Approaches to Codec Independent Speaker Identification in Voip Speech

Abstract

The performance of automatic speaker identification (ASI) systems on Voice over Internet Protocol (VoIP) speech varies with the type of codec used in the VoIP communication. The type of codec used depends on the service provider of the user. Thus there is a need for the codec-independent ASI systems to identify the speaker. Three modeling approaches based on UBM-GMM framework and i-vector framework are proposed to identify the speaker independent of codec used. These frameworks are also evaluated for the mismatch conditions with respect to the codec used in training and testing. The proposed approaches are evaluated on VoIP speech from four codecs with different bit rates along with uncoded speech.