The ins and outs of speaker recognition: lessons from VoxSRC 2020

TL;DR

This paper explores robust speaker recognition in challenging environments using ResNet variants, optimizing training for efficiency, and sharing lessons learned from the VoxSRC 2020 challenge.

Contribution

It introduces an efficient training framework for ResNet-based speaker recognition models and provides insights from the VoxSRC 2020 challenge.

Findings

Models outperform existing works with lighter architectures

Optimized training framework reduces resource requirements

Lessons learned improve robustness in challenging conditions

Abstract

The VoxCeleb Speaker Recognition Challenge (VoxSRC) at Interspeech 2020 offers a challenging evaluation for speaker recognition systems, which includes celebrities playing different parts in movies. The goal of this work is robust speaker recognition of utterances recorded in these challenging environments. We utilise variants of the popular ResNet architecture for speaker recognition and perform extensive experiments using a range of loss functions and training parameters. To this end, we optimise an efficient training framework that allows powerful models to be trained with limited time and resources. Our trained models demonstrate improvements over most existing works with lighter models and a simple pipeline. The paper shares the lessons learned from our participation in the challenge.