An Evaluation Study on Speech Feature Densities for Bayesian Estimation in Robust ASR

Bayesian estimators, especially the Minimum Mean Square Error (MMSE) and the Maximum A Posteriori (MAP), are very popular in estimating the clean speech STFT coefficients. Recently, a similar trend has been successfully applied to speech feature enhancement for robust Automatic Speech/Speaker Recognition (ASR) applications either in the Mel, log-Mel or in the cepstral domain. It is a matter of fact that the goodness of the estimate directly depends on the assumptions made about the noise and speech coefficients densities. Nevertheless, while this latter has been exhaustively studied in the case of STFT coefficients, not equivalent attention has been paid to the case of speech features. In this paper, we study the distribution of Mel, log-Mel as well as MFCC coefficients obtained from speech segments. The histograms of the speech features are first fitted into several pdf models by means of the Chi-Square Goodness-of-Fit test, then they are modeled using a Gaussian Mixture Model (GMM). Performed computer simulations show that the choice of log-Mel and MFCC coefficients is more convenient w.r.t. the Mel one from this perspective.