Drums to MIDI · TelkNet

MIDI / transcriptionGPU workloadGitHub 0ac2fe6a35a7968d70ecfb8866cad025c350f5bd / 2038fa6d9c509d824e66784db242134daf7340081d78d5202879bf4c20b4598e / msgld_obtb

Drums to MIDI

ADTOF Plus · ADTOF Direct · ADT_STR · MSG-LD. Full mixes go through the model’s drum-extraction pipeline first. Isolated drums go directly to transcription or drum-stem separation.

15 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

5 × WAV · MIDI · CSV

Use case

Edit notes, instrumentation and performance details.

Choose a model family

MSG-LDAvailableMSG-LD generates five drum stems and detects onsets; MIDI uses fixed velocity 100 and playback duration 0.1 seconds.ADTOFAvailableADTOF Plus adds drum separation, velocity mapping and open/closed hi-hat processing to ADTOF transcription and accepts full mixes.ADT_STRAvailableADT_STR detects onsets and velocities for 26 drum classes and exports MIDI.

Choose a version or processing mode

MSG-LD · OBTBTrained with onset and timbre auxiliary branches; generates five drum stems and detects hits.RecommendedAvailableMSG-LD · OBTrained with an onset auxiliary branch; generates five drum stems and detects hits.Available

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

6 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.

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：MDB · Overall · ENST · Overall

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

MSG-LD · Tables 3–4 ↗Checked: 2026-09-27

View chart data

Rank	Model	F1 (%)
1	ADTOF	75.2
2	MSG-LD OBUsed on this site	67.35
3	MSG-LD Vanilla	66.7
4	MSG-LD OBTBUsed on this site	66.3
5	MSG-LD TB	66.2
6	LarsNet	55.3

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.

Public leaderboard: Model comparison

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

ADTOF · MDBDrums++

ADTOF-PyTorch README

ADTOF-PyTorch README · PyTorch 88.51 · Keras 88.74

Public sources checked: 2026-09-27

Rank	Test setting	Model	F-measure (%)	Detailed metrics
#1	Original Keras runtime	ADTOF (original)Pinned README table	88.74	Precision (%)89.90Recall (%)88.68
#2	Converted PyTorch runtime	ADTOF Direct · PyTorchPinned README table	88.51	Precision (%)89.83Recall (%)88.26

Drums to MIDI

ADTOF Plus paper, Table 1

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: 2026-09-27

Rank	Test setting	Model	F1 (%)	Detailed metrics
#1	MDB · 5 classes · onset only · 50 ms	ADTOF PlusPaper Table 1	89.00	Precision (%)89.00Recall (%)89.00
#2	MDB · 5 classes · onset only · 50 ms	Original ADTOFPaper Table 1	87.00
#2	RBMA · 5 classes · onset only · 50 ms	ADTOF PlusPaper Table 1	63.00	Precision (%)66.00Recall (%)62.00
#1	RBMA · 5 classes · onset only · 50 ms	Original ADTOFPaper Table 1	65.00
#1	ENST · 5 classes · onset only · 50 ms	ADTOF PlusPaper Table 1	85.00	Precision (%)91.00Recall (%)81.00
#2	ENST · 5 classes · onset only · 50 ms	Original ADTOFPaper Table 1	84.00

Drums to MIDI

ADTOF Plus paper, Table 2

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: 2026-09-27

Rank	Test setting	Model	F1 (%)	Detailed metrics
#1	MDB · 8-class evaluation · onset only · 50 ms	ADTOF PlusPaper Table 2	84.00	Precision (%)84.00Recall (%)84.00
#2	MDB · 8-class evaluation · onset only · 50 ms	Vogl et al. 8-class baselinePaper Table 2	72.00
#2	RBMA · 8-class evaluation · onset only · 50 ms	ADTOF PlusPaper Table 2	56.00	Precision (%)60.00Recall (%)55.00
#1	RBMA · 8-class evaluation · onset only · 50 ms	Vogl et al. 8-class baselinePaper Table 2	58.00
#1	ENST · 8-class evaluation · onset only · 50 ms	ADTOF PlusPaper Table 2	76.00	Precision (%)81.00Recall (%)72.00
#2	ENST · 8-class evaluation · onset only · 50 ms	Vogl et al. 8-class baselinePaper Table 2	65.00

ADT-STR · ENST / MDB

ADT-STR · Table 2

Table 2 Setting-1 through Setting-4 are training experiments, not the selectable τ=0.4/0.6/0.8 weights. Scores use isolated drums and do not establish full-mix performance.

