MP3 to MIDI noise-tail check: what the detector keeps after the melody ends.
In this controlled WAV run, the four-note melody stayed at four events while a noise tail began after the final note; the silent control produced zero events.
Input condition
The test used mono PCM WAV at 11025 Hz with a deterministic four-note melody and a 35 ms-amplitude random-noise tail starting at 2.65 seconds. The exact detector and MIDI writer were evaluated in Node VM.
| Input | Observed output | What to inspect |
|---|---|---|
| melody-with-noise-tail.wav | Four MIDI events: pitches 60, 64, 67, 72 | Whether events appear after the final note |
| silence.wav | Zero output events | Whether the detector stays quiet on silence |
Outputs: melody-with-noise-tail.mid, silence.mid, and the full results.json.
What this does not prove
This tail experiment does not establish full-song accuracy, robustness to every noise type, or a universal noise threshold. It documents only the tested waveform, detector, and parameter set.
After conversion
Inspect the final note boundary, trim or gate unwanted tail audio, rerun the same detector, and compare the resulting MIDI events with the original control.
Questions about a noisy tail
Did the tested noise tail create extra MIDI notes?
The recorded run kept four melody events; it did not document an extra event after the final note.
What is the silent control for?
It checks that silence produces zero output events under the same detector setup.
Can this prove every noisy recording will convert cleanly?
No. It is one reproducible tail condition, not a full-song or all-noise benchmark.