Adaptation Strategy and Clustering from Scratch for New Domains of Speaker Recognition

This paper investigates the domain adaptation back-end methods introduced over the past years for speaker recognition, when the mismatch between training and test data induces a severe degradation of performance. The conducted analyses lead to suggest some ways, experimentally validated, for the task of collecting in-domain data and making the most of the first data at hand. The proposed strategy helps to quickly increase accuracy of the detection, without omitting to take into account the practical difficulties of the task of data collecting in real-life situations , and without the cost and delay for forming the expected large and speaker-labeled in-domain dataset. Moreover, a new approach of artificial speaker labeling by clustering is proposed, that dispenses of collecting a preliminary annotated in-domain dataset, with a similar gain of efficiency.