Results

• Setup: fbank80, num_frms200, epoch150, ArcMargin, aug_prob0.6, speed_perturb (no spec_aug)
• Scoring: cosine (sub mean of vox2_dev), AS-Norm, QMF
• Metric: EER(%)
• 🔥 UPDATE 2024.09.03: We support the SimAM_ResNet pretrained on VoxBlink2 and Finetuned on Voxceleb2!
• 🔥 UPDATE 2024.08.27: We support SSL models as the feature front-end, take a look at the WavLM recipe!
• UPDATE 2022.07.19: We apply the same setups as the winning system of CNSRC 2022 (see cnceleb recipe for details), and obtain significant performance improvement.
  • LR scheduler warmup from 0
  • Remove one embedding layer in ResNet models
  • Add large margin fine-tuning strategy (LM)

Model	Params	Flops	LM	AS-Norm	QMF	vox1-O-clean	vox1-E-clean	vox1-H-clean
XVEC-TSTP-emb512	4.61M	0.53G	×	×	×	1.989	1.950	3.412
			×	√	×	1.834	1.846	3.124
			√	×	×	1.749	1.721	2.944
			√	√	×	1.590	1.641	2.726
ECAPA_TDNN_GLOB_c512-ASTP-emb192	6.19M	1.04G	×	×	×	1.069	1.209	2.310
			×	√	×	0.957	1.128	2.105
			√	×	×	0.878	1.072	2.007
			√	√	×	0.782	1.005	1.824
ECAPA_TDNN_GLOB_c1024-ASTP-emb192	14.65M	2.65G	×	×	×	0.856	1.072	2.059
			×	√	×	0.808	0.990	1.874
			√	×	×	0.798	0.993	1.883
			√	√	×	0.728	0.929	1.721
			√	√	√	0.707	0.894	1.615
ResNet34-TSTP-emb256	6.63M	4.55G	×	×	×	0.867	1.049	1.959
			×	√	×	0.787	0.964	1.726
			×	√	√	0.718	0.911	1.606
			√	×	×	0.797	0.937	1.695
			√	√	×	0.723	0.867	1.532
			√	√	√	0.659	0.821	1.437
ResNet221-TSTP-emb256	23.79M	21.29G	×	×	×	0.569	0.774	1.464
			×	√	×	0.479	0.707	1.290
			√	×	×	0.580	0.729	1.351
			√	√	×	0.505	0.676	1.213
ResNet293-TSTP-emb256	28.62M	28.10G	×	×	×	0.595	0.756	1.433
			×	√	×	0.537	0.701	1.276
			√	×	×	0.532	0.707	1.311
			√	√	×	0.447	0.657	1.183
			√	√	√	0.425	0.641	1.146
RepVGG_TINY_A0	6.26M	4.65G	×	×	×	0.909	1.034	1.943
			×	√	×	0.824	0.953	1.709
CAM++	7.18M	1.15G	×	×	×	0.803	0.932	1.860
			×	√	×	0.718	0.879	1.735
			√	x	×	0.707	0.845	1.664
			√	√	×	0.659	0.803	1.569
ERes2Net34_Base	7.88M	3.43G	×	×	×	0.914	1.065	1.986
			×	√	×	0.803	0.976	1.787
			√	x	×	0.824	0.968	1.776
			√	√	×	0.744	0.896	1.603
Res2Net34_Base	4.68M	1.77G	×	×	×	1.351	1.347	2.478
			×	√	×	1.234	1.232	2.162
Gemini_DFResNet114	6.53M	5.42G	×	×	×	0.787	0.963	1.760
			×	√	×	0.707	0.889	1.546
			√	x	×	0.771	0.906	1.599
			√	√	×	0.638	0.839	1.427
SimAM_ResNet34 (VoxBlink2 Pretrain)	25.2M		√	x	×	0.415	0.615	1.121
			√	√	×	0.372	0.581	1.049
			√	√	√	0.372	0.559	0.997
SimAM_ResNet100 (VoxBlink2 Pretrain)	50.2M		√	x	×	0.229	0.458	0.868
			√	√	×	0.207	0.424	0.804
			√	√	√	0.202	0.421	0.795

PLDA results

If you are interested in the PLDA scoring (which is inferior to the simple cosine scoring under the margin based setting), simply run:

local/score_plda.sh --stage 1 --stop-stage 3 --exp_dir exp_name

The results on ResNet34 (large margin, no asnorm) are:

Scoring method	vox1-O-clean	vox1-E-clean	vox1-H-clean
PLDA	1.207	1.350	2.528

WavLM results

• Pre-trained frontend: the WavLM Large model, multilayer features are used
• Speaker model: ECAPA_TDNN_GLOB_c512-ASTP-emb192
• Training strategy: Frozen => Joint ft => Joint lmft

bash run_wavlm.sh --stage 3 --stop_stage 9

Training strategy	AS-Norm	QMF	vox1-O-clean	vox1-E-clean	vox1-H-clean
Frozen	×	×	0.595	0.719	1.501
	√	×	0.548	0.656	1.355
	√	√	0.489	0.619	1.224
Frozen => Joint ft	×	×	0.542	0.635	1.355
	√	×	0.521	0.594	1.237
	√	√	0.494	0.576	1.205
Frozen => Joint ft => Joint lmft	×	×	0.521	0.626	1.344
	√	×	0.495	0.588	1.247
	√	√	0.415	0.551	1.118