search

MTG / DeepConvSep

Deep Convolutional Neural Networks for Musical Source Separation
GNU Affero General Public License v3.0
472 stars 110 forks source link

how to improve separation between sources? #6

Closed ghost closed 7 years ago

ghost commented 7 years ago

No matter what mixture I try to separate, even if the separation gives me nice results, but I was wondering what are the main parameters and tweaks that can be done in the framework to improve the separation results, so that the vocals has really only the vocals, and the bass, drums, other have less vocal artifacts in them?

Is this maybe just a matter of how many training material is used? or/and is there a way to tweak the framework further to improve the results?

Thanks!

geekchen007 commented 7 years ago

I did a test according to the reaedme file.“Separate music into vocals, bass, drums, accompaniment in examples/dsd100/separate_dsd.py ”,the model is "model_dsd_fft_1024.pkl",But the results are not very good.Where did I go wrong? The sample is"100 - Young Griffo - Pennies",and the result is "http://pan.baidu.com/s/1bpaOpxt".

nkundiushuti commented 7 years ago

hi! well I suggest training with all the files in the dataset and doing some data augmentation. as any data driven approach, the more data you have, the better. model_dsd_fft_1024.pkl is trained only on the dev set and it is used to replicate the results in the sisecmus challenge. I am sorry I can't help you more with this. I am writing my thesis, supervising a few students and I have less time now. I think there is little insight you will get by playing with the parameters. research community will gain little from this. rather you should invest your effort in proposing different architectures, improving robustness.