If your voice still hisses after noise removal, the denoiser is usually fine and the stage AFTER it is undoing the work: loudness normalisation applies makeup gain to everything, residual noise included. Diagnose it with two numbers, gap separation and high-frequency tilt. The fix, in order: a gentle expander after the makeup gain, a low-pass at 15.5 kHz, and a makeup ceiling around 16 dB.
If your voice still hisses after you turn on noise removal, the denoiser is probably not the problem. In most editing chains the stage after it is quietly undoing the work: loudness normalisation raises everything by a fixed amount, and "everything" includes the noise the denoiser left behind.
We hit exactly this in Zella, measured it, and fixed it. Here is the diagnosis, because the technique transfers to any editor.
The symptom: noise removal makes hiss more noticeable
The confusing part is that noise removal genuinely works and the result still sounds worse. That happens when the chain looks like this:
- Denoise - pulls the noise floor down, but never to zero. What survives is usually a thin, high-frequency residue.
- Compress - squeezes the loud and quiet parts closer together, which lifts quiet material including that residue.
- Makeup gain - raises the whole track to hit a loudness target like -14 LUFS.
Steps 2 and 3 do not know what is voice and what is noise. So a denoiser that removes 20 dB of hiss, followed by 15 dB of makeup gain, nets you 5 dB of improvement - and the hiss is now sitting in a much quieter-sounding gap between words, where it is far easier to hear.
For the underlying concepts, Apple's audio fundamentals is a clear primer on levels, headroom, and why a mix that peaks correctly can still sound wrong.
What does a spectral denoiser actually remove?
Knowing what the tool is built to do explains why hiss is the case it handles worst.
A spectral denoiser chops your audio into short overlapping windows, runs each through an FFT to get an energy value per frequency bin, and compares every bin against an estimated noise floor. Bins well above the estimate are treated as signal and left alone. Bins close to it are attenuated. Then the frames are turned back into audio.
That design has a clear best case and a clear worst case:
| Noise type | Example | How well it comes out |
|---|---|---|
| Stationary, narrowband | 50/60 Hz mains hum, a fridge, a fan | Very well. The profile barely moves, so the estimate is accurate |
| Stationary, broadband | Mic self-noise, preamp hiss, room air | Partly. It sits under your voice at every frequency, so removing it all removes voice with it |
| Non-stationary | A door, a keyboard, a cough | Poorly with a classic profile-based denoiser; better with a learned model |
Broadband hiss is the awkward middle case. It is not concentrated anywhere the denoiser can carve out, and it overlaps the same frequencies as your consonants. Push the reduction hard enough to erase it and you take the air out of every "s" and "t" with it. So every sane denoiser leaves some behind on purpose. That residue is the thing your loudness stage then amplifies.
Why does a noise gate not remove hiss?
Because a gate works in time, not in frequency. It has one decision to make, many times a second: is the signal above the threshold or below it? Above, everything passes. Below, everything is attenuated.
Whenever you are speaking, your voice is far above the threshold, so the gate is wide open and the hiss underneath rides straight through with it. A gate only ever cleans the gaps between words. That is genuinely useful, but it means a gate can never be the answer on its own, and it is why gated audio with a bad noise floor sounds like it is breathing: silent between phrases, hissy the instant you talk.
Why does over-processing sound watery?
The chirping, underwater, phasey quality you get from an aggressive denoise setting has a name: musical noise. It comes from the same bin-by-bin decision described above.
When a bin's energy sits right at the noise estimate, tiny frame-to-frame fluctuations flip it between "keep" and "cut". The bins that survive are short-lived isolated tones scattered randomly across the spectrum, and each one lasts about the length of one analysis frame. Your ear does not hear them as noise. It hears them as little pitched blips, which is exactly the watery warble.
The fix is to stop asking for total removal. Back the reduction off until it leaves a low, steady noise floor rather than trying to reach silence, because a quiet consistent floor is far less distracting than an intermittent chirping one.
How to diagnose it in two measurements
You do not need golden ears. Two numbers tell you what is wrong.
1. Gap separation. Measure the RMS level of your loudest 200 ms and your quietest 200 ms. On a clean voice track the gap should be 40 dB or more below the speech. If it is only 10-15 dB below, there is no real silence anywhere in your file - the "quiet" parts are full of amplified noise.
2. High-frequency tilt. Compare the average energy in 4-8 kHz against 8-16 kHz during a pause. Room tone and most real-world noise roll off as frequency rises. If your top octave is louder than the band below it, something in your chain is adding high frequencies - and hiss lives there.
On the file that started our investigation, the numbers were a noise floor only 12 dB under program level, and 8-16 kHz sitting 10-19 dB above 4-8 kHz. That second number is the giveaway: no microphone produces that shape naturally.
Both are measurable in any editor with a level meter and a spectrum analyser. Find a pause of at least half a second, read the RMS there, then read it again over a normal sentence, and subtract. You are looking for one number, and it should be big.
The three fixes, in order of impact
1. Put a gate AFTER the makeup gain
This is the big one. A noise gate before loudness normalisation is pointless, because the gain stage re-amplifies whatever survives. Placed after it, a gentle downward expander - not a hard gate - pulls the gaps back down without chopping word endings.
