Good Noise Power Estimators Are Not Always Good

Abstract:
Methods for estimating the noise power in noisy speech signals have been developed and explored for many years. Thus, there are some good state-of-the-art noise power estimators, which were successfully tested on various noise types and different SNR conditions. Hence, a meeting scenario in an environment with high signal-to-noise ratio (SNR) of some close-talk speech recordings, where also some stationary sensor noise has to be estimated, should not pose a challenging problem for good noise estimators. However, we observed that high SNR conditions, as well as multi-talk situations can mislead state-of-the-art noise power estimators, and sometimes a more simple approach is to be preferred. In this contribution we present an analysis of three different noise power estimators when used in a meeting scenario with interfering speakers.