The most critical and difficult problem in speech analysis is reliable discrimination among Silence, Unvoiced and Voiced speech. Several methods have been proposed for making this three levels decision and most of them need Speech Activity Detection (SAD). In this study, we propose the Estimated Degree of Noise (EDON) to adjust the threshold of speech activity. To estimate the degree of noise, a function was previously prepared using the least-squares (LS) method, from the given (true) DON and the estimated parameter of DON. This parameter is obtained from the Auto-Correlation Function (ACF) of the noisy speech on a frame basis. Issues associated with this EDON for SAD approach are discussed, and experiments are done using the TIMIT database. Experimental result shows that using EDON improves the classification performance specially voiced and silent parts and the efficiency is compared with other existing published algorithms.