Ethical and Technical Limits of Deepfake Speech Datasets

2026-06-09

AI Summary

This study addresses critical limitations in the evaluation of fairness and generalization for deepfake speech detection, which are constrained by the quality and representativeness of existing datasets. For the first time, it conducts a systematic audit of 39 deepfake speech datasets, employing metadata analysis and source tracing to comprehensively assess key attributes including accessibility, documentation completeness, demographic and linguistic coverage, scale, and provenance of genuine speech sources. The findings reveal that the vast majority of datasets lack essential demographic metadata, and their underlying bona fide speech corpora exhibit substantial overlap. Consequently, meaningful fairness evaluation is rendered infeasible, and cross-dataset generalization performance is significantly overestimated, exposing structural limitations in current research practices within the field.

Abstract

Claims about the robustness and fairness of deepfake speech detectors are only as credible as the datasets used to train and evaluate those systems. We present a dataset-level audit of the deepfake speech landscape. We compile and analyze 39 deepfake speech datasets, examining key attributes including accessibility, documentation, demographic and language coverage, dataset scale, and the underlying bona fide speech sources. Our audit reveals two important takeaways. Firstly, fairness assessment is largely infeasible because most datasets lack demographic metadata, and only a few contain gender or language labels. This prevents any meaningful subgroup analysis and leaves other demographic attributes unaddressed. Secondly, we identify substantial overlap in underlying bona fide source corpora across datasets, which can undermine cross-dataset evaluation and lead to overstated generalization claims.

Problem

• deepfake speech
• dataset audit
• fairness
• demographic metadata
• generalization

Innovation

• deepfake speech
• dataset audit
• fairness evaluation
• demographic metadata
• source corpus overlap