Comparison between noise reduction algorithms for speech signals using auditory masking effects

Abstract

Auditory masking is aggressively exploited by algorithms used for the lossy compression of audio signals. In compression of audio signals, the intent is to hide the noise introduced by the coding below the masking threshold, thus making the noise inaudible. This will render the coding process transparent, enabling better compression without audible degradation of the signal. In this article, we show that using masking properties of the hearing system allows for improved noise reduction. A novel method for noise reduction in speech signals is proposed. This method is shown to outperform non-auditory based methods, and compares well with other perceptually motivated noise reduction methods. It is found that the proposed method, Soulodre's PNRF combined with the ITU's PEAQ auditory model, have more musical noise but less signal distortion than the method proposed by Tsoukalas, which obtain marginally better results in informal testing results.