Noise power spectrum estimation using constrained variance spectral smoothing and minima tracking

In this paper, we propose a new noise estimation algorithm based on tracking the minima of an adaptively smoothed noisy short-time power spectrum (STPS). The heart of the proposed algorithm is a constrained variance smoothing (CVS) filter, which smoothes the noisy STPS independently of the noise level. The proposed smoothing procedure is capable of tracking the non-stationary behavior of the noisy STPS while reducing its variance. The minima of the smoothed STPS are tracked with a low delay and are used to construct voice activity detectors (VAD) in frequency bins. Finally, the noise power spectrum is estimated by averaging the noisy STPS on the noise-only regions. Experiments show that the proposed noise estimation algorithm possesses a very short delay in tracking the non-stationary behavior of the noise. When the proposed algorithm is utilized in a noise reduction system, it exhibits superior performance over the other recently proposed noise estimation algorithms.