Adaptation of MVSep-MDX23 algorithm for Colab, with few tweaks:
Recent changes:
v2.3
- HQ3-Instr model replaced by VitLarge23 (thanks to MVSep)
- Improved MDXv2 processing (thanks to Anjok)
- Improved BigShifts algo (v2)
- BigShifts processing added to MDXv3 & VitLarge
- Faster folder batch processing
Full changelog & previous versions links :
- Improved MDXv3 chunking code (thanks to HymnStudio)
- D1581 demo model replaced by new InstVocHQ MDXv3 model.
v2.2.1
- Added custom weights feature
- Fixed some bugs
- Fixed input: you can use a file or a folder as input now
v2.2
- Added MDXv3 compatibility
- Added MDXv3 demo model D1581 in vocals stem multiband ensemble.
- Added VOC-FT Fullband SRS instead of UVR-MDX-Instr-HQ3.
- Added 2stems feature : output only vocals/instrum (faster processing)
- Added 16bit output format option
- Added "BigShift trick" for MDX models
- Added separated overlap values for MDX, MDXv3 and Demucs
- Fixed volume compensation fine-tuning for MDX-VOC-FT
- Updated with MDX-VOC-FT instead of Kim Vocal 2
- Updated with new Kim Vocal 2 & UVR-MDX-Instr-HQ3 models
- Folder batch processing
- Fixed high frequency bleed in vocals
- Fixed volume compensation for MDX models
Original work:
# MVSEP-MDX23-music-separation-model Model for [Sound demixing challenge 2023: Music Demixing Track - MDX'23](https://www.aicrowd.com/challenges/sound-demixing-challenge-2023). Model perform separation of music into 4 stems "bass", "drums", "vocals", "other". Model won 3rd place in challenge (Leaderboard C).
Model based on Demucs4, MDX neural net architectures and some MDX weights from Ultimate Vocal Remover project (thanks Kimberley Jensen for great high quality vocal models). Brought to you by MVSep.com.
python inference.py --input_audio mixture1.wav mixture2.wav --output_folder ./results/
With this command audios with names "mixture1.wav" and "mixture2.wav" will be processed and results will be stored in ./results/ folder in WAV format.
- Note 1: If you have not enough GPU memory you can use CPU (
--cpu), but it will be slow. Additionally you can use single ONNX (--single_onnx), but it will decrease quality a little bit. Also reduce of chunk size can help (--chunk_size 200000). - Note 2: In current revision code requires less GPU memory, but it process multiple files slower. If you want old fast method use argument
--large_gpu. It will require > 11 GB of GPU memory, but will work faster.
Quality comparison with best separation models performed on MultiSong Dataset.
| Algorithm | SDR bass | SDR drums | SDR other | SDR vocals | SDR instrumental |
|---|---|---|---|---|---|
| MVSEP MDX23 | 12.5034 | 11.6870 | 6.5378 | 9.5138 | 15.8213 |
| Demucs HT 4 | 12.1006 | 11.3037 | 5.7728 | 8.3555 | 13.9902 |
| Demucs 3 | 10.6947 | 10.2744 | 5.3580 | 8.1335 | 14.4409 |
| MDX B | --- | ---- | --- | 8.5118 | 14.8192 |
- Note: SDR - signal to distortion ratio. Larger is better.
- Script for GUI (based on PyQt5): gui.py.
- You can download standalone program for Windows here: zip1, zip2. Unzip archives and to start program double click
run.bat. - Program will download all needed neural net models from internet at the first run.
- GUI supports Drag & Drop of multiple files.
- Progress bar available.
@misc{solovyev2023benchmarks,
title={Benchmarks and leaderboards for sound demixing tasks},
author={Roman Solovyev and Alexander Stempkovskiy and Tatiana Habruseva},
year={2023},
eprint={2305.07489},
archivePrefix={arXiv},
primaryClass={cs.SD}
}