Audio-to-Score Transcription using Pre-trained Features, Data Augmentation, and the New SheetSage-A2S Dataset
This episode discusses the 'Audio-to-Score Transcription' paper, which introduces the SheetSage-A2S dataset—61 hours of audio paired with kern score encodings. The method uses data augmentation and MuQ features to improve robustness against real-world variations. Results show a 4.98% Symbol Error Rate (SER) on classical music and a 20.92% SER on popular music, providing an accessible tool for AI musical comprehension.