In order to anticipate criminals, researchers from the State University of Pennsylvania, USA, have dedicated themselves to seeking new ways in which people’s privacy can be violated.
Thus, during the Wisec 2025 security conference (July 2025) they showed that It is possible to train an AI to generate telephone call transcripts from radar measurements taken at a distance of up to three meters.
This is the “wireless interception”, which explores the possibility of deciphering distance conversations from the vibrations produced by the headset of a cell phone.
While precision remains limited (around 60% for a vocabulary of up to 10,000 words), the findings account for a future risk to privacy.
Same technology for daily use
The investigation was published in the Actas de la WiSec 2025: 18th ACM Conference on Security and Privacy in Wireless and Mobile Networks And it is based on a 2022 project on which the team used a radar sensor and voice recognition software to wirelessly identify 10 words, letters and predefined numbers with an accuracy of up to 83%.
The millimeter wave radar sensor used is the same technology that is used in autonomous vehicles, movement detectors and 5G wireless networkswhich can be as small as to install it in everyday objects like a pen. The vibrations detected are then decoded through a voice recognition model based on AI.
“When we talk on cell phone, We tend to ignore vibrations that arrive through the headset and that make the entire phone vibrate, ”said Suryoday Basak, leader of the project in a statement.“ If we capture these same vibrations through remote radars and apply automatic learning to understand what is said, Using context keys, we can determine complete conversations. By understanding the possibilities, we can help the public to be aware of potential risks, ”he added.
Technically feasible
As Basak explained, in the last three years there has been a boom in AI capacities and open source voice recognition models. As they are more oriented to analyze speech clean or cases of daily use, they had to adapt one to recognize low quality and noise radar data.
“The result were transcriptions of conversations, with the expectation of some errors, which represented a remarkable improvement with respect to our 2022 version, which only generates a few words,” said Mahanth Gowda, co -author of the study. “But even the detection of partial voice coincidences, such as keywords, is useful in a security context,” he said.
According to the researchers, Transcription accuracy could still be improved further incorporating manual corrections based on the context, such as adjusting certain words or phrases, when previous knowledge of the conversation is available.
“He Objective of our work was to explore if these tools could be used by malicious people to spy distance telephone conversations. Our findings suggest that this is technically feasible in certain circumstances, and we hope that this will raise the public so that people are more aware during sensitive calls, ”Basak emphasized.
News references
Artículo en Proceedings of WiSec 2025: 18th ACM Conference on Security and Privacy in Wireless and Mobile Networks. Wireless-Tap: Automatic transcription of phone calls using millimeter-wave radar sensing.
Nota de prensa Universidad Estatal de Pensilvania. Conversations remotely detected from cellphone vibrations, researchers report.
