You can remove background noise from audio in about 90 seconds with almost any modern editor: isolate a two-second noise-only sample, run the built-in noise-reduction tool, then lower the noise floor in small 3 dB to 6 dB increments instead of crushing it in one pass. The condition matters more than the button β€” capture a clean noise profile first, and never push reduction past the point where consonants start to smear. That two-part rule separates a fix that sounds professional from one that sounds underwater.

Capture the Cleanest Raw Recording First So Restoration Has Less to Undo

Every noise-reduction tool works by subtracting a predictable pattern from the whole file. If the pattern is inconsistent, the tool guesses, and the guess audibly fails. Record 5 to 10 seconds of room tone at the start of every take β€” nobody speaking, no movement, HVAC running exactly as it will for the rest of the session. That clip becomes your noise print later.

Microphone placement does more than any software slider. A dynamic mic like the Shure SM58 at 10 to 15 centimeters from the mouth rejects far more room sound than a condenser at 40 centimeters, because distance reduces direct-to-reverberant ratio. According to Shure's published polar pattern charts, the SM58's cardioid design rejects sound arriving from directly behind the capsule by roughly 20 dB across the vocal band.

Close the physical leaks before touching a plugin. Turn off the refrigerator circuit, mute phone notifications, and put the mic on a shock mount so desk bumps don't register as low-frequency thumps. A laptop fan is the most common culprit in home setups β€” move the machine at least one meter away or route the recording to a phone in airplane mode.

Gain staging sets the ceiling on what's recoverable. Peak speech around βˆ’12 dBFS to βˆ’6 dBFS, never near 0 dBFS, because clipped audio cannot be un-clipped and noise reduction on a hot signal amplifies distortion artifacts. If you're stuck with a file already recorded badly, skip to the spectral repair section rather than the broadband denoiser.

Identify the Noise Type Before Choosing a Tool, Since Each Fails Differently

Broadband hiss, hum, and intermittent clicks need three different fixes, and applying the wrong one adds artifacts you can hear. Hiss is steady high-frequency energy β€” think tape noise or cheap preamps. Hum is a narrow tone, usually 50 Hz or 60 Hz plus harmonics, tied to regional power grids. Clicks, plosives, and chair scrapes are transient.

Region dictates hum frequency. North American and parts of South America run 60 Hz mains; Europe, the UK, Australia, and most of Asia run 50 Hz. Getting this wrong means your notch filter lands one octave off and the buzz survives. Audacity's Noise Reduction dialog and Adobe Audition's DeHummer both let you type the frequency manually rather than guess.

The table below maps noise type to the right tool and the typical failure if misapplied.

Noise type Best tool class Starting setting Failure if misused
Steady hiss Spectral noise reduction 6–10 dB reduction Wispy "swishing" on consonants
50/60 Hz hum Notch or de-hummer Q of 20–30 at fundamental + harmonics Hollow, tinny voice
HVAC rumble High-pass filter 80–100 Hz cutoff (male), 100–120 Hz (female) Thin, telephone-like output
Keyboard clicks Spectral repair or de-click Manual region select Smearing adjacent syllables
Room reverb De-reverb, not denoise 20–40% mix Pumping, robotic tail

A high-pass filter costs nothing in CPU and cleans rumble in under five seconds, so run it before any spectral process. In Audacity it's Effect > Filter Curve EQ or the simpler High-Pass Filter under Effect. In Adobe Audition it's the built-in High Pass under Effects > Filter and EQ, with an adjustable rolloff that defaults to 48 dB per octave β€” steep enough to kill rumble without nibbling vocal fundamentals.

Run Noise Reduction in Audacity Using a Noise Profile, Not a Preset

Audacity is free, open source, runs on Windows, macOS, and Linux, and its Noise Reduction effect is a two-stage tool: capture a profile, then apply. According to the Audacity manual, the effect uses the profile to build a statistical fingerprint of the noise and subtracts matching energy, so the quality of the profile is the quality of the result.

