Tutorials Audio

Audio Restoration Beyond Noise Removal

By Infinite Creation · Advanced · ~35 min

Some links on this page are Amazon affiliate links. If you buy through them, we earn a small commission at no extra cost to you.

A 1960s mid-century comic book style editorial illustration showing two audio engineers at a mixing desk, with an audio waveform display visible in the frame.

Overview

Broadband noise reduction handles hiss and steady background sound, but it does nothing for hum, clicks, clipping, or dropouts. Each of these is a different kind of damage that needs a different, targeted repair. This guide covers the restoration problems that come up once you're past basic noise removal: fixing a ground-loop hum, repairing a distorted peak, and patching a digital glitch without making things worse.

What You Need

A DAW or dedicated restoration tool (Audacity for basics, iZotope RX for harder cases)
The damaged audio file, ideally the highest-quality source you have access to

Steps

Identify what you're actually dealing with

Zoom into the waveform and spectrogram before reaching for any tool. A steady low-frequency tone at 50/60Hz (or its harmonics) is hum. Sharp, narrow spikes are clicks. Flat-topped, squared-off waveform peaks are clipping. A sudden silent or garbled section is a dropout. Each needs a completely different fix, using the wrong tool wastes time and can make the artifact worse.

Remove hum and buzz with a targeted tool, not broadband noise reduction

Mains hum sits at a specific frequency (50Hz or 60Hz depending on region) plus harmonics. A narrow notch filter or a dedicated de-hum tool targeting exactly those frequencies removes it without touching the rest of the spectrum. Any notch still takes out wanted sound at those frequencies, a narrower one removes less hum and does less damage, and it only works if the hum sits exactly on the frequency you set, which real mains often does not.

De-click and de-crackle transient artifacts

Clicks (from vinyl transfers, damaged cables, or digital glitches) are short enough that a de-click tool can interpolate across them using the surrounding waveform, filling the gap convincingly. Run it as a separate pass from hum or noise removal, mixing repair types in one pass tends to produce worse results than sequential, targeted passes.

Repair clipped and distorted peaks

A de-clip tool reconstructs the missing waveform shape above the clipped ceiling using the surrounding signal as a model. This works convincingly for mild-to-moderate clipping but degrades as clipping gets more severe, since more of the original information is genuinely gone. Always keep the unprocessed original file. De-clipping is rarely perfect on the first pass.

Patch dropouts and digital glitches

A short dropout (a fraction of a second of silence, static, or garbled digital noise) can often be patched by borrowing a nearby moment of clean room tone or a repeated syllable from elsewhere in the same take. Longer dropouts in dialogue may need re-recording (ADR) rather than restoration, be honest about when a repair tool genuinely can't help.

Know when to stop

Every restoration pass risks introducing its own artifacts, over-aggressive de-clicking can add a subtle warble, heavy de-clipping can sound slightly synthetic. Compare against the original regularly, and stop once the result sounds natural rather than chasing a theoretically perfect measurement.

Pro Tips

Always work on a copy and keep the untouched original, restoration is rarely a one-pass process, and you'll want to re-try a step with different settings more often than you expect.
Fix problems in the right order: hum and clicks first, then clipping, then broadband noise reduction last, noise reduction on top of unaddressed hum or clicks tends to smear those artifacts rather than remove them.
Listen on more than one set of speakers or headphones before calling a restoration finished, artifacts that are inaudible on one system are sometimes obvious on another.

Tested: what each restoration filter fixes, and what it costs the voice

We tested the page's restoration claims on 2026-09-29 with FFmpeg 8.1.2 and numpy, on a synthetic voice with damage added and then repaired with FFmpeg's own filters: bandreject notches, afftdn for noise reduction, adeclick and adeclip. The voice is a harmonic source with a 125 Hz pitch, in phrases with pauses, at -20 dBFS. Every figure compares a result with the undamaged voice. "Voice error" is the difference from the original, in decibels relative to the voice, so a lower number is better. The filters delay their output, so results were aligned before comparing. These are FFmpeg's filters, not iZotope RX or Audacity, and nothing was listened to.

