Noise Power Spectral Density Estimation based on Maximum a Posteriori and Generalized Gamma Distribution

Noise power spectral density (PSD) estimation is a crucial part of speech enhancement system due to its contributory effect on the quality of the noise reduced speech. A novel estimation method for color noise PSD on the basis of an assumption of generalized Gamma distribution and maximum a posteriori (MAP) criterion is proposed. In the experiment, generalized Gamma PDF which is a natural extension of the Gaussian modeling of a nonwhite components distribution is found best fitting in four types of color noises compared with Laplace, Rayleigh distributions. After that MAP noise estimators based on the reported generalized Gamma PDF models are competed with Minimum Statistics (MS), minimum mean square error (MMSE) based PSD estimation and Maximum Likelihood estimation (MLE) noise tracking methods in evaluations. The performance of the proposed noise estimations are good as demonstrated by log error, segmental SNR and PESQ measures when they are integrated with the speech enhancement technique.