I-Vectors for Forensic Automatic Speaker Recognition

Forensic Automatic Speaker Recognition (FASR) will determine whether a voice recording (trace) in query belongs to a particular individual (suspected speaker). The aftereffect of voice disguise, Voice over IP (VoIP), and channel mismatches are important issues to be addressed in FASR. I-vectors have recently gained prominence in text-independent speaker verification, which can cope well with the above mentioned variability in speech. For this work, speech from 18 speakers through close speaking microphone, masked microphone, VoIP and telephone channel is collected to create a database. Utilizing this database, a speaker recognition framework using I-vectors is implemented and evaluated.