Original title: Audio Deepfake Detection Using Temporal Coherence Analysis
Justin D. Norman, Sarah Barrington · September 2026
Fake audio made by artificial intelligence threatens information integrity, causing voice fraud and music copyright disputes. Researchers built a system to spot these audio deepfakes across speech, instrumental music, and music with vocals. Their method cuts an audio clip into segments and converts them into mathematical representations using a model called CLAP. Next, it compares how similar these chunks are over time to measure temporal coherence, which means how naturally sounds connect. By feeding statistics from these similarity scores into small group models, the system reliably flags synthetic audio. It works efficiently and explains its decisions well. Surprisingly, twenty-one of twenty-nine statistical clues reversed direction between training and real-world deployment, and entropy worked in opposite ways for speech versus music.
Imagine a security guard who learns to spot counterfeit bills by weighing them—real bills weigh a certain amount. But when the guard starts working in a different city, counterfeiters there use paper that makes the bills weigh exactly the same as real ones. What worked during training now points the wrong way. In this paper, the statistical measurements that helped detect fake audio during development often reverse their meaning when the detector encounters audio from the real world, just like the weight test flips from useful to misleading.
Fake audio can impersonate your voice for fraud or copy songs without permission. Current detectors often fail when moved from labs to real life, and this work shows exactly why: the statistical patterns flip. Understanding these flips helps build more trustworthy detection systems that protect both people and artists.
The abstract does not compare this method to the best deep learning detectors on standard benchmarks, so it is unclear how much accuracy is sacrificed for interpretability. The findings on feature‑label inversion and direction reversal are empirical observations without a causal explanation, and may not generalize to other audio types (e.g., environmental sounds) or other feature sets.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Audio Deepfake Detection Using Temporal Coherence Analysis — Justin D. Norman, Sarah Barrington
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.