Use a gentle ratio (around 1.5:1), a threshold well below your softest consonant, and a slow release (200 ms or so). A hard gate that slams to silence is worse than the hiss: it makes breaths disappear mid-sentence and produces a pumping effect every time you pause.
2. Roll off above 15 kHz
There is nothing in a human voice above about 15 kHz that you will miss on a phone speaker, laptop, or set of earbuds. There is, however, plenty of microphone self-noise and denoiser residue up there. A gentle low-pass at 15.5 kHz removes it and is inaudible on the voice itself.
3. Lower your makeup ceiling
If your normaliser is allowed to apply +18 dB to hit its target, a quiet or noisy take will use all of it. Capping makeup gain at around 16 dB costs you very little loudness on normal takes and prevents the worst-case take from being blown up.
The trap: aggressive de-essing plus an air shelf
One more thing worth checking. Many "podcast" presets pair a de-esser (a dip around 6-7 kHz) with an "air" or "presence" shelf (a lift above 8 kHz). Individually both are reasonable. Together they carve a valley and then a peak right where hiss lives, which tilts your noise into the top octave - producing exactly the measurement in the section above.
If you are chasing hiss, try backing the air shelf off to +2 or +3 dB and the de-ess to -4 dB before you reach for anything more aggressive.
What order should the chain actually run in?
Every fix above is really the same instruction: put the stages in an order where nothing downstream can undo the stage before it.
- Denoise first, while the noise floor is still at its original level and easiest to estimate.
- De-ess next, before any shelf can exaggerate what you are about to dip.
- Compress, knowing it will lift whatever the denoiser left.
- Loudness-normalise, with a cap on how much makeup gain it may apply.
- Expand the gaps back down, after the last thing that could re-amplify them.
- Low-pass at the very end, so nothing later can put the top octave back.
Steps 5 and 6 are the ones almost every default preset gets wrong, because they are usually placed at the start of the chain where the gain stages simply undo them.
What Polish Voice does in Zella
Zella bundles that whole chain into one button. Polish Voice runs on-device and does noise removal, de-essing, compression and loudness normalisation in a fixed order, with the anti-hiss stages in the right place, so you are not assembling the chain by hand. Five studio tones sit on top of the same broadcast-clean foundation: Natural, Warm, Bright, Podcast and Cinematic. If a music bed is playing, Auto-Duck dips it under your voice and swells it back automatically, so you are not fighting the music for the same frequency space.
Auto-Polish already runs Remove Noise and Polish Voice as part of its pass, so on most recordings the audio is handled before you touch anything.
What we changed in Zella
Our anti-hiss stages - the post-makeup gate, the 15.5 kHz low-pass, and the makeup ceiling - were only running on Apple Silicon Macs. Intel Macs got the loudness stage without the protection, which is precisely the failure mode described above.
They now run on every Mac, and both architectures share the same gentler voice tuning. Measured on a synthetic take with a known hiss bed, gap separation went from failing to 42 dB and the high-frequency tilt from +3.8 dB to -0.2 dB - flat, which is what you want.
The lesson generalises past our codebase: when an audio "fix" makes things worse, look at what runs after it, not at the fix itself.
Frequently asked questions
Why does my audio sound worse after noise removal? Usually because a compressor and a loudness stage run after the denoiser and re-amplify the residue it left behind. The denoiser did its job; the gain applied afterwards put the hiss back, and now it sits in a quieter-sounding gap where it is easier to hear.
Can a noise gate remove hiss? Only in the gaps. A gate is a time-domain switch, so while you are speaking it is open and the hiss underneath your voice passes through untouched. Use a gentle downward expander for the gaps and fix the rest with denoising and a low-pass.
Why does my denoised voice sound watery or underwater? That is musical noise, caused by pushing the reduction so hard that bins near the noise estimate flicker between kept and cut, leaving isolated short tones. Back the reduction off until it leaves a low steady floor instead of chasing silence.
What frequency should I roll off to kill hiss? Around 15 to 15.5 kHz. There is nothing in a human voice above that you will miss on earbuds, a laptop or a phone speaker, but plenty of microphone self-noise and denoiser residue lives up there.
Is hiss a microphone problem or a software problem? Both are possible, and the high-frequency tilt measurement tells you which. If the 8-16 kHz band is louder than 4-8 kHz during a pause, no microphone made that shape and the cause is in your processing chain.
Does fixing hiss require paid software? No. In Zella, Remove Noise and Polish Voice are free, run entirely on your device, and already have the anti-hiss stages in the correct order.
The short version
Noise removal that leaves hiss is usually a chain-order problem, not a denoiser problem. Measure your gap separation and your high-frequency tilt. If the gaps are less than 30 dB below your speech, add a gentle expander after the makeup gain. If the top octave is louder than the band below it, roll off above 15 kHz and check whether a de-esser and an air shelf are fighting each other. In Zella, Remove Noise and Polish Voice do this for you, entirely on your device.
Make your next video with Zella.
Record, edit and ship on Mac, iPhone and iPad - local, private, free to start.
RELATED