Loading...
Split songs into vocal and instrument tracks for listening, mixing or practice backing tracks.
Input
Supports inputs in formats such as mp3 / wav / flac / ogg / m4a / aac / aif / aiff / mp4 / mov / mkv
Supported formats
Output
4 × WAV
Use case
CrowdioSet SCNet: M1 uses clean training, M2 adds reverb, M3 adds audience sounds and M4 adds both. M1/M2 output four stems; M3/M4 also output audience. Reverb is retained and sing-alongs may remain.
Choose a model for your material and purpose, and review the examples before submitting.
Cost, time, formats and parameters depend on the selected model.
Include the tool name, steps and the problem in your feedback. Screenshots can help us locate it.
CrowdioSet SCNet: M1 uses clean training, M2 adds reverb, M3 adds audience sounds and M4 adds both. M1/M2 output four stems; M3/M4 also output audience. Reverb is retained and sing-alongs may remain.
BS-RoFormer SW Fixed uses BS-Rofo-SW-Fixed.ckpt plus BS-Rofo-SW-Fixed.yaml to split a full mix into six WAV stems.
Separates a full mix into four standard MUSDB-style stems.
XLANCE-MSR extends a six-stem BS-RoFormer system into eight-class source restoration with separation, denoising, and dereverberation/restoration. The MIT checkpoint weights are public.
MVSep Mega 53 Stems separates a mix by the instruments it detects, helping you explore complex arrangements. Individual parts may be less clean than with specialist models.
BS-RoFormer SW Fixed uses BS-Rofo-SW-Fixed.ckpt plus BS-Rofo-SW-Fixed.yaml to split a full mix into six WAV stems.
Sony MIMO SCNet separates vocals, drums, bass and other instruments, with SCNet and SCNet Gen versions.
Sony MIMO SCNet separates vocals, drums, bass and other instruments, with SCNet and SCNet Gen versions.
MDX23C Phantom Centre extracts similar centre signals from stereo channels and outputs a side residual. Centred instruments may be extracted and panned vocals may remain; sources are not identified by instrument.
Extract bowed strings and export the remaining audio. Similar timbres and synthesized strings may bleed between tracks.
arXiv - 2026
Mean SDR from Tables 2 and 3: clean uses clean audio, noisy adds audience, and noisyrev adds audience plus reverb. Conditions are ranked separately; reverb-trained models do not dereverberate.
SDX23 music separation track - 2023
BS-RoFormer uses band-split spectrogram features plus RoPE Transformers; the SDX23 system ranked first in the music source separation track and a smaller version reported 9.80 dB average SDR on MUSDB18HQ.
ICASSP - 2023
HT Demucs replaces core Demucs layers with cross-domain Transformer attention and reports 9.20 dB SDR with extra data and fine-tuning.
ICASSP - 2025
Introduces the subband sparse-compression architecture behind the deployed Huge-SCNet-4stems checkpoint: the mixture spectrogram is split into bands compressed at different rates for efficient high-quality separation.
arXiv - 2026
Technical report for the XLANCE-MSR system submitted to the first Music Source Restoration Challenge.
arXiv - 2026
Challenge summary paper describing the MSR task, datasets, objective metrics, subjective evaluation, and final ranking tables.
arXiv - 2026
Pinned Sony README: MUSDB18-HQ, museval median-of-medians. Compare only within this protocol, separately from MVSEP and other paper scores.
4 output categories
Only records with verified public inference implementations and matching weights are shown; algorithms requiring no weights are also eligible. Closed services, unreleased weights and unmatched experiment versions are excluded. This is not an all-model leaderboard.
Each set of weights appears once per benchmark, using its best published result for the selected metric. An overall score comes from one evaluation. Expand to compare evaluations and settings.
Public weights do not establish a fully open-source license. Expand the data for declared weight licenses and access conditions. Undeclared licenses are marked separately.
