Finding out which stems are in which models

[sanderness66]

Hello! Am I right in thinking that only the demucs models separate into drums/bass/vocal/other stems and the others mostly into vocals/instruments? How can I find out which model separates in which stems?

[beveradb]

Sorta; there are a lot of models with different purposes, but the majority are two stem (instrumental and vocals) models.
Usually the name of the model will give you an idea of its purpose.

Run audio-separator -l and read through the list of models;
if it's not obvious from the name, it's probably a 2-stem (Vocals and Instrumentals) model.

Here are some examples of models which are not just vocals/instrumental:

• htdemucs_6s.yaml tries to separate 6 stems from the input audio (Vocals, Bass, Drums, Guitar, Piano, Other)

• mel_band_roformer_crowd_aufr33_viperx_sdr_8.7144.ckpt tries to separate crowd noise

• denoise_mel_band_roformer_aufr33_sdr_27.9959.ckpt and UVR-DeNoise try to remove background noise

• UVR-De-Echo-Aggressive.pth tries to remove echo

• deverb_bs_roformer_8_384dim_10depth.ckpt and Reverb_HQ_By_FoxJoy.onnx try to remove reverb

• UVR-BVE-4B_SN-44100-1 is designed to separate background vocals (BVE is "Backing Vocals Extraction")

• Any with Karaoke in the name e.g. 6_HP-Karaoke-UVR, Mel-Roformer-Karaoke-Aufr33-Viperx, UVR-MDX-NET Karaoke etc. will be a 2-stem model but if you give it only a vocal input it will output Lead Vocals and Backing Vocals separately.

• 17_HP-Wind_Inst-UVR.pth tries to separate Wind instruments specifically

• The various kuielab_ models are all 2 stem models but focus on a specific stem, e.g.

        "MDX-Net Model: kuielab_a_bass": "kuielab_a_bass.onnx",
        "MDX-Net Model: kuielab_a_drums": "kuielab_a_drums.onnx",
        "MDX-Net Model: kuielab_a_other": "kuielab_a_other.onnx",
        "MDX-Net Model: kuielab_a_vocals": "kuielab_a_vocals.onnx",

Hope this gives you an idea; overall though just experiment with different models till you get the result you want!

[sanderness66]

Thanks! There are some models in there that sound rather interesting!

[beveradb]

FYI I've now added a more comprehensive post explaining different use cases and recommending models:

#133

[sanderness66]

Very useful, thanks!