Voice to MIDI: Convert Singing or Humming into Notes
Short answer
Yes. And a voice is the easiest thing this converter is ever asked to read.
Not because anything here is tuned for singing, but because a voice already has the shape the tracker wants: one note at a time. In this build's own test on a C major scale with the ground truth known, 7 to 8 of every 8 notes come back at the correct pitch. A melody sung by one person is that same problem, just in a range you can name.
What breaks it is not the singing. It is the backing track, the vibrato, and the "s" sounds between the notes.
Most people start by throwing a full song at it and get junk back. Start with your own voice instead.
Test it on your own recording Runs in your browser, no upload, no sign-up.
Why singing is the case this tracker was built for
In Melody mode the tracker reads one fundamental pitch per short frame of audio. Hand it a finished mix and it has a choice to make at every instant: drums, bass, chords, or voice.
Which one does it take? Whichever is loudest. That is why a full mix comes back as the bass line.
A solo voice removes the choice entirely. There is nothing to pick between.
It also removes the second-hardest problem, which is polyphony. Can a one-note-at-a-time tracker hear five notes at once? No. And that is the single biggest reason dense piano and jazz convert badly.
A melody sung by one person is never five notes at once. It is one note, then the next one.
What we measured is 7 to 8 of every 8 notes on a clean single-line source. What we cannot tell you is how your particular voice will do, because that depends on the take.
Where a voice actually breaks
Three things. Only one of them is about the singing.
- The backing track. If the voice sits inside a mix, the tracker follows the loudest part rather than the singer. Use an a cappella take, or a stem.
- Vibrato. Measured on this build, vibrato shifted the first note's onset by a few hundred milliseconds. The pitches were still right. The attack moved.
- Sibilants. The "s", "sh", "t" and "k" sounds are broadband noise with no pitch in them. In the noisy version of the same test every correct note was still there and in order, but extra low notes appeared in the gaps between them. Sung consonants land in exactly those gaps.
All three are things you fix before you convert, not after.
One more case surprises people: a rap or spoken-word take. Speech has no sustained pitch for a tracker to follow. A tracker looking for one fundamental per frame will find pitch where a vowel is held, and nothing where the delivery is percussive.
This build has not been measured on spoken word, so treat that as reasoning from how the tracker works rather than a number. It is the wrong job for this tool, and no setting fixes it.
The one setting that matters most: Note range
For a voice, the highest-value control on the page is also the easiest one to use. Note range tells the tracker which span of pitches is allowed.
Set it to the singer's actual span and two things happen at once. The wrong octave stops being available, and false notes outside the range are removed rather than filtered out afterwards.
The octave error looks like a bug and is not. The pitch estimate for a short frame can lock onto a harmonic instead of the fundamental, so the note is real but written an octave above or below where you hear it.
It shows up most on bright, harmonically rich sources, and a trained voice is exactly that. If the singer never goes above A4, there is no reason to leave C7 available.
What comes back, and what is gone
You get a standard MIDI file you can edit. SMF format 1, one track holding the track name, a tempo event carrying the BPM you set, and the notes.
The division field in the header is 480 tpqn, so one beat is split into 480 steps and a sixteenth note is 120 ticks. Note times are rounded to the nearest tick as the file is written, so the grid itself can never move a note by more than half a tick.
Half a millisecond at 120 BPM. That is the whole timing error the file contributes. Everything you can actually hear comes from the performance instead.
What is gone is the words. MIDI stores note events, not lyrics. There is no field for a syllable, so a sung line arrives as pitches and durations with the words stripped out.
So where do the words go? Nowhere in the file. They stay in the recording and in your ears.
Also gone: breath, tone, and any sense of the vowel. You get the notes that were detected, not the performance that was given.
How to do it, in order
- Cut to the take you want. Solo voice, no backing, twenty to forty seconds.
- Trim the silence at the head and tail, and any reverb tail. Reverb glues one note to the next, and in this build's test it split one sustained note into two.
- Set Note range to the singer's real span. This is the highest-value step on the page.