Markers identify models or components used here. Scores and settings come from the cited tests, not a new evaluation of the complete site pipeline.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Deployable entries 4 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
SAMAudio †
Mean SDR from Tables 2 and 3: clean uses clean audio, noisy adds audience, and noisyrev adds audience plus reverb. Conditions are ranked separately; reverb-trained models do not dereverberate.
| Metric rank | Model | Test setting | vocals↑ | drums↑ | bass↑ | other↑ |
|---|---|---|---|---|---|---|
| - | CrowdioSet SCNet M1Used on this sitePublic weights · Inference code | clean | 1.46 ± 3.4 ( 10.05 ) | 7.04 ± 3.6 ( 10.62 ) | 4.15 ± 3.5 ( 8.97 ) | 1.43 ± 3.5 ( 6.69 ) |
| - | CrowdioSet SCNet M2Public weights · Inference code | rev | 1.10 ± 3.6 ( 9.01 ) | 8.23 ± 3.1 ( 9.86 ) | 5.36 ± 3.7 ( 7.77 ) | 3.63 ± 2.4 ( 5.98 ) |
| - | CrowdioSet SCNet M3Public weights · Inference code | noisy | 3.26 ± 2.9 ( 9.97 ) | 7.00 ± 3.3 ( 10.68 ) | 3.91 ± 3.6 ( 8.85 ) | 1.79 ± 3.2 ( 6.69 ) |
| - | CrowdioSet SCNet M4Public weights · Inference code | noisyrev | 3.15 ± 3.1 ( 9.07 ) | 8.19 ± 3.3 ( 9.98 ) | 5.66 ± 3.6 ( 7.86 ) | 3.55 ± 2.7 ( 5.88 ) |
Markers identify models or components used here. Scores and settings come from the cited tests, not a new evaluation of the complete site pipeline.
Datasets, versions and conditions remain separate, including sources with no eligible ranking rows.
Deployable entries 4 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
SAMAudio †
Deployable entries 2 / source rows 3
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
SAMAudio †
Scores apply to the listed dataset, metric and model version; results from different tests are not directly comparable.
The table includes only models with public weights or configurations that can run locally.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Deployable entries 4 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
SAMAudio †
Mean SDR from Tables 2 and 3: clean uses clean audio, noisy adds audience, and noisyrev adds audience plus reverb. Conditions are ranked separately; reverb-trained models do not dereverberate.
| Metric rank | Model | Test setting | vocals↑ | drums↑ | bass↑ | other↑ |
|---|---|---|---|---|---|---|
| - | CrowdioSet SCNet M1Public weights · Inference code | clean | 1.46 ± 3.4 ( 10.05 ) | 7.04 ± 3.6 ( 10.62 ) | 4.15 ± 3.5 ( 8.97 ) | 1.43 ± 3.5 ( 6.69 ) |
| - | CrowdioSet SCNet M2Public weights · Inference code | rev | 1.10 ± 3.6 ( 9.01 ) | 8.23 ± 3.1 ( 9.86 ) | 5.36 ± 3.7 ( 7.77 ) | 3.63 ± 2.4 ( 5.98 ) |
| - | CrowdioSet SCNet M3Public weights · Inference code | noisy | 3.26 ± 2.9 ( 9.97 ) | 7.00 ± 3.3 ( 10.68 ) | 3.91 ± 3.6 ( 8.85 ) | 1.79 ± 3.2 ( 6.69 ) |
| - | CrowdioSet SCNet M4Public weights · Inference code | noisyrev | 3.15 ± 3.1 ( 9.07 ) | 8.19 ± 3.3 ( 9.98 ) | 5.66 ± 3.6 ( 7.86 ) | 3.55 ± 2.7 ( 5.88 ) |
The table includes only models with public weights or configurations that can run locally.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Deployable entries 2 / source rows 3
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
SAMAudio †
Mean SDR from Tables 2 and 3: clean uses clean audio, noisy adds audience, and noisyrev adds audience plus reverb. Conditions are ranked separately; reverb-trained models do not dereverberate.
| Metric rank | Model | Test setting | audience isolation | |
|---|---|---|---|---|
| non-reverberant↑ | reverberant↑ | |||
| 1 | CrowdioSet SCNet M3Public weights · Inference code | noisy | 2.85 ± 1.3 | 1.25 ± 1.4 |
| 2 | CrowdioSet SCNet M4Public weights · Inference code | noisyrev | 1.55 ± 1.0 | 1.56 ± 0.9 |
Find processing methods and steps suited to your material.