Every noise remover — this one, Audacity, the expensive plugins — works the same way underneath: it learns what the noise sounds like, then subtracts that fingerprint from everything. Which means the quality of your result is mostly decided by one thing, and it isn't a slider. It's whether the tool gets a clean sample of the noise by itself.
1 – If you haven't recorded yet: leave a second of silence at the start. Press record, count one breath, then talk. That second contains pure noise — the room, the fan, the hiss — and it's worth more than any setting. In the cleaner, pick "first 0.7 s is noise only".
2 – Already recorded, no silence? Use auto mode — it fingerprints the quietest slices it can find. Works well unless you literally never stopped talking, in which case the pauses between sentences still carry enough noise to learn from.
3 – Clean at 70%, then listen. Not to the noise — to the voice. Flip the Before/After button mid-playback. Noise you can measure; damage you have to hear.
4 – Adjust in one direction only. Hiss survived? Nudge strength up. Voice went watery or robotic? Back it down and accept a whisper of floor — a faint, steady residue sounds natural; aggressive digital silence sounds like a broadcast dropout.
Maxing the strength. Past a point, the gate starts eating the quiet parts of consonants and the tails of words — the "underwater" sound. The measured readout may look glorious at −40 dB while the voice sounds like it's calling from a fish tank. Ears outrank numbers; that's why the A/B button is next to the export button.
Expecting it to remove events. A door slam, a cough, another conversation — those aren't noise, they're sounds, and a spectral gate can't tell them from your voice. Cut them out with the trimmer instead. Any tool claiming one-click removal of a barking dog is selling adjectives.
Order of operations, if you're doing a full cleanup: denoise → cut silences → normalize → export. Noise first, always — every later step works better on a clean floor.