What we tested
We generated three 22,050 Hz mono WAV files. Each contained the same isolated C4, D4, E4, and G4 sine tones, with every tone lasting 0.75 seconds and separated by 0.25 seconds. One file was clean. The other two added seeded white noise at signal-to-noise ratios of 20 dB and 0 dB, so the test can be reproduced exactly.
Each file went through the local converter at its default note sensitivity and minimum-note-length settings. All three browser runs returned HTTP 200, completed without console errors, and produced one-track MIDI files.
| Input condition | MIDI events | Pitch classes found | MIDI size |
|---|---|---|---|
| Clean | 8 | C4, D4, E4, G4 | 121 bytes |
| 20 dB SNR | 15 | C4, D4, E4, G4 | 179 bytes |
| 0 dB SNR | 29 | C4, D4, E4, G4 | 287 bytes |
What changed as noise increased
The clean result already divided each intended tone into a main event and a short edge event. At 20 dB SNR, C4 became six events, D4 became five, E4 stayed as one event, and G4 became three. At 0 dB SNR, the four tones became 7, 7, 8, and 7 events respectively.
No new pitch class appeared in this exact test. The failure mode was timing fragmentation: a held source note became a chain of shorter MIDI notes. A different instrument, noise spectrum, recording level, or mix can produce a different result.
How to prepare a noisy source
- Trim noise-only sections. Remove long room tone before and after the part you need.
- Reduce steady noise gently. Use a denoiser before transcription, then listen for watery or metallic artifacts.
- Expose one part. If possible, isolate the melody or instrument instead of feeding the full mix to the detector.
- Review note boundaries. Merge adjacent fragments of the same pitch only when they represent one held source note.
What this test proves—and what it does not
It records the current converter's output for one four-note synthetic melody under three controlled white-noise conditions. It does not claim a universal SNR cutoff, denoising gain, or accuracy level for public videos, speech, live rooms, hiss, hum, wind, or compressed music.
If noise has already created short fragments, use the measured extra-note cleanup guide. For dense mixes, read the instrument-separation boundary.