Why Two Instruments Collapse Into One MIDI Line

What “bleeding” looks like in a piano roll

The symptom is easier to recognise than to describe. Open the exported .mid and look at the note lane rather than the waveform:

• The line alternates. A phrase starts on the piano's register, jumps up to the accordion's sustained note for a beat, then drops back. Nothing overlaps, because a single line cannot overlap itself.
• Sustained notes get chopped. The accordion holds one pitch for two bars, but the MIDI breaks it into short fragments wherever the piano is momentarily louder.
• Chords lose their middle. A three-note voicing arrives as one or two notes, usually the top and the bass, because those are the most prominent partials.
• Notes appear in a register that is not in the recording. When two sources are close in pitch, the estimate can land between them instead of on either one.

If you see two or more of these at once, the converter is not misreading a clean performance. It is being asked a question it cannot answer: which of these simultaneous pitches is the note?

Pitch tracking follows one winner at a time

Monophonic transcription estimates a single fundamental frequency per short analysis frame — typically a few tens of milliseconds. In every frame, the algorithm has to commit to one answer. When a piano and an accordion sound in the same register, both produce strong harmonic series, and the harmonics overlap. The tracker picks whichever candidate is louder, more periodic, or closer to the previous frame's estimate, then moves on to the next frame and picks again.

What you hear as “two instruments” is, frame by frame, a sequence of single decisions. The output is therefore not two parts mixed together — it is one part that changes identity several times per second. That is why the result can sound recognisable in places and nonsensical in others: it is accurate about a pitch at each moment, and wrong about which instrument that pitch belongs to.

This is a property of the task, not a defect in a particular product. Any tool that returns a single melody line makes the same commitment. A polyphonic transcriber avoids it by allowing several simultaneous notes and then grouping them into parts — a different problem with its own failure modes, mainly spurious notes and octave errors.

Three checks that confirm it is bleeding, not a bad recording

Before changing your workflow, rule out the simpler causes. These three checks take about ten minutes and settle the question.

1. Solo test. Mute one instrument in the multitrack and convert the remaining part. If the MIDI is clean, the problem was overlap, not the converter. If it is still messy with one instrument playing alone, the input has a different issue — see why an audio to MIDI conversion can miss notes.
2. Register check. Look at the written ranges. Piano left hand around C2–C3 and an accordion melody around C4–C5 are far enough apart that a single tracker can often follow the melody without much damage. Two instruments inside the same octave — piano right hand and accordion both around C4 — is where collapse becomes near-total.
3. Density check. Count how often the exported line changes register within a bar. A melody that stays in one tessitura and moves stepwise is usually a real single line. A line that leaps between two stable registers every beat is two parts being sampled alternately.

One more signal: if the note count is far below what you can hear, the tracker is not “missing” notes so much as declining to choose among them. A dense mix and a quiet mix can both produce a sparse MIDI file for different reasons.

Fix it before conversion: four ways to isolate one part

The reliable fix is to give the converter one voice. In rough order of quality of result versus effort:

1. Re-record the part alone. Best result by a wide margin. If the performer is available, one clean take of the melody beats any amount of post-processing, and it removes the entire class of overlap errors.
2. Convert a single-voice excerpt. Often the melody you want only exists alone for eight or sixteen bars — an intro, a break, a solo passage. Convert that section, then assemble the rest in the DAW. This costs nothing and is frequently enough for a usable draft.
3. Run source separation first, then convert each stem. Separation tools split a mix into approximate instrument stems. Convert the stem you care about instead of the full mix. Expect artefacts: separation leaves bleed and smearing, especially on sustained reeds and on anything sharing a register, so the MIDI still needs editing. It is a real improvement over converting the mix, not a clean solution.
4. Edit a single channel of a stereo file. If the two instruments were panned apart in the mix, one channel may contain far less of the other. This only works when the mix actually has separation; a centred arrangement gives you two nearly identical channels. Extract the cleaner side in an audio editor, then convert that file.

What does not work reliably: EQ. Cutting the accordion's fundamental does not remove its harmonics, and those harmonics are exactly what the pitch tracker reads. Narrow EQ notches tend to damage the part you are trying to keep while leaving enough of the other instrument to keep confusing the estimate. Panning tricks applied after the fact have the same problem — the information is already summed in the file.

When you need real polyphonic transcription

If the goal is separate, editable parts for a full arrangement, a monophonic workflow will not get there no matter how the input is prepared. You need a tool that outputs several simultaneous notes. The common options:

• Open-source polyphonic models. Projects such as BasicPitch return multiple simultaneous notes from a mixed input. They handle chords and overlapping lines that a single-voice tracker cannot, at the cost of spurious notes that must be cleaned up.
• Polyphonic pitch editing in a DAW or dedicated editor. Some editors can detect and separate notes inside a polyphonic recording, letting you move individual notes within a chord. This is the closest thing to what most people mean by “split the instruments”, and it is also the slowest to work with.
• DAW audio-to-MIDI conversion. Several DAWs include a convert-to-MIDI command that can return harmony or melody as multiple notes. Quality varies widely by source material, and results usually need manual correction.

None of these is a one-click answer, and all of them produce output that needs review. The honest comparison is not “which is accurate” but “which failure mode can you fix faster” — extra notes you delete, or a single line you have to rebuild from scratch.

Quick answers

Why do two instruments become one track? Because the converter outputs one melody line. It estimates one pitch at a time, so simultaneous parts cannot be stored as separate tracks — they are sampled alternately into a single line.

Why does a chord come out as one note? The same reason. A single-voice estimate commits to one fundamental per frame, usually the most prominent one, so the remaining chord tones are not represented.

Can I split them without re-recording? Sometimes. Source separation followed by per-stem conversion, or isolating a passage where the part plays alone, both work. Results need editing, and heavily overlapping sustained instruments are the hardest case.

When should I switch tools? When you need separate editable parts rather than a melody sketch. If one melody line is the deliverable, this tool is faster; if you need the arrangement, use a polyphonic transcriber or a DAW's convert-to-MIDI.

Is a monophonic tool ever the better choice? Yes — for a single melodic instrument, it produces a cleaner result with fewer spurious notes than a polyphonic model, because it cannot invent notes that are not there.

Sources and scope

• MIDI Association: Standard MIDI Files — how note events are stored and what a track contains.
• MIDI Association: About MIDI — Part 4: MIDI Files — the distinction between MIDI instructions and recorded audio.
• AudioNote MIDI audio-to-MIDI tool — current scope: one monophonic melody line per conversion, processed in the browser.

This guide describes the behaviour of single-voice pitch tracking in general and the current AudioNote MIDI workflow in particular. It does not claim that any named third-party tool will separate a specific recording cleanly; test your own material before relying on the output.