README.md

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

• Test_trial: CNC-Eval-Avg.lst

• 🔥 UPDATE 2022.07.12: We update this recipe according to the setups in the winning system of CNSRC 2022, and get obvious performance improvement compared with the old recipe. Check the commit1, commit2 for details.

  • LR scheduler warmup from 0
  • Remove one embedding layer
  • Add large margin fine-tuning strategy (LM)

Model	Params	FLOPs	LM	AS-Norm	QMF	EER (%)	minDCF (p=0.01)
ResNet34-TSTP-emb256 (OLD)	6.70M	4.55 G	×	×	×	8.426	0.487
ResNet34-TSTP-emb256	6.63M	4.55 G	×	×	×	7.134	0.408
			×	√	×	6.747	0.367
			×	√	√	6.336	0.374
			√	×	×	6.652	0.393
			√	√	×	6.492	0.354
			√	√	√	6.119	0.361
ResNet221-TSTP-emb256	23.86M	21.29 G	×	×	×	5.965	0.362
			×	√	×	5.708	0.326
			√	×	×	5.886	0.362
			√	√	×	5.655	0.330
ECAPA_TDNN_GLOB_c512-ASTP-emb192	6.19M	1.04 G	×	×	×	8.313	0.432
			×	√	×	7.644	0.390
			√	×	×	8.004	0.422
			√	√	×	7.417	0.379
ECAPA_TDNN_GLOB_c1024-ASTP-emb192	14.65M	2.65 G	×	×	×	7.879	0.420
			×	√	×	7.412	0.379
			√	×	×	7.986	0.417
			√	√	×	7.395	0.372
RepVGG_TINY_A0	6.26M	4.65 G	×	×	×	6.883	0.399
			×	√	×	6.550	0.355