Improved Adaptive Wavelet Thresholding Based on Decomposition Layers for Noise Reduction

In practical application scenarios, sound acts as a vital information carrier. However, the process of sound signal acquisition is frequently perturbed by environmental interferences, causing noise to infiltrate the acquired signals. Hence, noise reduction processing becomes a pivotal aspect of signal processing. The wavelet threshold denoising method is extensively employed in the sound noise reduction realm owing to its simplicity and practicability. In this paper, by integrating the frequency traits of sound information with the characteristics of wavelet decomposition, an enhanced threshold function predicated on the wavelet decomposition layer is devised. Additionally, denoising experiments are executed by appending Gaussian noise of diverse signal-to-noise ratios (SNRs) to pristine sound signals. Concurrently, denoising experiments on speech signals gathered from real scenarios are also performed for comparative analysis. Three evaluation indices, namely SNR, mean square error (MSE), and waveform similarity, are utilized to appraise the performance. Experimental outcomes evince that within specific Gaussian noise environments, in contrast to traditional hard and soft threshold methods, the signal processed by the improved threshold function witnesses an SNR elevation ranging from 10% to 20%, an MSE diminution within 30% - 50%, and a waveform similarity augmentation between 15% and 25%, forcefully demonstrating its preeminent noise reduction prowess.