Multi-condition Gaussian probabilistic linear discriminate analysis in automatic speaker recognition

Automatic Speaker Recognition Systems (ASRSs) are being used in forensic investigations to help experts to appreciate the measure of confidence related to real vocal comparison cases. Accuracy is an important aspect in such kind of applications where the existing recognition approaches are known to be sensitive to recording conditions mismatch, the utterances lengths mismatch, and the linguistic variation effects. This paper proposes a recognition strategy based on the use of Multi-condition Gaussian Probabilistic Linear Discriminate Analysis (MGPLDA) to compensate the effects of this type of perturbation sources. Experiments using an Algerian Speaker Recognition Database collected in different recording conditions have shown that the proposed approach successfully improves the system performances in terms of discrimination and calibration quality.