What the noise floor costs you later

The cost of noise is paid later, not now

A recording with a quiet background hiss sounds acceptable on its own. The problem appears three steps downstream: you compress it and the hiss comes up with everything else, you cut between two takes and the background jumps, you add music and dialogue and the mix turns grey, and every attempt to fix it makes something else worse.

Noise is the constraint that limits what you can do to a recording afterwards. It is worth measuring rather than sensing.

Where it comes from, in the usual order of size

1. The room. Ventilation, a computer fan, a fridge, traffic, a light fitting’s transformer. Almost always the dominant source in a domestic recording, and almost always underestimated because you stop hearing it within a minute of entering.
2. The preamp. Its own thermal and circuit noise, amplified with the signal. The figure to look for is equivalent input noise, typically −125 to −129 dBu for a good preamp; a cheap one at high gain will be audibly worse.
3. The microphone. A condenser’s self-noise, published as an A-weighted figure. Under 15 dB-A is quiet; over 20 dB-A is audible on a soft voice.
4. The converter. Usually the smallest contributor in any modern interface.

The order matters because it tells you where to spend effort. Turning off the ventilation is free and typically wins 10dB. A better microphone might win 3.

Measure it in thirty seconds

Record ten seconds of the room with nobody speaking, then:

ffmpeg -i roomtone.wav -af "astats=measure_overall=RMS_level" -f null -

Then measure the speech. The difference is your signal-to-noise ratio.

• 60dB or more: very good, you can do anything to it.
• 45 to 60dB: normal for a decent home setup; usable with care.
• 30 to 45dB: the noise will be audible after compression and will rise in every quiet passage.
• Under 30dB: you are going to hear it, and denoising will cost you quality.

The inverse square law is a noise tool

Room noise is largely diffuse: it arrives from everywhere and its level does not change much as you move. Your voice obeys the inverse square law.

So halving the distance from your mouth to the microphone raises the voice by 6dB and leaves the noise where it was. Every halving of distance buys 6dB of signal-to-noise ratio, free. Going from 60cm to 15cm is 12dB: more than the difference between a budget microphone and a good one.

This is the same mechanism as the direct-to-reverberant argument, and it points the same way: distance is the most powerful control you have and it costs nothing.

Room tone, and why ten seconds saves an hour

Record the room in silence for at least ten seconds at the same gain, in the same position, before or after every setup.

It has two uses. First, edits. Cut a sentence out of an interview and the background stops for a frame; drop room tone underneath and the join disappears. Every dialogue editor does this and it is invisible when done and glaring when not.

Second, a noise profile. Every denoiser works better when given a clean sample of exactly the noise it is being asked to remove.

Ten seconds. Do it before anyone speaks, while everyone is still settling.

Why denoising leaves a strange texture

Classical noise reduction works by spectral subtraction. It measures the noise’s average level in each of several hundred frequency bands, then reduces each band of the signal by that amount.

The noise is random, so in any given moment a band’s actual noise level is above or below its average. Subtracting the average leaves a residual that is sometimes zero and sometimes a short isolated burst of energy in one band. Played back, those residuals are heard as brief tonal chirps appearing and vanishing: musical noise, and once you can name it you will hear it in every over-processed podcast.

Push the reduction harder and two further things happen. Consonants (s, t, k, f) are broadband and noise-like, so the algorithm treats them as noise and dulls them. And breaths and room reverb tails get chopped, which makes the voice sound as if it is being switched on and off.

Machine-learning denoisers, including the free RNNoise and the models built into several editors, are considerably better because they have learned what speech looks like rather than subtracting an average. They fail differently: they can remove non-speech content you wanted, smooth the voice into a slightly synthetic texture, and behave unpredictably on accents or languages under-represented in training.

The practical discipline is the same for both: use the least reduction that solves the problem. 6 to 10dB is usually enough and often inaudible. 20dB is nearly always audible and usually worse than the noise was.

Hum is a different problem

A steady tone at 50 or 60Hz with harmonics is mains hum, caused by a ground loop or an unbalanced cable near a power supply. It is not broadband noise and a denoiser is the wrong tool.

Fix it at the source: a balanced cable, a different socket, moving a power brick away from the microphone cable, or a ground-lift on a DI box. Failing that, a narrow notch filter at the fundamental and its harmonics removes it with far less collateral damage than broadband reduction.

The free path

Audacity has a noise-profile-based reduction: the classic spectral subtraction, with the residual characteristics described above. RNNoise is open source and available as an LADSPA plugin and inside several editors. ffmpeg has afftdn for spectral denoising and anlmdn for a non-local-means approach that is gentler on speech. DaVinci Resolve’s free edition includes its own voice isolation. iZotope RX is the paid industry standard and is named in job adverts; nothing here requires it.

The limitation

Denoising cannot add back what the noise masked. If a word sits at the same level as the hiss, removing the hiss removes the word with it. The information about what was said is not separable from the noise at that point: they are the same samples.

And the most advanced tools are the hardest to evaluate, because a machine-learning denoiser produces confident, clean output whether or not the result resembles what was said. On a quiet or accented passage it can smooth a word into something else entirely, and the result sounds plausible. Anything that will be quoted, transcribed or relied on should be checked against the original rather than trusted because it came out clean.