Figure out which kind of noise you actually have
Every denoiser is built for a specific problem, and pointing the wrong one at your audio is how you end up with a voice that sounds like it is being spoken through a wet paper bag. Spend two minutes identifying the noise before you touch a plugin:
- Constant low rumble — traffic, footsteps upstairs, desk thumps, HVAC noise below 100 Hz. Fix with a high-pass filter. No AI required.
- Steady broadband hiss — mic self-noise, preamp gain, distant fan wash. This is what noise reduction tools are designed for.
- Electrical buzz at a single pitch — 60 Hz in North America, 50 Hz in Europe and much of Asia, plus harmonics at 120, 180, 240 Hz. Needs notch filtering or a de-hum module, not a denoiser.
- Clicks and key presses — keyboard, mouse switches, chair squeaks. De-click tools or manual editing.
- One-off bangs — door slams, dog barks, notification chimes. Mute them by hand. Denoisers smear them into artifacts that are harder to fix than the original sound.
- Reverb and room bounce — not technically noise, but it ruins voiceovers more than hiss does. No denoiser fixes it properly. Move the mic closer or hang a moving blanket behind you.
Five minutes of prevention beats an hour of repair
If you can still re-record, do these first. They shrink the problem before any software has to guess:
- Get the mic 3–6 inches from your mouth, slightly off-axis so plosives miss the capsule. Closer mic plus lower gain means less room and less hiss captured in the first place.
- Kill the sources: AC, fans, mini-fridges, and anything plugged in nearby. Laptop fans count — record on battery if the machine runs quiet enough.
- Put something soft behind the mic and behind you. A closet, a corner full of clothes, or a $20 moving blanket does more for a voiceover than any plugin.
- Record 5–10 seconds of room tone at the start of every session — just silence, mic in position. You will need it as a noise profile later.
The processing chain that actually works
Order matters more than the tools. Run these in sequence, and check the result after each step instead of stacking all four and hoping.
1. High-pass filter (80–100 Hz)
This removes rumble, desk thumps, and plosive energy without touching the voice. In Audacity: Effect → High Pass Filter, 24 dB per octave, cutoff at 80 Hz for a deep male voice and 100–120 Hz for higher voices. Tradeoff: push it to 150 Hz and the voice thins out and loses body. If you are unsure, 80 Hz is safe for almost everyone.
2. De-hum
Find your fundamental — 60 Hz in the US, 50 Hz in most of the rest of the world — and notch it plus the first two or three harmonics. In Audacity, use Effect → Notch Filter with a narrow Q (10–20) applied at each frequency. In Reaper, a sharp bell cut in ReaEQ does the same job. Tradeoff: wide notches eat the warmth of a male voice, so keep them narrow and only target frequencies you can genuinely hear buzzing.
3. Broadband noise reduction
This is the step people get wrong. Take the room tone you recorded, or find the quietest half-second between sentences, select it, and let the tool learn the profile. In Audacity: select that section, then Effect → Noise Reduction → Get Noise Profile. Then select the whole track and open the same effect.
Starting settings that work for most voiceover: Reduction 9–12 dB, Sensitivity 6, Frequency smoothing 3. Two passes at 6 dB usually sound cleaner than one pass at 12 dB because the artifacts stay smaller. In Reaper, ReaFIR in Subtract mode does the same thing with a learned profile and often less smearing. Tradeoff: above roughly 15 dB you start hearing a metallic, underwater quality, especially on sibilants and word tails.
4. Gate or expander last
Only after the noise floor is low should you gate. Set the threshold just above the remaining floor, attack 1–5 ms, hold 100–200 ms, release 100–300 ms, and range no more than -40 dB. A gentle expander at roughly 2:1 is safer than a hard gate because it does not chop off breath and word tails. Tradeoff: aggressive gating is the single most obvious home-studio giveaway — you can hear the room vanish between words.
When an AI denoiser is the better call
If you are working with a phone recording, a Zoom call, a gaming clip with keyboard noise, or a room you cannot re-record in, the manual chain only gets you so far. AI denoisers trained on speech handle non-stationary noise — clatter, traffic, someone talking in the next room — that a static noise profile cannot learn.
Free options worth knowing: Adobe Podcast Enhance does a dramatic job on speech but can over-process and leave a synthetic edge. If you want something simpler for a quick turnaround, a free tool like ClearVoiceHQ (clearvoicehq.com) takes an upload, strips the background noise, and hands back a downloadable file without installing anything or learning a DAW. For paid control, iZotope RX Voice De-noise is still the standard for fine-tuning exactly how much gets removed and where.
Whichever one you use: export a copy of the original before processing, and never stack denoisers. Running a browser tool like ClearVoice and then a DAW reduction pass on top of it is a reliable way to make a voice sound waxy.
Real-time cleanup while you record or stream
For live work — OBS, Zoom, Discord, streams — you want a filter chain running before the audio leaves your machine:
- OBS: add a Noise Suppression filter to your mic. RNNoise gives better quality; Speex uses less CPU. Add a Noise Gate filter after it with a threshold just above your room floor.
- Zoom: Settings → Audio → Background noise suppression. Low is usually enough. High chews up consonants and makes you sound underwater whenever you laugh.
- GPU options: NVIDIA Broadcast (RTX cards) and Krisp both do impressive real-time suppression, but they add latency — fine for streaming, worse if you monitor yourself in headphones.
Tradeoff: real-time suppression cannot be undone. Pair it with a high-pass filter and record the cleanest possible source, because you are baking the cleanup in permanently.
How much noise reduction is too much
Listeners forgive faint hiss. They do not forgive a robot. Two tests to run before you export:
- Solo the silence. Play a gap between sentences and listen only to the noise floor. If it sounds like a metallic chirp or a pumping breeze, back the reduction off by 3 dB.
- Check the consonants. S, T, K, and F are the first sounds to break. If they lisp or ring, you have gone too far.
A good target is a floor that is audible when soloed but disappears the moment speech starts. Leaving a little room tone under the voice sounds more natural than absolute silence.
Quick checklist
- High-pass at 80–100 Hz.
- Notch 50 or 60 Hz plus harmonics if there is a buzz.
- Learn a noise profile from real room tone, then reduce 6–12 dB.
- Gate gently last — expander if you can, -40 dB maximum.
- Reach for an AI denoiser instead when the noise moves around, or the source is a phone memo or Zoom clip.
- A/B against the original before exporting. If the original is not distracting, leave it alone.