Audio denoising and dereverberation · XLANCE-MSR · TelkNet

Separation & stem splitGPU workloadxlance-msr · Source xlance-msr-ckpt · Weights XLANCE-MSR · Paper XLANCE-MSR · Paper

Audio denoising and dereverberation

Reduce noise, crowd sounds or reverb in recordings, compare with the original and download the processed audio.

40 credits/useFile inputMax 100 MBEstimated 900-3600 seconds

Input

Music audio or video with noise, room reverb, compression artifacts, or livestream degradation.

Supported formats

mp3wavflacoggm4aaacaifaiffmp4movmkv

Output

A cleaned full mix reconstructed from XLANCE-MSR restored sources.

Use case

Cleaning a complete mix before mastering, separation, transcription, or video remux.

Choose a model family

Music denoise & dereverb · XLANCE-MSRAvailableCleaning a complete mix before mastering, separation, transcription, or video remux.Music denoising · MelBand RoFormerAvailableClean noise from music recordings; compare standard and aggressive versions for remaining noise and lost detail.Vocal De-Echo & Dereverb · MelBand RoFormerAvailableReturn a dry vocal and the removed echo-and-reverb residual from an isolated vocal recording.Crowd Removal · Crowd MelBand RoFormerAvailableSeparate crowd sound from a live recording and return both the crowd and the crowd-removed programme audio.

Choose a model for your material and purpose, and review the examples before submitting.

Cost, time, formats and parameters depend on the selected model.

Quick feedback

Include the tool name, steps and the problem in your feedback. Screenshots can help us locate it.

Models, parameters, and sources

Music Denoise & Dereverb runs the open XLANCE-MSR source-restoration chain on a noisy or reverberant finished mix: denoise first, restore the eight source classes, then reconstruct one cleaned WAV mix. XLANCE-MSR ranked first in the MSR Challenge 2025 restoration results, which measure restoration quality rather than MVSep-style stem SDR.

XLANCE-MSRMSR Challenge 2025 official resultsxlancelabCUPAudioGroupAC_DCHachimicp-jkuXLANCE-MSR Technical Report for MSR Challenge 2025A Summary of the First Music Source Restoration ChallengeMelBand RoFormer · StandardOfficial checkpoint filenamesDenoise MelBand RoFormer

XLANCE-MSR

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.

XLANCE-MSR GitHub XLANCE-MSR MIT checkpoints MSR Challenge 2025 results XLANCE-MSR arXiv MSR Challenge summary

Targets eight source classes: vocals, guitar, keys, synth, bass, drums, percussion, and orchestral.
Official MSR Challenge 2025 results report objective rank #1 with MMSNR 4.4623, Zimt 0.0137, FAD 0.1988, and subjective overall MOS 3.4665. These are source-restoration metrics, not six-stem SDR.
Writes WAV stems only; the sound can be drier and closer to untreated material rather than preserving all record reverb, compression, EQ, mastering, and codec artifacts.
Score source: MSR Challenge 2025 ranked XLANCE-MSR first on objective and subjective source-restoration results; read MMSNR, Zimt, FAD, and MOS separation/restoration, not six-stem SDR.
Evaluation scope: this is eight-class source restoration / eight-stem restoration, scored on the MSR restoration task rather than merged into MVSep six-stem SDR rankings.

MelBand RoFormer · Standard

MelBand RoFormer · Standard: Clean up recordings with noise, echo or live interference.

MSST inference source (pinned)Official denoise checkpoint snapshot

MelBand RoFormer · Aggressive

MelBand RoFormer · Aggressive: Clean up recordings with noise, echo or live interference.

MSST inference source (pinned)Official aggressive denoise checkpoint snapshot

Vocal De-Echo MelBand RoFormer

MSST inference source (pinned)Official de-echo checkpoint (pinned)

Crowd MelBand RoFormer

MSST inference source (pinned)Official crowd checkpoint snapshot

Paper and benchmark notes

XLANCE-MSR Technical Report for MSR Challenge 2025

arXiv - 2026

Technical report for the XLANCE-MSR system submitted to the first Music Source Restoration Challenge.

arXiv:2602.09042 Official challenge page

A Summary of 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:2601.04343 MSR Challenge

Official / repositories

XLANCE-MSR GitHub XLANCE-MSR MIT checkpoints MSR Challenge 2025 results MSST inference source (pinned)Official denoise checkpoint snapshot Official aggressive denoise checkpoint snapshot Official de-echo checkpoint (pinned)Official crowd checkpoint snapshot

Papers / technical notes

XLANCE-MSR arXiv MSR Challenge summary

Parameter guide

XLANCE-MSR separates and restores eight source categories while treating noise and reverberation.