1: Open your file in Audacity. Select 2 to 3 seconds of pure room tone with no speech β€” click and drag across the waveform. 2: Go to Effect > Noise Removal and Repair > Noise Reduction. Click "Get Noise Profile." The dialog closes. 3: Select the entire track (Ctrl+A or Cmd+A). 4: Reopen Effect > Noise Removal and Repair > Noise Reduction. Set Reduction to 6 dB as a starting point. Leave Sensitivity at 6.00 and Frequency Smoothing at 3 bands. 5: Click Preview. Listen to a consonant-heavy passage β€” the letters s, t, k, and f expose over-processing first. 6: If artifacts appear, drop Reduction to 3 dB or raise Frequency Smoothing to 6 bands. If noise persists, raise Reduction by 2 dB and re-preview. 7: Apply, then re-listen to the full file before export.

If the profile is contaminated β€” you selected a segment where someone breathed β€” the effect will treat that breath as noise and remove it wherever it appears. Re-select a cleaner segment and re-capture the profile. Some Audacity versions label this effect "Noise Reduction" and others "Noise Removal"; both live in the same menu path.

Run reduction in two passes of 6 dB rather than one pass of 12 dB when hiss is heavy. Each pass is less aggressive on transients, and the Audacity manual notes that large single-pass reductions are more likely to introduce the metallic warble users describe as "underwater." Two passes cost roughly double the processing time β€” negligible for a 10-minute podcast at standard desktop speeds.

Apply Adobe Audition's Noise Reduction Process and DeHummer for Broadcast Files

Adobe Audition ships a dedicated Noise Reduction (process) effect under Effects > Noise Reduction / Restoration, which opens a graph of the noise floor and lets you drag the reduction curve per frequency band. It's a subscription product through Creative Cloud, and the exact menu labels can shift between releases β€” confirm on Adobe's site for your installed version before following screenshots from older tutorials.

Capture the profile the same way: select noise-only audio, then Effects > Noise Reduction / Restoration > Capture Noise Print (Shift+P on Windows). Then select the full clip and open the Noise Reduction process. The default reduction is around 6 dB with "Reduce by" set to a flat line; the graph lets you reduce less in the 2 kHz to 5 kHz range where sibilance and intelligibility live.

DeHummer sits under Effects > Filter and EQ > DeHummer. It auto-detects the base frequency in many cases, but typing 60 or 50 manually guarantees it lands. Enable the harmonic boxes β€” the third harmonic at 180 Hz and fifth at 300 Hz are what make cheap interfaces hum audibly even after you notch the fundamental.

Audition's advantage over Audacity is the frequency-painting diagnostic display, where you can see the noise floor as a distinct band and watch it drop. That visual confirmation shortens the guess-and-check loop. The trade-off is cost: Audition requires a paid Creative Cloud plan, while Audacity is free, so hobbyists rarely need to switch unless they're already in the Adobe ecosystem for video work.

Clean Voice Memos and Phone Recordings With Free Mobile Tools

Phone recordings carry three problems software can't fully reverse: heavy compression, aggressive automatic gain control, and limited frequency response below roughly 100 Hz. The Audio Engineering Society's published measurement work on consumer handsets consistently shows usable bandwidth caps well under studio standards, so don't expect broadcast quality from an iPhone Voice Memo no matter which app processes it.

Adobe Podcast's Enhance Speech, available free through a web upload with an Adobe account as of its public rollout, applies a trained speech model that removes room sound more aggressively than spectral subtraction. It works best on clean speech with steady noise and worst on music or overlapping voices, where it can thin instruments into artifacts.

On iOS, GarageBand's built-in EQ and the Voice Memos app's basic editing can't run noise profiles, but third-party apps like Hokusai Audio Editor offer spectral editing on-device. On Android, Lexis Audio Editor includes a noise-reduction module with an adjustable reduction slider. Both are limited by the phone's microphone and codec, so treat them as damage control.

