Historical Orchestral Recordings: An Unpaired Test Set for Music Restoration

If you use this dataset, please cite both the Zenodo dataset record and the accompanying paper.

Recommended citation: Cho, Steven (2026). Historical Orchestral Recordings: An Unpaired Test Set for Music Restoration (Version 1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22737610

This dataset contains the 149 full-length historical recordings forming the unpaired test set used to evaluate restoration systems in Full-Mix Historical Music Restoration in Latent Space. It comprises 70 Orchestra recordings and 79 Light Orchestra recordings sourced from the Internet Archive 78 rpm collection.

The Orchestra subset contains recordings performed by classical symphonic orchestras. Any retained recording not performed by a classical symphonic orchestra was assigned to the Light Orchestra subset. "Light Orchestra" identifies the performing ensemble and does not indicate a non-classical genre. Both subsets consist exclusively of instrumental classical music performed using instruments conventionally associated with classical and orchestral performance.

The historical recordings contain naturally occurring and differing amounts and types of degradation, including bandwidth limitation, surface noise, clicks, crackle, distortion, and other recording or preservation artifacts. Degradation severity was not used to assign categories and was not standardized across recordings. Musical genre, instrumental-only content, instrumentation, source metadata, and Orchestra/Light Orchestra assignments were manually verified for every recording.

This is an unpaired evaluation dataset: corresponding clean recordings of the same historical performances are unavailable.

All audio preprocessing steps are listed in the metadata.csv and README.md

For any additional questions, please contact steven{dot]cho{[at]}live[{dot}]com

Associated paper: arXiv URL forthcoming

Official code repository: https://github.com/SonyResearch/historical_music_restoration