ZFTurbo / Music-Source-Separation-Training

Repository for training models for music source separation.
MIT License
485 stars 67 forks source link

ultra hi fi vocal removal ensemble settings? #75

Open lizheng100 opened 2 months ago

lizheng100 commented 2 months ago

Looking to upgrade fidelity of vocal removal settings. Difficult to get close to mvsep premium. Their service uses additional training data which makes it higher quality.

Regardless, seems like might be a way to increase quality using ensemble. Best I can get is an ensemble of melbandroformer and bsroformer with the following settings:

python3 ensemble.py --files kimOut/mixture_vocals.wav vipermelOut/mixture_vocals.wav --weights 2 1 --type avg_wave --output vocalsExtracted.wav

Any recommendations for improving ensemble?