search

ZFTurbo / Music-Source-Separation-Training

Repository for training models for music source separation.
385 stars 51 forks source link

what dataset you use for vocals_mdx23c model？ #8

Open nukes opened 8 months ago

nukes commented 8 months ago

Hi, What dataset you use for vocals_mdx23c model training ? I want to reproduce your training and based on which to make improvement.

ZFTurbo commented 8 months ago

Hello. I used my own internal dataset.

nukes commented 8 months ago

Got it!

nukes commented 8 months ago

Hi, i just tried your bandit model, the pretrained model performance for speech is so fantastic! Unbelievable performance even compared to mdx23c. Did you use the Divide and Remaster (DnR) dataset to train it ?

nukes commented 8 months ago

Also, do you have the vocals_bandit_bsrnn_multi_mus64 checkpoint for sharing?

ZFTurbo commented 8 months ago

Hi, i just tried your bandit model, the pretrained model performance for speech is so fantastic! Unbelievable performance even compared to mdx23c. Did you use the Divide and Remaster (DnR) dataset to train it ?

It's weights provided by author of model. I didn't train on DnR

Also, do you have the vocals_bandit_bsrnn_multi_mus64 checkpoint for sharing?

Yes, I have checkpoint, but result is very poor. I will share later.

nukes commented 8 months ago

Hi, i just tried your bandit model, the pretrained model performance for speech is so fantastic! Unbelievable performance even compared to mdx23c. Did you use the Divide and Remaster (DnR) dataset to train it ?

It's weights provided by author of model. I didn't train on DnR

Also, do you have the vocals_bandit_bsrnn_multi_mus64 checkpoint for sharing?

Yes, I have checkpoint, but result is very poor. I will share later.

looking forward !

deyituo commented 8 months ago

dataset is essential..