Methods & Provenance
Sources, licences and how to verify a release
A result is only as good as the record behind it. This page lists every upstream source used in the PACC datasets, and describes how to check that what you downloaded is what we published.
Upstream sources
| Source | Used in | Licence | Citation |
|---|---|---|---|
| CSTR VCTK Corpus 0.92 | PACC-T, PACC-P | CC BY 4.0 | J. Yamagishi, C. Veaux, K. MacDonald. University of Edinburgh, CSTR, 2019. doi:10.7488/ds/2645 |
| AMI Meeting Corpus | PACC-T, PACC-P | CC BY 4.0 | University of Edinburgh et al. J. Carletta (2006), Announcing the AMI Meeting Corpus, ELRA Newsletter 11(1). groups.inf.ed.ac.uk/ami/corpus |
| MUSAN | PACC-P | CC BY 4.0 (only CC BY and public-domain files used) | D. Snyder, G. Chen, D. Povey (2015). arXiv:1510.08484 |
| FSD50K | PACC-P | CC BY (only CC0 and CC BY clips used) | E. Fonseca et al. (2022), IEEE/ACM TASLP 30. doi:10.5281/zenodo.4060432 |
Third-party noise, music and event clips mixed into PACC-P keep their own licences. Per-clip credits are published with the data in attribution.csv and should be credited alongside PACC-P.
What was left out, and why
Earlier internal builds also held RAVDESS and CREMA-D clips. They were removed so that PACC is CC BY only. Eight AMI clips with no usable speech are excluded from all of PACC and listed in excluded_sources.csv.
Verify a download
Each release publishes a checksum file, an index of its tarballs, and a manifest listing every file with its SHA-256. To check a set of tarballs after downloading:
sha256sum -c SHA256SUMSEvery condition also ships a per-clip record, params.csv, describing how each clip was made: for PACC-T the input and output hashes and the full command sequence, and for PACC-P the seeds, levels, gains and achieved SNR. Tarballs are deterministic, so the same inputs rebuild byte-identical files.
Reproducibility
Software versions are pinned and recorded in each datacard (PACC-T tools and versions, PACC-P how the material was made). Randomness is seeded per clip in PACC-P. For PACC-T, commands recovered after encoding were re-run on 60 clips per condition and reproduced the stored audio bit for bit in every case. Each datacard lists its known edge cases and limitations, and points to EDGE_CASES.txt in the release.
Datacard snapshots on this site
The datacards on this site are verbatim copies. Compare the SHA-256 below against the file you download to confirm you are reading the same text.