How to Clean Noisy Audio Recordings for Better Insights

Why Noise Matters More Than You Think

You have recorded an interview in a church hall, captured birdsong on a blustery afternoon, or pulled an old tape out of a drawer. The content is gold, but underneath it sits a constant background hiss, a low electrical hum, or the rumble of a passing lorry. It is tempting to ignore — surely the words are still intelligible? In practice, noise does more damage than most people realise. It masks quiet consonants, smears the timing of transients, and pushes your analysis towards false conclusions. Transcription accuracy drops. Automatic segmentation mistakes hiss for signal. Spectral measurements pick up a noise floor that has nothing to do with the sound you actually care about. Cleaning the audio is not vanity; it is a prerequisite for trustworthy insight.

Identify the Noise Before You Treat It

Different noise types need different tools, and applying the wrong one is how good recordings get ruined. Spend five minutes listening and looking before you change anything.

• Broadband hiss — the steady "shhh" of tape, cheap preamps or air conditioning. Broadly even across the spectrum and usually a few decibels below the quietest speech.
• Mains hum — in the UK this sits at 50 Hz with harmonics at 100 Hz, 150 Hz and beyond. Often caused by earth loops or unbalanced cables running near power supplies.
• Rumble and handling noise — energy below 80 Hz from traffic, footfall, desk bumps or microphone stands.
• Clicks, crackle and clipping — short, high-energy events from worn media, digital dropouts, or gain set too high.
• Intermittent noise — a fridge cycling on, a phone buzzing, a keyboard being hammered. These are usually handled by careful editing rather than blanket processing.

A spectrogram is your best friend here. Load the file, look at the frequency display, and you will typically see hum as horizontal lines, hiss as a uniform haze, and clicks as bright vertical streaks.

Prepare the File Properly First

Always work on a copy, and keep the original untouched. Process at the highest bit depth you have — 24-bit or 32-bit float — so repeated passes do not accumulate rounding errors. Then break long recordings into manageable sections of ten to thirty seconds. A single setting rarely suits a whole hour, and section-by-section work lets you tune the treatment to what is actually happening at any given moment.

Find a clean sample of noise. Ideally, ten seconds where nothing but the offending sound is present: a pause between sentences, or the room tone before the first word. That sample becomes your noise profile, and it is worth more than any preset.

Reducing Broadband Hiss in Stages

Gentle, repeated passes always beat one aggressive pass. Start with a reduction of 6 to 10 dB and listen carefully. If the result sounds watery, metallic, or as though the speaker is lisping, you have gone too far — undo, reduce the amount, or narrow the frequency range being treated.

• Learn from the noise profile, then apply spectral subtraction across the whole file rather than only the quiet passages.
• High-pass first. A 12 dB-per-octave filter at 80 Hz removes rumble and gives the noise reduction less to chew on.
• Use a light downward expander rather than a hard gate. A range of 3 to 6 dB with a slow release of 150–300 ms keeps breaths and room tone natural instead of chopping them into silence.
• Leave a little noise behind. Complete silence between words sounds unnerving and can confuse analysis tools that expect a continuous floor.

Tackling Hum, Clicks and Rumble Specifically

Hum is the easiest problem to solve and the easiest to over-solve. A narrow notch filter at 50 Hz plus its harmonics will remove most of it, but very narrow notches, with a Q of 20 or higher, preserve more of the surrounding speech. Sweep the frequency slightly to find the exact centre — mains is not always precisely 50 Hz. If the hum changes level over time, a dynamic notch or a narrow-band denoiser may work better than static filtering.

For clicks, use a dedicated de-clicker with short windows, then inspect the waveform. Manual redrawing is worth the effort for a handful of severe events. Crackle responds well to a de-crackle tool at moderate strength, applied twice rather than once at full power. Wind and rumble almost always yield to a high-pass filter; start at 60 Hz for speech and 40 Hz for music, raising it only until the sound begins to thin out.

Never process the same problem with two tools at once. If a notch filter is already handling hum, do not also ask the broadband denoiser to target that region — you will carve a hole in the low midrange that no amount of later work can fill.

Check the Result Before You Measure Anything

Never trust the display alone. Listen on headphones and on a small speaker — phone speakers reveal artefacts that studio monitors happily forgive. Then compare: invert one channel of the cleaned file against the original and sum them. What you hear is exactly what you removed. If you can hear speech in that difference signal, your processing has taken more than noise.

Finally, measure what matters. Noise floor, signal-to-noise ratio and loudness should all improve without the peaks falling. Take a fresh spectrogram and confirm the speech harmonics are intact. If a small amount of hiss remains but every consonant is present, you have made the right trade. Clean enough to analyse is the goal — not silent.