Multi-singer and choir separation · TelkNet

Separation & stem splitGPU workload

Multi-singer and choir separation

Separate multiple singers or choir parts with available vocal models. Review reference-audio requirements and the limits of overlapping voice separation.

25 credits/useFile inputMax 100 MBEstimated 60-1800 seconds

Input

Supports inputs in formats such as mp3 / wav / flac / ogg / m4a / aac / aif / aiff / mp4 / mov / mkv

Supported formats

mp3wavflacoggm4aaacaifaiffmp4movmkv

Output

WAV

Use case

Remix, practice with accompaniment and edit separate parts.

Choose a model family

Concat / FiLMAvailableConcat · dual lambda=0.1: Remix, practice with accompaniment and edit separate parts.Multi-Singer Separation · UNMIXXAvailableBlindly separate two or more concurrent singing voices into numbered singer stems with UNMIXX.Male & Female Vocal SeparationAvailableSplit overlapping male and female vocals with the released BS-RoFormer checkpoint.A Cappella Part Separation · SepACapAvailableSeparate alto, bass, finger snap, lead vocal, soprano, tenor and vocal percussion with SepACap.

Choose by input and output needs, then comparable tests, published evaluations and capabilities. Missing scores do not imply lower quality; more parameters or a newer release do not prove superiority. This is a suggested usage order, not an overall leaderboard.

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

Models in this field

10 entries

Published results include models on and off this site. The overall ranking opens first; switch the benchmark or category to see its ranking. Each benchmark keeps its own evaluation conditions.

Only models with verified public weights and inference code are included. Closed services and entries without verified weights are excluded.

BenchmarkRanking category

Overall rankingHigher is better · best first

The overall score is an equal mean of 2 results under matching conditions. Only entries with every component are included.

Score components：Singer 1 · Singer 2

MVSEP · Quality evaluation ↗Checked: 2026-09-28

View chart data

Rank	Model	Mean SDR (dB)
1	Vocals 163 model for Medley-Vox Public weights Inference code	4.1853
2	Vocals 135 model for Medley-Vox Public weights Inference code	4.1363
3	Vocals 200 model for Medley-Vox Public weights Inference code	4.0761
4	Vocals 188 model for Medley-Vox Public weights Inference code	3.9958
5	Multi singing LibriSpeech model for Medley-Vox Public weights Inference code	3.8013
6	Vocals 238 model for Medley-Vox Public weights Inference code	3.7333
7	BS-RoFormer (Male/Female by Sucial) Public weights Inference code	3.7128
8	Vocal 231 model for Medley-Vox Public weights Inference code	3.6515
9	Multi singing LibriSpeech 138 epoch model for Medley-Vox Public weights Inference code	3.2981
10	Mel-RoFormer (karaoke by becruily) Public weights Inference code	3.1797

All original results and sources

Scores apply only to the cited dataset, metric and model version. Leading this table does not mean a current world record or a result from this service. Capabilities and tests can justify prioritizing models without comparable scores.

Benchmark chart: Reference Singer Separation / Multi-Singer Separation · UNMIXX / Male & Female Vocal Separation / A Cappella Part Separation · SepACap

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

Reference Singer Separation

Singer-reference paper, Table 1

Scores come from the linked base checkpoint, not a new measurement of the model running here.

Public sources checked: 2026-09-10

Rank	Test setting	Model	SI-SDR (dB) ↑	Detailed metrics
#1	Duet target singer	Concat dual λ=0.1Corresponding tool model: Singer Reference · Concat dual λ=0.1Paper Table 1	5.58Best	Both singers · SI-SDR (dB) ↑0.59BestBackground · SI-SDR (dB) ↑-5.91BestInterfering singer + background · SI-SDR (dB) ↑7.98Best
#2	Duet target singer	FiLM layer-varyingCorresponding tool model: Singer Reference · FiLM layer-varyingPaper Table 1	3.39	Both singers · SI-SDR (dB) ↑0.17Background · SI-SDR (dB) ↑-10.14Interfering singer + background · SI-SDR (dB) ↑-2.99

Multi-Singer Separation · UNMIXX

UNMIXX paper, Table 3

Values come from the linked public source. Row labels identify the test set or checkpoint; models without a matching result are omitted.

Public sources checked: 2026-09-10

Rank	Test setting	Model	SDRi (dB) ↑	Detailed metrics
MedleyVox duet	MedleyVox duet	UNMIXXUNMIXX paper Table 3	17.52	SI-SDRi (dB) ↑16.47H/SSNR (dB) ↑15.96
MedleyVox unison	MedleyVox unison	UNMIXXUNMIXX paper Table 3	7.16	SI-SDRi (dB) ↑6.58H/SSNR (dB) ↑10.50

Male & Female Vocal Separation

Author model card · epoch 267

Values come from the linked public source. Row labels identify the test set or checkpoint; models without a matching result are omitted.

