Hiss and noise — denoising

The steady shhhhh behind a recording — tape hiss, a noisy microphone preamp, an air-conditioner, the electrical fuzz of a cheap interface — is the most common complaint in all of audio. Removing it is called denoising, and it’s the clearest example of the “analyse → modify → resynthesise” loop from chapter 2.

The core idea: subtract the haze

Recall that on a spectrogram, hiss looks like a faint, even haze sitting under everything, across many pitches at once, all the time. The voice, by contrast, is bright bands that come and go.

So denoising asks a simple question, pitch by pitch: “how much faint, ever- present haze is at this pitch?” That amount is the noise profile — a measurement of the hiss’s recipe. Once you know it, you go through the sound and, for each pitch in each moment, subtract the haze amount. Where the voice is loud, subtracting a little haze barely changes it. Where there’s only haze (the silences between words), subtracting the haze amount leaves… almost nothing. Silence. That’s the trick, and it has a name: spectral subtraction.

Two questions remain: how do you measure the haze, and how hard do you subtract.

Measuring the haze (the noise profile)

There are two ways, and cathar offers both:

• Learn it from silence. If your recording has a moment of “room tone” — a patch with no voice, just the background — you can point the tool at it and say “this is the noise; memorise its recipe.” Cathar calls this a noiseprint. It’s by far the more accurate method, because you’re showing the tool a clean example of exactly what to remove.
• Guess it automatically. If there’s no clean silence, the tool assumes the quietest moments at each pitch are mostly haze, and builds the profile from those. Cathar does this with a method called minimum statistics. It’s convenient and needs no setup, but it’s a guess, so it’s a little gentler and less surgical than a real noiseprint.

In practice: if you can spare even half a second of “just the room,” learn a noiseprint from it. It is the single biggest quality lever in denoising, in any program.

How hard to subtract (the aggressiveness knob)

Subtract too little and hiss remains. Subtract too much and you start eating into the voice — and you create a very recognisable artefact: a twinkly, watery, “underwater” or “musical noise” sound. (It happens because subtraction can leave isolated little flecks of pitch that warble.) So every denoiser has an aggressiveness control. In cathar it’s --alpha (gentle around 1.5, strong around 4–6) plus a floor (--beta) that refuses to ever fully silence a pitch, which keeps the result natural instead of glassy.

The whole game of denoising is this trade-off: hiss versus artefacts. There is no setting that removes all hiss and adds nothing; the skill is finding the spot where what’s left is less distracting than what you’ve added.

A gentler cousin: the Wiener filter

Instead of bluntly subtracting the haze, you can scale each pitch by how likely it is to be real sound versus noise: pitches that tower over the haze are kept almost fully, pitches barely above it are turned right down. This is the Wiener filter (cathar’s --wiener option). On steady, gentle hiss it often sounds smoother and less twinkly than plain subtraction. Same goal, slightly different maths for the “modify” step.

How the big tools do it

This is mature, well-understood territory, and the concepts are identical everywhere:

• Audacity (free) has “Noise Reduction,” which works exactly like cathar’s noiseprint method: you select a quiet bit, click “Get Noise Profile,” then apply. Same idea, same trade-off knobs.
• Adobe Audition offers both a learned-profile “Noise Reduction” and an adaptive “DeNoise” that guesses, like cathar’s two modes.
• iZotope RX is where the money shows. Its “Spectral De-noise” does classic profile-based subtraction very well, but its flagship “Voice De-noise” and the newer AI-powered modes use machine learning — models trained on thousands of hours of speech — to tell voice from noise far more cleverly than any “subtract the haze” rule. That’s the real frontier: the concept is the same, but a trained model makes a smarter decision in the modify step, pulling clean speech out of noise that classical subtraction would smear.

Cathar sits firmly in the classical camp: transparent, predictable, no weights, genuinely good on steady hiss — and honestly outclassed by RX’s ML on the hardest cases (heavy, non-steady background noise like a busy café). Knowing which problem you have tells you which tool you need.