VocalParse: AI Model for Unified Singing Voice Transcription
A new model called VocalParse has been developed by researchers to unify singing voice transcription (SVT) using a Large Audio Language Model (LALM). This innovative model tackles issues in automatic transcription, including the dependence on intricate multi-stage processes, challenges in aligning text with notes, and inadequate performance with unfamiliar singing data. VocalParse employs an interleaved prompting strategy that simultaneously captures lyrics, melody, and the relationship between words and notes, producing a sequence that correlates directly with a structured musical score. This method seeks to facilitate scalable, high-quality annotations for Singing Voice Synthesis (SVS) systems, minimizing the necessity for manual labeling. The research is available on arXiv with ID 2605.04613.
Key facts
- VocalParse is a unified singing voice transcription model.
- It is built upon a Large Audio Language Model (LALM).
- The model uses an interleaved prompting formulation.
- It jointly models lyrics, melody, and word-note correspondence.
- The generated sequence maps directly to a structured musical score.
- It addresses challenges in current automatic transcription systems.
- The paper is available on arXiv with ID 2605.04613.
- The model aims to enable scalable annotation for SVS systems.
Entities
Institutions
- arXiv