Why does polyphonic MP3 to MIDI add extra notes?

The 2003 TS-AudioToMIDI help described four recognition methods. One was assigned to polyphonic material. The other three were monophonic methods, each designed to follow one voice rather than report several simultaneous notes.

That split is useful because it shows that “audio to MIDI” was not one universal operation even inside a single historical product. Choosing a polyphonic method changed the task. Instead of stopping after one likely voice, the recognizer continued looking for more note candidates in the same slice of sound.

The archived explanation says three of the methods analyzed a spectrum from low pitches toward high pitches. When the system found a spectral maximum, it treated that peak as a possible fundamental. It then compared the surrounding spectrum with a user-set harmonic model. A candidate also had to clear a threshold before it was accepted as a note. After acceptance, the method subtracted the candidate's spectral components and continued its search.

This supplies a concrete route to an extra note. A recorded note does not contain only its fundamental frequency; its tone also includes higher components. In a chord or full mix, components from several sounds overlap. If the system interprets a remaining peak as another fundamental and the candidate passes its checks, the MIDI output gains a note that the listener may regard as false.

The vendor warned that its polyphonic method could produce unwanted notes and pointed users to the harmonic model and threshold as adjustments. That is a historical vendor instruction, not a measured success rate. The page provides no shared test set, error count, or comparison against a current converter, so none is inferred here.

The two controls addressed different questions. The harmonic model represented which frequency relationships were plausible for a note's spectrum. The threshold decided whether a candidate was strong enough to keep. Changing either control could alter which peaks survived, but the archive does not provide one setting that was correct for every instrument or recording.

The method choice also mattered. A monophonic route could deliberately return only one voice, while the polyphonic route searched onward for simultaneous notes. Fewer notes from the monophonic method would not prove that it understood the recording better; it would reflect a narrower task. Likewise, a denser polyphonic result could contain both real chord tones and false candidates.

A current open-source example shows why the idea remains testable without pretending that old and new algorithms are identical. Spotify Basic Pitch's inference source defines onset, frame, and minimum-note-length thresholds. These are not the same controls as TallStick's harmonic model, and this page does not equate them.

The shared lesson is narrower: transcription produces note events by applying decisions to uncertain evidence. A threshold can reject weak evidence, but raising it may also remove a quiet real note. Lowering it may recover that note while admitting more fragments. There is no unsupported “best” number here; the useful result is the tradeoff you can hear and inspect.

1. Use a short excerpt. Choose a passage where the unwanted note appears more than once, so you can compare the same event rather than guess from a whole song.
2. Place audio and MIDI together. Align the original recording with the converted MIDI in a DAW and solo one suspicious note at a time.
3. Look for a repeated relationship. If each false note appears with one source note and follows its timing, a harmonic may be getting accepted as a separate fundamental. Treat this as a hypothesis, not proof.
4. Check for transient fragments. Very short notes clustered at attacks suggest a different decision from a long, stable false pitch. Fix the category you can see instead of applying one blanket edit.
5. Compare a cleaner source if you have one. Convert the isolated instrument or stem containing the same phrase. If the false note disappears, overlapping material in the mix was relevant; if it remains, inspect the instrument's own spectrum and attacks.

This test does not reveal a model's private internals. It narrows the repair: remove an isolated false pitch, shorten transient fragments, or choose a source with less overlap. For the broader choice between a single-line, stem-first, and full-mix route, see Can you turn an MP3 into MIDI without getting a mess of notes?

Start with notes that repeat the same suspicious pattern. Audition them with and without the source audio. Delete or mute one group, then replay the phrase before touching legitimate chord tones. This keeps the correction reversible and separates false positives from music that is merely dense.

If the output is crowded across the entire range, changing the input route is usually more informative than deleting hundreds of events. A short isolated part gives you a clearer comparison and lets you determine whether overlap, the instrument's tone, or the detector's event decisions dominate the result.

If the unwanted events came from a beat rather than a pitched melody, read how drum audio and percussion note maps differ from melodic MP3-to-MIDI output.