A team of researchers employed a microwave radar sensor to detect smartphone vibrations and applied a large-scale AI-powered speech recognition model to convert those vibrations into intelligible speech.
The new “wireless-tapping” method explores remotely decoding conversations from subtle phone earpiece vibrations.
Radar Signals Could Be Exploited to Transcribe Phone Calls, Raising Privacy Concerns
Researchers at Penn State demonstrated that radar measurements from up to three meters away can generate phone call transcripts, raising privacy concerns.
Although the system currently achieves only about 60% accuracy with a vocabulary of up to 10,000 words, the findings highlight significant concerns about future privacy risks.
The researchers detailed the results in a study recently published in the Proceedings of WiSec 2025.
This research extends a 2022 study where radar and voice recognition detected 10 specific words, letters, and numbers with up to 83% accuracy.
“When we talk on a cell phone, earpiece vibrations travel through the device,” said Suryoday Basak, a Penn State doctoral student and lead author.
“By capturing vibrations with radar and using machine learning with context, we can reconstruct conversations.” By showing what’s possible, we aim to raise public awareness of potential risks,” the researcher explained.
Penn State Team Explores How Millimeter-Wave Radar Could Be Miniaturized for Everyday Devices
Basak and his advisor, Mahanth Gowda, tested a millimeter-wave radar to explore how they could shrink it to fit in everyday items like pens.
Self-driving cars, motion sensors, and 5G networks already use millimeter waves—microwaves between 300 MHz and 300 GHz.
The team emphasized their prototype was only for research, meant to anticipate future malicious uses.
Adapting AI Speech Models to Decode Noisy Radar-Based Signals
In the study, the team modified Whisper—an open-source, AI-based speech recognition system—to convert vibrations into readable speech transcripts.
“AI and open-source speech recognition have advanced rapidly over the last three years,” explained Basak. “We had to adapt these tools to handle the noisy, low-quality signals from radar.”
Instead of retraining Whisper from scratch, the researchers applied a technique known as low-rank adaptation. This method let them fine-tune the model for radar data by updating only about 1% of its parameters.
They used a millimeter-wave radar a few meters away to detect tiny vibrations from a smartphone’s earpiece.
Enhanced Transcription Accuracy Achieved Through Context-Aware Radar Signal Processing
The team fed radar signals into their adapted Whisper model, achieving up to 60% accuracy, with potential gains from context-based corrections.
“The system still errs, but it’s a big leap from the 2022 version that recognized only a few words,” said Gowda. “Even partial matches, like identifying keywords, hold value in security applications.”
The researchers compared the system to lip-reading, which catches only 30–40% of words yet remains understandable through context.
“Similarly, our model—when combined with contextual information—can reconstruct segments of phone conversations from several meters away,” Basak concluded.
Read the original article on: ZAP.aeiou
Read more: Google Will Restrict Nudity by Filtering Out Inappropriate Messages