Audio in, staff out · four cases measured on a real engraver
Audio to sheet music: the notes are solved, the staff is not.
Every tool that claims to turn a recording into a score does it in two steps: audio to notes, then notes to a staff. The first step is a solved problem — a detector writes down a pitch and a time in milliseconds. The second step is not, because a staff has no room for arbitrary times: it has a grid, and every note gets snapped onto it. Below are four rhythms pushed through a real engraver with the before and after numbers, then the five tools that rank for this search and what each one actually hands you. MP3 to sheet music is the same job with a different container; the container changes nothing about the two steps.
The short answer
If you want a score to read and print: Songscription Plus, $8.33/month billed yearly ($99.90), 60 minutes of transcription a month, and it exports PDF, MIDI, MusicXML and Guitar Pro. Klangio Transcription Studio is the same $8.33 at the same yearly billing for 50 transcriptions a month. Both give you a file you can edit rather than a picture of one.
If you only need the notes: stop at step one. A MIDI file is the honest output of a transcription, and it is the only output that does not silently round your rhythm. This site does that half, in the browser.
What no tool will give you: the rhythm you actually played, once it is on a staff. The measured section below is the reason.
The two steps
One of them loses nothing, the other loses the timing
Step one takes a waveform and produces events: this pitch started at 1.284 s and stopped at 1.531 s. Nothing about that is approximate in principle — the detector may be wrong about which note it heard, but the time it reports is the time it heard it, to the millisecond.
Step two takes those events and has to place them on a staff. A staff is a grid, and the grid of the engraver measured here is a sixteenth note: with 480 ticks per quarter, one cell is 120 ticks. Every onset is rounded to the nearest cell, every length is rounded up to at least one cell. That is not a bug in one product — a staff cannot draw a note at 1.284 s, only at a position in a bar.
So the losses are predictable, and they are all in the same family: anything that is not a multiple of a sixteenth gets moved. Triplets, swing, a human 16-millisecond late attack, a grace note, two notes closer together than a sixteenth — all of them are the same problem wearing different clothes.
The measurement
Four rhythms, before and after the grid
Four tiny MIDI files were written by hand, each one a single bar at 120 BPM with 480 ticks per quarter, then pushed through the engraver that draws the staff on this site — the same code that renders MIDI to sheet music — and converted back to MIDI so the result could be measured in ticks rather than read off a picture. All four files are published with this page, so you can run them yourself.
| Played | Onsets in | Grid cells | Written as | Onsets out |
|---|---|---|---|---|
| Three triplet eighths, then a quarter | 0, 160, 320 — 160 each | 0, 1, 3 | C D2 E F4 | 1, 121, 361 — 119, 239, 119 |
| Two swing eighths, twice | 0, 320 — 320 then 160 | 0, 3 | C3 E F3 A | 1, 361 — 359, 119 |
| Four eighths, each 15 ticks late | 15, 495, 975, 1455 | 0, 4, 8, 12 | C2 z2 D2 z2 E2 z2 F2 | 1, 481, 961, 1441 |
| Two notes 50 ticks apart | 0 and 50 — 40 each | 0, 0 | [=C=D] z3 E4 | 1 and 11 — 119, 109 |
1. Triplets do not survive. Three equal notes of 160 ticks are 1/3 of a beat each. The grid is 120 ticks, so they land on cells 0, 1 and 3 — not 0, 1.33 and 2.67. The engraver wrote a sixteenth, an eighth and a sixteenth. The second note moved 39 ticks earlier and the third moved 41 ticks later; at 120 BPM that is 41 and 43 milliseconds, and it is the difference between a triplet and a dotted rhythm. Unless the tool switches to a 3:2 tuplet — and this one has no tuplet code at all — a triplet cannot be written on a sixteenth grid.
2. Swing comes back as a dotted eighth and a sixteenth. A swung pair is played roughly 2:1 — 320 ticks then 160. On the grid it becomes 359 and 119, which is 3:1. That is the conventional way to write swing and a musician reading it will play it correctly, but the file now says something the recording did not: the second note moved 41 ticks, and a machine reading the score back will produce a stiff dotted rhythm rather than a swing.
3. A human performance is flattened onto the beat. Four eighth notes, each starting 15 ticks late — 16 milliseconds, about the amount a player is consistently behind — came back at 1, 481, 961 and 1441. The lateness is gone, and so is any evidence that this was played rather than typed. This is the loss that makes transcriptions sound mechanical: not the wrong notes, the missing push and pull.
4. Two notes closer than a sixteenth become a chord. Two notes 50 ticks apart — 52 milliseconds, a fast grace note or a flam — both round to cell 0. The engraver put them in the same cell and wrote one chord, [=C=D]. A 40-tick note came back as 119 ticks, three times its length, because a note cannot be shorter than one cell. Nothing warned; the score simply says something else.
The four inputs: triplets.mid (77 B), swing.mid (77 B), human.mid (80 B), fast.mid (67 B).
