Generation Provenance Before Behavior Attribution: Auditing Synthetic Speech Research Objects
A provenance contract audits synthetic Japanese care-handoff speech before attributing behavior to training data.
The paper argues that attributing model behavior to synthetic training data first requires knowing what produced each training item. It proposes a generation-provenance substrate binding source specification, content, waveform, target, fact requirements, quality signals, review lineage, and an immutable manifest identity. An audit of a private Japanese care-handoff pipeline covers 113 assets and 1.552 hours of synthetic speech across six scenario families. Exact upstream attribution remains blocked by floating generator aliases, missing per-clip TTS and code stamps, and an unversioned checking prompt; the authors claim neither causal attribution nor clinical validity.
- Audit covers 113 assets and 1.552 hours of synthetic speech.
- Two faithful-only manifests are seed-disjoint and immutably versioned.
- Missing TTS stamps and generator aliases block exact attribution.
- Provenance is necessary but not sufficient for causal attribution.
From the original article
1 min read at arxiv.orgAttributing model behavior to synthetic training data requires knowing what produced each training item before estimating what that item caused. A waveform-label pair does not preserve this knowledge.
By Sidi Chang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.