A New On-Line Model Quality Evaluation Method for Speaker Verification

Abstract

Abstract. The accurate selection of the utterances is very important to obtain right estimated speaker models in speaker verification. In this sense, it is important to determine the quality of the utterances and to establish a mechanism to automatically discard or accept them. In real-time speaker verification applications, it is decisive to obtain on-line measures to ask the speaker for more data if necessary. In this paper, we introduce a new on-line quality method based on a male and a female Universal Background Model (UBM). These two models act as a reference for new incoming utterances in order to decide if they can be used to estimate the speaker model or not. Text-dependent experiments have been carried out by using a telephonic multi-session database in Spanish. The database has been recorded by the authors and has 184 speakers.