WRAM
What should a machine sound like, here?
SubmittedICASSP 2027
Wiener-residual audio, frozen BEATs representations and normal-reference selection for noise-aware anomalous sound detection.
WRAM: Wiener-Residual Audio Matching with Normal-Reference Selection for Noise-Aware Anomalous Sound Detection
WRAM studies how normal-reference construction interacts with audio representation in noisy conditions. It combines Wiener-residual audio, a frozen BEATs encoder, relative-deviation pooling, compact reference selection and local-density-normalized matching.
Research status
This work has been submitted to ICASSP 2027 and is not accepted. This research overview and the earlier DCASE technical report are distinct research records.
Contact us to discuss the method or research collaboration.
Research record
WRAM: Wiener-Residual Audio Matching with Normal-Reference Selection for Noise-Aware Anomalous Sound Detection
JongWoo Sung · JeongSik Kim · HyeonJun Bae
2026 · Submitted · ICASSP 2027