New Beginning, New Inspiration
Pukyong National University, the first university in Busan,always paves a new path through specialized and converged research to lead the era of the fourth indutrial revolution.
| Dong-Heon Lee | Joint Research with Meta on Noise Cancellation for AI Glasses | |||
| 작성자 | 대외홍보센터 | 작성일 | 2026-11-18 |
| 조회수 | 14 | ||
| Dong-Heon Lee | Joint Research with Meta on Noise Cancellation for AI Glasses | |||||
![]() |
대외홍보센터 | ![]() |
2026-11-18 | ![]() |
14 |
Pukyong National University Professor Dong-Heon Lee Collaborates with Meta to Develop Noise-Cancellation Technology for AI Glasses
- Virtual audio augmentation technology delivers the performance of 16 microphones using just four … Paper accepted at NeurIPS 2026, a leading international AI conference

Professor Dong-Heon Lee of the Department of Information and Communications Engineering at Pukyong National University has developed Virtual-Flow, a high-performance noise-cancellation and spatial audio augmentation technology designed for integration into compact smart devices, including artificial intelligence (AI) glasses, through collaborative research with global technology giant Meta.
In recent years, wearable devices such as augmented reality (AR) glasses, wireless earbuds, and hearing aids have faced significant limitations in multichannel spatial audio processing, which requires precise removal of ambient noise and reverberation. Due to physical constraints on device size, these devices can accommodate only a limited number of microphones.
To address these challenges, virtual microphone estimation technology, which estimates signals at virtual microphone positions from actual microphone signals, has attracted considerable attention. However, conventional approaches have faced limitations, including the difficulty of large-scale supervised learning, which requires collecting extensive ground-truth data for numerous virtual microphones, as well as high computational complexity.
To overcome these limitations, Professor Dong-Heon Lee’s research team proposed Virtual-Flow, a novel semi-supervised flow-matching framework capable of generating high-quality virtual spatial audio information without requiring ground-truth virtual microphone data.
The research team utilized time-delay information from actual microphone signals to construct synthetic virtual targets and applied inference optimization techniques alongside a downstream beamforming objective function. As a result, even in a constrained environment with only four microphones, the technology achieved speech enhancement performance approaching that of a physical microphone array with 16 or more microphones. Furthermore, the system operates with just a single inference step, maximizing computational efficiency and enabling real-time processing on embedded devices.
Notably, the study demonstrated outstanding robustness not only in simulated environments but also with real-world recorded multichannel speech datasets and scenarios involving moving speakers. The technology is expected to provide a foundation for significantly improving speech recognition and call quality in compact consumer electronic devices with substantial physical constraints, such as AR glasses and advanced hearing aids.
The research, with Professor Dong-Heon Lee serving as the first author, has been officially accepted for presentation at the 40th Conference on Neural Information Processing Systems (NeurIPS 2026), one of the world’s most prestigious academic conferences in artificial intelligence, scheduled to begin on December 6 in Sydney, Australia. The achievement has attracted significant attention from both academia and industry. <Pukyong Today>


