How to convert audio to MIDI — and when it actually works

Audio-to-MIDI conversion promises something that sounds impossible: hand it a recording, get back the notes. It genuinely works — but how well it works depends almost entirely on one decision you make before you upload anything. This post covers what the conversion actually does, the input decision that makes or breaks it, and the cases where no converter is going to help you.

What "audio to MIDI" actually means

A MIDI file doesn't contain any sound. It contains instructions: this note, starting here, lasting this long, at this loudness. Converting audio to MIDI means listening to a recording and recovering those instructions — which pitches were played, and when.

That's a different job from separating a song into stems. Stem separation gives you back audio — an isolated bass track you can hear. Transcription gives you back data — the note the bass played, as a number your DAW can move, transpose, re-voice or print. The two are complementary, and as it turns out, one works much better after the other.

The whole game is what you feed it

Here's the part most "mp3 to midi" pages skip.

A pitch tracker analyses the frequency content of the audio over time and decides which notes are sounding. When one instrument is playing, that's a tractable problem: the fundamental and its harmonics are there, and the tracker follows them. When five instruments are playing at once, plus a vocalist, plus a drum kit spraying broadband noise across the spectrum, the tracker has to work out which overlapping partials belong to which note — and the harmonics of one instrument look a lot like the fundamentals of another.

So the same converter, on the same day, will hand you a clean bass line from a bass stem and an unreadable cloud of notes from the full mix it came out of. Nothing changed except the input.

This is the same argument we make about chord recognition in why stem separation makes chord detection better, and it holds for the same reason: the analysis is only ever as good as the signal it gets. Chord detection wants the harmony without the drums. Transcription wants one part, alone.

The workflow that actually works: separate, then transcribe

Given that, the reliable route from a finished song to usable MIDI is two steps, not one:

1. Split the song into stems. Run the track through the stem splitter and you get the vocals, drums, bass and harmonic instruments as separate files. (If you're new to this, how to split a song into stems walks through it.)
2. Transcribe the one part you want. From the stems view, each stem row has a To MIDI action that carries that stem straight into the audio-to-MIDI converter — no re-uploading, no exporting and re-importing.

If you already have an isolated part — a bass DI, a vocal comp, a synth bounce, a solo you recorded on your phone — skip step one and upload it directly. MP3, WAV, M4A, FLAC and AIFF all work.

The tool will also transcribe a whole track if you point it at one. It's there because sometimes the "track" is a single part, and it's honest about the trade-off when you do it on a dense mix.

What good input looks like

In rough order of how well it transcribes:

• Bass lines. One note at a time, low and clear, long note values. The easiest thing you can hand a transcriber — which is why the bass isolator and the MIDI tool are usually used back to back.
• Vocal melodies. Monophonic and expressive. The notes land; the slides and vibrato show up as extra short notes around the main pitch.
• Leads and solos. Synth leads, guitar solos, horn lines — single-line playing with clear attacks.
• Sparse keys. A piano playing block chords at a moderate tempo comes out as a rough but recognisable draft.

Where it doesn't work — read this before you're disappointed

No converter on the market is exempt from these. A tool that implies otherwise is setting you up:

• Dense polyphony. Six-note jazz voicings, orchestral tuttis, thick pad stacks. The more simultaneous pitches, the more the tracker has to guess, and the inner voices are what it loses first.
• Strummed guitar. A strum is six notes fired a few milliseconds apart, each with strong harmonics, often with two strings sounding the same pitch. It transcribes as a smear. If you want a strummed part, you almost certainly want the chords, not the individual notes — different tool, different answer.
• Heavy effects. Distortion adds harmonic content that was never played. Long reverb tails and delays keep old notes ringing under new ones. Chorus and heavy pitch modulation blur the fundamental the tracker is chasing.
• Drums and percussion. Transcription tracks pitch. A snare doesn't have one. You can point the tool at a drum stem, and it will dutifully return notes, but they're meaningless as a drum part.
• Very quiet or buried parts. If a part is barely audible in the stem, it's barely audible to the transcriber too.

The realistic expectation, even on a good input, is a faithful draft: correct notes in the right places, plus some extra short ones, plus the occasional octave that jumped. That's still an enormous head start over transcribing by ear from scratch — you're editing rather than starting from nothing.

Reading the result before you export it

The note grid isn't just a preview — it's how you tell a good transcription from a bad one in about ten seconds.

• The grid is chromatic and unquantised, so notes sit at their true pitch and their true time. Nothing is snapped into a key, which means bends and out-of-key passages stay where they belong — and it also means an out-of-tune or wandering source shows up honestly.
• Play it against the original. The result view has a two-channel mixer: your audio on one side, the synthesised MIDI on the other, each with its own level and mute. Mute the original, and if the MIDI plays back as the part you recognise, the transcription is good. Wrong octaves and ghost notes are obvious the second you hear them alone.
• Use the noise floor when there's clutter. Every detected note carries a loudness value, and the artefacts from stem bleed are reliably quieter than the real part. The noise-floor slider drops notes below a share of the track's loudest note — and it applies to the MIDI you download, not just the picture on screen. Nudging it up is the fastest way to clear ghost notes without touching a note editor.

When it looks and sounds right, Download MIDI gives you a standard .mid. There's also a copy-as-text action that puts every note — name, start, end, length, loudness — on the clipboard as tab-separated rows, if you'd rather paste the transcription into a spreadsheet than open a DAW.

What to do with the MIDI once you have it

The point of MIDI is that it's editable, so the last mile is in your DAW or notation app:

• Re-voice the part. Drop the bass line onto a synth bass, the vocal melody onto a pad, the guitar solo onto strings. This is the reason producers reach for audio-to-MIDI in the first place — you keep the phrasing of a reference and change everything else.
• Quantise and fix. The export keeps the exact timings from the performance, so it will not sit on your grid until you quantise it. Set your project tempo first — ChordSonic reads the song's BPM on a chord analysis, and how to find the BPM of a song covers getting that number right — then quantise, delete the strays, and correct any octave jumps. There's no note editor in the tool itself; fixes happen downstream.
• Learn the part. Slow the MIDI down without touching the pitch, loop four bars, read the notes off the piano-roll. For a bass line or a solo you're trying to learn, seeing the intervals laid out is often faster than any tab.
• Print it. Notation software will import the .mid and render a staff. Quantise before you do, or the notation will be technically accurate and completely unreadable.

MIDI notes and chords are different questions

One thing worth being clear about, because people arrive at audio-to-MIDI looking for the wrong answer: transcription tells you which notes were played. It doesn't tell you the song is in G major and the progression is I–V–vi–IV.

If what you actually want is the harmony — to play along, to learn the tune, to understand how the reference track is built — the chord finder is the tool for that, and how to find the chords of any song is the place to start. Notes and chords answer different questions, and the fastest workflow usually uses both: chords for the shape of the song, MIDI for the line you want to steal.