Sungkyunkwan University researchers develop voice security technology to detect spoofed audio

A research team at Sungkyunkwan University has developed a voice security technology designed to combat increasingly sophisticated voice deepfake crimes, the university said Wednesday.

The research team, led by Park Eun-il, a professor in the Department of Applied Artificial Intelligence, demonstrated that the technology can detect manipulated audio inserted into genuine voice recordings.

The technology is expected to have broad applications in areas related to everyday security, including voice phishing prevention, voice-based identity verification for financial services and digital evidence analysis.

The team used artificial intelligence (AI) to analyze subtle changes in audio signals and speech patterns that are difficult for the human ear to distinguish, pinpointing the exact locations of manipulated speech to within one second.

Two papers by the team have been accepted for presentation at Interspeech 2026, an annual conference of the International Speech Communication Association on speech science and technology, scheduled for Sept. 27 to Oct. 1 in Sydney, Australia.

The research team, known as the Data eXperience Laboratory, includes Kim Yun-su, Yang Seong-yeon, Park Sun-young, Yuqian Zheng, Lee Yu-been and Ahn Hyo-ju.

“Our efforts to develop technologies that people can use safely and effectively, rather than focusing solely on improving AI performance, have gained global recognition,” said professor Park, who led the research.

He added, “We will continue to develop next-generation AI technologies that can enhance the safety of our society and address real-world problems.”

The team also conducted a study examining how reliably people can trust the explanations AI provides for its decisions.

AI systems are often expected to help people make more accurate judgments by providing clear explanations of how they reach their decisions.

However, the team's experiments showed that even when users perceive such explanations as “useful,” their ability to detect errors may not improve and, in some cases, they may simply spend more time deliberating.

The findings are expected to provide an important benchmark for developing trustworthy AI systems that offer genuinely useful explanations rather than merely appealing ones, particularly in fields such as health care and finance where accurate decision-making is essential.

The research team also highlighted a diverse range of AI convergence studies related to everyday life, including a speech synthesis technology that naturally reproduces regional dialects and intonation using only photographs and location information.

The studies included a technology that detects malicious speech in real time based solely on voice tone and rhythm; a vision model capable of recognizing subtle emotional cues in images; and a medical imaging technology that accurately identifies diseased areas from limited information.

These studies highlight the breadth of the team’s AI convergence research aimed at addressing real-world challenges.

Six papers from these studies will be presented at the 35th ACM International Conference on Information and Knowledge Management, to be held in Rome, Italy, from Nov. 7-11.