Robust spectro-temporal speech features with model-based distribution equalization

Previously, we applied a distribution equalization on our HIerarchical Spectro-Temporal (HIST) features using distributions estimated from histogram of one or several utterances. Although a performance increase could be observed in both cases, we noticed low performance improvement when estimating the distribution only from one utterance. The aim here is to determine a parametric distribution from few data samples which gives the highest probability of producing the observed data considering different models. Afterwards, we perform a distribution equalization based on the estimated model after each feature extraction step of our HIST feature extraction framework. We compare the performance of the HIST features with those of conventional spectral features (RASTA-PLP), when a corresponding distribution equalization has been carried out using the TIDigits database and different noise types. By combining the HIST and RASTA-PLP features we achieved on the one hand better performance than spectral features (RASTA-PLP), on the other hand better performance than the same feature combination with distribution equalization obtained directly from the histogram without imposing a parametric distribution model.