What does an AI voice-detector score mean?

An AI voice-detector score is a probability at an operating point you usually cannot see, not the chance the voice is fake. The tool outputs how strongly a clip matches its synthetic training examples, and research reports that behavior as equal error rate, the threshold where false accepts and false rejects balance. The number is calibrated on a benchmark, not on your clip, which is why the same score can mean very different things from one recording to the next.

What is the detector actually outputting?

A single number that says how strongly your clip resembles the synthetic examples the model was trained on, mapped through a threshold into a label. It is not a measurement of reality. AASIST reaches about 0.83% equal error rate on ASVspoof 2019 (Jung, Heo, Tak et al., ICASSP 2022) and RawNet2 reaches 4.66% (Tak, Patino, Todisco et al., ICASSP 2021), figures that describe how each model behaves across a whole benchmark, not how sure you should be about one voice note.

What is equal error rate, and what is an operating point?

Equal error rate is the threshold where the false-accept rate and the false-reject rate are equal, a single figure for comparing detectors without picking a bias toward one kind of error. A deployed tool has to pick an actual operating point, and at any point other than the equal-error one the two error rates diverge: a threshold set to catch more fakes will also falsely flag more real voices. That tradeoff matters because the costs are asymmetric. In a workplace, a newsroom, or a family-scam case, a false accusation harms a real person; in a fraud case, a missed clone costs money or access. A score without its threshold, its false-positive rate, and its test condition is incomplete evidence, and the vendor rarely shows you which point they chose.

Does “94 percent AI” mean a 94 percent chance the voice is fake?

No. A model’s output is a similarity-to-training score passed through a function, not a calibrated probability that the voice is synthetic. Reading “94” as “94 percent likely fake” assumes the tool is calibrated on audio like yours, and it usually is not. The gap shows up the moment the clip leaves the benchmark distribution: a neural-vocoder detector scoring 0.13% equal error rate on its own vocoder degrades to between 4.6 and 45.35 percent on unseen vocoders (Sun, Jia, Hou, Lyu, CVPRW 2023). The same displayed confidence can sit on either side of that range.

Why does the same score mean different things on different clips?

Because the error rate behind the score depends on where the audio came from. On a benchmark, AASIST and RawNet2 read in the low single digits; in the wild, the survey by Yi, Wang, Tao et al. (2023) reports detector equal error rates climbing into the 35 to 55 percent range, with RawNet2 rising from 24.32% to 36.74% and AASIST from 19.77% to 34.81%. A “90 percent” on a clean studio clip from a known generator and a “90 percent” on a compressed phone clip from an unknown one are not the same claim, even though the interface prints one number.

What can make a high score misleading?

A high score can be reading the wrong thing entirely. Negroni, Cuccovillo, Bestagini et al. (ICASSP 2026) found that many detectors “base their decisions on unvoiced timesteps, frequency bands where no speech occurs, or even background noise, rather than the actual speech.” The score can be a statement about the silence at the start of a clip or the compression it carries, dressed up as a statement about the voice, which is exactly the failure examined in why did the voice detector flag this. It can also be a genuine vocoder-fingerprint match, which is useful only when the detector’s training covered your generator, and uncertain when it did not.

What can make a low score misleading?

A low score can mean the clip is real, or it can mean the detector never learned this generator, the evidence was stripped by compression, or the audio was too short. Codec effects matter most for voice notes and social clips: Negroni and colleagues report equal error rates of 21.8 percent for M4A, 29.2 percent for EnCodec, and 40.9 percent for MP3, so a low score on a heavily compressed MP3 is not a clean bill of health. First-party tools are narrow too. The ElevenLabs classifier is informative for ElevenLabs-style output, but its own documentation limits inspection to the first one minute and says it “does not reliably classify ElevenV3,” so a low score there does not clear other systems.

How should you read a score, then?

As a probability at an operating point you are usually not shown, and as an invitation to ask four questions: what file did you test, which model produced the score, what threshold was applied, and what is the false-positive rate on audio like this. A score without its operating point cannot be interpreted, and a score that might be reading background noise should not decide anything on its own. Preserve the original file, run a second method as corroboration, look for provenance, and weigh the source. Putting that reading to work on one clip is the practical question in is this voice AI-generated.

Sources

• Jung, Heo, Tak et al. (2022). AASIST: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks. ICASSP 2022.
• Tak, Patino, Todisco et al. (2021). End-to-End Anti-Spoofing with RawNet2. ICASSP 2021.
• 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.
• Negroni, Cuccovillo, Bestagini et al. (2026). Multi-Task Transformer for Explainable Speech Deepfake Detection via Formant Modeling. ICASSP 2026.
• ElevenLabs (2026). AI Speech Classifier.