• Feature: 80-dim fbank, mean normalization, speed perturb
• Training: lr [0.0001, 0.2], batch_size 256, 4 gpus(Tesla V100), additive angular margin
• Metrics: EER(%), MinDCF

• Train set: CNCeleb-dev + CNCeleb2, 2973 speakers
• Test set: CNCeleb-eval

Pretrained models are accessible on ModelScope.

• ECAPA-TDNN: speech_ecapa-tdnn_sv_zh-cn_cnceleb_16k

Here is a simple example for directly extracting embeddings. It downloads the pretrained model from ModelScope and extracts embeddings.

# Install modelscope
pip install modelscope
# ECAPA-TDNN trained on CNCeleb
model_id=damo/speech_ecapa-tdnn_sv_zh-cn_cnceleb_16k
# Run inference
python speakerlab/bin/infer_sv.py --model_id $model_id --wavs $wav_path

If you are using ECAPA-TDNN model in your research, please cite:

@article{desplanques2020ecapa,
  title={Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification},
  author={Desplanques, Brecht and Thienpondt, Jenthe and Demuynck, Kris},
  journal={arXiv preprint arXiv:2005.07143},
  year={2020}
}