A joint factor analysis model for handling mismatched recording conditions in forensic automatic speaker recognition

摘要

In forensics automatic speaker recognition (FASR), one of the most important factors that degrades its performance is the mismatch in recording conditions (session variability). Recently, joint factor analysis (JFA) combined with Gaussian mixture model (GMM) has become the state-of-the-art technique to cope with session variability in speaker recognition. Its ability relies on accurate estimation of session variability subspace for the operating conditions of interest. This paper integrates JFA into evaluation of the strength of evidence in FASR and analyzes the performance of JFA in simulated forensic caseworks where mismatch appears. It also investigates a JFA based compensation technique to cope with the mismatch in telephone transmission conditions. Experiments on the Polyphone IPSC-03 database demonstrate that such a compensation method improves performance of FASR.