aceability of 95 human listeners to distinguish natural speech from ElevenLabs-cloned speech after GSM, VoLTE, and VoIP transmission
Source article: Human Detection of Voice-Cloned Speech Under GSM, VoLTE and VoIP Conditions
The rapid progress of generative speech synthesis and voice-cloning technologies has enabled the creation of highly natural synthetic voices that pose a serious threat to telecommunication security. While most prior studies evaluate human ability to detect audio deepfakes using high-quality, studio-grade recordings, little is known about how real-world telecommunication channels affect perceptual detection. This study investigates the influence of three transmission scenarios—GSM (AMR-NB), VoLTE (AMR-WB), and Vo…
Contested: both sides are scored from claims and sources, not community votes.
"Taking this photograph got me a call from the Bienville Parish district attorney's office investigator" by Old White Truck is licensed under CC BY-SA 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/2.0/.
By June 2026, researchers tested how telecom transmission affects human detection of cloned speech. They created natural and ElevenLabs-synthesized utterances from nine speakers, processed them through simulated GSM, VoLTE, and VoIP codecs, and asked 95 participants to classify them as human or synthetic.
The work matters because voice cloning is being used to impersonate people over the phone for fraud and bypassing voice verification. Results showing near-chance performance overall and worst performance on high-quality VoLTE suggest current reliance on human ear alone is insufficient, though the small speaker set and simulated channels leave open how results generalize to live calls and diverse voices.
- Custom corpus used natural recordings from nine speakers and matching synthetic utterances from ElevenLabs voice cloning system.
- All samples were processed through simulated telecom channels using real codec implementations for GSM AMR-NB, VoLTE AMR-WB, and VoIP with packet-loss modeling.
- Listening test involved 95 participants performing binary human vs synthetic classification with confidence ratings.
- Overall detection accuracy was 54.8%, with VoLTE lowest at 44.0% and GSM highest at 63.7%.
Listeners detected ElevenLabs-cloned speech more accurately when audio was transmitted through narrowband GSM (AMR-NB) at 63.7%, compared to other telecom conditions.
95 listeners could barely distinguish natural speech from ElevenLabs-cloned speech after telecom transmission, with overall accuracy of 54.8% and only 44.0% on VoLTE, increasing susceptibility to voice spoofing.
The rundown
Researchers built a corpus from nine speakers and generated matched clones with ElevenLabs, then passed all clips through real implementations of AMR-NB for GSM, AMR-WB for VoLTE, and VoIP with packet-loss modeling.
In a controlled listening test, 95 participants labeled clips as human or synthetic and rated confidence, allowing comparison of detection across bandwidth and quality conditions.
Authors interpret that narrowband filtering may make prosodic irregularities of generative models more salient, while wideband high-quality channels mask artifacts and raise spoofing risk for systems relying on voice identity.
Findings are bounded by a small custom corpus of nine speakers and by use of simulated rather than live telecom transmission, which may not capture all real-world variability.
Sources
- Peer-reviewedAcoustics2026-06-17
How should this claim be treated?
The debate