Discriminative Capacity and Phonetic Information of Bottleneck Features in Speech

The impressive gain in performance obtained using deep neural network (DNN) for automatic speech recognition have motivated their application to other speech related tasks such as speaker recognition and language recognition, but there is still uncertainty about what is deep training strategy extracting, from the acoustic data, to make it such a powerful learning tool. This paper compares the discriminative capacity and the phonetic information conveyed by the feature-space maximum likelihood linear regression (fMLLR) before and after passing through a DNN trained to discriminate between tri-phone tied states. The proposed experimentation reflected the superiority of DNN bottleneck features regarding its information content.