Separate
Nonnegative Novelty Extraction (NNE) estimates normal and anomalous components using a dictionary learned from normal recordings.
The idea
Hear the sound behind an anomaly score.
An anomaly score flags a departure from everyday sounds. AnomSep also produces an audio waveform that lets a listener inspect the evidence behind that score. An anomalous sound deviates from an environment’s normal sound distribution; an anomaly does not necessarily indicate danger.
Refine
Flow matching in a pretrained audio latent space refines the initial estimates. Normal Region Exclusion selects surrogate anomalous audio for adaptation.
Detect and listen
The energy of the refined anomalous component provides an anomaly score. The same component can be played back as a listenable explanation.
Listening room
From a mixture to listenable explanations
Compare four methods on five mixtures containing anomalous sounds and five normal-only recordings. All method outputs below are estimates of the anomalous component.
Start with the mixture and references. Then compare the estimated anomalous sounds. Successful separation preserves the anomalous event and suppresses the normal background. A normal-only example should produce a quiet anomalous-sound estimate.
CLAP is audio embedding cosine similarity to the anomalous reference; SAJ is the automated SAM Audio Judge overall score. Higher is better for both. Scores are undefined for silent anomalous references in normal-only examples and are shown as N/A.
Loading listening examples…
Consistent playback. No track is normalized independently. A shared attenuation, when needed, keeps every track within an example below full scale and preserves relative amplitudes. Keep the player volume unchanged when comparing methods.
−80 to 0.0 dB · one spectrogram scale per example · 0.0–8.0 kHz
Spectrogram levels are relative to the maximum magnitude across all reference and estimated tracks in the selected example. Detection labels are the saved predictions from each method’s tuning-selected threshold.
Use AnomSep
Code and pretrained models
The release provides reusable AnomSep and Normal Region Exclusion classes, scene-specific model weights, and reproducible demo selection.
Original source code is available for noncommercial research only. See the LICENSE.