ZFTurbo/MVSEP-MDX23-music-separation-model

Hello teacher: What are the model combinations of version 1.0.1?

ybhka2022 opened this issue · 7 comments

ybhka2022 commented

Hello teacher: What are the model combinations of version 1.0.1?

ZFTurbo commented

Hello. Currently used:

  • Kim_Vocal_2.onnx
  • Kim_Inst.onnx
  • All 4 available Demucs4 models
ybhka2022 commented

Thank you teacher for answering. This combination model works very well. The only downside is that there are a few vocals in the accompaniment after separation. Is there any solution?

ZFTurbo commented

The problem with instrumental part is following:

  1. The goal of competition was to create best vocals, other, bass and drums. Instrumental didn't count.
  2. The source of problem is usage of vocals models with cut of high frequences - so this frequences are left untoched and bleed into instrumental after subtract from original mix.

As I know it was fixed in:
https://github.com/jarredou/MVSEP-MDX23-Colab_v2

  • ❤️1
ybhka2022 commented

Thank you teacher for pointing out: What are the combinations for extracting the best accompaniment models?

ybhka2022 commented

你好。目前使用：

  • Kim_Vocal_2.onnx
  • Kim_Inst.onnx
  • 所有 4 种可用的 Democs4 型号

Please update MVSEP-MDX23-Colab fork v2.2.1 desktop software, thank you!

ybhka2022 commented

你好。目前使用：

  • Kim_Vocal_2.onnx
  • Kim_Inst.onnx
  • 所有 4 种可用的 Democs4 型号

https://github.com/KimberleyJensen/kmdx-net_music-source-separation 老师，你用的与这个模型是一样吗

ybhka2022 commented

Teacher, are you using the same model as this one?