Closed xuancong84 closed 2 years ago
Thanks for using vocal-remover
for your project.
My dataset consists of 175 songs imported from my own CDs and 50 songs from DSD100. No synthetic songs are included.
The parameters used for training are described in the Train your own model
section of the README.
Thanks for sharing your data set info! In that case, would you share your training data set in bit-torrent (or so) so that others can replicate the training process? Data on public cloud tends to survive longer because our personal storage devices will break down one day or get lost.
I cannot share my dataset as it would be an infringement of copyright.
In that case, is it possible for you to list the song titles of your own collection so that other people can get a closer replication of your baseline model? Thanks!
Thank you for open-sourcing vocal remover. I have created a usable cross-platform Karaoke system using your model (see https://github.com/xuancong84/pikaraoke). Your vocal remover works excellent on almost all human voices regardless of any language.
However, on certain synthetic singing voice, it does not work very well. (e.g., https://www.youtube.com/watch?v=8vowBJU65zg ). One of the reasons could be due to that the singing voice synthesizer does not mimic human voice very well, and your model is not trained on those kind of data.
May I know what training set do you use? (Is it from LDC or some synthetic dataset) If possible, maybe you can include your training/evaluation script (together with all hyper-parameters) as well for reproducibility of your baseline model, so that others can make further improvements on top of your model. Thanks a lot!