Speech enhancement based on the structure of noise power spectral density

This paper proposes an adaptive averaging periodogram (AAP) spectral estimator based on the structure of noise power spectral density (NPSD) for speech enhancement, which will be herein referred to as NPSD-AAP. In the proposed spectral estimator, both the raw periodogram and the NPSD are smoothed over frequency to reduce their variances if the NPSD has a relatively flat spectrum. Otherwise no smoothing is performed so as to satisfy the high-frequency resolution demand for the non-flat spectrum of the NPSD. The NPSD-AAP provides a low-variance and adaptive-bandwidth estimate of the power spectral density, which could be applied to any frequency-domain speech enhancement algorithms. Especially, the NPSD-AAP is applied to spectral subtraction to suppress the musical noise without introducing audible speech distortion. Experimental results confirm the validity of the proposed algorithm.