Drum separation · TelkNet

Separation & stem splitGPU workloadGitHub 456656868538205ef756912c7cf5b0fd936de8af / 5856a9bee7c6d503842795756d238dc8470f6f3e010e9e4f33ede0362850cb4c

Drum separation

Separate drum recordings into editable instrument stems with DrumSep, IDM or MSG-LD; compare supported drums and outputs, then preview and download WAV files.

35 credits/useFile inputMax 100 MBEstimated 20-1800 seconds

Input

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

Supported formats

mp3wavflacoggm4aaacaifaiffmp4movmkv

Output

9 × WAV

Use case

Split drums into individual stems with IDM or MSG-LD.

Choose a model family

MDX23C DrumSepAvailableMDX23C DrumSep · 5 stems: Drum remixing, replacement, transient editing, practice tracks, and drum-focused analysis.Drum separation · IDMAvailableIDM exports nine drum stems without MIDI.MSG-LDAvailableMSG-LD generates five drum stems and detects onsets; MIDI uses fixed velocity 100 and playback duration 0.1 seconds.

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

2 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 5 results under matching conditions. Only entries with every component are included.

Score components：Kick · Snare · Toms · Hi-hat · Cymbals

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

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

View chart data

Rank	Model	Mean SDR (dB)
1	MDX23C DrumSep 5 stems by Aufr33 & JarredouUsed on this site Public weights Inference code	10.1846
2	MDX23C DrumSep 6 stems by Aufr33 & JarredouUsed on this site Public weights Inference code	8.4534

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: Model comparison

Scores come from the cited public sources; blue rows are models available in this tool.

Paper and benchmark notes

Pinned MSST pretrained-model table

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: 2026-09-23

Test setting	Model	SDR (dB) ↑	Detailed metrics
DrumSep test · official five-stem checkpoint	MDX23C DrumSep · 5 stemsMSST table · kick SDR is the headline value	16.66	Snare11.53Toms12.33Hi-hat4.04Cymbals6.36
DrumSep test · official six-stem checkpoint	MDX23C DrumSep · 6 stemsMSST table · kick SDR is the headline value; ride and crash are reported jointly	14.54	Snare9.79Toms10.63Hi-hat3.19Ride + crash6.08

IDM · StemGMD · 9 classes

IDM · Table III

IDM outputs nine drum WAV stems. nSDR, LSD and PES measure audio separation, not MIDI transcription, and are not ranked against MSG-LD FAD.

Ranks are computed within each table and evaluation context, retaining ties. Change the metric to reorder; scroll horizontally for every column. Missing values are unranked, never zero.

Public sources checked: 2026-09-23

Rank by metricTest setting

Showing 9/9 rows · 5 metrics

IDM · StemGMD · 9 classes
Rank	Test setting	Model	nSDR (dB)	LSD	PES	Synth LSD	Synth PES
—	kick	IDMTelkNet	17.47	2.58	-54.37	1.84	-55.46
—	snare	IDMTelkNet	17.29	1.73	-57.20	1.83	-58.87
—	hihat_closed	IDMTelkNet	1.66	1.90	-55.88	1.94	-56.83
—	hihat_open	IDMTelkNet	6.68	1.71	-59.81	1.86	-59.99
—	hi_tom	IDMTelkNet	8.50	1.16	-59.57	1.03	-59.85
—	mid_tom	IDMTelkNet	6.41	1.11	-59.62	0.90	-59.70
—	low_tom	IDMTelkNet	7.35	1.62	-59.30	1.32	-59.57
—	ride	IDMTelkNet	2.12	2.09	-59.57	2.24	-59.65
—	crash_left	IDMTelkNet	4.78	2.84	-59.88	2.99	-59.94

MSG-LD · mel-MSE

MSG-LD · Table 1

The paper reports OB, TB and OBTB separately; OB scores do not describe the default OBTB. Note F1 and audio mel-MSE/FAD are separate tables. The paper's detection pipeline has not been remeasured here.

Ranks are computed within each table and evaluation context, retaining ties. Change the metric to reorder; scroll horizontally for every column. Missing values are unranked, never zero.

Public sources checked: 2026-09-23

Rank by metricTest setting

Showing 12/12 rows · 1 metrics

MSG-LD · mel-MSE
Rank	Test setting	Model	mel-MSE
#1	Kick · MSG-LD Table 1	MSG-LD OBTB	0.78
#2	Kick · MSG-LD Table 1	MSG-LD OB	1.03
#1	Snare · MSG-LD Table 1	MSG-LD OB	1.33
#2	Snare · MSG-LD Table 1	MSG-LD OBTB	1.42
#1	Toms · MSG-LD Table 1	MSG-LD OBTB	1.41
#2	Toms · MSG-LD Table 1	MSG-LD OB	1.45
#1	Hi-Hats · MSG-LD Table 1	MSG-LD OB	3.46
#2	Hi-Hats · MSG-LD Table 1	MSG-LD OBTB	4.30
#1	Cymbals · MSG-LD Table 1	MSG-LD OB	1.92
#2	Cymbals · MSG-LD Table 1	MSG-LD OBTB	3.52
#1	Overall · MSG-LD Table 1	MSG-LD OB	2.12
#2	Overall · MSG-LD Table 1	MSG-LD OBTB	2.29

MSG-LD · FAD

MSG-LD · Table 2

Ranks are computed within each table and evaluation context, retaining ties. Change the metric to reorder; scroll horizontally for every column. Missing values are unranked, never zero.

