new models of various architectures

[Eddycrack864]

Models:

Roformers:

MelBand Roformer Kim | FT by unwa

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer | De-Reverb by anvuew

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer | De-Reverb Less Aggressive by anvuew

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer | De-Reverb-Echo by Sucial

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer | De-Reverb-Echo V2 by Sucial

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer Kim | SYHFT by SYH99999

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer Kim | SYHFT V2 by SYH99999

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer Kim | SYHFT V2.5 by SYH99999

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer Kim | SYHFT V3 by SYH99999

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer Kim | Big SYHFT V1 by SYH99999

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer Kim | Big Beta 4 FT by unwa

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer Kim | Big Beta 5e FT by unwa

Checkpoint Link: ckpt
Config Link: config

BS Roformer | Chorus Male-Female by Sucial

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer | Aspiration by Sucial

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer | Aspiration Less Aggressive by Sucial

Checkpoint Link: ckpt
Config Link: config

MelBand Roformer | Bleed Suppressor V1 by unwa & 97chris

Checkpoint Link: ckpt
Config Link: config

MDX23C:

MDX23C De-Reverb by aufr33 & jarredou

Checkpoint Link: ckpt
Config Link: config

VR Arch:

UVR-De-Reverb by aufr33 & jarredou

Checkpoint Link: pth
JSON Config: (In PR)
Code to be added to model_data_new.json as i said in #166 (comment):

"97dc361a7a88b2c4542f68364b32c7f6": {
        "vr_model_param": "4band_v4_ms_fullband",
        "primary_stem": "Dry",
        "nout": 32,
        "nout_lstm": 128,
        "is_karaoke": false,
        "is_bv_model": false,
        "is_bv_model_rebalanced": 0.0
    }

Another changes

I also renamed an existing model to give credit to its author.

These models were tested on my Windows 11 with CPU end. If you find any error lmk as I may have made some oppsie.

[beveradb]

Wow, this is amazing thank you! I'm excited to test a few of these later on; I'll upload all of the files to the model configs release and merge this as soon as I'm at a faster internet connection 🙏

[beveradb]

Released in version 0.27 🎉

Now I'm definitely getting to the point where I feel we need to provide users of audio-separator with more info about the models to help know what different models do or how well they do it, as there's so much choice but very little info to help decide what to use...

Do you or @Bebra777228 happen to know if there's an easy way to find the MVSEP leaderboard metrics (e.g. instrumental SDR, vocal SDR) for each of these models?

If we could reliably map model filenames or add additional metadata to our models.json to enable querying those metrics from MVSEP for every supported model, it would be great to then list the SDR in the model list output (audio-separator -l) to help folks choose models which perform better for specific tasks!

[Bebra777228]

I will leave the question about requesting metrics from MVSEP to you, but there is a way to find the average SDR. Some models, when opened as an archive, contain a folder whose name sometimes includes the average SDR value, such as in the models BS-Roformer-Viperx-1297 (avg SDR: 12.9755), Mel-Roformer-Viperx-1143 (avg SDR: 11.4360), and other similar ones.

For example, for the model MelBand Roformer Kim | FT by unwa, you can see the value 12.4604.

Screenshot:

I don't know if this will help in any way 🤷‍♂️

[Eddycrack864]

I can look up information about SDRs and ask directly the creators of many models. There is also a lot of information about the models, SDRs and user reviews here.

Although it will be difficult to get the SDR for older and rarely used models such as:

• UVR_MDXNET_Main.onnx
• UVR-MDX-NET-Inst_full_292.onnx
• 14_SP-UVR-4B-44100-2.pth
• MGM_LOWEND_A_v4.pth

Just to give a sample

[Bebra777228]

The model MelBandRoformerSYHFTV2.5.ckpt was not loaded in model-configs.

[Bebra777228]

The model MelBandRoformerSYHFTV2.5.ckpt was not loaded in model-configs.

What do you mean, sorry?

Isn't this the file? https://github.com/nomadkaraoke/python-audio-separator/releases/download/model-configs/config_vocals_mel_band_roformer_ft.yaml

model - MelBandRoformerSYHFTV2.5.ckpt
config - config_vocals_mel_band_roformer_ft.yaml

MelBand Roformer Kim | SYHFT V2.5 by SYH99999
Checkpoint Link: ckpt
Config Link: config

The model file is missing in model-configs.

[beveradb]

The model file is missing in model-configs.

Sorry, I deleted my comment when I re-read your message and realized I'd missed the actual model (my brain first somehow interpreted "missed the model config file" which I checked and was there 😅 )

Uploaded the actual model file now; thanks for catching that!

[beveradb]

Apologies, the VR UVR-De-Reverb-aufr33-jarredou.pth one was also missing from the repo and missing config in 0.27.0

I've just fixed it (among also implementing a model evaluation/scoring script which I intend to use to help folks choose better performing models more easily) in https://github.com/nomadkaraoke/python-audio-separator/pull/169/files

audio-separator release 0.27.1 🤞

[Bebra777228]

I just decided to check the models again and it turns out that the MDX23C-De-Reverb-aufr33-jarredou.ckpt file is also missing:

MDX23C De-Reverb by aufr33 & jarredou
Config Link: config

Checkpoint Link: ckpt ←

[beveradb]

Thanks, fixed!