Joint Training of Expanded End-to-End DNN for Text-Dependent Speaker Verification

We propose an expanded end-to-end DNN architecture for speaker verification based on b-vectors as well as d-vectors. We embedded the components of a speaker verification system such as modeling frame-level features, extracting utterance-level features, dimensionality reduction of utterance-level features, and trial-level scoring in an expanded end-to-end DNN architecture. The main contribution of this paper is that, instead of using DNNs as parts of the system trained independently, we train the whole system jointly with a fine-tune cost after pre-training each part. The experimental results show that the proposed system outperforms the baseline d-vector system and i-vector PLDA system.