Can noise reduction fix a recording made in a bad room?

Not fully. Room reverb is already inside the signal, so processing can only shorten it, usually with aggressive gating that sounds worse than the reverb did. A soft room and a shorter mic distance fix the problem at the source.

What order should noise reduction, EQ and compression go in?

Start from the quietest input level that captures the performance, run noise reduction on the voice, cut rumble below roughly 80 Hz, and apply compression last. Compression raises the noise floor, so using it before reduction leaves you with more noise to remove.

Why does my denoised voice sound underwater?

Reduction was set too high. Pushing until the hiss disappears completely also removes speech that resembles the noise, which is what smears consonants and flattens the voice. Back the strength off until the noise is inaudible but the voice stays natural.

Can noise reduction fix clipping?

No. A clipped recording is flat-topped at the peaks and the information is gone. Record at a lower input level; no effect chain can reconstruct a waveform that was never captured.

What does the noise floor reading in WaveCraft's denoise panel mean?

It is the level of the steady noise the app measured across the whole clip, in dBFS, before you press Reduce. WaveCraft estimates it by scanning for the quietest moments and taking a representative level per frequency; the panel disables Reduce when the floor is inaudible, because processing would only cost you a little high end.

Is it safe to denoise the only take of a recording?

WaveCraft writes a new file and swaps the clip onto it, leaving the clip's position and length untouched, and it registers as one undoable step in the editor. If you re-reduced it more than 50 edits later, undo is no longer available, so keep the original somewhere else before the first pass.

Why do some apps only listen to the first few seconds to learn the noise?

Because they expect you to hand-select a snippet of pure room tone, and a causal estimate of the opening seconds is wrong exactly when the recording starts on the first word. WaveCraft scans the whole clip and uses the quietest parts wherever they fall, so it does not need you to frame the noise in advance.

I am reducing hiss on a voice track. Should I denoise the music under it too?

No. Steady background hiss is what spectral subtraction is good at; a music bed is neither steady nor noise, and subtracting it dulls the track. Duck the music under the narration instead, which turns the background down only while the voice is present.

How to Denoise a Recording on iPhone: What the Tool Actually Measures

The short answer

Noise reduction is not a repair. It measures what the unwanted sound looks like and subtracts it, frequency by frequency. Applied at the right strength to a take whose noise is steady and whose voice is louder than it, it works well. Applied where the noise is louder than the voice, or the real problem is the room or a clipped peak, it recovers nothing — it only makes the track quieter and more artificial.

WaveCraft measures the noise in your clip before it processes anything, tells you how loud it found, and refuses to run when there is nothing worth removing. It is on the App Store if you would rather just try it. The rest of this page is about what denoising can and cannot do.

Why the searcher is here

The question behind "how do I denoise this" is usually not "how do I subtract noise". It is:

Can I make this take usable, or do I have to record it again?

Most people are standing in front of a take they cannot re-record — a voice memo from a noisy room, a podcast recorded on a train, the audio from a video someone shot. Steady background noise — mains hum, an air conditioner, a fan, a phone's own preamp floor — is the one problem a subtractive method solves cleanly. Sometimes the take has a problem denoising cannot touch at all, and the useful answer is "re-record, and here is what to change". The trap in between is believing the dial: any denoiser can be turned up until the hiss is gone. What differs between them is what they measured first, what they refuse to do, and how much of the voice they take with them.

What is actually in the take

Four different things get called "noise", and they need different tools:

Steady hum or hiss. Mains hum and its harmonics, air conditioning, a laptop fan, the preamp's own floor. Unchanging over time, so it can be learned and subtracted — the ideal case.

Room reverb. The echo of a voice in a reflective space. Partially fixable: heavy gating can shorten a room to a booth, but it cannot recover detail the room already destroyed.

Handling and mouth noise. Pops, plosives, chair creaks, breath. Transient and unpredictable, so broadband reduction handles them poorly; they need targeted editing.

Distortion already recorded. A take pushed until it clipped: the waveform is flat-topped and the information is gone. WaveCraft's recorder latches a clipping indicator the moment any channel passes −0.1 dBFS and keeps it after you stop, so an overload is something you see rather than discover later in the waveform.

What WaveCraft's Reduce Noise measures, and when it refuses

Tapping Reduce Noise opens a panel instead of processing anything. It scans the clip first — an "Analysing…" line — then shows the steady noise floor it found, in dBFS, next to a Strength slider (0–100 %, starting at 50 %) and a Reduce button. This is a measurement, not a progress animation: the biggest difference between a tool that works on your clip and one that only works on clips like the one it was demonstrated with.