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 8/8 rows · 6 metrics

ADT-STR · ENST / MDB
Rank	Test setting	Model	SUM F1 (%)	BD F1 (%)	SD F1 (%)	TT F1 (%)	HH F1 (%)	CY+RD F1 (%)
#1	ENST · ADT-STR Table 2	ADT-STR Setting-1	73.00	87.00	80.00	55.00	77.00	49.00
#2	ENST · ADT-STR Table 2	ADT-STR Setting-3	70.00	85.00	75.00	42.00	73.00	52.00
#3	ENST · ADT-STR Table 2	ADT-STR Setting-4	69.00	83.00	75.00	58.00	69.00	50.00
#4	ENST · ADT-STR Table 2	ADT-STR Setting-2	67.00	84.00	77.00	43.00	75.00	38.00
#1	MDB · ADT-STR Table 2	ADT-STR Setting-1	79.00	92.00	85.00	77.00	74.00	52.00
#1	MDB · ADT-STR Table 2	ADT-STR Setting-4	79.00	91.00	85.00	75.00	77.00	52.00
#3	MDB · ADT-STR Table 2	ADT-STR Setting-3	76.00	93.00	83.00	46.00	73.00	44.00
#4	MDB · ADT-STR Table 2	ADT-STR Setting-2	73.00	91.00	79.00	45.00	72.00	43.00

MSG-LD · MDB / ENST · 50 ms

MSG-LD · Table 3

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 36/36 rows · 3 metrics

MSG-LD · MDB / ENST · 50 ms
Rank	Test setting	Model	F1 (%)	Precision (%)	Recall (%)
#1	MDB · Kick · 50 ms · MSG-LD Table 3	MSG-LD OB	93.10	93.90	92.30
#2	MDB · Kick · 50 ms · MSG-LD Table 3	LarsNet	89.90	89.30	90.60
#3	MDB · Kick · 50 ms · MSG-LD Table 3	ADTOF	85.10	79.00	92.30
#1	MDB · Snare · 50 ms · MSG-LD Table 3	MSG-LD OB	76.00	79.10	73.20
#2	MDB · Snare · 50 ms · MSG-LD Table 3	ADTOF	75.20	82.80	69.00
#3	MDB · Snare · 50 ms · MSG-LD Table 3	LarsNet	75.10	82.10	69.20
#1	MDB · Toms · 50 ms · MSG-LD Table 3	ADTOF	52.00	52.00	52.00
#2	MDB · Toms · 50 ms · MSG-LD Table 3	MSG-LD OB	26.80	30.90	23.60
#3	MDB · Toms · 50 ms · MSG-LD Table 3	LarsNet	26.10	42.10	18.90
#1	MDB · Hi-Hats · 50 ms · MSG-LD Table 3	ADTOF	77.30	85.90	70.30
#2	MDB · Hi-Hats · 50 ms · MSG-LD Table 3	MSG-LD OB	65.00	55.00	79.40
#3	MDB · Hi-Hats · 50 ms · MSG-LD Table 3	LarsNet	41.00	40.50	41.50
#1	MDB · Cymbals · 50 ms · MSG-LD Table 3	ADTOF	84.90	80.20	90.20
#2	MDB · Cymbals · 50 ms · MSG-LD Table 3	MSG-LD OB	48.10	66.30	37.70
#3	MDB · Cymbals · 50 ms · MSG-LD Table 3	LarsNet	21.70	87.30	12.40
#1	MDB · Overall · 50 ms · MSG-LD Table 3	ADTOF	79.50	81.30	77.80
#2	MDB · Overall · 50 ms · MSG-LD Table 3	MSG-LD OB	70.70	70.90	70.40
#3	MDB · Overall · 50 ms · MSG-LD Table 3	LarsNet	61.30	70.00	54.50
#1	ENST · Kick · 50 ms · MSG-LD Table 3	MSG-LD OB	82.10	94.10	72.90
#2	ENST · Kick · 50 ms · MSG-LD Table 3	ADTOF	79.50	98.40	66.70
#3	ENST · Kick · 50 ms · MSG-LD Table 3	LarsNet	77.90	89.30	69.00
#1	ENST · Snare · 50 ms · MSG-LD Table 3	MSG-LD OB	64.20	69.80	59.40
#2	ENST · Snare · 50 ms · MSG-LD Table 3	LarsNet	60.60	78.70	49.30
#3	ENST · Snare · 50 ms · MSG-LD Table 3	ADTOF	59.90	93.20	44.10
#1	ENST · Toms · 50 ms · MSG-LD Table 3	ADTOF	58.10	66.90	51.30
#2	ENST · Toms · 50 ms · MSG-LD Table 3	MSG-LD OB	49.10	41.50	60.20
#3	ENST · Toms · 50 ms · MSG-LD Table 3	LarsNet	34.80	23.10	69.80
#1	ENST · Hi-Hats · 50 ms · MSG-LD Table 3	ADTOF	77.70	93.60	66.30
#2	ENST · Hi-Hats · 50 ms · MSG-LD Table 3	MSG-LD OB	63.00	64.30	61.80
#3	ENST · Hi-Hats · 50 ms · MSG-LD Table 3	LarsNet	34.00	69.90	22.50
#1	ENST · Cymbals · 50 ms · MSG-LD Table 3	ADTOF	65.70	78.80	56.30
#2	ENST · Cymbals · 50 ms · MSG-LD Table 3	MSG-LD OB	39.10	61.10	28.70
#3	ENST · Cymbals · 50 ms · MSG-LD Table 3	LarsNet	34.90	24.30	61.90
#1	ENST · Overall · 50 ms · MSG-LD Table 3	ADTOF	70.90	90.80	58.20
#2	ENST · Overall · 50 ms · MSG-LD Table 3	MSG-LD OB	64.00	69.50	59.30
#3	ENST · Overall · 50 ms · MSG-LD Table 3	LarsNet	49.30	50.40	48.30

