mp3tomidiflowOpen converter

HUM TO MIDI · SING TO MIDI · PRIVATE

Vocal to MIDI: Turn Singing or Humming into Notes

Record yourself singing or humming a melody, then convert the recording to MIDI on the homepage. The converter detects the pitches you sang and writes them as notes you can play back with any instrument in your DAW. A clean, dry, single-voice recording gives the best result. Lyrics are not transcribed — only notes.

Open the free converter

How to convert, step by step

  1. Open the converter. Go to the MP3toMIDIFlow homepage. The converter is at the top of the page; nothing needs to be installed.
  2. Choose your file. Select or drag in one vocal recording, such as a phone voice memo. Use a single, dry voice with no backing track.
  3. Convert in the browser. Select Convert to MIDI and keep the tab open. The note-detection model and decoder load from this site and run on your device; your audio is not uploaded.
  4. Download and edit. Download the .mid file and open it in a DAW or notation editor to set the tempo, fix wrong notes and choose an instrument sound.

What vocal to MIDI is good for

Humming a melody into your phone is often the quickest way to capture an idea. Converting that recording to MIDI gives you a starting point you can play on a synth or piano sound, harmonize, or tidy up into a lead line. It also works for a solo sung line from a recording you made yourself, for example to sketch a melody before writing parts around it.

How singing is turned into MIDI notes

The recording (a voice memo in M4A, MP3, WAV and other supported formats, or the audio track of an MP4 or MOV video) is decoded in your browser, mixed to mono and analysed by Spotify's open-source Basic Pitch model. The model estimates which pitch is sounding at each moment and where notes start, and the converter writes the detected notes into a .mid file.

The voice is not a keyboard: singers glide between notes, add vibrato, and drift slightly sharp or flat. MIDI notes in this file have fixed semitone pitches and no pitch-bend data, so the converter has to round what you sang to the nearest note. Slides, scoops and wide vibrato can therefore come out as a short extra note, a note split in two, or a neighbouring semitone.

Recording tips for cleaner vocal MIDI

Limits for vocals

After conversion

Load the .mid file onto a simple instrument such as a piano or a plain synth lead, play it alongside your original recording and correct wrong notes phrase by phrase. Because the voice wanders, pitch errors usually cluster around slides and long held notes, so check those first.

Limits that apply to every file

One 9.1-second sung take: 40 MIDI notes, not a perfect score

The 9.1 s vocadito take produced 40 unedited MIDI notes in Chrome on 10 Oct 2026, and only 14 of annotator A1's 35 reference notes matched at the strict 50 ms / 50-cent rule. That is a measured result from one solo vocal file, not a claim about every hummed or sung input.

The source is a 9.1-second solo vocal from the vocadito dataset, tested on 10 October 2026 in Chrome. The unedited output is a 40-note MIDI file. A successful download does not mean note-perfect transcription: 26 of the 40 output notes had no match in the A1 reference at the strict rule, and 21 of A1's 35 reference notes were missed.

Match counts for the 9.1 s vocadito take, unedited 40-note MIDI output, Chrome, 10 Oct 2026. Pitch difference allowed: 50 cents. Maximum one-to-one matching; duration not scored.
Onset windowReference setReference notesMatchedUnmatched referenceUnmatched output (of 40)
50 msA135142126
50 msA227101730
200 msA135221318
200 msA22722518

Files used: the source WAV at /guides/assets/real-vocal-20261010/vocadito_10.wav and the unedited MIDI at /guides/assets/real-vocal-20261010/vocadito_10-output.mid. The full write-up is the real vocal to MIDI example guide.

The same take improves from 14 to 22 matches when the onset window loosens from 50 ms to 200 ms

For annotator A1 the matched count rises from 14 at 50 ms to 22 at 200 ms, and for A2 from 10 to 22. This is a scoring-window change on the same 40-note output, not a better model run.

