Optimization of Threshold Selection for Wavelet-Based Denoising in Conference Scenarios
Duojia Li, Wenxiang Shen, Jun Wu · 2023
In this paper, a noise reduction scheme based on wavelet decomposition is advocated to solve the conundrum of smooth and non-smooth noise in conference scenes. Non-smooth noise is a difficult problem in the current speech processing of conference scenes because its noise characteristics vary with time, lack a priori knowledge and are difficult to model. To address this problem, we design a noise reduction scheme for non-smooth noise. A new optimization objective function is defined to optimize the threshold selection strategy for each layer. We use a genetic algorithm to optimize the objective function to obtain the optimal thresholds for each layer. Finally, we use wavelet thresholding noise reduction technique to reduce the noise of speech signals in conference scenes. The experimental results on the NOIZEUS corpus show that the scheme effectively removes the noise in the conference scenes. This study has practical implications for improving the quality of speech signals in conference scenes.