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 12/12 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 TB	69.80	70.60	69.10
#3	MDB Overall · MSG-LD Table 4	MSG-LD OBTBTelkNet	69.00	70.90	67.20
#4	MDB Overall · MSG-LD Table 4	MSG-LD Vanilla	68.90	70.10	67.80
#1	ENST Overall · MSG-LD Table 4	MSG-LD Vanilla	64.50	71.40	58.80
#2	ENST Overall · MSG-LD Table 4	MSG-LD OB	64.00	69.50	59.30
#3	ENST Overall · MSG-LD Table 4	MSG-LD OBTBTelkNet	63.60	70.30	58.00
#4	ENST Overall · MSG-LD Table 4	MSG-LD TB	62.60	67.70	58.20
#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 Vanilla	66.70	—	—
#3	Mean overall F1 (2 sets) · MSG-LD Table 4	MSG-LD OBTBTelkNet	66.30	—	—
#4	Mean overall F1 (2 sets) · MSG-LD Table 4	MSG-LD TB	66.20	—	—

Models, parameters, and sources

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

MSG-LD · OBTBddman1101/Separate-and-detectMSG-LD · checkpointMSG-LDADTOF-PyTorch READMEADTOF (original)Original Keras runtimePinned README tableADTOF Direct · PyTorchPyTorch converted runtimeADTOF Plus paper, Table 1ADTOF Plus

MSG-LD · OBTB

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

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.

ADTOF Plus

ADTOF Plus adds drum separation, velocity mapping and open/closed hi-hat processing to ADTOF transcription and accepts full mixes.

ADTOF Plus source (pinned)ADTOF PyTorch source (pinned)ADTOF Plus paper ADTOF ISMIR 2021 paper Pinned ADTOF Plus technical guide

Official ADTOF Plus input_is_mix control. Turn it on for a complete mix so MDX23C separation runs; turn it off only when the upload is already an isolated drum track.
Optional official ADTOF Plus default_threshold. Leave blank to keep the upstream -Infinity default; ADTOF PyTorch Direct does not use it.
After inference, Beat This final0 writes a valid fixed tempo or beat-by-beat tempo map into the MIDI while preserving every event's position in seconds. If valid beat evidence cannot be produced, the task fails instead of delivering MIDI with the model's default tempo.

