The MAP and cumulative distribution function equalization methods for the speech spectral estimation with application in noise suppression filtering

In this work we develop two statistical estimation methods of maximum a posterior probability (MAP) and cumulative distribution function equalization (CDFE) for the speech spectral component estimation approaches with the application in the noise suppression filters. In contrast to the histogram equalization approach, the CDFE is developed here based on speech and noise spectral modeling, which is also used in the MAP approach. Both of the conventional Gaussian and general gamma modeling of speech and noise spectral are investigated in this work. For the CDF estimation, we develop a flexible method for the non-Gaussian distribution by using the characteristic function. The advantage of proposed methods is that yields a flexible solution of the speech spectral estimation problem in general case of speech and noise modeling, which should be determined for each particular condition. Finally the systems of noise suppressed filters based on CDFE, MAP and MMSE are investigated via an experimental evaluation on the SNR improvement measurements. The performances of MAP and CDFE based systems are shown to be at least comparable or exceeded the conventional MMSE based in both the cases of Gaussian and gamma modeling. In the hearing test, the CDFE based system products a less musical noise level compared to the MAP and MMSE methods. . . .