A humming fridge, a laptop fan, preamp hiss: almost every home recording contains some background noise. The tools for reducing it are powerful, but used aggressively they produce an artificial-sounding voice, worse than the noise itself. It pays to know the available techniques and, above all, their limits.
First rule: prevent
No software gives you a recording as clean as one that started out clean. Before thinking about processing:
- switch off or move away noise sources: air conditioners, fans, appliances;
- bring the microphone closer to the mouth, to raise the signal relative to the noise;
- use a directional microphone and point its rear toward noise sources;
- always record a few seconds of silence at the start, with everyone still: it will serve as the sample for noise reduction.
The quantity to improve is the signal-to-noise ratio: the distance between the level of the voice and that of the noise. Every centimeter closer to the microphone improves it for free.
Identifying the type of noise
| Type | What it sounds like | Typical source | Suitable tool |
|---|---|---|---|
| Broadband hiss | Constant "sss" | Preamp, fans | Spectral reduction |
| Tonal hum | Constant low "mmm" | Electricity, appliances | Notch filters, de-hum |
| Impulsive noises | Clicks, pops | Mouth, cables, chairs | De-click, manual editing |
| Variable noise | Traffic, distant voices | Outside environment | Hard; prevention |
Each type calls for a different tool. Applying spectral reduction to a tonal hum works poorly, and the same goes the other way round.
Noise gates and expanders
A noise gate shuts the signal off when the level drops below a threshold. In the pauses between sentences the noise is silenced; when the person speaks, the gate opens and lets everything through, noise included.
Main parameters:
- Threshold: the level below which the gate closes; set it between the noise level and the quietest part of the voice.
- Attack: how quickly it opens; too slow cuts off the start of words.
- Hold: how long it stays open after the signal has dropped.
- Release: how quickly it closes; too fast cuts off the ends of words.
- Range: how much it attenuates when closed.
A gate that closes completely makes the contrast between silent pauses and a noisy voice obvious. That is why an expander, or a gate with range limited to 10-15 dB, often works better: noise in the pauses goes down without disappearing completely, and the effect sounds more natural.
Spectral reduction
Spectral reduction tools analyze a sample of noise alone, derive a frequency-by-frequency "fingerprint" and then subtract it from the whole recording.
The typical procedure:
- Select a portion of noise only, at least one or two seconds long.
- Let the plug-in learn the fingerprint.
- Apply the reduction to the whole track, starting with a moderate amount.
- Listen on headphones and compare with the original.
Artifacts to avoid
When reduction is excessive you get:
- "musical" artifacts: small metallic warbles, like gurgling, in the pauses;
- muffled voice: sibilant consonants are mistaken for noise and attenuated;
- underwater effect in the midrange.
A good rule is to reduce noise just enough that it stops being distracting, not until it is gone. Two light passes sometimes sound better than a single heavy one.
Machine learning tools
Recent tools separate voice from noise using trained models. They are often effective on variable noise such as traffic or background voices, where classic spectral reduction struggles. Here too, overdoing it produces an artificial voice, with altered tone or unnaturally erased breaths. Use the intensity control when available and always compare with the original.
Order in the chain
In my work I follow this order for a vocal track:
1. Manual cleanup of clicks and isolated noises
2. High-pass filter to remove rumble below the voice
3. Moderate spectral reduction
4. EQ and compression
5. Light expander, if needed
Noise reduction goes before compression, because a compressor raises the quiet parts, and therefore the noise in the pauses as well.
In short
- Prevention beats any processing: move the microphone closer and turn off noise sources.
- Always record a few seconds of silence as a noise sample.
- Gates and expanders work in the pauses; a limited range sounds more natural.
- Spectral reduction is effective on constant hiss but creates artifacts when overdone.
- Reduce noise before compression and always compare with the original.