The realistic ceiling matters. A noisy phone recording cleaned with a good tool typically improves intelligibility by a noticeable margin but never reaches what a $100 dynamic mic in a treated room would have given you. If the audio is for a paid client, re-record rather than salvage when travel allows.

Fix Hum, Hiss, and Clicks Separately Instead of One Heavy Pass

The most common amateur mistake is cranking one broadband denoiser to maximum and calling it done. That approach converts steady hiss into metallic warble while leaving the 60 Hz hum untouched, because hum is a tone, not broadband energy. Split the work into three targeted fixes.

For hum, add a parametric EQ notch at the fundamental with a narrow Q, then add notches at the second, third, and fifth harmonics. In Reaper's ReaEQ β€” bundled free with the DAW β€” you get unlimited bands and can solo each to confirm it kills the buzz. Set Q between 20 and 30; wider notches remove vocal fundamental energy and make the voice sound hollow.

For hiss, run spectral reduction at 6 dB, check a consonant, and stop. If residual hiss remains, one more 3 dB pass beats a single 9 dB pass. The harmonic distortion introduced by aggressive reduction scales nonlinearly, so the artifact curve steepens sharply above roughly 10 dB total reduction on typical speech.

For clicks and plosives, use spectral repair rather than a de-clicker where possible. iZotope RX's Spectral Repair (part of RX Standard and above) lets you select the click in the spectrogram and interpolate from surrounding audio, which preserves the syllable underneath. A broadband de-clicker smears the neighboring 20 milliseconds β€” audible on hard consonants like "t" and "k".

Accept That Some Noise Cannot Be Removed Without Damaging the Voice

There's a threshold past which restoration becomes voice replacement. Signal-to-noise ratios below roughly 10 dB β€” where the noise is nearly as loud as speech β€” defeat spectral subtraction because the tool can't separate what it needs to subtract from what it needs to keep. Knowing this boundary saves hours of futile processing.

Reverberant rooms are the classic unrecoverable case. A voice recorded in a bare room with hardwood floors carries a decay tail that de-reverb tools can shorten but never eliminate cleanly; the tool either leaves the tail or generates a pumping artifact as it tries to reshape the envelope. Treat the room, don't fight the recording.

Clipping is the other hard limit. Once a waveform's peaks are flattened at 0 dBFS, the harmonic content is baked in and no amount of spectral repair recovers the fundamental. If you see flat-topped peaks in the waveform view, the honest move is a re-record.

When noise is between 10 dB and 20 dB signal-to-noise, the trade-off becomes real: each additional decibel of reduction costs progressively more vocal fidelity. Set a ceiling before you start, listen on both speakers and headphones, and stop at the point where the voice still sounds like a person rather than a compressed radio signal.

Judge the Result With a Noise Floor Target and an Artifact Checklist

Numbers make this decision easier than ears alone. Measure the noise floor before and after using a loudness meter. In Audacity, Analyze > Plot Spectrum gives you a spectrum view for reading the residual noise level; dedicated loudness tools like the free Youlean Loudness Meter report integrated loudness and true peak. The benchmark to hit differs by delivery.

For podcast delivery, the commonly cited target from broadcast standards β€” derived from the EBU R128 recommendation of βˆ’23 LUFS for broadcast and its streaming adaptations β€” is a noise floor around βˆ’60 dBFS with speech peaking near βˆ’6 dBFS. For a voiceover placed under video, a βˆ’50 dBFS to βˆ’55 dBFS floor is often acceptable because music masks residual hiss.

Run this four-item checklist on every processed file before export. Play the quietest 10 seconds of speech β€” any residual hiss should sit below conscious attention. Check three consonant-heavy words for smearing. Sweep your headphones across the full file for pumping or breathing artifacts that follow speech onsets. Compare the processed and original side by side at matched loudness; if the processed version sounds thinner or duller, back off the reduction.

