search

sony / bigvsan

Pytorch implementation of BigVSAN
MIT License
198 stars 16 forks source link

Perfomance issue #5

Closed wh1t3tea closed 4 months ago

wh1t3tea commented 4 months ago

What's performance difference between "base" and other configs? Can I get results with bigvsan_base_22khz_80band.json like you provide with bigvsan_24khz_100band.json?

TakashiShibuyaSony commented 4 months ago

Thank you for having an interest in our work.

To be honest, we've never trained a model with base configs. We just inherited some config files from the BigVGAN repository (BigVGAN is the model that our BigVSAN is based on).

I believe that, even if we train a BigVSAN model with a base config, it will outperform its BigVGAN counterpart. But, from our experiences, BigVGAN and BigVSAN are data-hungry. If a dataset is not so large, a result might be unsatisfactory.

Does this answer your question?

wh1t3tea commented 4 months ago

Yeah, thank you