I am building a model using a supervised machine learning based on features I extract from speech signals.
The features include MFCC, auto correlation and energy derivatives.
According to this paper, a speaker-specific gaussinization can improve the model's performance. The normalization is based on an explicit normalization of feature distributions for each speaker prior to training, based on an additional sample of each speaker, as described in section 6 in the paper.
I do not fully understand the normalization process which is described in the paper, and the reference paper appears in the paper does not help as well.
Is anyone familiar with this kind of normalization of speech features? a clarification of this process will be highly appreciated.