AWA Long-Term Recorded Speech Corpus And Robust Speaker Recognition Method For Session Variability

Abstract

Session variability is one of the most important issues in the speaker recognition technology. On the other hand, our scientific interest lies in how individual voice changes as time progresses and where the limit of the changes. From these motivations, we have been constructing “AWA Long-Term Recorded speech corpus (AWA-LTR)” that contains one's same content speech recorded at morning, noon, and evening once a week for over 10 years using the same microphone in a soundproof chamber. AWA-LTR first version has been released by Speech Resources Consortium, National Institute of Informatics (NII-SRC), Japan in 2012. In addition, we will release AWA-LTR second version in 2018. Hence, in this paper, we describe the details of AWA-LTR and the data release schedule of this corpus. As an effective application example using the corpus, we propose a robust speaker recognition method for session variability and evaluate the proposed method by the speaker identification experiment in this paper.