ADTOF Direct · PyTorch

ADTOF Direct transcribes isolated drums with PyTorch ADTOF. For full-mix input, TelkNet first extracts drums with AudioSep-Hive; it does not apply Plus velocity or open/closed hi-hat post-processing.

ADTOF PyTorch source (pinned)Original ADTOF source ISMIR 2021 paper materials Signals 2023 paper

Five separate official ADTOF thresholds for kick, snare, hi-hat, tom and cymbal; no comma-separated text entry.
Advanced control from the official ADTOF CLI. The upstream default is 100 FPS; changing it alters time resolution and may affect quality.
After inference, Beat This final0 writes a valid fixed tempo or beat-by-beat tempo map into the MIDI while preserving every event's position in seconds. If valid beat evidence cannot be produced, the task fails instead of delivering MIDI with the model's default tempo.

ADT_STR · tau=0.8

ADT_STR detects onsets and velocities for 26 drum classes and exports MIDI.

pier-maker92/ADT_STR ADT_STR · checkpoint ADT_STR

Select an officially released checkpoint variant.
Maximum generated tokens per segment; the official default is 1024.
Official inference batch size: the number of segments processed together.
Full mixes first pass through drum extraction: official Demucs for MSG-LD and AudioSep for the other new models. Isolated drums skip this step.
Table 2 Setting-1 through Setting-4 are training experiments, not the selectable τ=0.4/0.6/0.8 weights. Scores use isolated drums and do not establish full-mix performance.

ADT_STR · tau=0.4

ADT_STR detects onsets and velocities for 26 drum classes and exports MIDI.

pier-maker92/ADT_STR ADT_STR · checkpoint ADT_STR

Select an officially released checkpoint variant.
Maximum generated tokens per segment; the official default is 1024.
Official inference batch size: the number of segments processed together.
Full mixes first pass through drum extraction: official Demucs for MSG-LD and AudioSep for the other new models. Isolated drums skip this step.
Table 2 Setting-1 through Setting-4 are training experiments, not the selectable τ=0.4/0.6/0.8 weights. Scores use isolated drums and do not establish full-mix performance.

ADT_STR · tau=0.6

ADT_STR detects onsets and velocities for 26 drum classes and exports MIDI.

pier-maker92/ADT_STR ADT_STR · checkpoint ADT_STR

Select an officially released checkpoint variant.
Maximum generated tokens per segment; the official default is 1024.
Official inference batch size: the number of segments processed together.
Full mixes first pass through drum extraction: official Demucs for MSG-LD and AudioSep for the other new models. Isolated drums skip this step.
Table 2 Setting-1 through Setting-4 are training experiments, not the selectable τ=0.4/0.6/0.8 weights. Scores use isolated drums and do not establish full-mix performance.

Model comparison

Model

Role

Why it is used

Output impact

Paper and benchmark notes

Enhanced Automatic Drum Transcription via Drum Stem Source Separation

arXiv - 2025

Paper Pinned implementation

ADTOF: A large dataset of non-synthetic music for automatic drum transcription

ISMIR - 2021

Paper Paper materials

High-Quality and Reproducible Automatic Drum Transcription from Crowdsourced Data

Signals - 2023

Paper

Official / repositories

ddman1101/Separate-and-detect MSG-LD · checkpoint ADTOF Plus source (pinned)ADTOF PyTorch source (pinned)Original ADTOF source pier-maker92/ADT_STR ADT_STR · checkpoint

Papers / technical notes

MSG-LD ADTOF Plus paper ADTOF ISMIR 2021 paper Pinned ADTOF Plus technical guide ISMIR 2021 paper materials Signals 2023 paper ADT_STR

Parameter guide

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.

Workflow

1Full mixes go through the model’s drum-extraction pipeline first. Isolated drums go directly to transcription or drum-stem separation.
2Download drum MIDI and review kick, snare, hi-hat and other kit parts in a drum editor.

Before you submit

Full mixes go through the model’s drum-extraction pipeline first. Isolated drums go directly to transcription or drum-stem separation.
Advanced control from the official ADTOF CLI. The upstream default is 100 FPS; changing it alters time resolution and may affect quality.

Quick feedback

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