Paderborn Researchers' Audio Diarization Detects Unknown Sounds in Real-Time
New AI technique spots never-before-heard audio events without any pre-training.
Traditional sound event detection (SED) relies on a fixed set of predefined classes—trained to recognize dog barks, glass breaking, or alarms. But what happens when you deploy a microphone in an unknown environment? You can't list every possible sound in advance. Now, researchers from Paderborn University have introduced a new paradigm called 'audio diarization' that tackles this exact blind-spot.
Their system, adapted from speaker diarization techniques, first pinpoints when any relevant sound event starts and stops (including overlapping events) without needing to know what the sound is. Only later does it attempt to classify or compare the event against known categories. This two-step architecture allows the system to achieve accuracy comparable to traditional closed-set SED while adding the critical ability to detect novel sounds—a feature current models lack. The team presented their findings at IWAENC 2026 and released the paper on arXiv.
- Defines audio diarization as detecting onset/offset times with overlap for an open set of classes without user prompts.
- Adapted from speaker diarization; requires no pre-defined event categories for detection.
- Matches closed-set SED performance while uniquely enabling detection of novel, unseen sound events.
Why It Matters
Enables real-time audio monitoring in unknown environments like wildlife tracking, security surveillance, and industrial safety without pre-labeled sound libraries.