detectai.media

Does voice liveness detection work against deepfake voice?

Liveness detection catches the playback event, not the synthesis, so it works on noisy replayed deepfakes and degrades sharply on clean or injected ones.

By The detectai.media team
4 min read
Contents

Partly, and the boundary is precise: liveness detection catches the playback event, not the synthesis. A deepfake voice played through a loudspeaker into a microphone acquires replay physics, and in noisy, reverberant conditions liveness detectors read those physics well. Against a clean, near-anechoic playback the same detectors degrade several-fold, and against a deepfake injected electrically, with no playback at all, they have nothing to read.

Two attacks, two detectors

A deepfake reaching a voice-verification system arrives one of two ways. Played acoustically, it is simultaneously a synthesis attack and a replay attack, and both detector types get a shot at it. Injected directly into the signal path, it is a synthesis attack only, the scenario the ASVspoof benchmarks call logical access as opposed to physical access (Wang, Yamagishi, Todisco et al., Computer Speech & Language 2020). Liveness detection only ever adjudicates the first case. That division of labor is why production systems run a liveness check and a synthesis check side by side, as described in how live deepfake detection works. The liveness check earns its place because speaker verification alone accepts 68% of replays at its own threshold (Villalba & Lleida, BioID 2011); a synthesis detector never sees that hole, because a real recording of a real person contains nothing synthetic to flag.

Replay hides the deepfake from one layer and exposes it to the other

When a deepfake is played through a speaker, the two layers trade places. Müller and colleagues at Fraunhofer AISEC (arXiv 2025) measured the synthesis side losing its grip: replaying synthetic speech through real hardware raised a state-of-the-art detector’s equal-error rate from 4.7% to 18.2%, while accuracy on genuine speech barely moved. But the same replay stamps loudspeaker distortion, low-frequency roll-off, and a second room signature onto the file, the exact features liveness classifiers are trained on (Ren, Fang & Liu, Multimedia Tools and Applications 2019). The launder that defeats the AI-voice detector is the evidence the liveness detector runs on.

How well the liveness layer reads a replayed deepfake

Neri and Virtanen (arXiv 2025) benchmarked replay-speech detectors under controlled acoustic conditions, and their numbers draw the working boundary of the whole approach:

Playback conditionALRAD EERM-ALRAD EER
Reverberant room6.6%2.4%
Anechoic (no reverb)18.2%14.9%

In a reverberant room, the multi-channel M-ALRAD detector operates at a 2.4% equal-error rate, because the doubled room acoustics give it a strong tell. Remove the reverberation and the error rate rises to 14.9%, a six-fold degradation for the same detector, since a clean playback in a dead room leaves far less of the double-acoustic signature to find. The engine research behind this cluster reads that asymmetry as the central limit: the detector depends on the messiness of the playback, and a careful attacker controls the messiness.

Where it fails outright

Two failure modes are measured rather than hypothetical. First, the training gap: Neri and Virtanen report that detectors trained on simulated replay channels and tested on real re-recorded audio reached equal-error rates above 50% in most environments, worse than chance. Synthetic training data does not yet capture real playback physics. Second, the injection hole: a deepfake delivered line-level, through a virtual audio device or a compromised input path, never passes through a loudspeaker or a room, so every acoustic feature in this article is silent. The forensic reading of that same boundary, what re-recording resets and what survives it, is drafted from the examiner’s side in does re-recording a deepfake through a microphone make it undetectable?.

What a layered read actually buys

The evidence supports a modest, specific claim. Liveness detection is a strong second layer against the common case, a deepfake played into a microphone under ordinary acoustics, precisely because ordinary acoustics are noisy. It thins out as the playback gets cleaner, and it vanishes when there is no playback. So a liveness pass is never proof a voice is real, and a liveness flag is never proof it is synthetic; the layer adjudicates delivery, not origin. The origin question belongs to the synthesis detectors, with their own measured gaps between benchmark and deployment (see is this voice AI-generated?), and the strongest read remains the combination: liveness, synthesis detection, and provenance, each covering a failure mode the others cannot see. How attackers probe the seams between those layers is measured in can voice liveness detection be bypassed?.

Sources

  • Neri, M. & Virtanen, T. (2025). Acoustic Simulation Framework for Multi-Channel Replay Speech Detection. arXiv:2509.14789.
  • Müller, N., Kawa, P., Choong, W.-H., Stan, A., Tirumala Bukkapatnam, A., Pizzi, K., Wagner, A. & Sperl, P. (2025). Replay Attacks Against Audio Deepfake Detection. arXiv:2505.14862.
  • 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.
  • Wang, Yamagishi, Todisco et al. (2020). ASVspoof 2019: A Large-Scale Public Database of Synthesized, Converted and Replayed Speech. Computer Speech & Language 2020.
  • Villalba, J. & Lleida, E. (2011). Detecting Replay Attacks From Far-Field Recordings on Speaker Verification Systems. BioID 2011.
#audio#voice#liveness#deepfake
Last updated
23 July 2026
Category
Reliability