Public sources checked: 2026-09-10

Test setting	Model	SDR (dB) ↑
opencpop + m4singer · 50-song validation · epoch 267	Chorus Male/Female BS-RoFormerTelkNetmodel_chorus_bs_roformer_ep_267_sdr_24.1275.ckpt	24.1275

A Cappella Part Separation · SepACap

SepACap paper, Table 3

Values come from the linked public source. Row labels identify the test set or checkpoint; models without a matching result are omitted.

Public sources checked: 2026-09-10

Rank	Test setting	Model	SDRi (dB) ↑
Alto	Alto	SepACapSepACap paper Table 3 · finger snap is not evaluated	14.6
Bass	Bass	SepACapSepACap paper Table 3 · finger snap is not evaluated	23.2
Lead vocal	Lead vocal	SepACapSepACap paper Table 3 · finger snap is not evaluated	13.0
Soprano	Soprano	SepACapSepACap paper Table 3 · finger snap is not evaluated	13.1
Tenor	Tenor	SepACapSepACap paper Table 3 · finger snap is not evaluated	17.0
Vocal percussion	Vocal percussion	SepACapSepACap paper Table 3 · finger snap is not evaluated	22.5

Models, parameters, and sources

Split overlapping male and female vocals with the released BS-RoFormer checkpoint.

Concat · dual lambda=0.1Singer-reference source and official checkpoint links (pinned)Singer-reference separation paperSinger-reference paper, Table 1Concat dual λ=0.1Duet target singerPaper Table 1FiLM layer-varyingConcat layer-varyingBaselineSinger-Informed Vocal Source Separation for Multi-Singer Music MixturesIWAENC

Concat · dual lambda=0.1

Concat · dual lambda=0.1: Remix, practice with accompaniment and edit separate parts.

Singer-reference source and official checkpoint links (pinned)Singer-reference separation paper

FiLM · Layer-varying

FiLM · Layer-varying: Remix, practice with accompaniment and edit separate parts.

Singer-reference source and official checkpoint links (pinned)Singer-reference separation paper

Model comparison

Model

Role

Why it is used

Output impact

Paper and benchmark notes

Singer-Informed Vocal Source Separation for Multi-Singer Music Mixtures

IWAENC - 2026

Extract the singer matching a separate enrollment recording with the official Concat dual λ=0.1 or FiLM layer-varying checkpoint.

Paper

UNMIXX: Untangling Highly Correlated Singing Voices Mixtures

ICASSP - 2026

Blindly separate two or more concurrent singing voices into numbered singer stems with UNMIXX.

Paper

Source Separation for A Cappella Music

arXiv - 2025

Separate alto, bass, finger snap, lead vocal, soprano, tenor and vocal percussion with SepACap.

Paper

Official / repositories

Singer-reference source and official checkpoint links (pinned)UNMIXX source and bundled checkpoint (pinned)MSST inference source (pinned)Official chorus gender checkpoint SepACap source (pinned)SepACap checkpoint (pinned)

Papers / technical notes

Singer-reference separation paper UNMIXX paper SepACap paper

Parameter guide

Official MSST test-shift count. Higher values cost more time and may improve stability.
Official test-time augmentation. It increases processing time and is disabled by default.

Workflow

1Upload overlapping male and female vocals.
2Download both vocal tracks and listen for bleed against the original.

Before you submit

Official MSST test-shift count. Higher values cost more time and may improve stability.
Official test-time augmentation. It increases processing time and is disabled by default.

Quick feedback

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

Multi-singer and choir separation

Models in this field

Benchmark chart: Reference Singer Separation / Multi-Singer Separation · UNMIXX / Male & Female Vocal Separation / A Cappella Part Separation · SepACap

Reference Singer Separation

Multi-Singer Separation · UNMIXX

Male & Female Vocal Separation

A Cappella Part Separation · SepACap

Models, parameters, and sources

Concat · dual lambda=0.1

FiLM · Layer-varying

UNMIXX

Chorus Male/Female BS-RoFormer

SepACap

Model comparison

Paper and benchmark notes

Singer-Informed Vocal Source Separation for Multi-Singer Music Mixtures

UNMIXX: Untangling Highly Correlated Singing Voices Mixtures

Source Separation for A Cappella Music

Official / repositories

Papers / technical notes

Parameter guide

Workflow

Before you submit

Quick feedback