A Divide-and-Conquer Noise Classification and Collaborative Noise Reduction for Speech Enhancement
Weiguang Liu, Zhaoji Ye, Bin Qin, Chenjia Li · 2025
In this paper, a novel Noise Classification and Collaborative Noise Reduction (NCCN) framework is proposed to address the challenges of noise diversity and dynamic speech enhancement. The system integrates a lightweight convolutional neural network (CNN)-based noise classifier with an encoder-decoder-based denoising network, leveraging the NOISE-92 and THCHS30 datasets for rigorous evaluation. The noise classifier identifies 15 common environmental noise types with 100% accuracy, enabling noise-specific denoising strategies. The denoising network employs residual learning and skip connections to enhance speech quality. In theoretical conditions, the system achieves an optimal noise energy reduction of approximately 30 dB.