Using regularized Fisher Discriminant Analysis to improve the performance of Gaussian Supervector in session and device identification

Abstract

In this paper, we propose Regularized Fisher Discriminant Analysis (RFDA) as a projection method applied on Gaussian Supervector (GSV). GSV was originally applied on speaker recognition and verification, and has exhibited good performance. Recently GSV has also been applied in audio forensics area, such as recording device identification. It has been shown that GSV can also capture useful information related to the recording device. In this paper, we show that GSV can also be applied in telephone session identification. However, although GSV can capture useful information for different identification purposes, the performance of the raw GSV may not be so good. Thus, we apply RFDA-based projection method on the raw GSV, and find that this projection method can significantly improve the performance of the raw GSV, in both telephone session identification and recording device identification tasks.