Contents
Every published method for telling a live speaker from a replayed recording keys on the same fact: a replay must pass through a playback chain, and the chain cannot be silent. A live voice arrives at the microphone directly. A replayed one arrives through a loudspeaker, a second room, and a second capture, and each stage of that detour feeds a distinct detection family. There are four.
Why the methods exist at all
Speaker verification by itself does not check for playback. Villalba and Lleida (BioID 2011) put a number on the consequence, testing far-field re-recordings against a verification system and reporting the result in plain terms: “if we would choose the EER operating point as the decision threshold, we would accept 68% of the spoofing trials.” A system that only asks “is this the right voice?” answers yes to a recording of the right voice. The four families below are the checks that ask the second question.
Family 1: loudspeaker-distortion classifiers
The playback loudspeaker is the one component a replay cannot omit, and small drivers distort in measurable ways. Ren, Fang and Liu (Multimedia Tools and Applications 2019) trained a support-vector machine on distortion features and reported a 97.78% true-positive rate with a 99.62% true-negative rate on their evaluation set. The same paper also measured the family’s main weakness: when the playback loudspeaker at test time differed from those seen in training, the true-positive rate fell to a floor near 89.7% for their method, and to 79.3% for the baseline they compared against. The classifier partly learns the devices, not just the physics.
Family 2: cepstral front-ends tuned to the replay band
The second family reads the spectrum. Constant-Q cepstral coefficients, introduced by Todisco and colleagues as a spoofing countermeasure with an equal-error rate of 1.85% on the RedDots-Replayed corpus (Todisco et al., 2017), became the standard physical-access front-end. Witkowski, Kacprzak, Żelasko, Kowalczyk and Gałka (INTERSPEECH 2017) then showed where the signal actually lives: limiting CQCC features to the 6 to 8 kHz band improved the ASVspoof 2017 development-set equal-error rate from 11.86% to 5.13%, because repeated digital-to-analog conversion stamps its anti-aliasing signature into the top of the band.
Family 3: far-field and channel detection
The third family models the recapture channel itself: the reverberation, noise, and frequency shaping that far-field re-recording adds. In Villalba and Lleida’s BioID experiments, a dedicated spoof detector reached equal-error rates between 0 and 9% depending on channel condition, with 0% when training and test channels matched and 7.32% under mixed conditions. Their companion study (IEEE ICCST 2011) measured why the channel is so informative: replaying through a handset degraded verification scores by 90.5 to 123.1% relative, meaning the channel dominates the evidence the recording carries.
Family 4: the dual-ENF forensic check
The fourth family comes from forensic audio rather than biometrics. A recording device picks up the local mains hum, the electrical-network-frequency signature, at capture time. The European Network of Forensic Science Institutes’ best-practice manual states that more than one ENF series in a single file indicates analogue copying (ENFSI FSA-BPM-002, §5.4.1): the second series is acquired during the replay pass. It is the only replay indicator named in a standards-grade forensic document, and an examiner-side view of what re-recording does to every other trace in a file is drafted in does re-recording a deepfake through a microphone make it undetectable?.
The four families side by side
| Family | Representative result | Known limit |
|---|---|---|
| Loudspeaker distortion | TPR 97.78% / TNR 99.62% (Ren 2019) | TPR floor near 89.7% cross-loudspeaker |
| Cepstral front-ends (CQCC) | EER 1.85% RedDots (Todisco 2017) | band-sensitive; corpus-tuned |
| Far-field / channel | EER 0 to 9% by channel (Villalba 2011) | 0% only when channels match |
| Dual-ENF | standards-named indicator (ENFSI) | ENF must be present in both passes |
What all four share
Each family reads a different stage of the same detour, which is why serious systems fuse them rather than pick one. They also share two structural limits. Every published accuracy above is an in-distribution number, and the cross-loudspeaker and cross-channel rows show what happens when the deployment device leaves the training set. And all four assume an acoustic playback event: a recording injected electrically, straight into the signal path, passes through no loudspeaker, no second room, and no second capture, and presents nothing for any of them to read. How attackers exploit exactly those seams is the subject of can voice liveness detection be bypassed?, and the mechanism behind the tells is unpacked in how does voice liveness detection work?.
Sources
- Villalba, J. & Lleida, E. (2011). Detecting Replay Attacks From Far-Field Recordings on Speaker Verification Systems. BioID 2011.
- Villalba, J. & Lleida, E. (2011). Preventing Replay Attacks on Speaker Verification Systems. IEEE ICCST 2011. DOI: 10.1109/CCST.2011.6095943.
- Ren, Fang & Liu (2019). Replay Attack Detection Based on Distortion by Loudspeaker for Voice Authentication. Multimedia Tools and Applications. DOI: 10.1007/s11042-018-6834-3.
- Todisco, M. et al. (2017). Constant Q Cepstral Coefficients: A Spoofing Countermeasure for Automatic Speaker Verification. Computer Speech & Language.
- Witkowski, M., Kacprzak, S., Żelasko, P., Kowalczyk, K. & Gałka, J. (2017). Audio Replay Attack Detection Using High-Frequency Features. INTERSPEECH 2017. DOI: 10.21437/Interspeech.2017-776.
- European Network of Forensic Science Institutes (2022). Best Practice Manual for Digital Audio Authenticity Analysis (ENFSI-FSA-BPM-002).