Damage	Repair	What it fixed	What it cost
Hum at 50 Hz and nine harmonics	10 notches, Q 30	Hum down 38.7 dB	Voice error -11.1 dB
Hum at 50 Hz and nine harmonics	10 notches, Q 60	Hum down 32.7 dB	Voice error -14.6 dB
Hum at 50 Hz and nine harmonics	10 notches, Q 200	Hum down 21.2 dB	Voice error -20.2 dB
Hum at 50 Hz and nine harmonics	afftdn, default	Hum down 0.0 dB	Voice error -41.0 dB
Hum at 50 Hz and nine harmonics	afftdn, noise profile taken from a pause	Hum down 19.6 dB	Voice error -20.4 dB
40 clicks	adeclick, default	Click error -12.1 to -33.6 dBFS, 6 of 40 clicks left	Voice error on a clean file -26.1 dB
40 clicks	adeclick, threshold 4	Click error -12.1 to -18.8 dBFS, 11 of 40 clicks left	Voice error on a clean file -62.4 dB
40 clicks	adeclick, threshold 8	Click error -12.1 to -18.1 dBFS, 12 of 40 clicks left	Voice error on a clean file -106.8 dB
Clipped at 0.9 of peak, 0.8% of samples	adeclip, default	Error at clipped samples -31.7 to -35.9 dBFS	Overall SNR 30.1 to 25.7 dB
Clipped at 0.7 of peak, 4.2% of samples	adeclip, default	Error at clipped samples -24.7 to -28.0 dBFS	Overall SNR 16.0 to 18.9 dB
Clipped at 0.5 of peak, 6.2% of samples	adeclip, default	Error at clipped samples -19.4 to -22.4 dBFS	Overall SNR 9.0 to 11.9 dB
Clipped at 0.3 of peak, 7.3% of samples	adeclip, default	Error at clipped samples -15.8 to -16.2 dBFS	Overall SNR 4.7 to 5.1 dB
Voice with noise, hum and clicks together	hum, clicks, then noise reduction (the page's order)	Hum -71.2 dBFS, noise -55.5 dBFS, clicks -33.1 dBFS	Overall SNR 13.2 dB
Voice with noise, hum and clicks together	clicks, hum, then noise reduction	Hum -70.5 dBFS, noise -55.2 dBFS, clicks -33.1 dBFS	Overall SNR 13.3 dB
Voice with noise, hum and clicks together	hum, noise reduction, then clicks	Hum -69.8 dBFS, noise -53.5 dBFS, clicks -30.2 dBFS	Overall SNR 13.3 dB
Voice with noise, hum and clicks together	noise reduction first, then hum, then clicks	Hum -67.5 dBFS, noise -51.1 dBFS, clicks -33.0 dBFS	Overall SNR 13.3 dB
Voice with noise, hum and clicks together	noise reduction first, then clicks, then hum	Hum -66.6 dBFS, noise -50.9 dBFS, clicks -33.2 dBFS	Overall SNR 13.0 dB

Step 2 is right that broadband reduction is the wrong tool for hum, and the comparison is closer than the page says. A default afftdn pass removed 0.0 dB of hum, which is none. Ten narrow notches at the hum's harmonics removed 33 dB at Q 60. But the notches cost the voice, because this voice's harmonics sit on or near several of the hum's harmonics and the notches cut them: the error was -11 dB at Q 30 and -20 dB at Q 200, and a narrower notch trades removal for damage, 21 dB removed at Q 200. Given a noise profile taken from a pause, afftdn removed 20 dB of hum for -20 dB of voice error, about the same as the Q 200 notch. So step 2 used to say "far less collateral damage", which did not hold here, and it now says what does. It depends on how close the wanted signal sits to the hum, and on the hum being an exact 50 Hz, which real mains is not.

Step 3 and step 6 both hold. The default adeclick took the click error from -12.1 to -33.6 dBFS and left 6 of 40 clicks. It was also the setting that changed the clean voice most, at -26 dB, so it altered material that was not a click. A less sensitive threshold left the voice almost alone, -62 dB, and fixed far less, 11 of 40 clicks left. That is the cost the page warns about, measured as error and not heard: sensitivity buys clicks and costs voice.

