Penn State Researchers Craft Radar-AI Technique That Can “Hear” Mobile Conversations from 10 Feet Away

UNIVERSITY PARK, Pa. — In a groundbreaking demonstration blending artificial intelligence and radar technology, researchers at Penn State University have shown it's possible to remotely detect and partially transcribe mobile phone conversations from as far as 10 feet (about 3 meters) away—a development raising both fascination and alarm.

Led by doctoral candidate Suryoday Basak and supervised by Associate Professor Mahanth Gowda of the Department of Computer Science and Engineering, the team introduced a novel system—dubbed "WirelessTap"—which uses millimeter-wave radar sensors to pick up tiny vibrations from a smartphone’s earpiece during a call, converting them into speech-to-text transcriptions .

How It Works

A millimeter-wave (mmWave) radar sensor—like those found in self-driving cars and 5G networks—is placed within line‑of‑sight of a phone in use.

The radar captures minute vibrations caused by speech traveling through the phone’s earpiece, vibrations imperceptible to human ears .

These signals are then processed using an adapted version of the AI speech recognition model “Whisper”, fine-tuned via a light-touch technique called low-rank adaptation (LoRA) to better interpret noisy radar-derived input .

Accuracy Over Distance

According to their June–July 2025 presentation at the ACM Conference on Security and Privacy in Wireless and Mobile Networks (WiSec 2025), the system’s transcription accuracy varies significantly with distance :

Distance Word‑level Accuracy

~20 inches (50 cm) Up to 59%
~2½–3 feet (human-held) Around 40.8%
~10 feet (300 cm) Only 2–4%, though still capturing isolated words


These results show that while full conversation transcription is not yet viable at longer ranges, even scant snippets of speech could potentially reveal sensitive information when combined with contextual understanding .

Privacy Implications & the Path Forward

Basak and Gowda warn that even partial overhearing of conversations—for example, hearing a name, a number, or a keyword—could pose serious privacy risks if leveraged maliciously . They liken the reconstruction of speech from fragmented data to lip-reading, where humans fill in context from partial cues .

Importantly, this method does not require access to encryption or hacking into the phone, only a clear, aligned view of the device—making it hard to detect and more difficult to guard against .

Potential for Good

The researchers also see benign applications for this technology—such as:

Smart home systems monitoring machinery wear and tear.

Health monitoring, where subtle vibrations might indicate emergencies and trigger alerts .

Summary
The Penn State team's WirelessTap experiment underscores a new frontier in surveillance—one where physical vibrations, not data breaches, become the vector. As mmWave sensors grow cheaper and AI models become more adept, the potential for misuse becomes more realistic. The work calls for proactive countermeasures and public awareness to safeguard privacy before vulnerabilities become widespread .