The single most useful move is an A/B against the original at equal loudness. Humans rate louder audio as better, so match levels before judging β€” and be suspicious of any processed version that only sounds superior because it's 3 dB hotter.

Export Settings, File Formats, and the Order of Processing That Preserves Quality

Process in a lossless format and export in a lossless format, or you're compounding codec artifacts with denoising artifacts. Work from a 24-bit WAV at the source sample rate β€” 44.1 kHz or 48 kHz depending on your delivery target β€” and keep a copy of the original untouched. Compressed MP3s already discard high-frequency information, and noise reduction on a lossy file amplifies the codec's own swishing.

The order of operations matters and is not arbitrary: high-pass filter first, hum notch second, spectral denoise third, de-click fourth, EQ and compression last. Each stage should be checked before the next. Compressing before denoising, a common mistake, raises the noise floor along with the voice and makes the job harder. For a 30-minute podcast episode, the whole chain typically runs 15 to 30 minutes of processing time on a modern laptop, with spectral repair being the slowest step per region edited.

Export at 24-bit WAV for editing handoff, 16-bit WAV at 44.1 kHz for CD-style delivery, or MP3 at 192 kbps or higher for web distribution. Keep the noise-reduced intermediate as a WAV so later revisions don't stack lossy generation loss. Name files descriptively with the process version, so you never re-edit a compressed copy thinking it's the master.

FAQ

Does noise reduction always make audio sound worse?

No β€” it degrades predictably with how hard you push it. Up to about 6 dB of reduction on steady hiss, most listeners can't tell. Every additional 3 dB raises the probability of audible artifacts, and past roughly 10 dB to 12 dB total reduction on speech, warbling is nearly guaranteed regardless of tool quality.

Can I remove background noise from a video file directly?

Yes, but only if the editor processes the embedded audio track. In Adobe Premiere Pro, right-click the clip and choose Edit in Audition, which round-trips the audio for noise-print reduction and returns it synced. Free alternatives like DaVinci Resolve's Fairlight page include a Noise Reduction module under the Fairlight effects rack, with reduction and focus controls that mirror a simpler version of Audition's tool.

Why does my 60 Hz hum survive noise reduction?

Because broadband denoisers subtract statistically random energy and hum is a pure tone with harmonics. You need a notch filter or de-hummer at the fundamental plus its odd harmonics β€” 60 Hz, 180 Hz, 300 Hz in North America. Spectral reduction barely touches it, which is why the hum remains after you "clean" the file.

What's the minimum signal-to-noise ratio for restoration to work?

Around 10 dB is the practical floor for spectral methods. Below that, the tool can't reliably distinguish voice from noise, and reduction produces the "underwater" effect. Between 10 dB and 20 dB, expect a visible fidelity trade-off: clearer noise floor at some cost to vocal presence, and always measure the result rather than trusting your ears on a single listen.

Is Adobe Podcast Enhance Speech free and does it replace a paid editor?

It's been available free through a web upload with an Adobe account during its public rollout, per Adobe's own announcements, though availability terms can change β€” check Adobe's site. It's excellent for spoken-word cleanup but not a substitute for an editor who can handle hum, clicks, spectral repair, and multi-track sessions. It's a fast first pass, not a mastering chain.

Can I fix audio already recorded at 0 dBFS clipped?

No. Clipping flattens waveform peaks and permanently introduces harmonic distortion. Some editors, including iZotope RX's De-clip module, reduce the audible distortion, but the original information is gone. Check the waveform for flat-topped peaks; if they're present, re-record whenever possible.

How long does noise reduction take on a one-hour podcast?

On a modern laptop, a single spectral pass in Audacity over a one-hour stereo file typically completes in 2 to 5 minutes; two passes double that. Manual spectral repair and click removal add minutes per region edited. Batch processing in Audition's Favorites panel compresses the per-file overhead if you're cleaning an entire season.