The other thing that gets invented
The key signature is a guess, and it can be wrong
Step one produces pitches, not a key. Nothing in a recording says “this is in F”, so the engraver fits the pitch content to the twelve major key signatures and picks the best match. On the four test files it did this:
- The triplet file plays
C D E Fand came back withK:F— one flat in the key signature, a note that never sounds in the bar. - The fast file plays
C D Eand came back withK:E— four sharps — which forced two naturals into the first chord:[=C=D].
Neither is a bug. With no key signature in the source there is nothing to read, so the engraver has to infer, and a three-note phrase does not contain enough information to infer from. What matters for you is the consequence: the first thing to check in a transcription is not the notes, it is the key signature, because a wrong key makes every accidental in the score a correction waiting to happen.
Who does it
Five tools that rank, and what each one hands you
These are five of the products on the first page of results for this search, with what they take in, what they give out, and what the free tier is. Prices are the ones printed on each vendor's own page on 11 November 2026; nothing here comes from a review site.
| Tool | You give it | You get | Free | Price |
|---|---|---|---|---|
| Songscription | An audio file, a YouTube link, or a PDF or photo of a score | PDF, MIDI, MusicXML, Guitar Pro | The first 30 seconds of each song; one page a day from a PDF or photo | $8.33/month billed yearly ($99.90), 60 minutes a month · Pro $24.99/month |
| Klangio Transcription Studio | Audio, video, a YouTube or TikTok link, or a recording | PDF, MIDI (quantized or not), MusicXML, Guitar Pro | A 30-second demo transcription | $8.33/month billed yearly ($19.99 monthly), 50 transcriptions a month |
| Soundslice | An audio or video file, or a YouTube link | Notation in the browser; MIDI and MusicXML export on Plus | 2 pages a month | Plus $5/month or $50/year, 100 pages a month |
| Nutif.ai | An audio file up to 50 MB | PDF and MIDI — no MusicXML | The service is free | Free; uploads may be used to train their models, and anonymous uploads are public |
| MuScriptor | An audio file, one instrument at a time or several | MIDI, sheet music PDF, guitar tabs | The demo is free | Free; code MIT, model weights CC BY-NC 4.0 (non-commercial) |
How to read that column of formats. MusicXML is the one that matters if you intend to edit the score: it opens in MuseScore, Dorico, Sibelius and Finale with the notes, bars and voices intact. A PDF is a picture of a score — fine to print, useless to edit. MIDI sits in between: editable, but it carries notes and times, not notation. If a tool only gives you a PDF, you are buying a printout, and the editing starts from zero.
Sources: songscription.ai/pricing, klang.io/transcription-studio, soundslice.com/plans, nutif.ai, muscriptor.kyutai.org — all read 11 November 2026.
When the score arrives
Five things to check before you trust it
- The key signature. Play the first bar and look at the accidentals. If the signature has flats or sharps that never appear, the engraver guessed and guessed wrong.
- The time signature. If the recording had no meter event, 4/4 was assumed. A waltz written in 4/4 has its bar lines in the wrong place, and every rhythm in it will look strange for a reason that has nothing to do with the rhythm.
- Any triplet you played. Look for a 3:2 bracket. If there is none, the triplets were rounded onto the sixteenth grid, and the notes are now a dotted rhythm.
- Notes shorter than a sixteenth. Grace notes, flams, fast pickups: they cannot be shorter than one grid cell, so they were lengthened. If two of them were closer together than a cell, they became one chord.
- The first bar. A pickup bar is the most common structural error, because the grid starts at zero whether the music does or not.
FAQ
Questions about audio and sheet music.
Is there a tool that goes from audio to sheet music in one step?
No, and the one-step appearance is a user-interface choice rather than a technical one. Every product above runs a note detector and then an engraver; some run both behind a single upload button. That matters because the two halves fail differently: a wrong note is the detector's fault, a wrong rhythm is almost always the engraver's grid.
Why did my triplets come out as straight notes?
Because the staff was using a sixteenth-note grid and a triplet is a third of a beat, which is not on that grid. Measured above: three equal notes at 160 ticks came back as 119, 239 and 119. A tool that supports tuplets will write a 3:2 bracket instead; if the export has no bracket, the rounding already happened.
Would a paid tool keep my timing?
The grid is a property of notation, not of price. What money buys is a better note detector and more editing tools — Klangio, for example, exports both quantized and unquantized MIDI, and the unquantized one is the one that keeps the performance. If timing is what you care about, export MIDI rather than PDF, and check whether the export offers an unquantized option.
Does this site convert audio to sheet music?
No. It converts audio to MIDI — the first half, without the grid. There is no notation output on this site and no upload box that produces one. If you need a score, use one of the five tools above, or take the MIDI from here into any notation program and engrave it yourself, where you can set the tuplets and the key.
Can I fix the timing after the fact?
Yes, and it is easier in MIDI than in a score. In MIDI the onset is a number you can nudge; in a score the onset is a rhythmic symbol, so fixing it means re-notating the bar. That is the practical argument for keeping the MIDI as the master file and treating the score as a readable copy of it.
Keep going