Forensic evidence reporting using GMM-UBM, JFA and I-vector methods: Application to Algerian Arabic dialect

摘要

Nowadays, under controlled conditions the speaker verification systems based on the GMM-UBM paradigm show very good performance. However, in forensic investigation activities the conditions; in which recordings are acquired; are uncontrollable, a naive use of the baseline GMM-UBM system without feature normalization, model transformation and score normalization techniques yields to unreliable forensic reporting. In this paper, we investigate forensic reporting using corpus-based likelihood ratio evaluation; which gained popularity in recent years; using two state-of-the-art speaker recognition systems: The JFA system which models explicitly the speaker and session variability during training stage and the I-vector paradigm which models the total variability and use compensation techniques to handle session mismatch. The GMM-UBM, Joint Factor Analysis and I-vector systems are compared in verification performance using Half Total Error Rates (HTER) and in forensic reporting using TIPPET plots. Experimental results on an Algerian Arabic dialect under different telephonic recording conditions confirm the robustness of I-vector and JFA systems in handling cross-channel mismatch and highlight clearly the drastic deterioration of the performance of the GMM-UBM system.