How to turn a voice memo into sheet music

Can an app turn humming or singing into sheet music?

Yes, when the recording holds one clear melody. The software does two separate jobs. First it listens for notes: each pitch, when it starts and when it stops. Then it writes those notes down, which means choosing a key, a meter, note values and where the bar lines go. If the tune is still only in your head, sing it into your phone first; everything below applies to that recording.

In ScoreStarling, for one instrument or voice, the first job is done by Spotify’s open-source Basic Pitch model, which turns audio into MIDI on our server. Spotify says Basic Pitch works on “almost any instrument, including voice.” That is their claim. Our own measurements so far cover guitar recordings, one band excerpt and synthesized piano, not singing. The second job goes to MuseScore Studio 4.7.5’s MIDI importer, and Verovio engraves the page.

Timing is often the hardest part of a voice memo. ScoreStarling detects the tempo or uses the BPM you supply, but for one instrument or voice it still writes to a straight grid starting in 4/4, so drift, silence before the tune and a wrong tempo can produce misplaced ties and rests. Solo piano follows the playing’s own pulse and meter, but it is meant for piano, not sung memos. The advice here helps give a melody a steady pulse.

How to record a voice memo for transcription

Give the software one line of music at one steady tempo, close to the microphone, with nothing else playing.

• Record one voice or one instrument: no backing track, no second singer, no guitar under the vocal. Busy mixes, heavy reverb and several instruments at once make the first draft harder to read.
• Keep the phone close and the room quiet. Apple’s guide sets the recording level by distance: move the phone closer to or farther from the sound. Close, but not so close that your loudest notes distort. A room with soft furnishings echoes less than a bathroom or a stairwell.
• Use a click that you can hear and the phone can’t: a metronome in headphones, played from a second device. Apple’s guide says Voice Memos stops recording if another app on the iPhone starts playing audio, so the click shouldn’t come from the recording phone.
• Write down the BPM. If you recorded without a metronome, tap along afterward with a tap-tempo metronome, or count beats for 15 seconds and multiply by four.
• Count in silently. A spoken “one, two, three, four” is sound like any other and can end up written as notes, so count in your head or in the headphones, or trim the count-in off afterward.
• Sing on a syllable such as “da” or “doo” rather than a closed hum, so repeated notes don’t blur into one. That’s our suggestion, not a measured result.
• Keep each take under five minutes: only the first five minutes of a recording are transcribed, from a file of up to 100 MiB. Record long songs in sections beginning at a bar line.
• Leave the extras off. On Pixel 9 and later, Google’s Recorder can add AI background music to a recording, and added music is more sound to transcribe. If you sang over an instrumental with Voice Memos’ layered recording, separate the layers and upload only the vocal.

Then trim the start. Bar 1 begins at 0:00 of the file, so cut any silence or count-in until the file starts right on the first note. If the tune begins with a pickup, you can mark the first bar as a pickup later. In Voice Memos, the editor’s yellow trim handles do this, and you can pinch to zoom for precision. Google’s Recorder has Crop & Remove, which saves a trimmed copy.

What format do iPhone and Android voice memos use?

An iPhone exports a Voice Memos recording as an .m4a file by default; Android recorders differ by maker, so check the file extension after you share one.

Apple’s iPhone User Guide (iOS 27, checked October 3, 2026) says default exports are .m4a, which flattens layered recordings and turns Spatial Audio into stereo. To get the file off the phone, use Save to Files from the recording’s share options. Skip the Editable option: it exports QuickTime Audio, a format for adjusting effects and layers later. For a layered recording, use Separate Layers first; each layer then becomes its own .m4a.

On Pixel phones, Google’s Recorder shares a recording as an audio file: touch and hold it, then choose Share, File and Audio. Google’s help pages don’t name the format (checked October 3, 2026), and other Android brands ship their own recorder apps. Whatever arrives, ScoreStarling accepts WAV, MP3, FLAC, M4A, OGG and AIFF. You can also skip the file and record in the browser; that recorder stops at 4:59, just inside the five-minute limit.

Step by step: voice memo to sheet music in ScoreStarling

Upload the memo, give its tempo, transcribe, then correct the score by ear before you export it. Transcribing one instrument or voice, editing and PDF downloads are free on ScoreStarling; unlocking the score’s MusicXML and MIDI uses credits (checked October 7, 2026). Open the workspace to start.

