Contents
A flag is not an explanation. It means the recording resembled the synthetic examples the detector was trained on, scored past some threshold, and nothing more. It does not tell you which part of the audio moved the needle, and the research on these models keeps returning the same uncomfortable answer: the deciding feature is often not the voice at all, but the silence around it, the codec that compressed it, or the noise behind it.
What does a “flag” actually mean?
A synthetic-voice detector outputs a score for how closely a clip matches the traces it learned to associate with generated speech, and a threshold turns that score into a label; what that underlying number does and does not mean is covered in what does an AI voice-detector score mean. AASIST, one of the strongest research detectors, reaches about 0.83% equal error rate on the ASVspoof 2019 benchmark (Jung, Heo, Tak et al., ICASSP 2022), and its authors describe the giveaway artifacts as able to “reside in spectral or temporal domains.” The flag tells you the clip looked synthetic to a particular model at a particular threshold. It does not tell you the model was right, or what it was looking at.
Can the flag come from something other than the voice?
Yes, and this is the central finding of the explainability work. Negroni, Cuccovillo, Bestagini et al. (ICASSP 2026), studying what speech detectors actually attend to, report that many models “base their decisions on unvoiced timesteps, frequency bands where no speech occurs, or even background noise, rather than the actual speech.” Salvi and colleagues showed the same thing from the other direction: they split each recording into its speech and noise components, trained a separate detector on each, and the noise-only detector beat the full-signal and speech-only detectors on every dataset they tested, reaching 0.71 balanced accuracy on FakeOrReal against 0.53 for the full signal and 0.69 on the In-the-Wild set against 0.53 (Salvi, Balcha, Bestagini, Tubaro, ICASSPW 2024). The fake signal can live in the parts of the file that are not speech at all, which means a high score may be reading the recording conditions, the microphone, or the quiet at the start of a clip, not the speaker.
Does compression alone trip the flag?
It can. Negroni and colleagues measured how much a detector’s equal error rate rises under everyday codecs, from 21.8 percent under M4A to 29.2 percent under EnCodec and 40.9 percent under MP3. A codec reshapes the signal in ways that look, to a narrow model, like the fingerprint of a generator. So a clip that was only ever compressed, never synthesized, can drift toward a fake reading purely because the codec left traces the detector associates with its synthetic training set. This is also why you should test the original file rather than a re-encode, a point covered in how to tell if audio is AI-generated.
What would a trustworthy flag be reading?
The useful contrast is a detector that keys on the voice itself. Formant structure, the resonances of the vocal tract, is a property of real speech, and Negroni’s method is built around modeling it so the decision rests on speech rather than silence. Phoneme-level models push the same way: Zhang, Hua and Lan (AAAI 2025) reach 4.27% equal error rate by comparing feature discrepancies across phonemes, pooling the whole utterance rather than a convenient edge. And a speaker-specific forensic method that models a person’s own vowels reached 4.4% equal error rate against 27.5% for a standard MFCC baseline (Yang, Sun, Lyu, Rose, 2025). A flag grounded in voiced, speech-bearing regions is worth more than one that could be reading the quiet.
Why can two voice detectors disagree about the same clip?
Because they may be reading different evidence. One model can lean on vocoder artifacts, another on noise or codec traces, a third on self-supervised speech embeddings, and each is exposed to a different kind of clip. That specialization is measurable: a neural-vocoder detector scoring 0.13% equal error rate on the vocoder it trained on rises to between 4.6 and 45.35 percent on unseen ones (Sun, Jia, Hou, Lyu, CVPRW 2023), so a clip from a new generator can look clean to one model and synthetic to another. When Yi, Wang, Tao et al. (2023) moved detectors trained on ASVspoof 2021 onto in-the-wild audio, equal error rates jumped from 20 to 30 percent up to 35 to 55 percent, with RawNet2 rising from 24.32% to 36.74% and AASIST from 19.77% to 34.81%. Disagreement is not a glitch to average away; it is a sign the clip sits outside at least one detector’s comfort zone, the same way two image detectors split on one picture (see why do AI detectors give different results).
So how should you treat a flag?
As a prompt to look closer, not a finding. Ask what the detector could have keyed on: a long silent lead-in, heavy compression, obvious background noise. Test the original file, since compression alone can move the score. Run more than one tool and read agreement as weak corroboration rather than proof. And where you can, prefer a method that is explicitly reading voiced speech. A flag says “this resembled my fake examples.” Turning that into “this voice was synthesized” is your job, not the detector’s. Doing that for one clip is the practical question in is this voice AI-generated.
Sources
- Negroni, Cuccovillo, Bestagini et al. (2026). Multi-Task Transformer for Explainable Speech Deepfake Detection via Formant Modeling. ICASSP 2026.
- Salvi, Balcha, Bestagini, Tubaro (2024). Listening Between the Lines: Synthetic Speech Detection Disregarding Verbal Content. ICASSPW 2024.
- Jung, Heo, Tak et al. (2022). AASIST: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks. ICASSP 2022.
- Zhang, Hua, Lan (2025). Phoneme-Level Feature Discrepancies: A Key to Detecting Sophisticated Speech Deepfakes. AAAI 2025.
- Yang, Sun, Lyu, Rose (2025). Forensic Deepfake Audio Detection Using Segmental Speech Features.
- Sun, Jia, Hou, Lyu (2023). AI-Synthesized Voice Detection Using Neural Vocoder Artifacts. CVPRW 2023.
- Yi, Wang, Tao et al. (2023). Audio Deepfake Detection: A Survey.