Vocals to MIDI · ROSVOT / GAME · TelkNet

MIDI / transcriptionGPU workloadSOME · Source SOME v1.0.0-baseline · Weights SOME v1.0.0-baseline · Changelog SOME · Paper

Vocals to MIDI

Convert singing, humming or vocal stems to MIDI. Compare ROSVOT, GAME, SOME and VOCANO results before correcting the melody.

25 credits/useFile inputMax 100 MBEstimated 45-1800 seconds

Input

Upload solo singing, humming or an isolated lead vocal. If you only have a finished song, choose “Full song” and the lead vocal will be separated before transcription.

Supported formats

mp3wavflacoggm4aaacaifaiffmp4movmkv

Output

MIDI

Use case

Copying a sung melody into a DAW for editing, doubling, synth programming or a notation draft.

Choose a model family

Vocals to MIDI · ROSVOTAvailableRobust path for singing with noise or accompaniment leakage; the current official checkpoint is trained mainly on Mandarin singing.Vocals to MIDI · GAMEAvailableTurn singing into MIDI, with controls for singing language, segmentation and note detection.Vocals to MIDI · SOMEAvailableSOME is the fast, lightweight singing-to-MIDI route.Vocals to MIDI · VOCANOAvailableVOCANO extracts vocal pitch with Patch-CNN and segments notes with PyramidNet-110. Its released inference entry point has no additional public model controls.

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

SOME is the fast, lightweight singing-to-MIDI route.

ROSVOTROSVOT paper, Table 1ROSVOT + RWBDVOCANO: A Note Transcription Framework for Singing Voice in Polyphonic MusicRobust Singing Voice Transcription Serves SynthesisGAME: Generative Adaptive MIDI ExtractorSOME: Singing-Oriented MIDI ExtractorGAMESOMEVOCANO

ROSVOT

Robust path for singing with noise or accompaniment leakage; the current official checkpoint is trained mainly on Mandarin singing.

ROSVOT source (pinned)ROSVOT project page ROSVOT paper

ROSVOT note-boundary threshold: 0.85 by default; accepts 0–1.
ROSVOT word-boundary threshold: 0.5 by default; accepts 0–1.

GAME

Turn singing into MIDI, with controls for singing language, segmentation and note detection.

GAME v1.0.3 source (pinned)GAME v1.0.3 weights GAME v1.0.3 technical guide (pinned)GAME v1.0.3 ONNX guide (pinned)

Leave this blank to use GAME's official language ID 0. This is not automatic language detection; select a language when you know what is being sung.
GAME boundary-segmentation confidence threshold.
GAME boundary search radius in seconds.
Starting noise time used by GAME’s official D3PM schedule.
GAME denoising steps: 8 by default; accepts 1–128.
GAME note-estimation confidence threshold.
Analyze the original beats after transcription and write the tempo to MIDI, preserving note timing.
Separate evaluation results for GAME v1.0.3 are not yet listed.

SOME

SOME source (pinned)SOME baseline weights SOME technical guide (pinned)

Analyze the original beats after transcription and write the tempo to MIDI, preserving note timing.

VOCANO

VOCANO extracts vocal pitch with Patch-CNN and segments notes with PyramidNet-110. Its released inference entry point has no additional public model controls.

VOCANO source and official weights VOCANO ISMIR 2021 paper

Choose “Solo vocal / vocal stem” for a cappella or already separated vocals. Choose “Full song” to separate the lead vocal first.

Paper and benchmark notes

VOCANO: A Note Transcription Framework for Singing Voice in Polyphonic Music

ISMIR - 2021

VOCANO extracts vocal pitch with Patch-CNN and segments notes with PyramidNet-110. Its released inference entry point has no additional public model controls.

Official source and citation

Robust Singing Voice Transcription Serves Synthesis

ACL - 2024

Robust path for singing with noise or accompaniment leakage; the current official checkpoint is trained mainly on Mandarin singing.

Paper

GAME: Generative Adaptive MIDI Extractor

Official technical guide - 2026

Turn singing into MIDI, with controls for singing language, segmentation and note detection.

v1.0.3 technical guide (pinned)v1.0.3 ONNX guide (pinned)v1.0.3 weights

SOME: Singing-Oriented MIDI Extractor

Official technical guide - 2023

Technical guide (pinned)Baseline weights

Official / repositories

ROSVOT source (pinned)ROSVOT project page GAME v1.0.3 source (pinned)GAME v1.0.3 weights SOME source (pinned)SOME baseline weights VOCANO source and official weights

