Speech variability in automatic speaker recognition systems for forensic purposes

It is becoming increasingly usual to find audio physical traces (telephone calls, recorded tapes, security surveillance recordings, etc.) while committing crimes, forcing the speech research community to find reliable methods that allow the association of an unknown voice sample with a determined person identity. Regarding speech variability in forensic approaches, some of these variability sources highly degrade the speaker verification process, namely: channel influence, inter-session variability and emotional state. In this contribution, channel and inter-session variability are explored in order to accomplish real automatic systems for forensic speaker recognition. Results are presented making use of the ´GAUDVAHUMADA´ large speaker recognition-oriented database in Spanish.