ReDimNet2: Scaling Speaker Verification via Time-Pooled Dimension Reshaping

This is an official implementation of a neural network architecture presented in the paper: ReDimNet2: Scaling Speaker Verification via Time-Pooled Dimension Reshaping.

Overview

ReDimNet2 extends the ReDimNet dimension-reshaping framework by introducing pooling over the time dimension in the 1D processing pathway. This allows significantly more aggressive channel scaling without proportional compute increase, yielding a strictly better accuracy–efficiency Pareto front at every scale point.

Results

Trained on VoxCeleb2-dev

EER (%) on VoxCeleb1 cleaned protocols:

Model	Params	GMACs	Vox1-O	Vox1-E	Vox1-H
ReDimNet2-B0	1.1M	0.33	1.04	1.16	1.97
ReDimNet2-B1	2.1M	0.56	0.78	0.96	1.72
ReDimNet2-B2	3.6M	0.95	0.57	0.76	1.41
ReDimNet2-B3	4.1M	2.70	0.42	0.66	1.22
ReDimNet2-B4	6.6M	4.62	0.37	0.58	1.07
ReDimNet2-B5	8.9M	9.62	0.33	0.56	1.07
ReDimNet2-B6	12.3M	13.05	0.29	0.52	0.99

Trained on VoxBlink2-based mixtures

EER (%) on VoxCeleb1 cleaned protocols and out-of-domain cross-dataset benchmarks:

Model	Train set	Weights	Vox1-O	Vox1-E	Vox1-H	Vox1-B	SITW	VOICES	CN-Celeb
b6	vb2+vox2	lm	0.23	0.35	0.67	0.89	0.74	2.55	7.49
b6	vb2+vox2	ptn	0.29	0.50	0.94	1.27	1.01	3.25	8.50
b3	vb2+vox2+cnc2	lm	0.51	0.60	1.14	1.72	1.25	3.87	4.78
b3	vb2+vox2+cnc2	ptn	0.69	0.82	1.47	2.16	1.91	5.96	5.29
b6	vb2+vox2+cnc2	lm	0.26	0.41	0.79	1.15	0.87	3.03	3.59
b6	vb2+vox2+cnc2	ptn	0.42	0.58	1.07	1.51	1.15	4.30	4.32

Quick Start

Via torch.hub

import torch
import torchaudio

# VoxCeleb2-only variant (default)
model = torch.hub.load("PalabraAI/redimnet2", "redimnet2",
                      model_name="b6", train_type="lm", pretrained=True)

# VoxBlink2 + VoxCeleb2 + CN-Celeb variant
model = torch.hub.load("PalabraAI/redimnet2", "redimnet2",
                      model_name="b6", train_type="lm",
                      dataset="vb2+vox2+cnc2_v0", pretrained=True)
model.eval()

waveform, sr = torchaudio.load("audio.wav")
# waveform: (batch, samples), 16 kHz
with torch.no_grad():
    embedding = model(waveform)  # (batch, emb_dim)

Available model names: b0, b1, b2, b3, b4, b5, b6. Available train types: ptn (pretraining), lm (large-margin fine-tuning, recommended), and dis (distilled; available for B6 vb2+vox2_v0). Available datasets: vox2 (default), vb2+vox2_v0 (B6), vb2+vox2+cnc2_v0 (B3, B6).

Model Zoo

Pretrained weights are available as GitHub Release assets.

model size	train set	ptn weights	lm weights
b0	vox2	b0-vox2-ptn.pt	b0-vox2-lm.pt
b1	vox2	b1-vox2-ptn.pt	b1-vox2-lm.pt
b2	vox2	b2-vox2-ptn.pt	b2-vox2-lm.pt
b3	vox2	b3-vox2-ptn.pt	b3-vox2-lm.pt
b4	vox2	b4-vox2-ptn.pt	b4-vox2-lm.pt
b5	vox2	b5-vox2-ptn.pt	b5-vox2-lm.pt
b6	vox2	b6-vox2-ptn.pt	b6-vox2-lm.pt
b6	vb2+vox2	b6-vb2+vox2_v0-ptn.pt	b6-vb2+vox2_v0-lm.pt
b3	vb2+vox2+cnc2	b3-vb2+vox2+cnc2_v0-ptn.pt	b3-vb2+vox2+cnc2_v0-lm.pt
b6	vb2+vox2+cnc2	b6-vb2+vox2+cnc2_v0-ptn.pt	b6-vb2+vox2+cnc2_v0-lm.pt

Roadmap

We are actively working on improving model stability and expanding the release. Planned updates:

• Release a new version of ReDimNet2 with improved normalization unified for both 1D and 2D feature map representations (agg_gnorm + dual_agg options, first used by the vb2+vox2_v0 B6 release)
• Release the training pipeline (VoxCeleb2)
• Release VoxBlink2 + VoxCeleb2 pretrained model (B6)
• Release VoxBlink2 + VoxCeleb2 + CN-Celeb2 pretrained models (B3, B6)
• Release training recipes for VoxBlink2 + VoxCeleb2 + CN-Celeb2 training

Citation

@inproceedings{redimnet2_2026,
  title={ReDimNet2: Scaling Speaker Verification via Time-Pooled Dimension Reshaping},
  author={Yakovlev, Ivan and Okhotnikov, Anton},
  booktitle={Proceedings of Interspeech 2026},
  year={2026},
  note={Accepted to Interspeech 2026},
  eprint={2603.11841},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  url={https://arxiv.org/abs/2603.11841}
}

License

MIT