Spectro-temporal features with distribution equalization

We could show in the past that Hierarchical SpectroTemporal (HIST) features improve the performance of Automatic Recognition Systems (ARS) of speech in difficult environments when they are combined with conventional speech spectral features. The target here is to improve the noise robustness of the HIST features by investigating a channel distribution equalization in our feature hierarchy. Thereby, we determine the empirical cumulative distribution of the speech training data set, which is referred to as reference distribution. Afterwards, a distribution adjustment of the training as well as test data is performed with respect to the reference distribution. We carry out the above mentioned distribution equalization in the preprocessing step as well as after each feature extraction step of our HIST feature extraction framework. We evaluate the benefits of such an equalization in the HIST feature extraction process with different noise types.