Song to MIDI: Why the Whole Song Is the Hardest Input
Short answer
Yes. But a whole song is the worst case for this kind of tool, and the reason is structural rather than technical.
A song is a finished mix: vocals, harmony, bass, drums and pads sounding at the same moment. This converter reads one pitch at a time, so from a full mix it returns whichever part is loudest. That is usually the bass.
In this build's own test on a melody over bass and percussion, every recovered note came from the bass part.
The fix is not in the settings. Split the song into stems first, then convert one stem at a time. The same tracker then reads a clean single-line part and returned 7 to 8 of every 8 notes at the correct pitch.
Try it on one stem, not the whole song Runs in your browser, no upload, no sign-up.
A song is the hardest input, and not because the file is long
Length is not the problem here, and neither is the file size. A three-minute MP3 and a three-second WAV fail in exactly the same place.
The problem is that a song is polyphonic. Several notes sound at the same instant, and this converter is a melody tracker, which means it follows one pitch at a time. When more than one pitch is present it has to choose one, and the rule it follows is loudness.
That is why a mix of five instruments does not come back as five tracks. It comes back as one line, and the line belongs to whatever sits loudest.
Worth saying plainly: this is a limit of reading notes out of a waveform, not a defect in this page. Any tool doing the same job faces the same choice.
What you actually get back from a whole song
Four things were measured on this build, and together they describe what a finished mix will give you.
The loudest part wins. On a melody over bass and percussion, every recovered note came from the bass part. The melody was in the audio and it was not in the output.
Reverb splits notes. A sustained note with a long tail was read as two notes. The tail looks like a second attack to a tracker that is listening for the start of a pitch.
Noise adds notes nobody played. In the noisy version of the same test, all eight correct notes were still there and still in the right order. Extra low notes appeared in the gaps between them.
Vibrato moves the start of a note. Measured on this build, vibrato shifted the first note's onset by a few hundred milliseconds. That is a timing error, and it survives every setting on the page.
So the honest description of song to MIDI for a finished mix is this: you get one line out of the arrangement, at a timing that drifts where the performance drifts.
Why one part at a time works better
The difference between those two results is not a setting. It is the input.
A stem is closer to a single line than a finished mix is. Take the vocal, or the bass, or the melody instrument on its own, and the tracker stops choosing between parts. There is only one part left to choose.
How much of a typical song survives the whole-mix route is not something this build has measured, so no percentage is quoted here.
Two settings matter more than the rest once you are working one stem at a time. Note range tells the tracker where to look, and narrowing it to the real span of the part removes the octave errors that come from searching too wide. Minimum note length drops the short blips that come from breath and string noise.
Chords mode does exist, and it reads spectral peaks instead of tracking a single pitch. It is labelled experimental for a reason: on dense material it loses notes. For a song, the reliable route is still one stem at a time in Melody mode.
So how do you get more than one line out of a song? You convert each part separately and stack the results.
What the file does not carry
A song has a key and a time signature. The MIDI file this page writes has neither.
The file carries note pitches, start times, lengths, a velocity taken from loudness, a track name and a tempo event. That is the whole list. Read out of the source, the writer emits a track-name event, a tempo event and an end-of-track event, and nothing about key or metre.
MIDI also cannot say whether a note is C sharp or D flat, because it stores a note number rather than a spelling. If the MIDI is going into notation software, that decision comes back to you.
Can you tell a wrong note from a wrong key signature? On paper they look nothing alike, and only one of them came from the audio.
None of this is a defect. It is what this converter writes, and knowing it in advance saves a confusing first import.
How to do it, in order
- Separate the song into stems. However you do that separation, do it before you convert rather than after. Vocals, bass, melody instrument, drums.
- Take one stem and cut it to thirty to sixty seconds. A short clip tells you whether the settings work before you commit the whole song.
- Trim the silence at both ends and any reverb tail. Reverb glues one note to the next, and in this build's test it split a sustained note in two.
- Set Note range to the real span of that part. This is the highest-value step on the page.
- Leave Mode on Melody, set Analyse to the first 30 seconds, and convert once. If the result holds up, run the whole stem.
- Repeat per stem, then stack the MIDI tracks in your DAW. Each stem arrives as its own file, and the arrangement is rebuilt where it belongs.
Step three is the one people skip. Which step do you think costs the most notes?
FAQ
Can I convert a whole song to MIDI?
Yes, and you will get one line back rather than the whole arrangement. This converter reads one pitch at a time, so from a finished mix it follows whichever part is loudest.
In this build's test on a melody over bass and percussion, every recovered note came from the bass. Split the song into stems and convert one at a time for a usable result.
Why does my song come out as just the bass line?
Because the bass is usually the loudest single pitch in a mix, and loudness is the rule this tracker uses to choose. It is not following the bass on purpose.
The melody is still in the audio. Narrowing Note range to the melody's real span, or converting the melody stem on its own, is what brings it out.
Should I use Chords mode for a song?
Not as a first move. Chords mode reads spectral peaks, and the page labels it experimental because dense material loses notes.
Melody mode on one stem at a time is the reliable route. Treat Chords mode as an experiment to try after the stem-by-stem pass, not instead of it.
Does the MIDI file include the key and the time signature?
No. The file carries note pitches, start times, lengths, a velocity, a track name and a tempo event, and nothing about key or metre.
So notation software has to guess both. Set them by hand once the file is open, and expect to fix the enharmonic spelling too, because MIDI stores a note number rather than a sharp or a flat.
Can it separate the vocals from the backing track?
No. Separating a mix into parts is a different job from reading pitches out of a waveform, and this page does not do it.
Whatever you use to separate stems, do that first. Each stem then converts on its own, which is where the usable results come from.
Can it convert drums to MIDI?
No, and the reason is worth knowing. Drums have no stable pitch to track, and this converter estimates pitch, so it has nothing to measure.
A kick or a tuned tom may come through as a note, because those do have a pitch. Cymbals and snare will not. Drum transcription is a separate problem from note transcription.
Is my song uploaded anywhere?
No. Your browser decodes the file and runs the pitch tracker on your own device, and the MIDI is built in memory. There is no account, no upload and no e-mail.
That matters more with songs than with practice clips, because an unreleased track is exactly the kind of file you should not be handing to a server. You can check it yourself: open the network panel, convert something, and watch. Nothing moves.
How long should the clip be?
Thirty to sixty seconds of a single stem. Set Analyse to the first 30 seconds, convert once, and judge the result before running the whole file.
Length itself is not the limiting factor. A three-minute stem converts the same way a thirty-second one does. The input, not the duration, is what decides the outcome.
Judge it on one stem, not on the whole song. Thirty seconds of a single part, no backing, no reverb tail. That tells you more than any page can, including this one.