EMD-Based Method for Windowed Filtering of Speech Signals

Speech signal filtering is used to increase speech intelligibility and reduce noise therein. In practice, speech signals are noisy to some extent, and depending on its intensity, noise can significantly distort informative speech parameters. The article presents a novel method for windowed filtering of speech signals based on the improved complete ensemble empirical mode decomposition with adaptive noise. The method uses segmentation of noisy speech into fragments, their decomposition into intrinsic mode functions (modes), formation of a set of mode signals, assessment of residual noise, and detection of the filtered signal. Formation of mode signals is a combination of noisy and informative empirical modes containing minimum noise fraction and a maximum of useful information. The research into the proposed method was carried out on speech signals with white, pink, and brown noises at a signal-to-noise ratio of -5 to 15 dB with a step of 5 dB. The research results have evidenced that the proposed method provides favorable effects of windowed filtering of speech signals, including those at low signal-to-noise ratio values, and can be effectively tested on real noisy speech signals in noise suppression systems.