Contents
You usually cannot tell by ear anymore, and a detector will hand you a probability rather than a verdict. The reliable way to check whether a clip is AI-generated is a short procedure: test the original file, read what the detector actually measured, and corroborate the result with provenance and the source of the recording. A high score is a reason to look closer, not a finding on its own.
What a detector is listening for
A synthetic-voice detector does not recognize a fake the way you might catch an odd cadence. It measures faint statistical traces that a vocoder, the component that turns a model’s output into a waveform, leaves in the signal. As the authors of the AASIST detector (Jung, Heo, Tak et al., ICASSP 2022) note, the artifacts that separate spoofed from genuine speech “can reside in spectral or temporal domains,” and their model reaches an equal error rate, the point where false accepts and false rejects balance, of about 0.83% on the ASVspoof 2019 benchmark. That is strong, on that benchmark. What “on that benchmark” hides is the rest of the story.
Test the original recording, not a re-encode
Audio traces are fragile. Re-encoding a clip to MP3, forwarding it through a messaging app, or screen-recording a video re-compresses the waveform and strips much of the high-frequency evidence a detector relies on. Whenever you can, run the original file, the untouched upload or voice note, not a re-shared copy. A clean “not AI” result on a heavily compressed clip carries little information, the same way a re-saved screenshot defeats an image detector, a problem covered in how to check an image for AI.
Read the score for what it is
A detector score says how strongly a clip carries the traces the tool was trained on, not whether a machine made it. A high score means the recording resembles the synthetic examples in the tool’s training set, and a low score means it does not. Neither is proof. The number is only as meaningful as the match between your clip and the data the detector learned from, which is why the same tool can be confident and correct on one generator and confident and wrong on the next.
Where audio detectors fall apart
The gap between benchmark and reality is the whole problem. When Müller et al. (2022) tested detectors on their In-the-Wild benchmark, real audio collected outside the lab, the field learned how far these tools slip. The survey by Yi, Wang, Tao, Zhang, Zhang and Zhao (2023) puts numbers on it: trained on ASVspoof 2021 and tested in the wild, detector equal error rates jump from 20 to 30 percent up to 35 to 55 percent, and end-to-end models fall the same way, with RawNet2 rising from 24.32% to 36.74% and AASIST from 19.77% to 34.81%. The survey names the cause directly as the “poor generalization of existing detection methods to unknown fake attacks.” The collapse is sharpest across model families: a detector that scores 0.13% equal error rate on the vocoder it trained on rises to between 4.6 and 45.35 percent on unseen vocoders (Sun et al., CVPRW 2023). A detector that has never heard your generator is close to guessing, so treat a confident verdict on an unfamiliar clip with suspicion.
Check for provenance and known tools
Provenance beats detection whenever it exists. A Content Credential or an embedded generator watermark, when present, is far stronger evidence than any classifier score, though most clips in the wild carry no mark at all. Some vendors ship a classifier for their own output: the ElevenLabs speech classifier is built to recognize audio from ElevenLabs models, but on its own terms it inspects only the “first 1 minute” of a clip and “does not reliably classify ElevenV3” audio from its newest voices. That is the shape of a first-party tool. It is informative for the model that made it and silent about everything else.
When a reference voice is available
There is one setting where confidence rises sharply: when you hold clean audio of the real person to compare against. Yang, Sun, Lyu and Rose (2025) built a forensic method that models a speaker’s own vowels and reached a 4.4% equal error rate detecting ElevenLabs v2 speech, against 27.5% for a standard MFCC baseline. The gain is real but conditional. Their approach is speaker-specific and needs reference audio of the genuine voice, so it is a forensic comparison for a suspected clone of someone you have recordings of, not a screening tool for a stranger’s clip.
What actually settles it
So how do you tell? You combine signals rather than trust one. Test the original file, run more than one detector and read agreement as confidence rather than proof, look for provenance, and weigh the source and context of the clip. When the file is degraded, the generator is unfamiliar, and no provenance exists, the correct output is “undetermined,” not a percentage dressed up as certainty. A detector earns its place as the prompt that makes you look harder, and the answer comes from everything you gather after that. For the narrower question of judging one clip, see is this voice AI-generated; for how far these tools can be trusted overall, see are AI detectors accurate.
Sources
- Jung, Heo, Tak, Shim, Chung, Lee, Yu, Evans (2022). AASIST: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks. ICASSP 2022.
- Müller, Czempin, Dieckmann, Froghyar, Böttinger (2022). Does Audio Deepfake Detection Generalize? Interspeech 2022.
- Yi, Wang, Tao, Zhang, Zhang, Zhao (2023). Audio Deepfake Detection: A Survey.
- Sun et al. (2023). AI-Synthesized Voice Detection Using Neural Vocoder Artifacts. CVPRW 2023.
- Yang, Sun, Lyu, Rose (2025). Forensic Deepfake Audio Detection Using Segmental Speech Features.
- ElevenLabs. AI Speech Classifier. Vendor detection tool for ElevenLabs-generated audio.