search

modelscope / 3D-Speaker

A Repository for Single- and Multi-modal Speaker Verification, Speaker Recognition and Speaker Diarization
Apache License 2.0
1.02k stars 89 forks source link

Inquiry about training large-scale data #79

Closed axuan731 closed 5 months ago

axuan731 commented 5 months ago

Hello,

Thank you very much for your contributions to the code and models. I am currently attempting to train a speaker verification model using a substantial amount of data. However, I have observed that conventional training strategies result in a low accuracy for the model. Therefore, I would like to inquire about how you adjusted parameters such as epochs and learning rates when training a speaker model with 200k speakers. Your insights would be highly appreciated.

yfchenlucky commented 5 months ago

The code's training configurations for large-scale and small-scale data remain consistent. To adjust, consider tweaking the learning rate or lowering the margin.