ReDimNet2: Scaling Speaker Verification via Time-Pooled Dimension Reshaping

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.

Figures (2)

• Figure 1: Computational Cost vs. Average EER. EER is averaged over Vox1-O, Vox1-E, and Vox1-H protocols. Marker area is proportional to parameter count, color indicates model family. Dashed line: ReDimNet, solid line: ReDimNet2. GMACs measured on 2-second input.
• Figure 2: ReDimNet2 architecture overview for a single stage. Top: ReDimNet v1 processes 1d input through a to2d reshape operation, block2d (frequency stride only), a to1d reshape operation followed by a block1d, and stage-wise weighted aggregation ("stack and weight"), preserving the time dimension $T$ throughout. Bottom: ReDimNet2 applies an additional stride over the time axis inside the 2D block, halving both $F$ and $T$. The 1D block then operates on the shorter sequence ($T/2$), and an upsample-to-initial-$T$ step restores the original temporal resolution before aggregation.