Audacity Noise Reduction: The Settings That Work, and When It Can't Win

Audacity's Noise Reduction effect is free, offline, and genuinely good at steady noise — fan, AC, mains hum, tape hiss. It is also the effect people most often get wrong, because the three sliders aren't labelled in plain English and the one control that tells you whether you've gone too far is hidden in plain sight. Here's what each setting actually does, where to start for each kind of noise, and the point at which no amount of tuning will save the file.

First, the two passes

The step that trips people up is that Noise Reduction is a two-pass effect. Pass one shows Audacity what the noise sounds like; pass two removes that profile from the whole recording. Skip pass one and the effect does nothing at all.

Audacity's two-pass noise reduction

1. Select a noise-only region. Find 1–2 seconds with no speech — just the hum or hiss — and drag to select it. This is the sample the effect learns from, so pick a clean patch.
2. Get the noise profile. Effect ▸ Noise Removal and Repair ▸ Noise Reduction…, then Get Noise Profile. The window closes — that's expected, not a crash.
3. Select the whole track. ⌘A on Mac, Ctrl+A on Windows.
4. Reopen Noise Reduction, set the sliders, Preview, Apply. The rest of this guide is about that step.
5. Export. File ▸ Export Audio.

What each slider actually does

These three are the whole effect. Audacity's Noise Reduction manual page documents the ranges; what follows is what they mean when you're listening.

Slider	Range	Default	What it changes
Noise reduction (dB)	0–48	12	How much quieter the identified noise gets
Sensitivity	0–24	6.00	How much of the audio gets identified as noise in the first place
Frequency smoothing (bands)	0–12	3	How much the reduction is blurred across neighbouring frequencies

Noise reduction (dB) — the volume knob for noise

The most intuitive of the three: how far down the noise gets pushed. 12 dB is a real, audible reduction and is where you should start. Going to 30 or 40 dB does not make the result "cleaner" — it makes the gaps between words unnaturally silent, which the ear reads as processed. Most recordings sound best somewhere between 8 and 18 dB.

Raise it when the noise is still clearly audible and the voice still sounds natural. Lower it when the voice has gone thin or the silences sound hollow.

Sensitivity — the one that causes the damage

This is the slider people reach for when the dB slider isn't working, and it's the one most likely to wreck the recording. Sensitivity decides the threshold at which a given frequency at a given moment counts as noise rather than signal. Raise it and Audacity treats more of the audio as noise — including quiet consonants, breaths and the tails of words, which live at roughly the same level as the noise floor.

Symptoms of sensitivity set too high: lisping or missing "s" and "f" sounds, clipped word endings, a voice that sounds like it's being gated on and off. Keep it near 6 unless you have a specific reason, and move it in steps of 1.

Frequency smoothing — the artifact control

Smoothing spreads each reduction decision across neighbouring frequency bands instead of applying it to one band in isolation. That costs a little noise removal and buys a large reduction in the tinkly, bell-like artifacts described below. 0 removes the most noise and sounds the worst; 3 is the default; 6–12 is where you go when the result is artifacty and you'd rather keep some hiss than keep the sparkle.

Reduce vs Residue — the preview nobody uses

Below the sliders is a pair of radio buttons labelled Noise: Reduce / Residue. Almost everyone leaves it on Reduce and never finds out what Residue is for. It is the single most useful control in the dialog.

• Reduce previews the result — the audio with the noise taken out. This is what you export.
• Residue previews only what would be removed. Everything you hear in Residue is what you're about to throw away.

Why it sounds metallic, watery or robotic

This is the most common complaint about Audacity's noise reduction, and it isn't a bug or a settings mistake you can fully tune away — it's inherent to how the effect works.

Noise Reduction uses spectral subtraction: it chops the audio into short overlapping frames, splits each frame into frequency bands, and subtracts the profiled noise level from each band. Where a band falls below the threshold it gets attenuated; where it stays above, it's kept. Because the noise in a real recording fluctuates randomly from frame to frame, some bands cross that threshold in one frame and not the next. What survives is a scatter of short-lived, isolated tones at random frequencies — and the ear hears that scatter as tinkling, bubbling or bell-like sound over the silences.

It has a name in the audio literature: musical noise. Every tool built on spectral subtraction produces some of it, which is why learned models trained on speech — the approach behind RNNoise, Apple's Voice Isolation and the engine in Soundproof — became the standard approach: they estimate what the speech should look like rather than subtracting what the noise looked like.

