BUTSpeechFIT / DiaPer

MIT License
42 stars 2 forks source link

8k pretrained model #2

Closed xiangzai0115 closed 3 weeks ago

xiangzai0115 commented 8 months ago

Hi,

Thanks for this amazing work and open-sourcing the code!

I saw there are several 16k pretrained models available. Is that possible to provide 8k ones as well?

Cheers!

fnlandini commented 8 months ago

Hi, the 16k models we released were all trained with public and free data. Unfortunately, the data used to train the 8k models is not free and there could be problems with licenses. I know it is not the best scenario, but if you have some telephone data yourself, you can try to fine-tune the 16k model to your (upsampled) data and it might still give you something reasonable. We did some comparison (Figure 1) with a similar model before and the results were not too bad after fine-tuning.

I hope this helps. Federico

fnlandini commented 3 weeks ago

Closing due to inactivity