GAP-URGENet: A Generative-Predictive Fusion Framework
for Universal Speech Enhancement

Xiaobin Rong1,2, Yushi Wang1,2, Zheng Wang1,2, Jing Lu1,2
1Key Laboratory of Modern Acoustics, Nanjing University
2NJU-Horizon Intelligent Audio Lab, Horizon Robotics
🏆 Challenge Leaderboard 📄 ArXiv Paper
Challenge Leaderboard
Figure: Rankings in the objective stage of the final leaderboard, with our team (WR) achieving first place.
Audio Demos from the Validation Set
fileid_1.flac: moderate noise with clipping distortion.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_2.flac: Strong noise.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_35.flac: Mild noise with frequent packet loss.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_64.flac: Medium-level babble noise.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_110.flac: Medium-level noise with medium-gap packet loss (max burst ~0.2s).
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_123.flac: Mild noise with medium-gap packet loss (max burst ~0.18s)
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_124.flac: Mild noise with codec artifact.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_133.flac: Milde noise with long-gap packet loss (max burst ~0.3s)
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_218.flac: Multiple noise types with packet loss.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_224.flac: Mild noise with reverberation.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_269.flac: Strong noise.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_481.flac: Syren noise.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_584.flac: Music noise.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
fileid_748.flac: Extreme noise.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
Clean Signal
Audio Demos from the Blind Test Set
fileid_538.flac: Real-recorded music noise.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
fileid_658.flac: Vocal separation.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
fileid_672.flac: Vocal separation.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output
fileid_830.flac: Whispered speech.
Noisy Signal
Pred. Branch Output
Gen. Branch Output
GAP-URGENet Output