Maximizing Speech Denoising Effectiveness with CNNs: A Focus on Key Architectural Strategies

Audio signal denoising is important in speech enhancement. Most traditional methods cannot remove noises thoroughly and models proposed recently still achieve poor performance in denoising under low SNR conditions. This paper proposed a model whose speech denoising performance is higher than that of most models. The model is based on convolutional neural network and is added several strategies to improve its metrics. The importance of strategies like residual connection, batch normalization, Dropout regularization and "encoder-decoder" processing mode in the model is illustrated in the paper. The metrics used in the paper is MSE, SNR, PESQ and STOI. Studies show that the proposed model performs better in the four metrics and over-high or over-low Dropout ratios can affect the performance negatively.