Establish fine-grained retrieval-support independence

Establish whether the retrieval and profile support used by the HuBERT-large nearest-neighbor evidence channels is independent of prompt, channel, and recording source beyond the available corpus-level source tags.

Background

The speech deepfake decision record incorporates retrieval and speaker-profile evidence from a support set whose synthesis families are held out during evaluation. The identifier-overlap audit reports no reuse of utterance IDs, speaker IDs, or audio paths, but the available source metadata identifies only broad corpus-level tags. Consequently, the evaluation does not establish independence with respect to finer-grained properties such as recording prompt, acoustic channel, or recording source, leaving the degree of support-set independence unresolved.

References

The available source tag is shared only at corpus granularity, so prompt, channel, and recording-source independence remain unresolved.

— From Scores to Evidence: Auditable Decisions Can Improve Speech Deepfake Detection  (2609.08899 - Geng et al., 8 Sep 2026) in Section 4, Experimental Protocol, subsection “Models and scores”