Exploring the Use of an Unsupervised Autoregressive Model as a Shared Encoder for Text-Dependent Speaker Verification

In this paper, we propose a novel way of addressing text-dependent automatic\nspeaker verification (TD-ASV) by using a shared-encoder with task-specific\ndecoders. An autoregressive predictive coding (APC) encoder is pre-trained in\nan unsupervised manner using both out-of-domain (LibriSpeech, VoxCeleb) and\nin-domain (DeepMine) unlabeled datasets to learn generic, high-level feature\nrepresentation that encapsulates speaker and phonetic content. Two\ntask-specific decoders were trained using labeled datasets to classify speakers\n(SID) and phrases (PID). Speaker embeddings extracted from the SID decoder were\nscored using a PLDA. SID and PID systems were fused at the score level. There\nis a 51.9% relative improvement in minDCF for our system compared to the fully\nsupervised x-vector baseline on the cross-lingual DeepMine dataset. However,\nthe i-vector/HMM method outperformed the proposed APC encoder-decoder system. A\nfusion of the x-vector/PLDA baseline and the SID/PLDA scores prior to PID\nfusion further improved performance by 15% indicating complementarity of the\nproposed approach to the x-vector system. We show that the proposed approach\ncan leverage from large, unlabeled, data-rich domains, and learn speech\npatterns independent of downstream tasks. Such a system can provide competitive\nperformance in domain-mismatched scenarios where test data is from data-scarce\ndomains.\n