What we tested
We generated a 22,050 Hz, 16-bit mono WAV containing isolated C4, E4, G4, and C5 sine tones. Each note lasted 0.900 seconds. Their relative amplitudes stayed at 25%, 50%, 75%, and 100% of the file peak.
We rendered the same sequence at peaks of 0.12, 0.80, and 0.98. The 0.80 version represents peak normalization: every sample was multiplied by the same gain, so the note balance and waveform shape stayed unchanged. We ran all three files through the current converter at its default controls without changing its logic.
| Source peak | MIDI events | Intended pitches retained | Main-event velocities |
|---|---|---|---|
| 0.12 (quiet) | 5 | C4, E4, G4, C5 | 66, 75, 78, 76 |
| 0.80 (normalized) | 5 | C4, E4, G4, C5 | 73, 77, 79, 81 |
| 0.98 (near full scale) | 5 | C4, E4, G4, C5 | 74, 79, 80, 82 |
Peak normalization changed gain, not dynamics
The quiet file's four steady-note RMS values were 0.02121, 0.04242, 0.06365, and 0.08484. Peak normalization to 0.80 multiplied them to 0.14142, 0.28281, 0.42430, and 0.56558. Their 1:2:3:4 relationship stayed intact.
This differs from dynamic range compression. A compressor can reduce the gap between quiet and loud sections and reshape attacks or releases. Peak normalization applies one constant gain to the entire file.
The event count stayed fixed, but one boundary moved
Every version returned five events in a 97-byte MIDI file. The quiet source produced a 0.070-second E4 tail fragment. The 0.80 and 0.98 files instead produced a 0.070-second C4 tail fragment.
The normalized and near-full-scale outputs had the same pitches, event starts, and durations. Their velocities differed by only one or two MIDI steps. That small difference is still enough to show that higher gain is not automatically a transparent preprocessing step for this transcription model.
When to compare a normalized copy
- Keep the original. Make normalization a reversible comparison, not the only source.
- Avoid clipping. Set the target below full scale and inspect the rendered peak.
- Change only gain. Do not add compression, limiting, EQ, or noise reduction to the same test.
- Compare note boundaries. Count intended pitches, short fragments, start times, durations, and velocities.
What this test does not prove
It does not define an ideal normalization target or a minimum usable level. It does not test clipped audio, lossy encoding, mixed instruments, mastered music, or streamed video audio. It records this tool's current output for three controlled local WAV files.