[Paper Review] The SpeakIn System for VoxCeleb Speaker Recognition Challange 2021

Miao Zhao, Yufeng Ma+2 more|arXiv (Cornell University)|Sep 5, 2021

Speech Recognition and SynthesisComputer Science25 references36 citations

TL;DR

The SpeakIn team presents a VoxSRC 2021 system using offline and online data augmentation, RepVGG and ResNet backbones, MQMHA pooling, domain-based large-margin fine-tuning, and back-end calibration to achieve top performance with a 0.1034 minDCF and 1.846% EER on VoxSRC 2021 tracks 1 and 2.

ABSTRACT

This report describes our submission to the track 1 and track 2 of the VoxCeleb Speaker Recognition Challenge 2021 (VoxSRC 2021). Both track 1 and track 2 share the same speaker verification system, which only uses VoxCeleb2-dev as our training set. This report explores several parts, including data augmentation, network structures, domain-based large margin fine-tuning, and back-end refinement. Our system is a fusion of 9 models and achieves first place in these two tracks of VoxSRC 2021. The minDCF of our submission is 0.1034, and the corresponding EER is 1.8460%.

Motivation & Objective

• Motivate robust speaker verification for VoxSRC 2021 using VoxCeleb2-dev data only.
• Explore data augmentation strategies (offline and online) to expand training diversity.
• Develop novel backbones and pooling to improve utterance-level embeddings.
• Apply large-margin fine-tuning and score calibration to maximize verification performance.

Proposed method

• Use VoxCeleb2-dev as the sole training set for Tracks 1 and 2.
• Apply 3-fold speed augmentation plus offline Kaldi-based augmentation (RIRs, MUSAN) and an online augmentation chain.
• Experiment with RepVGG and ResNet backbones, combined with attention-based pooling (MQMHA).
• Enhance losses with subcenter AM/Arc variants and Inter-TopK penalties, plus domain-based large-margin fine-tuning.
• Incorporate AS-Norm and Quality Measure Functions (QMF) on a back-end with cosine scoring and length normalization.
• Fuse 9 sub-systems to obtain final VoxSRC 2021 submission results.

Experimental results

Research questions

• RQ1 How do offline and online data augmentation strategies affect speaker verification performance on VoxCeleb-derived datasets?
• RQ2 What is the impact of using RepVGG backbones and the MQMHA pooling scheme on embedding quality?
• RQ3 Can domain-based large-margin fine-tuning and back-end calibration (AS-Norm, QMF) optimally improve verification metrics under VoxSRC conditions?

Key findings

• Fusion of 9 models achieves first place on VoxSRC 2021Track 1 and Track 2.
• Final minDCF of 0.1034 and EER of 1.846% on VoxSRC 2021 evaluation.
• Sub-system analyses show larger models (e.g., RepVGG-B1, ResNet-101) outperform smaller baselines, but very large models may overfit on VoxCeleb2-dev.
• Online augmentation complements offline augmentation but offline models yielded strongest performance.
• Domain-based large-margin fine-tuning and AS-Norm/QMF calibration significantly reduce EER and minDCF compared to baseline.