- Set Analyse to the first 30 seconds and convert once. If the settings hold, run the whole file.
- Look at the piano roll before you download, and press play to hear what was detected. If an octave looks wrong, fix the range rather than editing the notes afterwards.
Step two is the one everybody skips, and it is the one that costs the most notes.
FAQ
Can you convert a voice recording to MIDI?
Yes, and a solo voice is the easiest input this converter gets. A voice is one note at a time, which is exactly what a melody-mode tracker reads. In this build's own test on a C major scale, 7 to 8 of every 8 notes came back at the correct pitch. Singing a melody is the same problem in a narrower range.
Does humming work better than singing?
Humming is usually cleaner, and that is reasoning from how the tracker works rather than a measurement: a hum is a steady vowel with no consonants, so nothing hisses in the gaps between notes. Singing adds words, and those consonants are broadband noise with no pitch in them. If all you need is the melody, hum it.
Why did my converted voice come out an octave wrong?
Because the pitch estimate for a short frame can lock onto a harmonic instead of the fundamental. The note is real, but it is written an octave above or below where you hear it. It shows up most on bright, harmonically rich sources, which is what a trained voice is. Narrow the Note range to the singer's actual span and the wrong octave stops being available.
Can it convert a song where the voice is mixed with instruments?
Not reliably. Melody mode follows one pitch at a time, so in a full mix the tracker follows whatever is loudest at each moment instead of the singer. In this build's own test on a melody over bass and percussion, every recovered note came from the bass part. Use an a cappella take or a stem.
Does vibrato affect the conversion?
The pitches survive. The timing moves. Measured on this build, vibrato shifted the first note's onset by a few hundred milliseconds. At 120 BPM that is a few hundred ticks, far larger than the half-tick rounding the file itself introduces. Trimming to a phrase with a steadier opening avoids it.
Can it convert rap or spoken word to MIDI?
This is the wrong job for this tool, and no setting fixes it. Speech has no sustained pitch for a tracker to follow. There is pitch where a vowel is held, and nothing where the delivery is percussive. That is reasoning from how the tracker works rather than a measurement, because this build has not been tested on spoken word. Use a sung or hummed take instead.
Are the lyrics included in the MIDI file?
No. MIDI holds note events, not words. There is no field for a syllable, so a sung line arrives as pitches and durations with the words stripped out. The file is written as SMF format 1 at 480 tpqn, one track, with a tempo event carrying the BPM you set. Nothing is uploaded, there is no sign-up and no watermark.
Is my recording uploaded anywhere?
No. Your browser decodes the file and runs the pitch tracker on your own device, and the MIDI file is built in memory. There is no account, no upload and no e-mail. You can check it yourself: open the network panel, convert something, and watch. Nothing moves.
Can I convert vocals to MIDI?
Yes. A single vocal line is the same job as a solo voice, because it is one note at a time and that is what melody mode reads. In this build's test on a C major scale with the ground truth known, 7 to 8 of every 8 notes came back at the correct pitch.
Check what the word means in your own session before you convert. A lead vocal on its own is one line. A doubled vocal, a harmony stack, or a group of singers is two or more pitches at the same time.
That last case is not a voice problem any more. The tracker reads one pitch per frame, so a harmony comes back as one of its parts instead of all of them.
That is reasoning from how the tracker works rather than a measurement. Export one vocal line and convert that.
Is there a free voice to MIDI converter?
Yes. This one is free, and free here does not mean a reduced version of something else. You get the same tracker, the same controls, and the same SMF format 1 file at 480 tpqn.
The reason it can be free is where it runs. The conversion happens on your own device, so there is no server time to charge for. There is no account, no upload and no watermark.
The one thing that limits a run by default is the Analyse setting. It starts at 30 seconds, and you can set it to 60 seconds, 2 minutes, or the whole file.
The ceiling that is not ours to lift is your device's memory. Decoding a long recording is what it costs you, not a plan.
Judge it yourself in a minute. Hum eight bars into your phone, put the file through, and look at what comes back. That tells you more than any page can, including this one.