Reproduce “Blind Estimation of Audio Processing Graph” 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
Sungho Lee, Jaehyun Park, Seungryeol Paik, Kyogu Lee
This paper presents a deep learning framework for blindly estimating an entire audio processing graph and its parameters from a processed audio signal. Researchers at Seoul National University developed a two-stage system, using a Transformer-based architecture, to reconstruct human-interpretable conventional digital signal processing graphs.
Key insights
- Blindly estimating a heterogeneous audio processing graph is achievable with a two-stage, token-by-token Transformer-based approach, ensuring structurally valid (acyclic, connected) graph reconstructions.
- Separating the estimation of graph structure (prototype) from continuous parameters provides better performance compared to a single-stage method that attempts both simultaneously.
- Providing the dry source audio alongside the processed reference significantly improves the model's ability to generalize and accurately estimate the processing graph, especially for unseen input sources.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.