Neural Discriminant Analysis for Deep Speaker Embedding

TL;DR

This paper introduces Neural Discriminant Analysis (NDA), a nonlinear extension of PLDA using invertible neural networks, which better models non-Gaussian speaker embeddings and improves recognition performance.

Contribution

The paper proposes NDA, a novel nonlinear discriminant analysis method employing invertible neural networks to handle non-Gaussian data in speaker recognition.

Findings

NDA outperforms PLDA on speaker recognition datasets.

NDA effectively models non-Gaussian distributions of speaker embeddings.

Experimental results show improved accuracy with NDA.

Abstract

Probabilistic Linear Discriminant Analysis (PLDA) is a popular tool in open-set classification/verification tasks. However, the Gaussian assumption underlying PLDA prevents it from being applied to situations where the data is clearly non-Gaussian. In this paper, we present a novel nonlinear version of PLDA named as Neural Discriminant Analysis (NDA). This model employs an invertible deep neural network to transform a complex distribution to a simple Gaussian, so that the linear Gaussian model can be readily established in the transformed space. We tested this NDA model on a speaker recognition task where the deep speaker vectors (x-vectors) are presumably non-Gaussian. Experimental results on two datasets demonstrate that NDA consistently outperforms PLDA, by handling the non-Gaussian distributions of the x-vectors.