search

clovaai / voxceleb_trainer

In defence of metric learning for speaker recognition
MIT License
1k stars 272 forks source link

Is there anynone tried large margin finetune? #135

Closed llearner closed 2 years ago

llearner commented 2 years ago

I tried, but did not achieve good results like the reference.

Reference: The IDLAB VoxSRC-20 Submission: Large Margin Fine-Tuning and Quality-Aware Score Calibration in DNN Based Speaker Verification https://arxiv.org/abs/2010.11255

lhanzl commented 2 years ago

I try larger margin fune-tuning in speechbrain's recipe: The initial loss was higher than the original, and it seemed that more epoch training was needed to make valError lower than the original. What are the phenomena of your training?

llearner commented 2 years ago

I try larger margin fune-tuning in speechbrain's recipe: The initial loss was higher than the original, and it seemed that more epoch training was needed to make valError lower than the original. What are the phenomena of your training?

The training loss rise first and then fall, the valid loss fall gradually, but the test EER always higher than the original EER.

Jungjee commented 2 years ago

Closing this issue since it has been inactive for six months