Audio separator (UVR / MDX / Demucs)
Splits a video's sound into voices and background with UVR, MDX-Net, BS-Roformer or Demucs models, on your own computer. It is an alternative to the built-in ffmpeg separator, which relies on a stereo trick; each model downloads once, the first time you use it.
Settings Add-ons audio-separator (UVR / MDX / Demucs) Install
Subtitld downloads it, checks its SHA-256, and keeps it updated.
- Models
- 5
- Add-on download
- 295–432 MB
- RAM, at least
- 4 GB
What it does
Every model gives Subtitld two files: the voices, and everything else. Demucs splits drums, bass and the rest; the add-on mixes them back into one background.
Export audio with Include background audio and your dub is mixed over the background. It stays in stereo, so music and ambience keep their width.
When a voice-cloning add-on needs a sample of a speaker, Subtitld cuts it from the voices file, so the music stays out of it.
Subtitld separates a video when you open it and keeps both files, so reopening the project reuses them.
Choose it under Global settings › General › Audio separator. The built-in ffmpeg separator stays in the list.
Models
Start with Kim Inst, the default. Try BS-Roformer when the split matters more than the time it takes. Karaoke 2 keeps only the main voice and leaves backing vocals with the music. Demucs works in four stems, which the add-on mixes back into one background.
| Model | Download | Good for |
|---|---|---|
| Kim_Inst.onnx Kim_Inst.onnx | Voices against everything else; the default | |
| UVR-MDX-NET-Inst_HQ_3.onnx UVR-MDX-NET-Inst_HQ_3.onnx | A fast, balanced split | |
| UVR_MDXNET_KARA_2.onnx UVR_MDXNET_KARA_2.onnx | The main voice alone; backing vocals go with the music | |
| model_bs_roformer_ep_317_sdr_12.9755.ckpt | The best split, when time allows | |
| htdemucs_ft.yaml | Four stems, mixed back into two |
Each model downloads the first time you use it. Subtitld keeps the stems of every video it has separated and reuses them, so a different model applies only to videos that haven't been separated yet. To redo one, delete its files from the audioseparation folder in Subtitld's cache.
Where models are kept
In a folder of their own, outside the add-on, so reinstalling or updating it doesn't download them again. To keep them elsewhere, set AUDIO_SEPARATOR_MODEL_DIR before you start Subtitld.
| System | Folder |
|---|---|
| Linux | ~/.cache/subtitld/addons/audio-separator/models |
| macOS | ~/Library/Caches/subtitld/addons/audio-separator/models |
| Windows | %LOCALAPPDATA%\subtitld\subtitld\Cache\addons\audio-separator\models |
Options
In Settings › Add-ons › audio-separator (UVR / MDX / Demucs) › Configure.
Kim_Inst.onnxWhich model to use. See the table above.
cpuCPU works everywhere. The released builds don't include a CUDA runtime, so choosing CUDA doesn't make them use an NVIDIA card.
FLACPlatforms
| System | Download | Tested |
|---|---|---|
| Linux x86-64 | 432.4 MB | In CI, start-up only |
| macOS (universal) | 312.2 MB | In CI, start-up only |
| Windows x86-64 | 295.5 MB | In CI, start-up only |
In CI, each build is started once to check that it answers Subtitld; no separation runs there.
Troubleshooting
Subtitld doesn't show separation errors yet. When a separation fails, the Music/Voice control stays off, and Include background audio stays greyed out when you export audio. Check these first:
- ffmpeg on your PATH. The add-on calls
ffmpegfrom your PATH to read the video's sound and to convert both stems. It doesn't use the copy inside Subtitld, which on Windows and macOS isn't on the PATH, so install ffmpeg so that it runs from a terminal. - A connection on first use. Each model downloads the first time you use it.
License
The add-on and python-audio-separator are MIT. Each model has its own license and the add-on doesn't list them; the catalog notes that some UVR and MDX models are for non-commercial use only, so check a model's card before you use it commercially.
Versions
- 0.0.5
Automated release for audio-separator v0.0.5.