The measurement is a minimum-statistics estimate: the clip is cut into about 0.37 s segments, and since speech and music are intermittent, whatever remains in the quietest segments is the noise. Three details make it trustworthy:

A bias correction of ×1.3 compensates for the minimum across segments always sitting at the low tail of the distribution. They over-estimate slightly: over-estimating removes a little more, under-estimating leaves audible hiss.
A median filter along the frequency axis (±4 bins) rejects narrow tonal peaks: noise is broadband and flat, a tone is a few bins wide. Without it, a sustained note surviving in the quietest segment is treated as noise across the whole clip — measured, that cost 8.5 dB of the tone. It also means 50/60 Hz mains hum, whose harmonics sit tens of hertz apart, is folded into the broadband profile rather than targeted; a high-pass at 80 Hz does hum better.
One profile for both channels. Noise is a property of the room, so left and right share one estimate; estimating them separately makes the stereo image drift with the noise.

Below about −60 dBFS the noise is inaudible: WaveCraft shows "No steady noise detected" and disables Reduce, because on a clean take the pass would only cost you a little high frequency.

What the strength slider actually does

The reduction is spectral subtraction in the classical Berouti form: clean² = max(mix² − α·noise², β·noise²), where α is an over-subtraction factor and β is a spectral floor. Strength maps onto both.

α > 1 compensates for noise being random: in any one frame its energy is above the long-term average, so subtracting only the average leaves patches of hiss.
β, a floor proportional to the noise, is what keeps the result from being "sparkly": a zero floor lets isolated bins jump between present and absent, which is the "musical noise" you hear as ticks and bells.

A strength that removes too much also removes parts of the voice, so the usable range is bounded from both sides. And the analysis is frequency-bounded, so a hum partial or squeal narrow enough to sit inside one bin can slip past; a vocal's fundamentals share bins with traffic rumble, so low-end cleanup is EQ territory. Denoise first, then set the level.

Order, and the two cases that need a different tool

The order that works for a voice track:

Cut rumble with EQ. A high-pass around 80 Hz removes handling knocks, HVAC and building vibration before the subtractive stage sees them; both undo in the same edit, but the EQ is cheaper and more predictable.
Run Reduce Noise. The result replaces the clip's audio with a new file of the same length and position, in one undoable step.
Compress last. Compression lifts the noise floor along with the voice; compressing first means the denoiser has more noise to remove.
Match loudness afterwards. Match Loudness scales the clip so its peak lands at 0.95 and Auto Mix levels each track to about 0.9 (both capped at ×2 gain); measuring a noisy peak before cleaning it is how a clean track ends up quiet.

Two cases where the denoiser is the wrong tool:

Noise that is not steady. A passing car, a dog, a cough. Subtraction assumes the noise persists; intermittent noise is left in or half-removed, which sounds worse than leaving it. Edit it by hand — or use Remove Silence, which cuts gaps where the level sits between −60 and −20 dB (default −40 dB, minimum 0.4 s).
A music bed under narration. Music is not noise. WaveCraft's Duck Music finds where the voice track is audible and turns the music down only there, by an amount and fade times you set. That is the right tool here, and it is part of Pro.

What WaveCraft does differently

It measures before it destroys. The noise floor is on screen, and the button is disabled when there is nothing to fix.
It reports a measured fact, not a promise. How many dB will be removed depends on how entangled the noise and signal are, so no tool can quote it in advance.
It is honest about artifacts. The 8.5 dB tone loss and the 27 % flux penalty above are numbers the algorithm is built to avoid; they are what you hear as "underwater" or "sparkly".
Reduce Noise itself is free. Multitrack recording, every effect and preset, WAV/AAC/ALAC/FLAC export and audio from video stay free; Pro is stem separation, auto-mix and ducking, one-tap mastering, stem export and batch export. Everything runs on-device, with no account and nothing uploaded — the same pass on a train with no signal.
Honest limits. It is destructive and undo expires after 50 later edits, so export or back up the project (a .wcproj package) first. No denoiser fixes clipping, room reverb, or a level too hot at capture.

What to do next

If you have not heard the take yet: watch the clip indicator while recording, aim for peaks around −6 dBFS rather than full scale, and put 15–20 cm between the phone and your mouth in a room with soft surfaces. That is worth more than any denoiser. If you already have the take: open it in WaveCraft on iOS 17 or later, tap Reduce Noise, and read the number before deciding. An inaudible floor means a level problem — fix it with EQ or Normalize, not reduction. A high floor with the voice well above it means reduce at the default strength, listen on headphones, and go higher only if the hiss is audible, not just visible.

For a singer competing with a full mix rather than a room, removing vocals from a song is the separation path and mastering on your phone is the last stage after the mix is clean.

How to Denoise a Recording on iPhone: What the Tool Actually Measures | WaveCraft

Frequently asked questions