ResNet34 Result

Training Config

• Feature info: using 80 dims fbank, no cmvn, speed perturb(0.9, 1.0, 1.1)
• Train info: lr 1e-4, batch_size 64, 1 gpu(Tesla V100), acc_grad 1, 300000 steps, clip_gradient_norm 3.0, weight_l2_regularizer 0.01
• Loss info: additive angular margin softmax, feature_scaling_factor=8, margin 0.25
• Model info: ResNet34, global statistics pooling, Dense
• Train config: conf/train_sv_resnet34.yaml
• Model size: 5.60 M parameters

Results (EER & minDCF)

• Test set: Alimeeting-test, CN-Celeb-eval-speech