ERes2NetV2: Boosting Short-Duration Speaker Verification Performance with Computational Efficiency

TL;DR

This paper introduces ERes2NetV2, a computationally efficient model that significantly improves short-duration speaker verification accuracy by enhanced feature fusion and pruning techniques, outperforming previous models on VoxCeleb datasets.

Contribution

The paper proposes ERes2NetV2, an improved version of ERes2Net with expanded channels and pruning, achieving better short-duration speaker verification performance with reduced complexity.

Findings

Achieves 0.98% EER on 3s trials

Reduces model parameters and computational cost

Outperforms previous models on VoxCeleb datasets

Abstract

Speaker verification systems experience significant performance degradation when tasked with short-duration trial recordings. To address this challenge, a multi-scale feature fusion approach has been proposed to effectively capture speaker characteristics from short utterances. Constrained by the model's size, a robust backbone Enhanced Res2Net (ERes2Net) combining global and local feature fusion demonstrates sub-optimal performance in short-duration speaker verification. To further improve the short-duration feature extraction capability of ERes2Net, we expand the channel dimension within each stage. However, this modification also increases the number of model parameters and computational complexity. To alleviate this problem, we propose an improved ERes2NetV2 by pruning redundant structures, ultimately reducing both the model parameters and its computational cost. A range of experiments conducted on the VoxCeleb datasets exhibits the superiority of ERes2NetV2, which achieves EER of 0.61% for the full-duration trial, 0.98% for the 3s-duration trial, and 1.48% for the 2s-duration trial on VoxCeleb1-O, respectively.