How to convert audio to MIDI
- Drop an audio file (MP3, WAV, M4A, FLAC, OGG) onto the box, or press Record and sing, hum or play into your microphone.
- Wait while the notes are transcribed. The first run loads a 0.9 MB model from this site; after that it is cached. With graphics acceleration a three-minute song usually takes well under a minute; on the CPU expect roughly the length of the song or more.
- Check the piano roll and press Play to hear the MIDI, the original, or both together.
- Adjust the note settings if needed, set the tempo, then press Download MIDI and drag the .mid file into your DAW or notation app.
The transcription uses Basic Pitch, an open-source neural network from Spotify’s Audio Intelligence Lab that detects multiple notes at once and follows pitch bends. Here it runs entirely inside your browser with TensorFlow.js: your recording is never uploaded.
What works well (and what doesn’t)
- Great: solo piano, acoustic or electric guitar, bass lines, singing, humming and whistling, flute, violin, synth leads. Chords are fine; the model is polyphonic.
- Okay: a voice or instrument with light accompaniment. Raise “Note confidence” to keep only the strongest line.
- Poor: full band mixes, heavy reverb, distortion and drums. Drums have no pitch, so they turn into random notes. If you only need a song’s tempo, use the BPM finder instead.
For a clean result, record close to the microphone in a quiet room. If your source is a video, pull out the soundtrack first with the audio extractor. If a recording comes out silent, check the input with the microphone test.
Understanding the settings
- Note splitting (onset threshold): how strong an attack must be to start a new note. Lower it when fast repeated notes merge into one long note; raise it when sustained notes are chopped up.
- Note confidence (frame threshold): how sure the model must be that a note is sounding. Lower it to catch quiet notes; raise it to remove ghost notes and harmonics an octave above the real note.
- Minimum note length: removes blips shorter than this. 100 to 150 ms suits most melodies; lower it for fast runs.
- Lowest and highest note: restrict the pitch range, for example E1 to G4 for a bass guitar, to drop stray notes outside the instrument’s range.
- Pitch bends: adds pitch-bend events so slides and vibrato survive. MIDI bends apply to a whole channel, so they are only written for notes that play alone.
Getting the MIDI into your DAW
Notes keep their real timing in seconds. The tempo you set decides where bar lines fall, so use the song’s tempo (press Detect or measure it with the BPM finder) and keep “line up the detected beats” on, so beat one of your DAW’s grid matches the music. Quantize only if you want robotic, on-grid timing; for expressive playing, leave it off and quantize selected parts in your DAW. The file is a standard type 1 MIDI file with one track, so it opens in Ableton Live, FL Studio, Logic Pro, GarageBand, Cubase, Reaper, MuseScore and others.
Frequently asked questions
How do I convert an MP3 to MIDI?
Drop the MP3 onto the converter and wait a few seconds while the notes are transcribed. Check the piano roll, adjust the settings if some notes are missing or extra, then press Download MIDI to save a .mid file you can open in any DAW or notation app.
Can I convert my voice or humming to MIDI?
Yes. Press Record, sing or hum a melody for up to two minutes, then stop. Voice works best recorded close to the microphone in a quiet room. Turn on pitch bends if you want slides and vibrato kept in the MIDI file.
Does it work on full songs with drums and vocals?
Only roughly. The model transcribes every pitched sound it hears, so a full mix produces many overlapping notes, and drums turn into random notes. For clean MIDI, use a recording of a single instrument or voice, such as an isolated stem.
Is my audio uploaded to a server?
No. The Basic Pitch model is downloaded once from this site and runs inside your browser with TensorFlow.js. Your file or recording is decoded and transcribed on your device and never leaves it.
Why are there extra notes an octave above the real ones?
Strong harmonics, especially in voice, brass and bright guitars, can be heard as extra notes. Raise Note confidence a little, or limit the Lowest and Highest note to the instrument's range, and the ghost notes disappear without re-running the transcription.
What tempo does the MIDI file use?
The converter detects the tempo automatically when the recording has a steady beat, and you can type any tempo. Notes keep their real timing either way; the tempo only decides where bar lines fall in your DAW, so matching it to the song makes editing much easier.
Is this the same as Spotify's Basic Pitch?
It uses the same open-source Basic Pitch model and note-detection settings, running locally in your browser. On top of that it adds microphone recording, a piano-roll preview with playback, tempo detection, quantizing and instrument selection.