Blind score normalization method for PLDA based speaker recognition

TL;DR

This paper introduces a novel blind score normalization technique for PLDA-based speaker recognition that improves accuracy without requiring additional development data, especially when enrollment data varies.

Contribution

The paper proposes a new blind PLDA score normalization method that eliminates the need for extra data and enhances detection performance in variable enrollment scenarios.

Findings

Improved speaker recognition accuracy on NIST SRE 2014 data.

Normalization method is optimal for detection cost function.

No additional development data needed for normalization.

Abstract

Probabilistic Linear Discriminant Analysis (PLDA) has become state-of-the-art method for modeling $i$-vector space in speaker recognition task. However the performance degradation is observed if enrollment data size differs from one speaker to another. This paper presents a solution to such problem by introducing new PLDA scoring normalization technique. Normalization parameters are derived in a blind way, so that, unlike traditional \textit{ZT-norm}, no extra development data is required. Moreover, proposed method has shown to be optimal in terms of detection cost function. The experiments conducted on NIST SRE 2014 database demonstrate an improved accuracy in a mixed enrollment number condition.