Turn a recording into MIDI notes
This audio to MIDI converter listens to a recording and writes down the notes it hears as a Standard MIDI File. Drop an MP3, WAV, M4A or a voice memo; a few moments later you get a piano roll of the notes, a .mid file to open in any DAW or notation program, and a preview you can play back with real instrument sounds.
It runs Basic Pitch, the open-source transcription model from Spotify’s Audio Intelligence Lab, inside your browser. Basic Pitch is polyphonic, so it can pick out chords as well as melodies, and it is not tied to one instrument. Be realistic about what it can do, though: it is good on one instrument or one voice at a time — a piano piece, a guitar part, a sung or hummed melody, a flute solo — and weak on full mixes, where drums, bass, vocals and effects blur into extra and missing notes. For a song, separate the part you want first with the Vocal Extractor.
How to use the Audio to MIDI Converter
- Drop an audio file on the box (up to 10 minutes), or press Try a generated sample to see how it works.
- Press Transcribe to MIDI. The first time, TensorFlow.js and the model (about 2.5 MB) download and are cached.
- Look at the piano roll. If there are too many stray notes, raise Note confidence; if repeated notes are merged, lower Note splitting; drop blips with Shortest note. The notes update as you move the sliders.
- Choose the MIDI instrument and the tempo for the bar grid, then press Play under “Listen to the MIDI” to hear it.
- Press Download MIDI, or Notes as CSV for a list of start, end, pitch and velocity.
Audio to MIDI Converter features
Polyphonic transcription
Detects several notes at once — chords on piano or guitar — across the full 88-key range.
Live note settings
Note splitting, confidence, shortest note and pitch range re-run only the final note-picking step, so changes show at once.
Piano roll and playback
An overview of every note follows your recording as it plays; the MIDI preview plays the result with a General MIDI sound set.
Clean MIDI files
Format 0, 480 ticks per beat, velocities from the note strength, your choice of General MIDI instrument and tempo, optional pitch bends.
Private
The recording is decoded and analysed on your device. Nothing is uploaded.
Free and open
Basic Pitch (code and trained model) and TensorFlow.js are Apache-2.0 open source.
What works well and what does not
Overtones are the most common error: a strong, bright note can add a quieter note an octave or a twelfth above it. These extra notes are usually faint — raise Note confidence or set a Highest note to remove them. Clean, dry recordings with little reverb give the best results.
| Recording | Typical result |
|---|---|
| Solo piano, guitar, flute, violin, sax | Good: most notes and chords come out right; check fast runs and very low bass notes |
| Singing, humming or whistling | Good for the melody; add pitch bends to keep slides and vibrato |
| A single synth or bass line | Good, if it is not heavily distorted |
| Full band or a finished song | Weak: drums and overlapping parts add many wrong notes. Isolate a part first |
| Drums and percussion | Not supported — they have no pitch to transcribe |
How it works
The audio is mixed to mono and resampled to 22,050 Hz, then cut into two-second windows. For each window the Basic Pitch neural network, published by Bittner and colleagues at ICASSP 2022, estimates three things for every semitone from A0 to C8, 86 times a second: how likely a note is sounding, how likely one is starting, and a finer pitch contour in thirds of a semitone. The network runs on your graphics chip through WebGL when the browser allows it, otherwise on the processor.
Notes are then picked from those activations exactly as the official library does: a note starts where the onset estimate peaks above Note splitting and lasts while the note estimate stays above Note confidence; the “extra notes” option then follows remaining strong energy to recover notes without a clear attack. Long files are processed in about 30-second pieces cut at quiet moments, so this step stays fast.
Tips for better MIDI
- Trim silence and talking from the start with the Audio Cutter so the MIDI begins on the first note.
- The MIDI keeps the timing you played. To line notes up with bars in your DAW, set the song’s tempo here or detect it with the BPM Detector, then quantise in the DAW.
- Pitch bends apply to a whole MIDI channel, so they are only added to notes that do not overlap others.
- The Key Finder tells you the key of the recording, which helps when fixing wrong notes.
Audio to MIDI Converter FAQ
How do I convert MP3 to MIDI for free?
Drop the MP3 on this page and press Transcribe to MIDI. When the piano roll appears, adjust the settings if needed and press Download MIDI. There is no sign-up and no upload.
Can it convert a whole song to MIDI?
It will try, but results from full mixes are weak: drums and overlapping instruments produce many wrong notes. Extract the vocal or one instrument first, or use a solo recording.
Can I turn my singing or humming into MIDI?
Yes. A clear, dry recording of one voice works well. Tick Add pitch bends to keep slides and vibrato, or leave it off for clean, quantisable notes.
Why are there extra notes an octave higher?
Bright sounds have strong overtones that the model sometimes hears as notes. Raise Note confidence, set a Highest note, or delete them in your DAW.
Is my audio uploaded?
No. Decoding and the neural network run in your browser. Only the model and libraries are downloaded.
How long does it take?
It depends on your computer. Without graphics acceleration our test machine needed about half a second per second of audio; with a graphics chip it is usually much faster.
Which DAWs can open the file?
Any program that reads Standard MIDI Files: FL Studio, Ableton Live, Logic, GarageBand, Cubase, Reaper, MuseScore and others.