Why Audio to MIDI Adds a Note an Octave Below

What the extra note looks like in a piano roll

A documented case makes the pattern clear. A user played the interval C4 and G4, a perfect fifth, into a transcription tool. The output contained C3, C4 and G4. The two notes that were played are both present and correctly placed. The third note is an octave below the lower one, and it was never played.

In a piano roll, this reads as a stack rather than a melody: a long low note under a pair of shorter ones, often starting at the same tick as the lower note of the interval. If you transpose the whole excerpt up an octave, the ghost usually follows. If you delete the ghost and the remaining two notes still match the recording, you are looking at a false fundamental rather than a missing note or a shifted pitch.

It helps to separate this from two neighbours. A missing note is an event that should exist and does not. Bleeding is two real lines collapsing into one. A false fundamental is the opposite of bleeding: one line is invented where the source only implied it.

Why a fifth produces a ghost an octave below

Start with the numbers, using equal temperament and A4 at 440 Hz. C4 is about 261.63 Hz. G4 is about 392.00 Hz. C3 is about 130.81 Hz. Now list the harmonics of that C3: the first is 130.81 Hz, the second is 261.62 Hz, and the third is 392.43 Hz. The second harmonic is C4. The third is G4.

That is the whole problem. A real C3 played on a piano contains strong energy at C4 and G4 whether or not anyone plays those keys, because those frequencies are its second and third harmonics. A microphone recording of C4 and G4 together therefore produces a spectrum that is also consistent with a single low C3. The tracker has to choose. When it chooses the lower root, you get three notes; when it chooses the two upper notes, you get two.

This ambiguity has a name in signal processing: the missing fundamental. It is the same effect that lets a small speaker suggest a bass note it cannot physically reproduce. The ear infers the low pitch from the spacing of the harmonics; a pitch tracker does the same thing with a mathematical model instead of a cochlea.

It also explains why the ghost is C3 and not G3. G3 is about 196.00 Hz. Its second harmonic is G4 at 392 Hz, which fits, but its third harmonic would be near 588 Hz, close to D5 — not C4. A G3 root cannot account for the C4 energy, so it is a weaker explanation than C3, which accounts for both notes with two adjacent harmonics.

Two conditions make this more likely. The first is register: the lower the interval, the more harmonics land inside the band where pitch estimators are confident. The second is timbre: bright instruments with strong low harmonics — piano, accordion, distorted guitar, some synth patches — supply more of the evidence a false fundamental needs. A pure sine tone has only one harmonic, so it cannot be misread this way.

Three ways to confirm it is a false fundamental

1. Replace the notes with sine tones. Play the same two pitches as pure sines, or generate them in a DAW, and convert again. Sines carry almost no harmonic energy, so the harmonic explanation disappears. If the ghost vanishes, the cause was harmonic inference rather than a timing or decoding fault. If it survives, look elsewhere — the problem may be in the file, the export, or the tracker's onset logic.
2. Move the interval up an octave. Convert C5 and G5 instead of C4 and G4. A false fundamental should move with the input: you would expect C4 as the ghost, and you would expect it to be weaker, because the harmonic spacing now sits higher and overlaps less with the estimator's confident range. A ghost that stays at C3 regardless of the input octave is not following the harmonics, which points to a different bug.
3. Count events and read the note numbers. Open the exported file in a piano roll and read the MIDI note numbers rather than the names. C3 is 48, C4 is 60, G4 is 67. If you see 48, 60 and 67, the file is internally consistent and the question is only whether 48 belongs there. Do not rely on the octave label alone: different programs disagree about whether middle C is called C3 or C4, which is why the note number is the safer reference.

Run the tests in that order. The sine test isolates the cause, the octave test confirms the mechanism, and the event count tells you what is actually in the file. Changing two variables at once — a new converter and a new recording — produces a result you cannot interpret.

Fix it in five steps

1. Convert one note at a time. A monophonic sketch is designed for a single line. If you feed it a fifth, you are asking it to resolve an ambiguity it was not built to resolve. Track the melody alone, then add the harmony by ear or with a polyphonic tool.
2. Raise the register of the test phrase. Move the same musical idea up an octave and compare. If the ghost weakens or disappears, you have found a practical workaround for that part: play or transpose it higher, convert, then transpose the MIDI back down.
3. Remove the ghost deliberately, not blindly. Select the suspect note by pitch, confirm it against the waveform, then delete it. Keep a copy of the unedited MIDI so you can compare. If the ghost overlaps a real bass note that you also need, do not delete the whole pitch class — split the note at the point where the real bass enters.
4. Check the note length before you blame the detector. AudioNote MIDI discards detected segments shorter than 0.15 seconds, and it keeps notes from C2 through B5. A very short low note can vanish for a reason unrelated to harmonics, which makes the remaining stack look stranger than it is.
5. Re-run a fixed seven-note test. Convert the same short phrase you used before the change and count the events. A stable count across two runs tells you the workflow is reproducible; a count that changes between runs tells you the input, not the tool, is the variable.

None of these steps requires editing the audio. The goal is to reduce the number of competing explanations the tracker has to choose between, then verify what came back.

When a different tool is the better fit

If the part is genuinely two or more simultaneous notes, a monophonic converter will keep producing one of two failures: a merged line or an invented root. The honest fix is a polyphonic transcriber, or separating the part into stems before conversion. The same applies when you need to keep vibrato and slides: a note sketch that stores one pitch per event cannot represent a bend, which is why pitch bends need explicit support rather than a better setting.

If you only need to check a melody before committing to an arrangement, the reverse is true: a short local pass is faster than opening a DAW project. AudioNote MIDI runs in the browser, keeps the audio on your machine, and accepts files up to 100 MB and 120 seconds. Those limits are deliberate — a two-minute excerpt is enough to answer the question "is this the right melody", and it keeps memory use bounded. Try it on a short solo phrase and inspect the note numbers before you build anything on top of the result.

Quick answers

Is the extra note a bug in the converter? Not necessarily. The spectrum of C4 plus G4 is genuinely consistent with a low C3, so a tracker that prefers the lower root is making a defensible choice on ambiguous evidence. It becomes a bug when the same input reliably produces a note that the tool's own documentation says it cannot represent.

Why does it happen on fifths and power chords more than on thirds? A perfect fifth is close to the 3:2 frequency ratio, which is exactly the spacing between the second and third harmonics of a root an octave below. Major and minor thirds sit at ratios that do not line up as neatly with adjacent low harmonics, so they offer the tracker a weaker low-root explanation.

Can I just delete the ghost and move on? Yes, if you have confirmed it against the recording. Count the events first: if the file has one more note than you played, deleting the extra one restores the count. If the count matches but a pitch is wrong, you are dealing with a different problem and deletion will not fix it.

Sources and scope

The C4/G4 to C3/C4/G4 case is documented in the Basic Pitch issue tracker: spotify/basic-pitch issue #205, "False fundamentals", opened on October 5, 2026. The frequencies above are calculated from equal temperament with A4 at 440 Hz; they are not measurements of a specific recording.

This guide describes a diagnostic method, not a guarantee. It has not been validated across every instrument, microphone, or codec, and the harmonic explanation does not cover cases where the extra note comes from room noise, edit points, or a decoding error. The tool limits quoted here — monophonic output, C2 to B5, a 0.15-second minimum segment, 100 MB and 120 seconds — are the documented limits of AudioNote MIDI at the time of writing.