Contents
An AI voice detector does not recognize a fake the way a person catches an odd cadence. It measures faint statistical traces that the synthesis pipeline, especially the vocoder that turns a model’s output into a waveform, leaves in the signal, and then maps those traces to a probability at a chosen threshold. It is strong on the systems it was trained on and much weaker on everything else, so understanding what it actually measures is the difference between reading a score well and trusting it blindly.
What a voice detector actually measures
The thing a detector reads is a production fingerprint, not “AI-ness.” When a model synthesizes speech, the vocoder stage leaves regularities that a real recording does not have, and analysis of these systems shows the differences concentrate in the higher frequencies. Frank and Schönherr (2021), building the WaveFake dataset across six generator architectures, found the generators deviate from real audio “specifically among the higher frequencies.” The AASIST authors describe the artifacts more generally, noting they “can reside in spectral or temporal domains” (Jung, Heo, Tak et al., ICASSP 2022). Other tells sit in the fine detail of articulation and in the parts of a clip where no one is speaking: Negroni, Cuccovillo, Bestagini and colleagues (ICASSP 2026) note that many detectors “base their decisions on unvoiced timesteps, frequency bands where no speech occurs, or even background noise, rather than the actual speech.” In other words, the signal a detector keys on is often subtle, high-frequency, and not where a human would listen.
The main families of detector
“AI voice detector” is not one method but a family of them, and they differ in what they feed the classifier, which is a large part of why two of them can split on the same clip (why do two voice detectors disagree):
- Handcrafted spectral features. Older systems compute cepstral coefficients, such as LFCC or CQCC, from the spectrum and hand them to a statistical classifier. Competitive in the lab, they lean on exactly the fine spectral detail that processing erases.
- Raw-waveform, end-to-end networks. RawNet2 learns directly from the waveform rather than a hand-designed feature, reaching about 4.66% equal error rate on ASVspoof (Tak, Patino, Todisco et al., ICASSP 2021).
- Spectro-temporal graph attention. AASIST models artifacts across both the spectral and temporal structure of a clip at once, reaching about 0.83% equal error rate on ASVspoof 2019 (Jung, Heo, Tak et al., ICASSP 2022).
- Self-supervised front-ends. The current strongest approach starts from a large speech model pretrained on ordinary audio, such as wav2vec 2.0 or XLS-R, then fine-tunes it to detect synthesis. Tak, Todisco, Wang et al. (2022) reported the “lowest equal error rates reported in the literature” with a wav2vec 2.0 front-end and data augmentation, around 0.82% on ASVspoof 2021, and these learned representations generalize better across generators than handcrafted features do.
- Forensic, speaker-specific comparison. A different regime models one known speaker’s own vowels and compares a suspect clip against reference audio of the real person. Yang, Sun, Lyu and Rose (2025) reached a 4.4% equal error rate detecting ElevenLabs v2 speech this way, against 27.5% for a standard baseline, but it needs clean reference audio and is a forensic tool, not a screen for a stranger’s clip.
From a waveform to a score
Whatever the family, the pipeline is the same shape: the front-end turns the audio into features, a classifier scores how strongly those features resemble the synthetic examples it was trained on, and a threshold converts that score into a label. Research reports the behavior as equal error rate because it is threshold-neutral, a single figure that describes the whole benchmark. A deployed tool has to pick an actual operating point, and there the two kinds of error diverge: a threshold set to catch more fakes will also flag more real voices. This is why a raw “94 percent” is not a 94 percent chance the voice is fake, a point covered in what does an AI voice-detector score mean.
Why lab accuracy does not survive contact with the wild
The sub-1 percent figures above are real and also fragile, because they are measured on the systems the detector learned. Move off that distribution and detection collapses. A neural-vocoder 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, Jia, Hou, Lyu, CVPRW 2023). Tested on real audio collected outside the lab, the survey by Yi, Wang, Tao and colleagues (2023) reports equal error rates climbing from the 20-to-30 percent range up to 35 to 55 percent, with RawNet2 rising from 24.32% to 36.74% and AASIST from 19.77% to 34.81%. The cause, in the survey’s words, is the “poor generalization of existing detection methods to unknown fake attacks.” A detector that has never heard your generator is close to guessing.
What this means for reading a result
A voice detector is a narrow, capable reader of the fingerprints it was taught, and a weak one everywhere else. Treat a flag as a strong first-pass signal when the clip is a clean original from a system the tool likely covers, and as weak evidence when the clip is short, compressed, or from an unfamiliar generator. The reliable way to use one is to preserve the original file rather than a re-encode, run more than one method and read agreement as confidence rather than proof, and look for provenance where it exists. For the single-clip question this feeds into, see is this voice AI-generated.
Sources
- Frank, Schönherr (2021). WaveFake: A Data Set to Facilitate Audio Deepfake Detection. NeurIPS 2021 Datasets and Benchmarks.
- 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.
- Tak, Todisco, Wang et al. (2022). Automatic Speaker Verification Spoofing and Deepfake Detection Using wav2vec 2.0 and Data Augmentation.
- Sun, Jia, Hou, Lyu (2023). AI-Synthesized Voice Detection Using Neural Vocoder Artifacts. CVPRW 2023.
- Yang, Sun, Lyu, Rose (2025). Forensic Deepfake Audio Detection Using Segmental Speech Features.
- Negroni, Cuccovillo, Bestagini et al. (2026). Multi-Task Transformer for Explainable Speech Deepfake Detection via Formant Modeling. ICASSP 2026.
- Yi, Wang, Tao et al. (2023). Audio Deepfake Detection: A Survey.