with separate SATB tracks.
• It uses range prior and ordered continuity to model pitch ranges and melodic structure.
• Achieves 0.225 macro F1 on YouChorale, outperforming baselines by 36.4%.
• Shows strong cross-dataset generalization on CSD and Cantoria.

Written only from the official arXiv abstract by an open-source model, then checked automatically (every number and name must appear in the abstract). The paper is authoritative.