Skip to main content
QUICK REVIEW

[论文解读] ReDimNet2: Scaling Speaker Verification via Time-Pooled Dimension Reshaping

Ivan Yakovlev, Anton Okhotnikov|arXiv (Cornell University)|Mar 12, 2026
Speech Recognition and SynthesisComputer Science被引用 0
一句话总结

ReDimNet2 在 ReDimNet 的 1D 处理路径中引入时间池化,使在相同计算预算下实现更激进的通道缩放和更好的人声验证精度。在 VoxCeleb1 基准测试的 7 个配置(B0–B6)上实现了前沿的效率-精度权衡。

ABSTRACT

We present ReDimNet2, an improved neural network architecture for extracting utterance-level speaker representations that builds upon the ReDimNet dimension-reshaping framework. The key modification in ReDimNet2 is the introduction of pooling over the time dimension within the 1D processing pathway. This operation preserves the nature of the 1D feature space, since 1D features remain a reshaped version of 2D features regardless of temporal resolution, while enabling significantly more aggressive scaling of the channel dimension without proportional compute increase. We introduce a family of seven model configurations (B0-B6) ranging from 1.1M to 12.3M parameters and 0.33 to 13 GMACS. Experimental results on VoxCeleb1 benchmarks demonstrate that ReDimNet2 improves the Pareto front of computational cost versus accuracy at every scale point compared to ReDimNet, achieving 0.287% EER on Vox1-O with 12.3M parameters and 13 GMACS.

研究动机与目标

  • Motivate scaling of dimension-reshaping architectures for speaker verification.
  • Propose time-pooling within the 1D processing pathway to unlock wider channel dimensions at fixed compute budgets.
  • Demonstrate improved accuracy at matched GMACs across a Pareto front over ReDimNet.
  • Show strong out-of-domain generalization and competitive efficiency relative to large SSL models.

提出的方法

  • Extend ReDimNet with time-pooling in the 1D pathway using the same strided conv2d layer used for frequency downsampling.
  • Maintain residual connections by nearest-neighbor upsampling at the aggregation point to align different temporal lengths.
  • Allow aggressive channel scaling (increase C) because time-pooling reduces compute in 1D and compacts 2D representations.
  • Evaluate seven configurations (B0–B6) spanning 0.33 to 13 GMACs and 1.1M to 12.3M parameters.
  • Train with VoxCeleb2 using a two-stage pipeline (pretraining with augmentation and large-margin finetuning).
  • Report EERs on VoxCeleb1 cleaned protocols (Vox1-O, Vox1-E, Vox1-H) and GMACs measured on 2-second inputs.

实验结果

研究问题

  • RQ1Does time-pooling in the 1D pathway enable more efficient scaling of the channel dimension without harming representation quality?
  • RQ2Can ReDimNet2 improve accuracy at equivalent compute budgets across the Pareto front compared to ReDimNet?
  • RQ3How does time-pooling affect out-of-domain generalization for speaker verification?
  • RQ4What is the impact of time-pooling on training stability and model size across small to large configurations?

主要发现

  • ReDimNet2 consistently achieves lower EER than the corresponding ReDimNet at matched compute budgets across all B0–B6 configurations.
  • ReDimNet2-B6 attains 0.29% EER on Vox1-O with 12.3M params and 13 GMACs, a 28% relative improvement over ReDimNet-B6.
  • Time-pooling enables more aggressive channel scaling, yielding better discriminability at the same GMACs.
  • Across out-of-domain tests (SITW, VOiCES, VoxCeleb1-B), ReDimNet2-B6 shows lower EER than ReDimNet-B6, indicating preserved generalization.
  • Compared to SSL-like large models, ReDimNet2 offers competitive accuracy with far fewer parameters (e.g., approaching WavLM/W2V-BERT 2.0 in some regimes while being substantially smaller).
  • Training stability is strong for smaller models but shows more variability for larger ones, suggesting extra regularization may help at scale.

阅读本论文

在 Nubint 阅读器中打开原文,边读边就任意段落提问。

在 Nubint 中阅读本论文

本解读由 AI 根据论文原文生成,引用前请核对原文。