Demonstration of deep binaural multi-frame Wiener filter for impulsive noise reduction. Please use headphones to perceive binaural cues. The noisy covariance matrix and the noisy-target cross-covariance vector are adaptively estimated using a deep neural network (DNN) based on the Full and Sub-Band Long Short-Term Memory (FSB-LSTM) architecture. The algorithm runs in realtime on a single CPU core (MacBook Air M1).

Presenters: Wiebke Middelberg & Daniel Fejgin

Related References:
- Deep Multi-Frame MVDR Filtering: [ Ссылка ]
- Extension Towards Binaural Listening Scenarios: [ Ссылка ]
- DNN Architecture:
[ Ссылка ]

Shoutout to our friends at the University of Hamburg:
[ Ссылка ]

свернуть