Step 4 and the FAQ hold, and the mild case is a surprise. As clipping got worse, adeclip's improvement in overall SNR went +3.0 dB at 0.7 of peak, +2.9 at 0.5 and +0.4 at 0.3, and the peak it reconstructed was only 0.54 of the original at the worst level. At the mildest, 0.9 of peak, with under 1 percent of samples clipped, the overall error got worse, 30.1 to 25.7 dB, even though the error at the clipped samples fell. The repair changed samples that were not clipped. Light clipping is not necessarily best left to a de-clipper.

The order of operations matters, modestly. Doing hum, then clicks, then noise reduction left hum at -71.2 dBFS and noise at -55.5. Starting with noise reduction, then hum, then clicks, left -67.5 and -51.1, so 3.7 dB more hum and 4.4 dB more noise, because the noise profile was learned with the hum present. Click removal was the same in every order, and overall SNR moved by less than 0.2 dB, since it is dominated by what the notches cost the voice.

The hum notch chain that was measured at Q 60, ten notches from 50 Hz: ffmpeg -i in.wav -af "bandreject=f=50:width_type=q:w=60,bandreject=f=100:width_type=q:w=60,bandreject=f=150:width_type=q:w=60,bandreject=f=200:width_type=q:w=60,bandreject=f=250:width_type=q:w=60,bandreject=f=300:width_type=q:w=60,bandreject=f=350:width_type=q:w=60,bandreject=f=400:width_type=q:w=60,bandreject=f=450:width_type=q:w=60,bandreject=f=500:width_type=q:w=60" out.wav

Validated. Measured 2026-09-29 with FFmpeg 8.1.2 and numpy on a synthetic voice. The hum is an exact 50 Hz with fixed harmonics and the clicks and clipping are generated, so real recordings will differ, and the voice's pitch sits close to the hum harmonics, which makes the notch cost look larger than it may be on another voice. We did not test iZotope RX, Audacity, dropouts (step 5), or how any result sounds. The command is the notch chain that was measured.

Knowledge Base

Why "Just Add More Noise Reduction" Doesn't Work

Broadband noise reduction tools work by learning a noise profile (typically from a quiet section) and subtracting that spectral signature across the whole file. Hum, clicks, and clipping don't fit that model, hum is a specific tonal frequency rather than broadband noise, clicks are transient rather than continuous, and clipping is a shape distortion rather than an additive noise layer. Applying more broadband reduction to any of these either does nothing or damages the wanted signal trying to remove something it isn't built to target.

The Order of Operations Actually Matters

Restoration passes interact with each other. Removing hum and clicks before noise reduction gives the noise-reduction algorithm a cleaner signal to learn a profile from, producing a more accurate and less destructive result. Doing noise reduction first can smear hum and click artifacts into the broadband noise floor, making them harder to isolate and remove cleanly afterward.

Some Damage Is Genuinely Not Fully Recoverable

Restoration tools are very good at plausible reconstruction, not magic. Severely clipped audio has lost real information permanently. A repair tool is making an educated guess about what probably was there, not recovering the original waveform. Setting realistic expectations up front avoids wasted time chasing a "perfect" fix that the remaining data simply can't support.

Where This Fits

Restoration repairs damage after the fact, and Audio Recording Fundamentals covers the recording decisions that prevent most of it.

FAQ

Q: Can you actually fix clipped or distorted audio?
A: Partially. De-clipping tools reconstruct the missing waveform peaks using the surrounding signal as a reference, and modern spectral tools do this convincingly for mild-to-moderate clipping. Severe, heavily-clipped distortion loses information that no tool can fully invent back. The more clipped a signal is, the less convincing the repair.

Q: Do I need iZotope RX for this, or can free tools do it?
A: Free tools cover the basics reasonably well. Audacity has a Notch Filter and Click Removal effect, for instance. iZotope RX's paid tiers add specifically-trained de-clip, de-hum, and spectral repair modules that handle harder cases with noticeably better results, particularly on material where free tools produce audible artifacts. For occasional light restoration work, free tools are enough. For frequent or difficult restoration, RX pays for itself in saved time.