Audio to MIDI converter that separates the stems first

Unveil's Midify tool turns an audio file into MIDI by splitting it into four stems and transcribing each one separately. You download the result as a ZIP of standard MIDI files that open in any DAW. Running it needs a free Unveil account, and signed-out visitors are sent to sign-up.

How audio-to-MIDI conversion works

MIDI does not contain sound. It is a list of instructions: which note, when it starts, how long it lasts and how hard it was played. Converting audio to MIDI therefore means estimating those instructions from a waveform. A transcription model looks at the audio over time, works out which pitches are active in each short slice, and then groups consecutive slices of the same pitch into notes.

This is much easier for a single clear melody than for a full mix. When many instruments play at once their overtones overlap, so the model has to guess which energy belongs to which note. That is why good tools do not transcribe the whole mix in one pass. They separate it first.

What Unveil does, step by step

Midify follows a fixed pipeline. Your file is uploaded, then split into four stems with the Demucs htdemucs separation model: drums, bass, other and vocals. Each stem is transcribed to MIDI with Spotify's Basic Pitch model. The raw MIDI is then cleaned with the music21 library.

• Stem separation with Demucs htdemucs into drums, bass, other and vocals.
• Note transcription of each stem with Basic Pitch. If one stem fails, it is skipped and the rest continue.
• Cleanup with music21: filtering short artifact notes, quantizing note timing and lengths, and key-aware note spelling.
• A structural analysis pass that prunes outlier notes.
• A master MIDI is built from the cleaned stems and everything is packed into one ZIP.

Preparing audio for better results

Transcription quality follows source quality. Use the highest quality file you have: WAV, AIFF or FLAC avoids the extra artifacts that heavy MP3 compression adds. Avoid files that are clipped or heavily limited, because crushed dynamics blur note onsets.

Material with clear, separate parts converts best. A sparse arrangement with a defined bass line and a single lead melody will give cleaner MIDI than a dense wall of layered synths and reverb. If you only care about one part, an isolated recording of that part will always beat extracting it from a full mix.

Using the tool

1. Create a free Unveil account or sign in. Signed-out visitors are sent to sign-up.
2. Open Midify and drop in an MP3, WAV, AIFF or FLAC file.
3. Choose a quantization setting: Auto (Smart), 1/8 notes or 1/16 notes. Auto uses a sixteenth-note grid that also allows triplets.
4. Choose Key Correction. Auto Detect finds the key from the music, or you can force one of the listed keys such as C Major, A Minor or F# Minor.
5. Press Convert to MIDI and wait while the stems are separated and transcribed. This is heavy processing, so expect it to take a while on longer tracks.
6. Review the key, tempo, note counts and piano roll for each stem, then download the full ZIP or a single stem.

Where audio-to-MIDI still goes wrong

Automatic transcription is a starting point, not a finished score. Expect to see these problems and plan to fix them by hand.

• Polyphony: chords and layered parts are the hardest case. Some notes of a chord are missed, and ghost notes appear where overtones were mistaken for real pitches.
• Drums: percussion has no pitch in the usual sense, so drum MIDI is best treated as a rhythm sketch that you re-map to your own drum kit.
• Octave errors: a note can land an octave above or below where you hear it, especially on bass and low vocals.
• Timing: quantization snaps notes to a grid. That tightens loose playing but can also move intentional swing or rubato.
• Vocals: pitch bends, vibrato and sung words become approximate notes, and breaths or consonants can create stray short notes.
• Tempo and key: detection can be wrong on tracks with tempo changes or ambiguous harmony, and the results screen shows what was detected so you can check it.

Cleaning the MIDI in your DAW

Drag the master.mid or individual stem files onto instrument tracks. Set your project tempo to the detected value first so the notes line up with the grid. Then play each part back against the original audio and fix what you hear: delete stray notes, shift octave errors, extend or shorten notes that end too early, and re-draw chords that came out incomplete.

Common mistakes are importing MIDI into a project at a different tempo, trusting the detected key without listening, and keeping every very short note. Short notes under a few milliseconds are almost always artifacts. Finally, check that you have the right to transcribe and reuse the source recording before you release anything based on it.

What you get

• A ZIP containing one .mid file per stem (drums, bass, other/harmony, vocals) plus a combined master.mid.
• Detected key and tempo shown on the results screen, with total note count and note range for each stem.
• A piano roll preview for every stem, and a separate download button for each individual MIDI file.
• Quantization and key options you can set before converting, including forcing a specific key.

Questions

Related

• Free online LUFS checker
• Free online true peak meter
• Online AI mastering with an analysis-driven mid/side chain
• Mix analysis guides
• Dynamic range analysis: DR, crest factor & LRA