Public sources checked: 2026-09-23

Rank by metricTest setting

Showing 12/12 rows · 1 metrics

MSG-LD · FAD
Rank	Test setting	Model	FAD
#1	Kick · MSG-LD Table 2	MSG-LD OBTB	0.23
#2	Kick · MSG-LD Table 2	MSG-LD OB	0.32
#1	Snare · MSG-LD Table 2	MSG-LD OBTB	0.09
#2	Snare · MSG-LD Table 2	MSG-LD OB	0.10
#1	Toms · MSG-LD Table 2	MSG-LD OBTB	0.13
#2	Toms · MSG-LD Table 2	MSG-LD OB	0.16
#1	Hi-Hats · MSG-LD Table 2	MSG-LD OB	0.38
#2	Hi-Hats · MSG-LD Table 2	MSG-LD OBTB	0.48
#1	Cymbals · MSG-LD Table 2	MSG-LD OB	0.71
#2	Cymbals · MSG-LD Table 2	MSG-LD OBTB	0.82
#1	Mean · MSG-LD Table 2	MSG-LD OB	0.33
#2	Mean · MSG-LD Table 2	MSG-LD OBTB	0.35

MSG-LD · MDB / ENST · 50 ms

MSG-LD · Table 3

Ranks are computed within each table and evaluation context, retaining ties. Change the metric to reorder; scroll horizontally for every column. Missing values are unranked, never zero.

Public sources checked: 2026-09-23

Rank by metricTest setting

Showing 12/12 rows · 3 metrics

MSG-LD · MDB / ENST · 50 ms
Rank	Test setting	Model	F1 (%)	Precision (%)	Recall (%)
—	MDB · Kick · 50 ms · MSG-LD Table 3	MSG-LD OB	93.10	93.90	92.30
—	MDB · Snare · 50 ms · MSG-LD Table 3	MSG-LD OB	76.00	79.10	73.20
—	MDB · Toms · 50 ms · MSG-LD Table 3	MSG-LD OB	26.80	30.90	23.60
—	MDB · Hi-Hats · 50 ms · MSG-LD Table 3	MSG-LD OB	65.00	55.00	79.40
—	MDB · Cymbals · 50 ms · MSG-LD Table 3	MSG-LD OB	48.10	66.30	37.70
—	MDB · Overall · 50 ms · MSG-LD Table 3	MSG-LD OB	70.70	70.90	70.40
—	ENST · Kick · 50 ms · MSG-LD Table 3	MSG-LD OB	82.10	94.10	72.90
—	ENST · Snare · 50 ms · MSG-LD Table 3	MSG-LD OB	64.20	69.80	59.40
—	ENST · Toms · 50 ms · MSG-LD Table 3	MSG-LD OB	49.10	41.50	60.20
—	ENST · Hi-Hats · 50 ms · MSG-LD Table 3	MSG-LD OB	63.00	64.30	61.80
—	ENST · Cymbals · 50 ms · MSG-LD Table 3	MSG-LD OB	39.10	61.10	28.70
—	ENST · Overall · 50 ms · MSG-LD Table 3	MSG-LD OB	64.00	69.50	59.30

MSG-LD · Vanilla / OB / TB / OBTB

MSG-LD · Table 4

Ranks are computed within each table and evaluation context, retaining ties. Change the metric to reorder; scroll horizontally for every column. Missing values are unranked, never zero.

Public sources checked: 2026-09-23

Rank by metricTest setting

Showing 6/6 rows · 3 metrics

MSG-LD · Vanilla / OB / TB / OBTB
Rank	Test setting	Model	F1 (%)	Precision (%)	Recall (%)
#1	MDB Overall · MSG-LD Table 4	MSG-LD OB	70.70	70.90	70.40
#2	MDB Overall · MSG-LD Table 4	MSG-LD OBTB	69.00	70.90	67.20
#1	ENST Overall · MSG-LD Table 4	MSG-LD OB	64.00	69.50	59.30
#2	ENST Overall · MSG-LD Table 4	MSG-LD OBTB	63.60	70.30	58.00
#1	Mean overall F1 (2 sets) · MSG-LD Table 4	MSG-LD OB	67.40	—	—
#2	Mean overall F1 (2 sets) · MSG-LD Table 4	MSG-LD OBTB	66.30	—	—

Models, parameters, and sources

IDM exports nine drum stems without MIDI.