1. Choose New score. Drop in the file, or choose Record to sing straight into the browser. Nothing is kept until you choose Create score.
2. Under What’s playing?, keep One instrument or voice, which runs Basic Pitch on our server for free. Solo piano is free too. Band or several instruments uses credits, 40 a minute with the price shown before it starts, and sends the audio to our band partner Mirelo; it runs only if you pick it.
3. Open More options and enter Beats per minute, in quarter notes from 30 to 240, or leave it on Auto to have the tempo detected. It sets how the score is written, not the recording. If you can’t decide between 70 and 140, either will do: 140 writes every note at twice the value.
4. Choose Create score. The score appears with playback in sync with your recording.
5. Check it by ear, two bars at a time, as the next section describes.
6. Fix what you hear. Select a note to raise or lower it a semitone (the up and down arrow keys work too), change its length, or delete it; deleting leaves a rest, so the beat stays put. Undo, redo and saved versions are there if a fix goes wrong.
7. Finish in Score settings: key and clef, time signature, a pickup bar if the tune starts before the downbeat, and transposition by interval. Transpose last. It moves every note and the key, but your recording keeps its original key, so comparing by ear gets harder afterward.
8. Download what you need: PDF to print, MusicXML for a notation app, MIDI for a DAW, page images, separate parts, or an MP3 or WAV preview. MusicXML and MIDI need the score unlocked (15 credits, once); until then the PDF, page images and parts carry a small footer line. The audio is a General MIDI rendering of the score, not your voice. Our guide to MusicXML, MIDI and PDF explains which file to pick.

How to check a transcription against the recording

Work two bars at a time: select them, turn on looping, and listen three ways before you change anything. The player switches between Score, Both and Recording while it plays. Start with the recording alone to fix the melody in your ear, then the score alone, then both together. A wrong pitch clashes with your voice; a note in the wrong place sounds doubled, like an echo.

Fix things in order. Check the bar lines first, because if the first beat is wrong, every bar is wrong: mark a pickup bar, or re-trim the file and transcribe again. Then fix pitches, then lengths, and leave expression and layout for last.

What to expect from the first draft

Expect a draft to correct, not a finished part. We haven’t measured singing yet, so this list comes from how the One instrument or voice choice works and from our measurements on 12 solo and accompaniment guitar recordings from the public GuitarSet dataset, run on October 2, 2026 and again on October 4.

• One instrument or voice starts in 4/4 on a straight grid at the detected or supplied tempo; triplets and swing are not inferred. Check the rhythm and change the time signature in Score settings where needed. Solo piano can follow 3/4, 4/4, 6/8 and 12/8, but it is for piano.
• Note ends are less reliable than note starts. On those guitar recordings on October 2, Basic Pitch scored a macro onset F1 of 0.744, but 0.508 when the end of each note had to match too. An October 4 re-run, after we began retuning off-pitch takes and writing a melody one note at a time, scored 0.763 and 0.522. Onset F1 counts a note as found when its pitch matches and its start lands within 50 milliseconds, then balances missed notes against invented ones; 1.0 is perfect. Our accuracy measurements explain the method.
• An out-of-tune take is corrected first. A recording that is steadily 10 cents or more sharp or flat of A440 is retuned before transcription, so its notes are written at concert pitch; a voice whose pitch drifts is left as it is.
• Written lengths are simplified for reading. MuseScore Studio wrote 26% of the guitar notes longer than they were played (after snapping to the grid), by up to three sixteenths, and never shorter. Playback and the MIDI export keep your timing; change a note’s length if the page should show what you sang.
• Repeated notes need a look. Two notes on one pitch with no gap between them give the model little to separate, so check for one long note where you sang two. To fix it, shorten the note, then fill the rest that appears with the same pitch.
• The key is a guess. It’s estimated from the notes; on the guitar set it was wrong for 2 of the 12 recordings, both times a fifth away. Set the right one under Key & clef.
• Check the melody line. A detected single-note line is written as Melody on one staff, keeping the strongest of notes struck together and dropping faint echoes, while a part playing chords keeps them on one staff. For a single-part recording, Score settings → Layout can apply One melody line to an older score too, using the existing notes. It removes extra notes from playback; Undo restores them. Check that it kept the notes you sang.
• You get notes only. ScoreStarling doesn’t write lyrics or chord symbols from a recording.

How to write a melody down by ear instead

If no tool suits the job, you can write the melody down yourself with a keyboard (a phone app will do) and a player that loops a passage. It’s slower, and it trains your ear.

1. Find home. Hum the note the tune wants to end on and find it on the keyboard. That is usually the key’s home note, which suggests a key signature to try.
2. Find the beat. Tap your foot to the recording and count. A strong beat every three taps suggests 3/4, every four taps 4/4; if each beat splits into three, think 6/8.
3. Write the rhythm first, one bar at a time, as marks under the counts.
4. Add the pitches. Start from a note you know, judge each next note as a step, a skip or a leap, and check it on the keyboard.
5. Play your version back against the recording and sing along. Anything that rubs is a note to fix.

Other apps take singing too. ScoreCloud’s Express app for iPhone and iPad notates singing, whistling or humming, and Klangio’s Sing2Notes runs in a browser and on iOS and Android, with PDF, MIDI and MusicXML downloads (both as described on their own sites, checked October 3, 2026). We haven’t tested either.