What Is Audio Cleaning Processing? A Plain Guide
Updated: May 2026
The full processing stack and what each stage targets
Audio cleaning is not one thing. It is a sequence of distinct processing stages, each targeting a different type of problem. Understanding which stage does what helps you decide which stages to apply and in what order -- and more importantly, which stages to skip when they would do more harm than good.
The standard order for voice recordings:
- Noise reduction
- De-reverb
- Filler and stutter removal
- Silence trimming
- EQ and spectral shaping
- Dynamic compression
- Loudness normalization
Noise reduction: the first pass
Noise reduction removes broadband background noise -- HVAC, fan hum, traffic, line noise, ambient room sound. It operates in the spectral domain, comparing the observed signal against a learned model of what voice should look like and suppressing frequency content that does not match. Output: a quieter noise floor with the voice preserved.
Noise reduction should run first because subsequent stages (EQ, compression, normalization) will amplify the background noise if it is not removed first. A 6 dB boost in an EQ makes your voice louder -- and it makes your HVAC 6 dB louder too. Clean the noise before any gain operations.
De-reverb: treating the room
Reverb and room reflections are separate from noise. A recording can have a low noise floor and still sound distant and echoey if it was made in a reflective room. De-reverb reduces the late reflections without removing the direct signal from the voice. The result is a tighter, more forward-sounding voice that feels closer to the listener.
De-reverb is the trickiest stage to calibrate. Too little and the room is still audible. Too much and the voice sounds unnaturally dry, lacking the slight early reflections that make voices sound natural and present. The right setting depends heavily on the specific recording environment.
Filler and stutter removal
Filler word removal (um, uh, like, you know, and similar) and stutter removal (false starts, repeated syllables) are speech-specific editing steps. They require speech recognition to locate fillers and classify them in context -- not all instances of "like" are fillers, for example. Aulio Studio's on-device speech model timestamps each filler and presents them for review before any cuts are applied, so you can skip fillers that are contextual rather than habitual.
Loudness normalization: matching platform targets
Loudness normalization brings a recording's perceived loudness to a target level measured in LUFS (Loudness Units Full Scale). Different platforms have different targets: Apple Podcasts and Spotify use -16 LUFS, YouTube uses -14 LUFS, broadcast television uses -23 LUFS. A recording that is louder or quieter than the platform's target will be turned down or up by the platform's playback normalisation, which can affect perceived dynamic range.
Normalization should run last, after all editing and processing steps, because any subsequent cut or gain change will alter the loudness again. Run it once at the end and export.
When to skip stages
Not every recording needs every stage. A recording made in a treated studio with a good microphone may not need noise reduction at all -- applying it adds a small processing tax for no benefit. A recording with perfect pacing and no filler habit does not need filler detection. Over-processing is a real risk: each stage introduces small amounts of artifact, and stacking unnecessary stages compounds those artifacts. Apply what the recording needs, skip what it does not.
Aulio Studio home · Pricing · For podcasters · Alternatives · Support