MSST-WebUI: a browser interface for music source separation models, on its way out

The first thing the README says is to use something else

The most important line in the README is an important notice, that this project will soon stop being maintained and that the recommendation is pymss-desktop, which supports more comprehensive models, provides faster inference than MSST-WebUI, and has a better-looking GUI. Three successor projects are linked, a cross-platform desktop GUI for pymss music source separation, the pymss Python package itself, and a set of ComfyUI custom nodes. That framing should govern the rest of this article. MSST-WebUI is a project in its sunset period, and the interesting question for a reader is not whether to adopt it but what it shows about the shape of this kind of tool and what migrating away involves.

What the app actually does, underneath the web interface

The introduction explains the scope precisely. This is a webUI for Music-Source-Separation-Training, a repository for training models for music source separation, and you use the webUI to infer with the MSST models and the VR models, where the preset process page lets you customise the processing flow. Models can be installed from an Install Models interface. The useful detail for anyone who has used Ultimate Vocal Remover before is that you do not need to download the VR models again, because you can go to the Settings page and select your UVR5 model folder directly, so the two tools share weights rather than duplicating them. The WebUI also bundles convenience tools, among them the Singing-Oriented MIDI Extractor, and offers an advanced ensemble mode.

Installation has four shapes, and the Colab route skips the setup

The usage section covers four deployment paths. On Windows you download the installer from Releases or clone the repository and run from source. On Linux and macOS you clone and run from source. On Google Colab there is a notebook, webUI_for_colab.ipynb, linked from the README with a Colab badge. And for Chinese users there is a one-click UCloud deployment, on the Compshare platform, linked with a UCloud badge in the header. The release distribution is split across two hosts, with GitHub Releases carrying only the installer and no models, and Hugging Face carrying the installer and all the available models, so a user who wants the weights should look at the model repository rather than the code repository.

Running from source means fighting numpy and librosa

The source install is a four-step process and the third step is the memorable one. You clone the repository and change into the directory, create a conda environment with Python 3.10 which the README recommends, then install torch, torchvision and torchaudio from the CUDA 12.1 wheel index followed by the requirements file.

conda create -n msst python=3.10 -y
conda activate msst
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt --only-binary=samplerate

Then you have to edit an installed library. The instructions say to go into the site-packages folder, open librosa/util/utils.py at line 2185, and change np.dtype(complex): np.dtype(np.float).type to np.dtype(complex): np.dtype(float).type, because numpy removed the np.float alias. The README offers a shortcut for people who would rather not edit a site-packages file.

pip uninstall librosa -y
pip install tools/webUI_for_colouds/librosa-0.9.2-py3-none-any.whl

That is a shipped patched wheel, and the requirements.txt repeats the same instruction as a comment for anyone who installs it another way. Finally the app is started with python webUI.py.

The dependency list tells you this is a research stack

The pyproject.toml pins around fifty packages, and the list is the clearest statement of what the tool does. The deep learning stack is torch at 2.7.1 or newer, torchaudio, torchvision, transformers at 4.53 or newer, lightning 2.4.0, asteroid 0.7.0 for audio source separation, demucs 4.0.0, segmentation-models-pytorch, neuraloperator, safetensors, rotary-embedding-torch and the torchseg and torch-log-wmse packages. The audio side brings pedalboard, samplerate, soxr, spafe for speech enhancement, mido for MIDI, and librosa 0.9.2 pinned in requirements.txt. The interface is Gradio 4.38.1 with FastAPI 0.111.0 underneath and pyside6-fluent-widgets for the desktop styling, plus a ComfyUI/ directory in the repository root. The presence of asteroid and demucs means this is not only a wrapper around one model but a hub for several separation architectures.

The tooling is set up for uv, with regional mirrors

The pyproject.toml configures uv with two indexes, a default pointing at the USTC PyPI mirror and an explicit index for the CUDA 12.8 PyTorch wheels at the Nanjing University mirror, with a tool.uv.sources entry mapping torch to that index. The repository also carries a uv.lock and a requirements-uv.txt alongside the main requirements.txt, so there are two supported installation paths, and the pyproject version is 1.8.0pre.1 with the project name msst-webui and a requires-python of 3.10 or later. The rest of the tree is a working application, with webUI.py and webview-gui.py at the root, an inference/ directory, train/, modules/, utils/, scripts/, tools/, pretrain/, webui/, docs/, and backup directories for configs and data.

A known failure with swin_upernet, and a deprecated tool inside

The README flags a specific error users hit, with the model_type swin_upernet producing a ValueError about the channel dimension of the pixel values not matching the one set in the configuration, and points at an issue in the tracker for the fix. That kind of note is the mark of a project where people actually run it. There is also an important notice about the bundled tool: SOME, the Singing-Oriented MIDI Extractor, has been deprecated, and the new audio to MIDI project GAME, the Generative Adaptive MIDI Extractor, will be adapted shortly after stable testing. So the MIDI-to-audio half of the feature set is in transition even within this tool. The release history shows 1.8.0pre.1 in March 2026, 1.7.0 in January 2025 and 1.7preview.4 in late 2024, and the last push was on 2026-09-16.

Editorial conclusion

MSST-WebUI fits someone who already has the MSST or UVR model weights and wants a browser interface for inference rather than a command line, since the model directory can be pointed at an existing UVR install and the Colab notebook removes the local install entirely. It is a poor fit for a new project, because the repository itself recommends pymss-desktop as the supported path, with broader model support and faster inference. Verify first which successor you will actually use, since three are listed, the desktop app, the Python package and a ComfyUI node set, and note that the source install needs a patched librosa because of a numpy dtype change, which is a maintenance tax you avoid by taking the published installer instead.

Frequently asked questions

Official sources

1. Issues
2. License: AGPL-3.0
3. README
4. Releases
5. SUC-DriverOld/MSST-WebUI on GitHub