SERA Noise Estimation for Speech Enhancement

Abstract

A practical single channel speech enhancement system consists of two major components, estimation of noise power spectrum and the estimation of speech. Therefore, a crucial component of any algorithm is the estimation of the noise power spectrum for highly non stationary noise environments. The performance of noise estimation algorithm is usually a tradeoff between speech distortion and noise reduction. In existing methods, noise is estimated only during speech pauses and these pauses are identified using Voice Activity Detector (VAD). This paper describes novel noise estimation method SERA (Spectral Entropy Recursive Averaging) to estimate noise in highly non stationary noise environments. In SERA, noise estimation is updated in both speech pauses and also speech present frames. Speech presence is determined by computing the ratio of the noisy speech power spectrum to its local minimum, which is computed by averaging past values of the noisy speech power spectra with a look-ahead factor. Environmental noise is present in all the frames of the noisy speech signal and if the speech/silence detection is not accurate, then it yields speech echoes and residual noise in the enhanced speech. In this paper, noise estimation is updated by dividing speech signal into pure speech, quasi speech and non-speech frames based on adaptive multiple thresholds without using of VAD. The proposed method is compared with weighted average noise estimation method in terms of segmental SNR. The simulation results of the proposed algorithm shows better performance over a system that uses VAD in noise estimation.