What Is Noise Reduction?

What types of noise it targets

Noise reduction tools are designed for stationary or semi-stationary noise — noise that doesn't change much over time and exists throughout the recording at a relatively consistent level and frequency profile.

• Microphone hiss — the broadband high-frequency noise from a mic's self-noise
• Electrical hum — 50 Hz or 60 Hz interference from power lines or ground loops
• Air conditioning / HVAC noise — constant low-frequency rumble and mid-frequency hiss
• Fan noise — computers, cameras, equipment in the recording space
• Room tone — the ambient acoustic fingerprint of the recording space

How traditional noise reduction works

Classic spectral noise reduction (used in tools like iZotope RX and Audacity) works in two stages:

Stage 1 — Noise profiling: You identify a section of the recording that contains only noise — no speech or music — and the tool analyses it. It builds a model of the noise's frequency spectrum: what frequencies are present, at what levels, varying over time.

Stage 2 — Attenuation: The tool applies that noise model to the entire recording. At any moment, frequencies that match the noise profile are attenuated (turned down) by a set amount. Frequencies carrying signal (speech, music) that differ from the noise profile are left alone.

The limitation: the distinction between "noise" and "signal" is probabilistic, not certain. When the algorithm is wrong — particularly during quiet passages where signal and noise are close in level — it attenuates some of the signal too, creating artifacts.

AI-based noise reduction

More recent tools use machine learning models trained on large datasets of clean and noisy recordings. Tools like Krisp, NVIDIA RTX Voice, Adobe's AI Enhance, and iZotope RX's Voice De-noise can separate speech from noise in real time without requiring a manual noise profile.

These models are often dramatically better on voice recordings specifically — they've been trained to recognise human speech patterns and preserve them while attenuating everything else. For non-voice content (music, environmental recordings), traditional spectral tools are still often more appropriate.

The trade-off: artifacts vs noise

This is the most important practical insight about noise reduction: it introduces its own artifacts if pushed too hard. The most common noise reduction artifacts are:

• Warbling or metallic shimmer:The noise reduction algorithm creates modulation artifacts in the frequency domain — the audio sounds as if it's underwater or being processed through a metallic filter.
• Loss of natural room tone:Silence between words or phrases becomes unnaturally dead — the listener can hear the noise reduction "switching on and off."
• Thinning of the audio:The tool attenuates noise and some signal together, leaving the voice or instrument sounding thinner or hollow.

The goal isn't eliminating all noise — it's getting the noise below the threshold of distraction without introducing artifacts that are themselves distracting. Subtle reduction (3–6 dB of attenuation) often sounds more professional than aggressive reduction (12+ dB).

What noise reduction cannot fix

• Dynamic noise: A car horn, a dog barking, a slamming door — these vary too much for a static noise profile to capture. They need manual editing or multiband gating.
• Noise that overlaps the signal: If the noise occupies the same frequencies as the voice, removing the noise removes part of the voice.
• Clipping: Noise reduction doesn't repair distorted waveforms.
• A bad room: Heavy reverb and echo aren't noise — they're reflections of the signal itself. Noise reduction makes them worse, not better.

Prevention is better than treatment

The most reliable noise reduction is not recording the noise in the first place. Recording in a quiet room, using a directional microphone close to the source, turning off fans and air conditioning before recording, and managing gain staging properly all reduce the need for noise reduction in post-production.

A recording with light noise that's been carefully reduced will almost always sound better than a recording with heavy noise that's been aggressively treated.