Portuguese 
English 
Spanish 
4 min of reading 
0 comments 
Penn State Researchers Created a System That Combines Radar and Artificial Intelligence to Transcribe Phone Conversations Up to Three Meters, Raising an Alert About Privacy Risks.
The faintest vibration from your phone can reveal what you say. Penn State researchers developed a radar technique to intercept conversations through subtle vibrations emitted by a mobile phone’s speaker.
The method uses millimeter-wave radar combined with a speech recognition system based on artificial intelligence.
With this technology, it was possible to capture and transcribe phone calls from up to three meters away, with about 60% accuracy.
-
In Cuba, where blackouts last for hours and charging an electric tricycle takes 10 hours with no guarantee of power, an entrepreneur inaugurated the country’s first free solar charging station; residents are already coming from 70 kilometers away just to get power.
-
Goodbye allergy and fine dust: Xiaomi launches purifier that removes 99.99% of particles in one hour, eliminates formaldehyde in two hours, reduces 98.9% of pollen, and delivers 11,666 liters of clean air per minute
-
Comet 3I/ATLAS comes from outside the Solar System, carries “heavy” water at an unprecedented level, and leaves scientists intrigued about where this alien visitor truly originated.
-
While refrigerators and air conditioners still rely on gases that can leak, a Cambridge startup raises $10 million to create refrigeration with plastic crystals that cool when squeezed.
The most important aspect is that the result was achieved without physical contact with the device. It is sufficient to detect the vibrations generated by the voice for the system to reconstruct parts of the dialogue.
Evolution of a Previous Experiment
This was not the first time the group worked on the topic. In 2022, they had already achieved up to 83% accuracy in recognizing ten predefined words, using a similar approach.
Now, the goal was more ambitious: to create continuous speech transcription. The accuracy dropped due to the difficulty of interpreting noisy data captured by the radar, but the technological advancement was clear.
According to the lead author, Suryoday Basak, when someone speaks on the phone, vibrations spread throughout the device. Capturing them with remote radars and using machine learning to interpret them allows for conversation reconstruction.
How the System Was Assembled
The radar used is the same type of sensor applied in autonomous cars, motion detectors, and 5G networks. It measures tiny surface vibrations caused by speech on the earphones.
To interpret the signal, the researchers adapted Whisper, an open-source speech recognition model designed for clean audio.
A low-level adaptation technique was used, retraining only 1% of the model’s parameters to handle radar data.
This specific adjustment avoided the need to train a new system from scratch and improved the quality of transcription results, even with low-quality data.
Results Obtained
In practice, the sensor was positioned three meters away from a mobile phone. The measured vibrations were processed by the AI model, which generated transcriptions with 60% accuracy over a vocabulary of up to 10,000 words.
Though far from ideal, the scientists emphasize that even partial matches of keywords can be dangerous. According to associate professor Mahanth Gowda, the technology already represents a significant advancement over the previous version.
He compared the performance to lip reading, which identifies 30% to 40% of the words. Even incomplete, the information can be sufficient to deduce the content of a conversation, especially based on context.
Privacy at Risk
For Basak, the central concern is that the system could be used for espionage. Just as a lip reader combines visual cues and context to understand dialogues, the combination of partial transcriptions with additional information can expose confidential data.
The work was conducted to assess technical feasibility and alert about potential risks before they are maliciously exploited. “We hope people become more vigilant during sensitive calls,” said the researcher.
Next Steps and Protection Measures
The research received support from the National Science Foundation of the United States. The group plans to develop ways to protect conversations against this type of remote capture.
Among the possibilities are methods to reduce or mask the vibrations of the phone. The idea is to make it difficult for external radars to register clear signals for the AI to interpret.
Additionally, the scientists believe that device manufacturers may adopt design solutions to limit the propagation of these micro-vibrations.
Alert About Emerging Technologies
This study shows how existing resources — such as millimeter-wave radars — can gain unexpected functions when combined with artificial intelligence.
The most important thing is to understand that advances like these bring both benefits and risks. While radar and AI can help in areas such as security and accessibility, they can also be used to violate privacy.
The researchers highlight that even without achieving 100% accuracy, the system proves that it is possible to capture sensitive data remotely. And, therefore, that privacy in environments with these technologies needs to be rethought.
The complete work was published in the Proceedings of WiSec 2025: 18th ACM Conference on Security and Privacy in Wireless and Mobile Networks, reinforcing the academic and technological relevance of the topic.
As AI and wireless communications advance, the need for protection against forms of espionage that, until recently, seemed unlikely grows. After all, even the faintest vibration from your cell phone can reveal more than you imagine.
Be the first to react!