An Insight Into the Rise Time of Exponential Smoothing for Speech Enhancement Methods

Abstract

Exponential smoothing is a widely used averaging function to estimate the speech and noise statistics. However, the setting of the smoothing constant has been inconsistent or to a certain extent arbitrarily set. This paper aims to fill the gap by formulating the smoothing constant as a function of rise time to better reflect the variability of the signal to be smoothed. Experimental results with real world noise reveal that the performance is very sensitive to the rise time of the short term averaging function, whilst less so for the longer term averaging function. The results provide a guideline for speech enhancement methods to set the smoothing constant for the estimation of speech and noise statistics.