MP3 to MIDI Without Uploading Your File

What "no upload" actually means here

Most audio to MIDI services work the same way: you pick a file, it is sent to a machine somewhere, a model runs there, and the MIDI comes back. That is a reasonable design if the model is large. It also means your recording sits on someone else's disk for as long as their retention policy says.

This page does not do that. There is no account, no queue and no backend. When you choose a file, the browser reads it from your disk into memory and the analysis runs right there. If you open your browser's network tab while a conversion runs, you will not see the file going out. The only requests the page makes are for the page itself and its analytics tag.

How the analysis runs without a server

Everything has to fit inside a browser, so the method is deliberately narrow. The chain is:

• The file is decoded with your browser's own audio decoder - the same one it uses for playback.
• The result is mixed to mono and resampled to 16 kHz. Pitch tracking does not need the full band, and a lower rate keeps the analysis fast enough to run on a laptop.
• The signal is walked through in short overlapping frames. Each frame is scored with YIN pitch detection - a difference function followed by a cumulative mean normalised difference - with the lag search restricted to the range that matches 60-2000 Hz.
• The first lag that drops below the threshold is taken as the fundamental for that frame. Runs of frames that agree on a semitone are merged into one note.
• The notes are written out as a standard MIDI file: format 0, one track, 480 ticks per quarter note.

Nothing in that chain needs a trained model, which is exactly why it can run on your side of the wire.

What the 60-2000 Hz window means for your file

The lag search is bounded on purpose, and the bounds are worth knowing before you convert something.

Range	What falls inside
60-2000 Hz	The tracked range. Covers bass lines up to the top of most melodic instruments and voices.
Below 60 Hz	Not tracked. Very low sub content has no note to report here.
Above 2000 Hz	Not tracked. Harmonics up there are ignored rather than mistaken for the fundamental.

A 2000 Hz ceiling is above the fundamental of every common melodic instrument, so for a single line it is not the thing that limits you. What limits you is polyphony - see below.

Where this approach stops

The trade-off for running locally is that the method tracks one pitch at a time. That is a property of the technique, not a setting:

• Chords and full mixes collapse to a single note - there is one fundamental per frame by design.
• Instruments are not separated, and chords are not detected.
• Percussion and other unpitched sound has no fundamental to track, so it produces nothing useful.
• One file at a time; there is no batch mode.
• Files above 80 MB are turned away before decoding starts.

Where a paid service offers chord recognition or one MIDI track per instrument, it is running a trained model on a server. That is a different method with a different cost, and it is the reason those services need your file. This page does not do that, and does not claim to.

What tends to convert well

• A hummed or sung melody
• A solo instrument recorded on its own - piano, guitar, flute, violin, bass
• A whistle, a synth lead, a bass line

These are all single lines, which is the case the method is built for. A clean recording of one of them usually comes out usable; anything denser needs cleaning up in a DAW afterwards.

Checking the result instead of trusting it

After a conversion the page reports how many notes it found, how long the analysis took and the pitch range it landed on. Those three numbers are the quickest way to tell whether the run went well - an implausibly low note count or a range that sits outside what you played usually means the source audio was too dense or too noisy for a single-line tracker.

Open the converter and try your own file