Hyperion Toolkit Speaker Verification pre-trained Model
Model Configuration
This model was trained using recipe voxceleb/v1.1
The configuration for this modeis is defined in config_fbank80_stmn_lresnet34_arcs30m0.3_adam_lr0.05_amp.v1.sh
This is an x-vector model with:
- 80 logMel filter-banks with short-time mean normalization.
- ThinResNet34 (aka Light ResNet34) encoder.
- Mean+Stddev pooling
- AAM-softmax loss (m=0.3, s=30)
- Mixed prec. training.
- Downloads last month
- 2
Unable to determine this model's library. Check the
docs
.
Evaluation results
- EER Vox1-O on Voxceleb1self-reported2.110
- Minimum DCF Vox1-O prior=0.05 on Voxceleb1self-reported0.135
- Minimum DCF Vox1-O prior=0.01 on Voxceleb1self-reported0.208
- EER Vox1-E on Voxceleb1self-reported1.930
- Minimum DCF Vox1-E prior=0.05 on Voxceleb1self-reported0.121
- Minimum DCF Vox1-E Original prior=0.01 on Voxceleb1self-reported0.204
- EER Vox1-H on Voxceleb1self-reported3.210
- Minimum DCF Vox1-H prior=0.05 on Voxceleb1self-reported0.190
- Minimum DCF Vox1-H Original prior=0.01 on Voxceleb1self-reported0.298