Use cases

Remove noise and room reverberation from music
Prepare a cleaner mix for separation or MIDI transcription
Repair compressed, livestream, or room-recorded material

Workflow

1Upload music audio or video
2XLANCE-MSR denoises, restores separated sources, and reconstructs one cleaned mix
3Preview or download the cleaned mix, or remux it into the source video

What to check

Prefer the highest-quality source available
Restoration cannot recreate detail that is permanently missing
Compare original and processed audio at matched loudness

Models in this field

3 entries

Includes public test results for models used here and elsewhere. The complete source table is shown by default; conditions and metrics can also be selected for comparison. Scores from different test conditions are not directly comparable.

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.

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.

BenchmarkRanking category

MSR Challenge 2025 official results

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 3 / source rows 5

Not ranked: 2

A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.

Hachimi · cp-jku

Official results rank XLANCE-MSR first on objective metrics and listening tests; these are eight-class source-restoration scores, not six-stem SDR scores.

Rank by metric

MSR Challenge 2025 official results
Metric rank	Model	Overall MMSNR↑	Overall Zimt↓	Overall FAD↓	MMSNR Rank↓	Zimt Rank↓	FAD Rank↓	Macro Rank↓
1	xlancelabUsed on this site Public weights · Inference code Inference codeWeight license not declared	4.4623	0.0137	0.1988	1	1	1	1.00
2	CUPAudioGroup Public weights · Inference code Inference codeWeight license not declared	2.3405	0.0164	0.2253	2	2	2	2.00
3	AC_DC Public weights · Inference code Inference codeWeight license not declared	1.4520	0.0182	0.2907	4	3	3	3.33

Original records (3)

Markers identify models or components used here. Scores and settings come from the cited tests, not a new evaluation of the complete site pipeline.

MSR 2025 · Objective ranking ↗Checked: 2026-10-25

All evaluation sources (2)

Datasets, versions and conditions remain separate, including sources with no eligible ranking rows.

MSR Challenge 2025 official results ↗
Deployable entries 3 / source rows 5
Not ranked: 2
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Hachimi · cp-jku
MSR Challenge 2025 official results ↗
Deployable entries 3 / source rows 5
Not ranked: 2
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Hachimi · cp-jku

All original results and sources

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.

MSR Challenge 2025 official results

Deployable entries 3 / source rows 5

Not ranked: 2

A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.

Hachimi · cp-jku

Official results rank XLANCE-MSR first on objective metrics and listening tests; these are eight-class source-restoration scores, not six-stem SDR scores.

Rank by metric

MSR Challenge 2025 official results
Metric rank	Model	Overall MMSNR↑	Overall Zimt↓	Overall FAD↓	MMSNR Rank↓	Zimt Rank↓	FAD Rank↓	Macro Rank↓
1	xlancelab Public weights · Inference code Inference codeWeight license not declared	4.4623	0.0137	0.1988	1	1	1	1.00
2	CUPAudioGroup Public weights · Inference code Inference codeWeight license not declared	2.3405	0.0164	0.2253	2	2	2	2.00
3	AC_DC Public weights · Inference code Inference codeWeight license not declared	1.4520	0.0182	0.2907	4	3	3	3.33

Original records (3)

MSR Challenge 2025 official results ↗Checked: 2026-10-25

The table includes only models with public weights or configurations that can run locally.

MSR Challenge 2025 official results

Deployable entries 3 / source rows 5

Not ranked: 2

A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.

Hachimi · cp-jku

Official results rank XLANCE-MSR first on objective metrics and listening tests; these are eight-class source-restoration scores, not six-stem SDR scores.

Rank by metric

MSR Challenge 2025 official results
Metric rank	Model	MOS Sep↑	MOS Rest↑	MOS Overall↑	Sep Rank↓	Rest Rank↓	Overall Rank↓	Macro Rank↓
1	xlancelab Public weights · Inference code Inference codeWeight license not declared	4.2358	3.3892	3.4665	1	1	1	1.00
2	CUPAudioGroup Public weights · Inference code Inference codeWeight license not declared	3.8360	2.9173	2.9253	2	2	2	2.00
3	AC_DC Public weights · Inference code Inference codeWeight license not declared	3.5425	2.4768	2.5412	5	4	4	4.33

Original records (3)

MSR Challenge 2025 official results ↗Checked: 2026-10-25

Related use-case guides

Find processing methods and steps suited to your material.

BS-RoFormer six-stem separation · online tool and model guide