Hanwha Vision Wins DCASE 2026 Acoustic AI Challenge

Acoustic AI has long struggled with a fundamental flaw: models forget previously learned sounds when trained on new ones. Hanwha Vision and GIST just solved that problem — and won a world-leading competition proving it.

For security and industrial safety systems, an AI model that loses knowledge every time it adapts to a new environment is a critical liability. Recording equipment changes, ambient noise shifts, and locations vary — yet an urban-noise classifier that suddenly forgets those sounds after learning airport acoustics is not deployable in the real world. This challenge, known as catastrophic forgetting, has held back practical acoustic AI in surveillance for years.

What was announced

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, organized by the IEEE. The team took the top spot in the Domain-Agnostic Incremental Learning for Audio Classification category, competing against 21 teams in that track from 135 total participants worldwide.

The mission required identifying 10 different sounds — including baby cries, dog barks, and fire alarms — in environments where the source of the audio was completely unknown. The Hanwha Vision AI Lab and Professor Hong Kook Kim's team at GIST achieved 79.62% average accuracy, outperforming the average of the other 20 participating teams (approximately 68%) by more than 10 percentage points. The team's final ensemble system and three additional single systems swept the 1st through 4th places in the official rankings.

Why this matters

For CIOs, CISOs and physical security leaders, acoustic AI that adapts without forgetting is a game-changer for deployment economics. Traditional models require costly retraining and data collection every time an environment changes. Continual Learning eliminates that overhead, enabling AI cameras and edge devices to accumulate knowledge across diverse sites — retail, transportation, industrial, healthcare — while maintaining prior classification accuracy.

The result is more reliable event detection with lower total cost of ownership, and a clear path toward video-plus-audio AI security systems that understand context far better than vision alone.

Technical details

• Competition: DCASE 2026 Challenge, organized by IEEE, 135 teams globally.
• Category: Domain-Agnostic Incremental Learning for Audio Classification, 21 teams.
• Result: 1st place with 79.62% average accuracy vs. approximately 68% peer average.
• Core method: Continual Learning with DeepInversion-based Generative Replay.
• Data efficiency: Mathematically reconstructs virtual audio mimicking prior learned sounds without new physical data collection.
• Robustness: Ensemble method aggregates predictions from multiple AI models for stable accuracy.
• Task scope: Classification of 10 sound classes including baby cries, dog barks, and fire alarms.
• Next step: Findings will be presented at the DCASE 2026 Workshop in Boston, USA, October 28–29.

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

Softprom and Hanwha Vision

Softprom is the official distributor of Hanwha Vision. Partners and integrators can access the full Hanwha Vision portfolio of AI-powered cameras, NVRs and video management systems through Softprom, including advisory, licensing and technical enablement.

This content was prepared as part of the Softprom DistriFlow project — an automated system for monitoring and adapting vendor news. Original source: original article.