MDX23C DrumSep · 5 stemsOfficial 5-stem checkpointOfficial 5-stem configPinned immutable mirror used by TelkNetPinned MSST pretrained-model tableDrumSep test · official five-stem checkpointMSST table · kick SDR is the headline valueMDX23C DrumSep · 6 stemsDrumSep test · official six-stem checkpointMSST table · kick SDR is the headline value; ride and crash are reported jointlyA Semantically Consistent Dataset for Data-Efficient Query-Based Universal Sound SeparationICML

MDX23C DrumSep · 5 stems

MDX23C DrumSep · 5 stems: Drum remixing, replacement, transient editing, practice tracks, and drum-focused analysis.

Official 5-stem checkpoint Official 5-stem config Pinned immutable mirror used by TelkNet

MDX23C DrumSep · 6 stems

MDX23C DrumSep · 6 stems: Drum remixing, replacement, transient editing, practice tracks, and drum-focused analysis.

Official 6-stem checkpoint Official 6-stem config Pinned checkpoint bytes used by TelkNet

IDM

bernardo-torres/inverse-drum-machine IDM · checkpoint ddman1101/Separate-and-detect MSG-LD · checkpoint IDM

wiener uses the official Wiener mask; none disables masking.
Power exponent of the Wiener mask; the official default is 1.
Full mixes first pass through drum extraction: official Demucs for MSG-LD and AudioSep for the other new models. Isolated drums skip this step.
IDM outputs nine drum WAV stems. nSDR, LSD and PES measure audio separation, not MIDI transcription, and are not ranked against MSG-LD FAD.

MSG-LD · OBTB

MSG-LD generates five drum stems and detects onsets; MIDI uses fixed velocity 100 and playback duration 0.1 seconds.

ddman1101/Separate-and-detect MSG-LD · checkpoint MSG-LD

msgld_ob uses the onset branch; msgld_obtb uses both onset and timbre branches.
Official DDIM sampling steps, default 200; unsupported timestep combinations produce an error.
Full mixes first pass through drum extraction: official Demucs for MSG-LD and AudioSep for the other new models. Isolated drums skip this step.
The paper reports OB, TB and OBTB separately; OB scores do not describe the default OBTB. Note F1 and audio mel-MSE/FAD are separate tables. The paper's detection pipeline has not been remeasured here.

MSG-LD · OB

MSG-LD generates five drum stems and detects onsets; MIDI uses fixed velocity 100 and playback duration 0.1 seconds.

ddman1101/Separate-and-detect MSG-LD · checkpoint MSG-LD

msgld_ob uses the onset branch; msgld_obtb uses both onset and timbre branches.
Official DDIM sampling steps, default 200; unsupported timestep combinations produce an error.
Full mixes first pass through drum extraction: official Demucs for MSG-LD and AudioSep for the other new models. Isolated drums skip this step.
The paper reports OB, TB and OBTB separately; OB scores do not describe the default OBTB. Note F1 and audio mel-MSE/FAD are separate tables. The paper's detection pipeline has not been remeasured here.

Model comparison

Model

Role

Why it is used

Output impact

Paper and benchmark notes

A Semantically Consistent Dataset for Data-Efficient Query-Based Universal Sound Separation

ICML - 2026

Extract an open-vocabulary sound with AudioSep-Hive, FlowSep-Hive or CLAPSep, or isolate a source that matches a 10-second reference with Banquet.

arXiv:2601.22599

The Inverse Drum Machine: Source Separation Through Joint Transcription and Analysis-by-Synthesis

arXiv - 2025

IDM

Separate-and-Detect: Unified Drum Transcription and Stem Generation via Latent Diffusion

arXiv - 2026

MSG-LD generates five drum stems and detects onsets; MIDI uses fixed velocity 100 and playback duration 0.1 seconds.

MSG-LD

Official / repositories

Official 5-stem checkpoint Official 5-stem config Pinned immutable mirror used by TelkNet Official 6-stem checkpoint Official 6-stem config Pinned checkpoint bytes used by TelkNet bernardo-torres/inverse-drum-machine IDM · checkpoint ddman1101/Separate-and-detect MSG-LD · checkpoint

Papers / technical notes

IDM MSG-LD

Parameter guide

wiener uses the official Wiener mask; none disables masking.
Power exponent of the Wiener mask; the official default is 1.
Full mixes first pass through drum extraction: official Demucs for MSG-LD and AudioSep for the other new models. Isolated drums skip this step.
IDM outputs nine drum WAV stems. nSDR, LSD and PES measure audio separation, not MIDI transcription, and are not ranked against MSG-LD FAD.

Workflow

1Full mixes first pass through drum extraction: official Demucs for MSG-LD and AudioSep for the other new models. Isolated drums skip this step.
2Choose IDM or MSG-LD and submit the recording, then preview and download the WAV stems; MSG-LD also provides onset CSV and MIDI.

Before you submit

IDM produces nine drum classes and MSG-LD five; the MIDI represents drum onsets, not melodic pitches.

Quick feedback

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