GenTraceBench: A Benchmark for Tracing Audio Deepfakes Across Pre- and Post-training Stages

ACL ARR 2026 January Submission5319 Authors

Keywords: audio deepfake, text-to-speech, detection, attribution, benchmark

Abstract:

Modern text-to-speech (TTS) models increasingly rely on foundation pretraining followed by post-training adaptation, creating new challenges for audio deepfake detection and attribution in the wild.
Prior benchmarks mainly test against fixed generators and thus under-estimate the impact of adaptation-induced shifts.
We present GenTrace, a benchmark that tracks TTS evolution from foundation pretraining to diverse adaptation strategies, with controlled prompts and speakers to isolate model-induced differences (16 variants, 49,728 synthesized utterances).
Using GenTrace, we find that alignment-based adaptation typically preserves detection accuracy, while architecture and pretraining data have a substantially larger effect on attribution performance.
GenTrace supports reproducible evaluation of detection and attribution robustness under realistic model adaptation scenarios. GenTrace will be publicly released upon acceptance.

Paper Type: Short

Research Area: Speech Processing and Spoken Language Understanding

Research Area Keywords: Speech Recognition, Text-to-Speech and Spoken Language Understanding

Contribution Types: Data resources, Data analysis

Languages Studied: English, Chinese

Submission Number: 5319