Hanwha Vision and GIST Win First Place at Prestigious International Audio AI Competition

Screenshot of the DCASE 2026 challenge results page on the official website

Hanwha Vision, a global vision solution provider, has presented a breakthrough solution to one of the most persistent challenges in acoustic artificial intelligence (AI), securing first place at a prestigious international competition. The advanced acoustic AI technology developed through this initiative will be integrated with Hanwha Vision’s video surveillance technologies to develop next-generation AI security systems.

On July 14, Hanwha Vision announced that its joint research team with the Gwangju Institute of Science and Technology (GIST) won first place at the Detection and Classification of Acoustic Scenes and Events (DCASE) 2026 Challenge. The team achieved this top spot using their newly developed "Continual Learning" technology.

This award-winning technology was co-developed by Hanwha Vision AI Lab and Professor Hong Kook Kim’s team at GIST. The project has garnered significant industry attention for successfully addressing a critical AI vulnerability: the tendency of AI models to forget previously learned sounds when training on new ones.

Organized by the Institute of Electrical and Electronics Engineers (IEEE), the DCASE Challenge is the world’s leading competition for acoustic AI, bringing together top academic and corporate research teams globally. This year, 135 teams took part.

Hanwha Vision and GIST won first place out of 21 teams in the "Domain-Agnostic Incremental Learning for Audio Classification" category. The mission was to identify 10 different sounds—like baby cries, dog barks, and fire alarms—in environments where the source of the audio was completely unknown.

Solving a Major AI Bottleneck with 'Continual Learning'

The joint research team focused on solving "Catastrophic Forgetting."

Traditional audio recognition AI models suffer from performance degradation when exposed to new environments, such as different recording equipment, locations, or ambient noise. More critically, when these models are trained on new data, they tend to overwrite and lose previously acquired knowledge. For example, an AI model highly proficient at classifying urban noises might struggle to recognize those same sounds after being retrained on airport acoustic environments.

This phenomenon is particularly pronounced in specialized AI models, which are relatively smaller in scale compared to ultra-large foundation models (though large-scale models also inherently experience this issue to a lesser degree). In fields like security and industrial safety—where external environments change constantly—resolving this issue is considered a core milestone for practical AI deployment.

To address this, the team's "Continual Learning" technology incorporates a "DeepInversion-based Generative Replay" method to reconstruct and utilize previously learned acoustic features. Without requiring additional physical data collection, the DeepInversion technique enables the audio classification model to mathematically back-calculate and recreate virtual audio data that mimics previously learned sounds.

For instance, if an AI in an indoor security system learns highway traffic noises, it can track back the decibel, amplitude, and other acoustic characteristics of previously learned indoor sounds. This minimizes the corruption of existing knowledge during new training phases, allowing the AI to retain its prior classification capabilities.

The team also utilized an "Ensemble Method"—which aggregates predictions from multiple AI models to make a final decision—to deliver high accuracy and robust stability. The final system submitted by the team achieved a top-ranking average accuracy of 79.62%, outperforming the average of the other 20 participating teams (approx. 68%) by over 10 percentage points. Furthermore, the team's final ensemble system, along with three individual single systems they additionally submitted, swept the 1st through 4th places in the official task rankings, proving the overwhelming technical maturity of their approach.

Paving the Way for Next-Generation Video-Acoustic AI Security Solutions

This technology, which stably accumulates acoustic data across diverse environments, is expected to be actively integrated into Hanwha Vision’s next-generation AI security portfolio.

"This recognition proves that Hanwha Vision is leading the way with world-class technology and research," said Jeong Eun Lim, Head of the AI Lab at Hanwha Vision. "Continual Learning is key to helping AI security cameras adapt to real-world environments. We will keep building on this technology to deliver practical, future-ready solutions."

The team will present their findings and receive their award at the DCASE 2026 Workshop, held in Boston, USA, from October 28 to 29.