Reproduce “Improving Choral Music Separation through Expressive Synthesized Data from Sampled Instruments” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ke Chen, Hao-Wen Dong, Yi Luo, Julian McAuley, Taylor Berg-Kirkpatrick, Miller Puckette, Shlomo Dubnov
Researchers from UC San Diego and Tencent AI Lab developed an automated pipeline to generate 8.2 hours of expressive synthetic choral music data from sampled instruments. Training models on this synthesized data improved their generalization to real-world choral recordings, leading to up to 2.31 dB higher average median SDR compared to models trained without pretraining.
Key insights
- Frequency-domain models generally achieve better performance than time-domain models for choral music separation, demonstrating higher SDRs on both piano and vocal synthetic datasets.
- Expressive synthesized data, incorporating elements like legato and dynamic velocity, significantly improves the generalization capability of separation models to real-world choral music, outperforming simpler synthetic data and models trained without pretraining.
- The proposed synthesis methodology is extensible and could address data scarcity in other homologous source separation tasks, such as string quartet separation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.