Analyzing speaker verification embedding extractors and back-ends under language and channel mismatch

TL;DR

This paper investigates how speaker verification systems perform under language and channel mismatch, analyzing embedding architectures and back-end models, and proposing methods to improve robustness and scoring efficiency.

Contribution

It provides a comprehensive analysis of speaker embedding and back-end scoring under domain mismatch, introducing optimized configurations and scoring methods for better performance.

Findings

Reduced temporal stride improves embedding performance.

Language-dependent PLDA and nuisance projection significantly enhance accuracy.

Efficient scoring and fusion methods boost speaker verification results.

Abstract

In this paper, we analyze the behavior and performance of speaker embeddings and the back-end scoring model under domain and language mismatch. We present our findings regarding ResNet-based speaker embedding architectures and show that reduced temporal stride yields improved performance. We then consider a PLDA back-end and show how a combination of small speaker subspace, language-dependent PLDA mixture, and nuisance-attribute projection can have a drastic impact on the performance of the system. Besides, we present an efficient way of scoring and fusing class posterior logit vectors recently shown to perform well for speaker verification task. The experiments are performed using the NIST SRE 2021 setup.