Why this research matters
Historical music recordings are both cultural documents and listening experiences. Many carriers are damaged, noisy, band-limited, or affected by clicks, hum, clipping, and distortion. Restoration can make these recordings easier to study, reissue, and enjoy, but it can also remove evidence or impose a modern sound that the original never had. We study restoration as a balance between recovering what is missing and remaining faithful to the source.
Research directions
- Noise and defect removal: Separate persistent noise, impulsive defects, and musical content so that restoration reduces interference without erasing performance details.
- Bandwidth extension and dereverberation: Recover plausible high-frequency content and reduce room effects while keeping the result consistent with the original recording.
- Source separation for historical recordings: Explore ways to isolate instruments or voices in dense, low-quality mixes for study, remixing, and archival access.
- Score- and metadata-informed restoration: Use available scores, session information, and musicological knowledge as constraints rather than relying only on generic audio models.
- Generative and self-supervised models: Investigate models that can fill gaps or reconstruct missing content while exposing uncertainty and preserving musical structure.
- Evaluation and listening: Combine objective measures with listening tests and musicological criteria, and compare restoration alternatives rather than hiding the choices inside one automatic output.
How we approach the work
We connect signal processing and machine learning with archive and musicology expertise. Reproducibility, uncertainty, and the distinction between “cleaner” and “more faithful” are central to our work. We prefer methods that make their assumptions visible and let a restorer inspect and compare different results.
What we aim to enable
Tools that help archives, researchers, musicians, and listeners work with historical recordings: not a single automatic “fix”, but controllable restoration that supports access, scholarship, and careful reuse.