Unmatched output notes fall from 26 of 40 (A1) and 30 of 40 (A2) at the strict window to 18 of 40 for both annotators at 200 ms. The pitch rule stays at 50 cents in every row; only the allowed onset difference changes.

Effect of the onset window on the same unedited 40-note MIDI output. Pitch difference allowed: 50 cents in all rows. 50 ms is the strict rule; 200 ms is diagnostic.
Reference setMatched at 50 msMatched at 200 msUnmatched output at 50 msUnmatched output at 200 ms
A114222618
A210223018

A looser window is more forgiving, not more accurate. If two hummed notes are closer than 200 ms, a looser check will count them as hits, and a DAW user will still need to split or merge them by ear. Read the strict 50 ms row when you want the harsher view; read the 200 ms row as a diagnostic.

A leftover MIDI 82 note and a short pre-flight for hummed takes

The unedited output includes a MIDI pitch 82 event that is not in either human reference range, so inspect or delete it before you use the file. It is one visible sign that the 40-note output is raw material, not a finished part.

Pre-flight checklist for a hummed or sung take:

The same example is documented in the real vocal to MIDI example guide, with the WAV and unedited MIDI downloads. For other audio sources, see the audio to MIDI converter page.

Match rate and note density on the same 9.1-second take

On this 9.1-second vocadito take, the strict 50 ms / 50-cent rule matched 14 of 35 A1 reference notes (40.0%) and 10 of 27 A2 reference notes (37.0%), while the unedited output still carried 40 note events, or 4.40 notes per second.

Loosening only the onset window to 200 ms raises those match rates to 22 of 35 for A1 (62.9%) and 22 of 27 for A2 (81.5%). The output does not change; 18 of the 40 MIDI notes (45.0%) still have no one-to-one reference match at 200 ms, down from 26 of 40 (65.0%) for A1 and 30 of 40 (75.0%) for A2 at 50 ms.

The two human references are themselves a different density: A1 labels 35 notes (3.85 per second) and A2 labels 27 notes (2.97 per second) on the same 9.1 s file. The converter output at 4.40 notes per second sits above both, which is consistent with extra split notes and the leftover MIDI 82 event rather than with a missing-note pattern alone.

Match rate and note density for the 9.1 s vocadito take (counts from the 10 Oct 2026 Chrome run; duration not scored)
SetNotesNotes per secondMatch rate at 50 msMatch rate at 200 ms
Annotator A1353.8514/35 = 40.0%22/35 = 62.9%
Annotator A2272.9710/27 = 37.0%22/27 = 81.5%
Unedited MIDI output404.40unmatched 26/40 = 65.0% (vs A1), 30/40 = 75.0% (vs A2)unmatched 18/40 = 45.0% (vs A1 or A2)

These percentages are arithmetic on the published counts from one monophonic example. They are not a product-wide accuracy score and should not be compared as if Jukeblocks, Basic Pitch, or Mirelo had been measured on the same file. Under the strict 50 ms rule, A1 is short by 21 missed reference notes and the output still has 26 unmatched events, so the disagreement is not only deletions or only insertions. At the 200 ms window, A1 and A2 both reach 22 matched notes, even though A1 started from 35 reference notes and A2 from 27.

Questions

Can I hum a melody and get MIDI?

Yes. Record the humming (a phone voice memo works), choose the file on the converter and download the detected notes as MIDI. Clear, separated notes give the best result.

Does it work with phone voice memos?

Yes, if the memo is in a supported format such as M4A, MP3, WAV or AAC, or is the audio track of an MP4/MOV video, and it is under 50 MB and 3 minutes.

Can it extract the vocal melody from a song?

Not on its own. The converter transcribes everything pitched in the recording and does not separate the voice from the instruments.

Why does my vibrato become several notes?

The MIDI file stores fixed semitone notes without pitch bend, so wide pitch movement is rounded into separate notes. Sing with a straighter tone or delete the extras.

For mixed audio (not only dry voice), the same pipeline lives on the audio to MIDI converter; format notes are collected under audio tools.

Related pages

Open the free converter