What we tested
We generated three mono 16-bit PCM WAV files at 22,050 Hz. Every file used the same isolated C4, E4, G4, and C5 sine tones, with 0.700-second notes, 0.450-second gaps, and a normalized 0.72 peak amplitude.
The dry file had no added reflections. The other files used deterministic dense reflection taps with 0.35-second or 1.10-second decay windows. We ran each file through the current local converter at its default controls and inspected the exported MIDI.
| Condition | MIDI events | Detected pitches | Last note end | File size |
|---|---|---|---|---|
| Dry | 5 | C4, E4, G4, C5 | 4.403 s | 97 bytes |
| 0.35 s decay | 7 | C4, E4, G4, C5 | 4.449 s | 113 bytes |
| 1.10 s decay | 6 | C4, E4, G4, C5 | 5.424 s | 108 bytes |
What changed in the MIDI
The dry file produced four main notes plus one 0.070-second C4 edge fragment. Short reverb added a 0.070-second G4 fragment and a 0.104-second C5 fragment. Long reverb produced a separate 0.499-second C5 event beginning at 4.925 seconds, after the main C5 had ended.
All three runs retained the four intended pitch classes, returned HTTP 200, used one musical track, and produced no console errors. This test observed extra boundaries and a late tail note; it did not measure transcription accuracy for real rooms.
How to handle reverberant source audio
- Prefer the driest source you have. A close microphone, direct recording, or isolated stem gives the note detector fewer overlapping tails.
- Compare late notes with the waveform. A note after the played attack may be a decaying reflection rather than a new performance event.
- Edit short fragments after export. Remove or merge only the events that do not match an audible attack.
- Test cleanup cautiously. De-reverb can change the audio too, so compare the resulting MIDI with the original performance.
Reverb is not the same as a discrete echo
This test used many closely spaced, decaying reflections. The earlier echo test used clearly spaced repeat delays. Both can extend note energy, but their timing patterns and MIDI side effects are different.
What this test does not prove
It tests one synthetic four-note sequence and two deterministic reflection patterns. It does not establish a universal decay limit or predict concert halls, phone recordings, mixed music, direct YouTube-link input, or other converters.