A Novel Framework for Speech Signal Denoising using PSO Optimized ICA-DWT Algorithm

Abstract

Objectives: In this paper a hybrid approach of ICA-DWT algorithm optimized using Particle Swarm Optimization (PSO) is proposed to deal with problems of aperiodic and period noises in industrial noise environment. Method/analysis: The feature of Independent Component Analysis (ICA) for separating the signals of various channels is exploited to separate noise peaks from the speech channel. To reserve the original signal and discern the noise, the speech is segmented in various levels of frequencies via discrete wavelet transform. The adaptive filtration through wavelet filters has been a powerful tool for signal segmentation into various frequencies. The output of ICA is sourced to Discrete Wavelet Transform (DWT) and is optimized using PSO to generate threshold value and number of wavelets for it. Findings: The results indicate that additional overhead computation of DWT has a better Signal-to-Noise Ratio (SNR) value compared to clean fast ICA algorithm and thus validate the improvement in speech signal intelligibility and quality.For the range of input signals and noise environment, the optimization of PSO to filter the speech signals has best SNR compared to conventional algorithm. Novelty/ Improvement: In the proposed denoising model two stages optimized filtering is presented. The number of wavelet levels and value of threshold is depicted using objective function to minimize Spectral Noise Density (SND). The objective function is optimized with PSO in constraints of SND to generate the best possible levels of DWT and thus maximize the SNR ultimately.