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 with a 0.300-second gap. The centered waveform ranged from -0.55 to +0.55.
We made two copies by adding a constant value to every sample. The +0.15 file ranged from -0.40 to +0.70, and the +0.30 file ranged from -0.25 to +0.85. No file clipped, and pitch, timing, amplitude, envelopes, and sample rate otherwise stayed fixed. We converted all three at the site's default controls without changing its logic.
| Measured mean | MIDI events | Intended pitches retained | Added output |
|---|---|---|---|
| 0.00 | 6 | C4, E4, G4, C5 | None after the final C5 |
| +0.15 | 10 | C4, E4, G4, C5 | A0, A-sharp 0, F-sharp 1, G-sharp 1 |
| +0.30 | 12 | C4, E4, G4, C5 | One early C4 plus five low tail notes |
The intended notes survived
Every output contained the four designed pitches. Their main events stayed close: C4 began at 0.255 seconds in all files, E4 at 1.451 or 1.462 seconds, G4 at 2.648 seconds, and C5 at 3.856 or 3.868 seconds.
Main-note velocities also stayed within a narrow range. The centered file produced 83, 83, 79, and 79; the +0.30 file produced 81, 83, 79, and 77. The clearest difference was not a lost intended pitch but additional short events.
The offset created a low-note tail cluster
The +0.15 file added four notes beginning at 4.925 seconds: A0, A-sharp 0, F-sharp 1, and G-sharp 1. Their durations ranged from 0.093 to 0.151 seconds. The +0.30 file added those pitches plus D-sharp 2 at the same boundary, and it also added a 0.070-second C4 fragment before the designed first note.
The test does not establish why the model produced those exact pitches. It does show that a constant sample bias can change this converter's event output even when the audible sine frequencies are unchanged.
How to check a source for DC offset
- Inspect the waveform center. Silence and symmetric passages should sit around the zero line rather than consistently above or below it.
- Measure the sample mean. A value far from zero is evidence of a constant bias, although the appropriate tolerance depends on the source.
- Make a corrected copy. Use a DC-offset removal or high-pass operation, then keep the original for comparison.
- Compare the MIDI drafts. Check intended pitches, short edge fragments, low-note clusters, timing, and velocities.
What this test does not prove
It does not define a safe offset limit, recommend a high-pass frequency, or test mastered music, lossy streaming audio, mixed instruments, or clipped samples. It records the current tool's output for three controlled local WAV files.