Unbiased Noise Estimator for Q-Spectral Subtraction based Speech Enhancement

摘要

We present a technique for denoising speech using a modification of spectral subtraction method based on Tsallis statistics, q-SS. This technique works by assuming that noise power spectra are unknown but could be estimate. Since the target and interfering signal are not available, power spectra of both signal must be esimate from the mixed noisy signal. In this paper, we estimate the noise power spectra under MMSE criteria. Implementation of soft decision speech presence probability with fixed prior ensure that the estimator to be unbiased. Our experiment showed that the unbiased estimator signicantly improve q-SS filter performance.