Papers / technical notes

ROSVOT paper GAME v1.0.3 technical guide (pinned)GAME v1.0.3 ONNX guide (pinned)SOME technical guide (pinned)VOCANO ISMIR 2021 paper

Parameter guide

Choose “Solo vocal / vocal stem” for a cappella or already separated vocals. Choose “Full song” to separate the lead vocal first.
Leave this blank to use GAME's official language ID 0. This is not automatic language detection; select a language when you know what is being sung.
GAME boundary-segmentation confidence threshold.
GAME boundary search radius in seconds.
Starting noise time used by GAME’s official D3PM schedule.
GAME denoising steps: 8 by default; accepts 1–128.
GAME note-estimation confidence threshold.
Analyze the original beats after transcription and write the tempo to MIDI, preserving note timing.
ROSVOT note-boundary threshold: 0.85 by default; accepts 0–1.
ROSVOT word-boundary threshold: 0.5 by default; accepts 0–1.

Use cases

Turn a lead-vocal line into editable notes.
Compare four transcription models when the recording has bleed, noise or unclear note boundaries.
Use the MIDI as a guide track in a DAW, then correct pitches and note lengths by ear.

Workflow

1Choose the upload type, then add solo singing, a vocal stem or a full song. Full songs are separated first.
2Start with the default model; use the four-version comparison when one result is hard to judge.
3Switch between the original vocal and each piano roll, correct the chosen MIDI, then download it.

What to check

This tool follows the sung lead melody; it does not rebuild drums, chords or a complete arrangement.
GAME exposes language and segmentation controls. ROSVOT's released checkpoint is mainly trained on Mandarin singing.
The four-version comparison runs four separate conversions and returns four independent MIDI files; it does not fuse the models.

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

The selected model has no verified matching public result and has not been ranked.

GAME 1.0 · Clean vocals

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 -.

Author’s private data, 100 held-out recordings; each size uses its reported best K. OA is overall frame accuracy, not note F1, and cannot be pooled with ROSVOT scores.

Rank by metric

GAME 1.0 · Clean vocals
Metric rank	Model	Overall within this protocol↑OA (%)	QER (%)↓	Presence F1 (%)↑	RPA (%)↑	Pitch RMSE↓	Chamfer (frames)↓
1	GAME 1.0 Medium · K=8 Public weights · Inference code Public weights Inference code`GAME-1.0.3-medium-onnx.zip`Weight license (repository declaration): MIT	95.03	12.3	99.23	94.86	0.552	1.44
1	GAME 1.0 Large · K=1 Public weights · Inference code Public weights Inference code`GAME-1.0.3-large-onnx.zip`Weight license (repository declaration): MIT	95.03	12.87	99.15	94.92	0.684	1.44
3	GAME 1.0 Small · K=2 Public weights · Inference code Public weights Inference code`GAME-1.0.3-small-onnx.zip`Weight license (repository declaration): MIT	94.99	13.13	99.14	94.99	0.653	1.5

Original records (3)

GAME · Technical Report §10 ↗Checked: 2026-09-27

All evaluation sources (1)

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

ROSVOT paper, Table 1 ↗
Deployable entries 2 / source rows 10
Not ranked: 8
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
TONY · VOCANO · MusicYOLO · Yong et al. (2023) · ℳ (conformer) · ℳ (conv) · ℳ (w/o wbd) · ℳ (w/o noise)

All original results and sources

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

ROSVOT paper, Table 1

Deployable entries 2 / source rows 10

Not ranked: 8

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

TONY · VOCANO · MusicYOLO · Yong et al. (2023) · ℳ (conformer) · ℳ (conv) · ℳ (w/o wbd) · ℳ (w/o noise)

Robust path for singing with noise or accompaniment leakage; the current official checkpoint is trained mainly on Mandarin singing.

Rank by metric

ROSVOT paper, Table 1
Metric rank	Model	COn (F) ↑		COff (F) ↑		COnPOff (F) ↑		Pitch (AOR) ↑		Melody (RPA) ↑
Metric rank	Model	clean	noisy	clean	noisy	clean	noisy	clean	noisy	clean	noisy
1	ℳ (ours) Public weights · Inference code checkpoints.zip · rosvot + rwbd + rmvpe Inference codeWeight license not declared	94.0	93.8	94.5	94.4	77.4	77.0	97.0	97.1	87.6	87.4
2	ℳ (w/ E W ) Public weights · Inference code checkpoints.zip · rosvot + rwbd + rmvpe Inference codeWeight license not declared	93.3	93.5	93.2	92.9	77.1	77.0	96.5	96.2	87.5	87.2

Original records (2)

ROSVOT paper, Table 1 ↗Checked: 2026-10-25