Noise estimation using an MVDR-like approach for acoustic signal enhancement

In this paper, we present a novel algorithm to accurately estimate time-variant noises without signal activity detection for acoustic signal enhancement. This is the first algorithm that does not require nor assume noise-only at the beginnings of recordings. This is the first algorithm that can directly estimate noises during the signal activity periods instead of by smoothing noises from neighbouring noise-only periods. To do so, we propose a minimum variance based approach along the time to estimate time-variant noises in spectral domain. The main advantages of the proposed algorithm over minimum statistics are the smoothness of the estimated noise spectra and the robustness to the analysis window length. Without the assumption of noise-only at the beginning of a recording, the proposed algorithm can be applied to speech enhancement and beyond. Experimental results show that the proposed algorithm can estimate time-varying noises accurately and speech enhancement algorithms using the proposed noise estimator perform better than their counterparts using a VAD-based noise estimator.