Vocal Melody to MIDI · TelkNet

MIDI / transcriptionGPU workload

Vocal Melody to MIDI

One model: 25 credits · four-model comparison: 250 creditsFile 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

One editable note MIDI from each selected model. Lyrics, vibrato curves and the rest of the arrangement are not included.

Use case

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

Models, parameters, and sources

Turn an isolated lead vocal, a cappella take, or humming into editable MIDI notes. ROSVOT is the default; the four-version comparison runs GAME, ROSVOT, SOME, and VOCANO separately so you can listen before choosing. One model costs 5 credits; all four cost 45.

GAMEROSVOTSOMEVOCANOGAME v1.0.3 source (pinned)GAME v1.0.3 weightsGAME v1.0.3 technical guide (pinned)GAME v1.0.3 ONNX guide (pinned)ROSVOT source (pinned)ROSVOT project pageROSVOT paperSOME source (pinned)

GAME

GAME converts singing to MIDI with its official boundary, D3PM and estimation controls. Leaving language blank uses official ID 0; it neither detects the language nor forces Chinese.

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 D3PM denoising steps. Upstream defaults to 8 and only sets a minimum of 1; 128 is TelkNet's public-task limit, not an upstream maximum.
GAME note-estimation confidence threshold.
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.

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 confidence threshold; upstream defaults to 0.85. The 0–1 input range is TelkNet validation, not an upstream parser limit.
ROSVOT word-boundary confidence threshold; upstream defaults to 0.5. The 0–1 input range is TelkNet validation, not an upstream parser limit.

SOME

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

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

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.

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

GAME converts singing to MIDI with its official boundary, D3PM and estimation controls. Leaving language blank uses official ID 0; it neither detects the language nor forces Chinese.

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

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

Technical guide (pinned)Baseline weights

Official / repositories

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

Papers / technical notes

GAME v1.0.3 technical guide (pinned)GAME v1.0.3 ONNX guide (pinned)ROSVOT paper 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 D3PM denoising steps. Upstream defaults to 8 and only sets a minimum of 1; 128 is TelkNet's public-task limit, not an upstream maximum.
GAME note-estimation confidence threshold.
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.
ROSVOT note-boundary confidence threshold; upstream defaults to 0.85. The 0–1 input range is TelkNet validation, not an upstream parser limit.
ROSVOT word-boundary confidence threshold; upstream defaults to 0.5. The 0–1 input range is TelkNet validation, not an upstream parser limit.

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.

Before you submit

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 for 45 credits per file. It returns four independent MIDI files; it does not fuse the models.

Related use-case guides

Working on a similar task? These guides list the matching tools, model boundaries, sources, and FAQs for each scenario.

Audio to MIDI · online tool and model guide

Quick feedback

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