Hanwha Vision, GIST team tops global audio AI competition

Hanwha Vision and a research team from the Gwangju Institute of Science and Technology (GIST) announced Tuesday that they won first place at the DCASE 2026 Challenge, an international audio AI competition, using jointly developed continual learning technology.

The achievement came from Hanwha Vision's AI Research Lab, part of its R&D Center, working alongside a team led by Kim Hong-guk, a professor in GIST's School of Electrical Engineering and Computer Science. The collaboration addressed a critical weakness in AI systems: the tendency to forget previously learned sounds when trained on new ones.

The DCASE Challenge, organized by the Institute of Electrical and Electronics Engineers, is the world's most prestigious competition for audio AI technology, drawing researchers from academia and industry. This year, 135 teams participated.

The joint team placed first among 21 competitors in the "incremental learning for domain-independent audio classification" task — one of seven categories in the competition. The team successfully identified 10 types of sounds, including a baby crying, a dog barking and a fire alarm, in environments where the source of the audio recordings was unknown.

The team focused on solving what researchers call "catastrophic forgetting." Existing audio recognition AI systems lose performance when recording conditions change — such as differences in equipment, location or background noise. This occurs largely because AI models built for specific purposes tend to be smaller than large general-purpose models, though even large models experience the same problem, just less frequently. The issue is considered particularly critical in fields such as security and industrial safety, where external conditions can change rapidly.

The continual learning technology the team developed applies a technique called "DeepInversion-based generative replay," which reconstructs and reuses acoustic features the AI learned in the past. DeepInversion allows each sound classification model to independently reproduce and back-calculate audio data similar to what it previously learned, without collecting additional data. Through this method, the team preserved what the AI had learned at each stage, minimizing the degradation of existing knowledge during new training.

The team also achieved high accuracy and stable performance by applying an ensemble method that aggregates multiple model predictions to reach a final decision. The submitted system recorded a mean accuracy of 79.62% — the highest among all competing teams — compared with an average of about 68% for the other 20 teams.

The research was supported by Hanwha Vision's internal research program and a project funded by the Ministry of Trade, Industry and Energy. The results will be presented and winning teams recognized at the DCASE 2026 Workshop, a two-day event in Boston beginning Oct. 28.

"This award proves that the direction and technological level of Hanwha Vision's research are the best in the world," said Lim Jeong-eun, head of Hanwha Vision's AI Research Lab. "Continual learning technology is directly linked to the ability of security cameras to adapt to changing environments, and we will develop this new technology into future solutions."