Model for Sound demixing challenge 2023: Music Demixing Track - MDX'23. 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.
--input_audio- input audio location. You can provide multiple files at once. Required--output_folder- output audio folder. Required--cpu- choose CPU instead of GPU for processing. Can be very slow.--overlap_large- overlap of splitted audio for light models. Closer to 1.0 - slower, but better quality. Default: 0.6.--overlap_small- overlap of splitted audio for heavy models. Closer to 1.0 - slower, but better quality. Default: 0.5.--single_onnx- only use single ONNX model for vocals. Can be useful if you have not enough GPU memory.--chunk_size- chunk size for ONNX models. Set lower to reduce GPU memory consumption. Default: 1000000.--large_gpu- it will store all models on GPU for faster processing of multiple audio files. Requires at least 11 GB of free GPU memory.--use_kim_model_1- use first version of Kim model (as it was on contest).--only_vocals- only create vocals and instrumental. Skip bass, drums, other. Processing will be faster.
- 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). - 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. - There is Google.Collab version of this code.
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 (~730 MB). Unzip archive and to start program double click
run.bat. On first run it will download pytorch with CUDA support (~2.8 GB) and some Neural Net models. - Program will download all needed neural net models from internet at the first run.
- GUI supports Drag & Drop of multiple files.
- Progress bar available.
executing web-ui.py with python will start the web interface locally on localhost (127.0.0.1).
You'll see what port it is running on within the terminal output.
- Browser-Based user interface
- Program will download all needed neural net models from internet at the first run.
- supports Drag & Drop for audio upload (single file)
- Settings in GUI updated, now you can control all possible options
- Kim vocal model updated from version 1 to version 2, you still can use version 1 using parameter
--use_kim_model_1 - Added possibility to generate only vocals/instrumental pair if you don't need bass, drums and other stems. Use parameter
--only_vocals - Standalone program was updated. It has less size now. GUI will download torch/cuda on the first run.
@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}
}