Deepfake speech datasets lack fairness, overlap in sources

A new audit of 39 deepfake speech datasets reveals significant limitations in their fairness and technical robustness. Researchers found that most datasets lack crucial demographic metadata, making fairness assessments nearly impossible and preventing subgroup analysis. Additionally, a substantial overlap in the source corpora used for bona fide speech across these datasets could lead to overstated generalization claims and undermine cross-dataset evaluations. AI

IMPACT Highlights critical data limitations that could hinder the development and evaluation of fair and robust deepfake speech detection systems.

RANK_REASON The cluster contains an academic paper detailing a dataset audit. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Deepfake speech datasets

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →