Abstract

Speech is most likely the simplest and efficient type of human-human communication, as well as the most intuitive and effective way of human-machine interaction. Human voice is often damaged in real-world contexts by both reverberation and noise from the surroundings, which has a detrimental impact on speech intelligibility and quality. In terms of denoising, a model-based approach has been thoroughly researched, and several practical solutions have been created. In comparison, study on dereverberation has been sparse. Significant advances have been achieved in the study of a model-based strategy for dereverberation. The resultant approach may be used to any deep neural network that provides masks in the time-frequency domain with just a few extra variables that can be trained and an overhead of computation that is low for state-of-the-art neural networks. A deep learning-based approach in this article is developed that eliminates early reverberations, late reverberations, and noise from speech signals in order to enhance speech signal quality. The method is tested using data from three simulated rooms—a conference room, a seminar hall, and a room from reference paper number seven—with Reverberation Time (RT60) of 0.3 s and variety of noise like Additive White Gaussian Noise (AWGN), realistic noise such as babble, restaurant and a variety of signal-to-noise ratio values. The proposed technique outperforms baseline multichannel dereverberation and denoising algorithms as well as a cutting-edge multichannel dereverberation and denoising algorithm, resulting in a considerable improvement.