14thAnnual Conference of the International Speech Communication Association

Lyon, France
August 25-29, 2013

Effective Estimation of a Multi-Session Speaker Model Using Information on Signal Parameters

Konstantin Simonchik, Andrey Shulipa, Timur Pekhovsky

Speech Technology Center Ltd., Russia

The paper deals with the problem of estimation an optimal i-vector based speaker voice model using several sessions of his or her voice recordings, each of which has different signal parameters: speech duration and SNR. Our aim is to minimize inter-session variability so as to achieve minimal EER in the task of speaker recognition. We examine the influence of the main signal parameters on intersession variability and propose a model for multi-session i-vector estimation based on minimizing inter-session variability.

Full Paper

Bibliographic reference.  Simonchik, Konstantin / Shulipa, Andrey / Pekhovsky, Timur (2013): "Effective estimation of a multi-session speaker model using information on signal parameters", In INTERSPEECH-2013, 1604-1608.