Evaluation of I-vector and GMM Based Speaker Verification Systems for Forensic Application

Abstract

In addition to visual and auditory evaluation, forensic audio experts use automatic speaker verification systems to analyze forensic voice evidence. Gaussian Mixture Model is widely used in such systems. Recently, the usage of i-vector, derived from factor analysis, has gained popularity. In this work, performance comparison of Gaussian Mixture Model and i-vector methods on a speaker verification system is conducted on short recordings of speech. The experimental results revealed that the system that uses i-vector produces a much lower equal error rate.