What actually helps in Audacity, in order:

1. Lower the dB. Less subtraction, fewer threshold crossings, less musical noise. Try 8–10 dB and accept some residual hiss — a quiet, steady hiss is far less distracting than sparkle.
2. Raise Frequency smoothing to 6, then 9. This averages the decision across neighbouring bands and is the control specifically designed for this problem.
3. Lower Sensitivity. Fewer bands classified as noise means fewer of them flickering across the threshold.
4. Run two gentle passes instead of one aggressive one. Two 8 dB passes generally sound better than one 16 dB pass, because each pass re-profiles against a quieter noise floor.

When there's no clean noise-only sample

Noise Reduction cannot run without a profile, and plenty of real recordings never contain 1–2 seconds of noise on its own — someone starts talking immediately, or the interview is wall-to-wall speech. Options, in the order worth trying:

• Check the very start and end of the file. Room tone before the first word and after the last is the most commonly missed sample; it's often there in the second before "action".
• Use a gap between sentences. Shorter than ideal works: even a few tenths of a second of a breath-gap will produce a usable profile, though it'll be less accurate.
• Profile from a different file recorded in the same room with the same mic and settings. The profile lives in the session, not the file, so you can load a second recording, profile its room tone, then switch to the one you care about.
• Skip Noise Reduction entirely for rumble. For low-frequency traffic or air-handling rumble, Effect ▸ EQ and Filters ▸ High-Pass Filter at around 80 Hz removes a lot with no profile and no artifacts. Speech fundamentals sit above it.
• Use Noise Gate for gaps only. It needs no profile, but it only silences the pauses — the noise under the speech stays exactly where it was. Useful for a voiceover, misleading for an interview.
• Use a tool that doesn't need a sample. Model-based denoisers learn what speech sounds like in advance, so there's nothing to profile. That's the structural difference, not a marketing one.

Starting points by noise type

Starting points, not prescriptions — the right numbers depend on how loud the noise is relative to the voice. Always preview.

What you have	dB	Sens.	Smooth	Note
Laptop fan / air conditioning	10–15	6	3–6	The best case for this effect
Broadband hiss (preamp, tape, cheap mic)	12–18	6	3	Watch the "s" sounds
Mains hum (50/60 Hz buzz)	8–12	5–6	3	A Notch Filter at the hum frequency often beats this outright
Low rumble / traffic	—	—	—	High-Pass Filter at ~80 Hz instead
Quiet room tone	6–10	5	3	Go gentle; the noise is already near inaudible
Background voices, café chatter	Not a fit — see below

When Audacity's Noise Reduction can't win

Some of these are tuning problems and some are not. Being able to tell them apart saves a lot of time:

• "It barely did anything." Tuning. Your noise sample probably contained speech, or the dB is too low. Re-select a cleaner patch and add a few dB.
• "It sounds robotic / tinkly." Tuning, up to a point — see musical noise above. There's a floor you can't tune below.
• "Background voices are still there." Not tuning. The effect removes steady noise. Another person's voice has the same statistics as your voice, so a noise profile can't separate them. No slider fixes this — see what noise can (and can't) be removed.
• "The room echo is still there." Not tuning. Reverb is your own voice arriving late; subtracting a noise profile does nothing to it. De-reverb is a different class of tool — Hush does it, Audacity and Soundproof do not.
• "The noise changes through the recording." Not tuning. A single profile describes one noise. If a fan kicks in halfway, profile and treat the two halves separately.
• "I have to redo this for every file." Correct, and unavoidable — the profile is per-recording. This is the real cost of the manual approach, and it's what makes Audacity painful at 20 files rather than 1.

The faster alternative

Audacity is the right answer when you want free, you're working on one file, and you don't mind the steps — and it's a genuinely capable editor well beyond this one effect. Download it free if you haven't.

It becomes the wrong answer at volume, on video, or when there's no clean sample to profile. Soundproof uses a trained speech model instead of a subtracted profile: nothing to sample, nothing to tune, and none of the musical-noise sparkle that comes with spectral subtraction. It cleans video files as well as audio — remuxing the cleaned track back into the original container, same picture, no re-edit — batches a whole folder, and levels the result to a consistent LUFS target. It's a one-time purchase, it runs entirely on your Mac, and the benchmark numbers are published, including where it's weak.

Frequently asked questions