Saxophone separation · Demucs v3 Saxophone / Mega 53 · TelkNet

Separation & stem splitGPU workloaddemucs_v3_saxophone_separation · Weights

Saxophone separation

The specialist Demucs v3 checkpoint outputs saxophone and accompaniment stems; Mega 53 can render an A/B result for the same file.

20 credits/useFile inputMax 100 MBEstimated 45-900 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

2-stem saxophone/accompaniment · Demucs v3 SaxophoneAvailableExport saxophone and accompaniment as two stems for practice tracks or separate solo/background editing.Saxophone-only extraction · Mega 53AvailableExport only the saxophone stem for solo extraction; choose the Demucs two-stem version when accompaniment is needed.

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

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

Models, parameters, and sources

Demucs v3 Saxophone: Remix, practice with accompaniment and edit separate parts.

Demucs v3 SaxophoneRiley and Dixon, Table 2LALAL.AI Wind StemMega 53 · Saxophone

Demucs v3 Saxophone

Demucs saxophone weights Mega 53 reference

Mega 53 · Saxophone

Mega 53 · Saxophone: Remix, practice with accompaniment and edit separate parts.

Demucs saxophone weights Mega 53 reference

Official / repositories

Demucs saxophone weights Mega 53 reference

Papers / technical notes

No paper links are provided.

Workflow

1Upload a music recording containing saxophone.
2Preview and download the extracted saxophone track after processing.

Models in this field

2 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.

Published evaluation conditions

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.

Benchmark

Riley and Dixon, Table 2

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 1 / source rows 2

Not ranked: 1

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

LALAL.AI Wind Stem

The sources use different training or task settings, so compare numbers only within the stated protocol.

Rank by metric

Riley and Dixon, Table 2
Metric rank	Model	SDR (dB)↑
-	Ours (Demucs + FiloSax)Used on this site Public weights · Inference code Inference codeWeight license not declared	14.22 ± 2.10

Original records (1)

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

Riley and Dixon, Table 2 ↗Checked: 2026-10-25

All evaluation sources (2)

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

MVSEP · Wind ↗
Deployable entries 5 / source rows 26 · Checked: 2026-10-25
Not ranked: 21
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
MVSep Wind BS Roformer (2025.08) · sami-bytedance · MLSLABS WCJ · Wind BS Roformer (2026.03) · SCNet logic · MVSep Wind Ensemble (SCNet + Mel) (2024.10.20) · tt · Wind BS Roformer (2026.04) · w4 · t2 · OtherMix · t4 · MVSep Wind SCNet Large (2024.10.18) · MVSep Wind MelBand Roformer (2024.10.18) · Bs Roformer Step 2 · SCNet+Bs2 · SCNet Large (wind) · test_wind_0908 · BS Roformer SW Wind · X-Minus Wind (Mel-RoFormer)
Riley and Dixon, Table 2 ↗
Deployable entries 1 / source rows 2
Not ranked: 1
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
LALAL.AI Wind Stem

All original results and sources

Scores apply to the listed dataset, metric and model version; results from different tests are not directly comparable.

Riley and Dixon, Table 2

Deployable entries 1 / source rows 2

Not ranked: 1

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

LALAL.AI Wind Stem

The sources use different training or task settings, so compare numbers only within the stated protocol.

Rank by metric

Riley and Dixon, Table 2
Metric rank	Model	SDR (dB)↑
-	Ours (Demucs + FiloSax) Public weights · Inference code Inference codeWeight license not declared	14.22 ± 2.10

Original records (1)

Riley and Dixon, Table 2 ↗Checked: 2026-10-25

Saxophone separation

Models, parameters, and sources

Demucs v3 Saxophone

Mega 53 · Saxophone

Official / repositories

Papers / technical notes

Workflow

Riley and Dixon, Table 2

Riley and Dixon, Table 2

Quick feedback