Turning a recording into sheet music is really five jobs in a row. First, if the recording has several instruments, the part you want is separated out. Second, a model listens for notes and reports each one with a pitch, a start time, a length and a loudness.
Third, the rhythm is quantized: start times in seconds become eighth notes, quarter notes and triplets. Fourth, the key and time signature are estimated, so a note becomes an F♯ rather than a G♭ and the bar lines fall in the right places. Fifth, the notes are engraved into a score with stems, beams, ties and rests.
Each step can be adjusted after the fact, which is why the settings panel exists. Changing the time signature re-bars the score; changing the key respells the accidentals; changing the hand split moves notes between staves. You do not need to transcribe again.
Why solo piano gives the cleanest score
Piano has a hard, clear attack on every note, which makes start times easy to find. It is also the most studied instrument in music transcription research, and the piano-specific model in HQ mode comes out of that work. Put those together and solo piano is the case where AI transcription is most useful today.
That shapes what a song to piano conversion can and can’t do. A recording of someone playing piano becomes a two-hand score. A full band song becomes a score of whatever is audible, not a piano arrangement: the melody, bass and chords all land on the same staves. Separate the part you want first, or start from the piano stem.
The hard cases are the ones a human transcriber also finds slow: heavy sustain pedal that blurs repeated notes, very fast runs, and a band where the piano hides under guitars and drums. For a song, separate the piano first, then transcribe the stem on its own.
MIDI, MusicXML or PDF: which export do you need?
A PDF is for reading and printing. It looks like a finished score, but you cannot change the notes in it without retyping them.
MusicXML is the notation file. It keeps the things a score needs and MIDI lacks: the key signature, how each note is spelled, which staff and voice it belongs to, ties, beams and the time signature. Open it in MuseScore, Dorico, Sibelius or Finale and keep editing.
MIDI is the performance file. It keeps exact timing and velocity, which is what a DAW wants, but it has no idea whether a note is an A♯ or a B♭. Use MIDI for production and MusicXML for notation.
If MIDI is all you need, MP3 to MIDI gives you a tempo-aligned .mid without engraving. If you already have a MIDI file, MIDI to Sheet Music turns it into notation with no transcription step, and Transpose Sheet Music moves the finished score into another key or for a B♭, E♭ or F instrument.
Reading and fixing an AI transcription
Most mistakes fall into a few groups. Octave errors put a note one octave too high or low, usually in the bass. Rhythm spelling shows a dotted rhythm as tied sixteenths, or a triplet as straight eighths. And quiet inner voices sometimes go missing under louder notes.
Start with the bars marked as uncertain, play them against the recording, and fix them in your notation app. Then check the rhythm: if many bars look crowded, try a coarser quantize grid or a different time signature before editing note by note.
If you’ve ever had to transcribe music by ear, the workflow will feel familiar, just faster. A transcription is a fast first draft, not a finished edition. Treat it the way you would treat a friend’s handwritten copy: useful, mostly right, and worth checking against your ears.