• Feature: 80-dim fbank, mean normalization, speed perturb
• Training: lr [0.00005, 0.2], batch_size 512, 8 gpu(Tesla V100), additive angular margin, speaker embeddding=192
• Metrics: EER(%), MinDCF(p-target=0.01)

• Train set: 3D-Speaker-train
• Test set: 3D-Speaker-test

Pretrained models are accessible on ModelScope.

• ERes2net-base: speech_eres2net_base_sv_zh-cn_3dspeaker_16k
• ERes2Net-large: damo/speech_eres2net_large_sv_zh-cn_3dspeaker_16k
• 200k labeled speakers: speech_campplus_sv_zh-cn_16k-common

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

# Install modelscope
pip install modelscope
# ERes2Net trained on 3D-Speaker
model_id=damo/speech_eres2net_large_sv_zh-cn_3dspeaker_16k
# ERes2Net trained on 200k labeled speakers
model_id=damo/speech_eres2net_sv_zh-cn_16k-common
# Run inference
python speakerlab/bin/infer_sv.py --model_id $model_id --wavs $wav_path

If you are using ERes2Net model in your research, please cite:

@article{eres2net,
  title={An Enhanced Res2Net with Local and Global Feature Fusion for Speaker Verification},
  author={Yafeng Chen, Siqi Zheng, Hui Wang, Luyao Cheng, Qian Chen, Jiajun Qi},
  booktitle={Interspeech 2023},
  year={2023},
  organization={IEEE}
}