Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 7·5 min read
Audio-to-Score Transcription using Pre-trained Features, Data Augmentation, and the New SheetSage-A2S Dataset
TL;DR
A new dataset, SheetSage-A2S, enables audio-to-score transcription for popular music, achieving a 20.92% symbol error rate. This model significantly outperforms existing classical music systems, demonstrating improved generalization through data augmentation and pre-trained features.
✦ Why It Matters
Explore the SheetSage-A2S dataset to enhance your audio-to-score transcription projects today.
Key Takeaways
- Utilize the SheetSage-A2S dataset for popular music A2S research.
- Implement data augmentation to improve model generalization.
- Leverage MuQ for effective feature extraction in music audio.
- Aim for lower symbol error rates in your A2S applications.
Full Summary
Existing audio-to-score (A2S) systems have largely focused on classical music, leaving a gap in popular music applications. To address this, a new dataset called SheetSage-A2S was introduced, containing 61 hours of audio and score encodings for 9,468 clips from 6,066 unique songs.
The study enhances A2S performance by employing data augmentation techniques and utilizing MuQ, a pre-trained feature-extraction model for music audio. The results indicate that the proposed A2S model achieves a symbol error rate (SER) of 4.98% on classical music, a significant improvement over the previous state-of-the-art rate of 15.3%.
For popular music, the model achieves a SER of 20.92%, establishing a strong benchmark for future research. The dataset, model, and code are publicly available for further exploration.