Self-Adaptive noise distribution network for Speech Enhancement with heterogeneous data of Cross-Silo Federated learning

• We propose a SASE model with adaptive noise distribution, which achieves state of the art results on the VioceBank+DEMAND dataset.
• We simulated the federated learning setting of a real environment and verified the robustness of the proposed SASE noise reduction model in a real environment through experiments and visualization.
• The proposed SASE model is computed based on the complex domain, and the TF-GA block is used to extract richer information of speech distribution and noise distribution, while SA-GOEA and SA-GUEA are adaptive to learn the distribution mask of noise.
• In this paper, we propose a model aggregation optimization weighting strategy that is more applicable to FLbased speech enhancement tasks.

Dependencies

• python >=3.6 (3.8.5 was used in the experiments)
• PyTorch == 1.10.0+cu113
• flwr == 2.0.1

Github:

View Github