The CLIR-CLSP System for the IberSPEECH-RTVE 2020 Speaker Diarization and Identity Assignment Challenge

Abstract

This paper describes the Speaker Diarization system jointly developed by the Computational Learning and Imaging Research (CLIR) laboratory of the Universidad Autónoma de Yucatán and the Center for Language and Speech Processing (CLSP) of the Johns Hopkins University for the Albayzin Speaker Diarization and Identity Assignment Challenge organized in the IberSPEECH 2020 conference. The Speaker Diarization system follows an x-vector-PLDA-VBx pipeline built with the Kaldi toolkit. It uses a Time Delay Neural Network (TDNN)-based Speech Activity Detector (SAD), with x-vectors as acoustic features, clustered with Agglomerative Hierarchical Clustering (AHC) as initialization for variational Bayes clustering. The system was only evaluated in the Speaker Diarization condition.