DCT-BASED DENOISING of SPEECH SIGNALS

Abstract

This paper considers a traditional task of additive white Gaussian noise removal from speech signals. An opportunity to use denoising based on discrete cosine transform is investigated. Both conventional signal-to-noise ratio and PECS metric are used in analysis of filtering efficiency. It is shown that, for the analyzed test signals, it depends on, at least, three factors: input signal-to-noise ratio, type of threshold, and parameter β used in threshold setting. The main observations are the following: 1) improvement of signal-noise ratio due to denoising is the largest for low input SNR; 2) there are optimal values of β that have the tendency to increase if input SNR decreases; 3) the combined threshold that is the first time tested for speech signals performs better than the hard threshold. The directions of further studies are discussed.