Fixing Bad Audio in Short-Form Video

Viewers forgive soft focus, a slightly crooked frame and a mediocre background. They do not forgive audio they have to strain to hear. In short-form video the tolerance is even lower, because the viewer has a thumb hovering and no investment in you yet.

Most phone-recorded audio has the same four defects, and they have a repair order. Doing them out of order is why amateur audio work often sounds worse after processing than before.

Diagnose before you process

Listen once, on a phone speaker, at a normal volume. Then write down what is actually wrong. The usual suspects:

• Too quiet. You have to turn the phone up past halfway.
• Clipped. The loudest words crackle or distort; the waveform hits the ceiling.
• Rumbly. Air conditioning, traffic, handling noise, a table thump.
• Hissy. A constant high-frequency wash under the voice.
• Uneven. Loud laugh, quiet sentence, loud again.
• Sibilant. The "s" sounds pierce.
• Muddy. Voice sounds boxed, like it is coming through a wall.

Each has a different fix, and some are unfixable. Applying noise reduction to a clipping problem, which is the single most common mistake, removes detail and leaves the distortion in place.

Step one: set the level

Work to a target rather than to taste. A commonly used target for streaming and social platforms is an integrated loudness around minus fourteen LUFS, with true peaks no higher than minus one decibel. Exact numbers matter less than consistency: if every clip in a series sits at the same loudness, the series feels professional and nobody can say why.

Two rules:

• Normalise, do not simply amplify. Normalisation to a target listens to the whole clip and raises it as a unit; amplifying by a fixed number of decibels will clip the loud parts.
• If the source is already clipping, stop. Clipping is distortion baked into the file. No amount of level work recovers it. Re-record, or accept it and move on.

Step two: remove the low rumble

A high-pass filter — cutting everything below roughly 80 to 100 hertz — is the single highest-value move on phone audio. Almost nothing useful in a human voice lives down there, and almost all room noise does. The result is an immediately cleaner, less muddy sound with no audible change to the voice itself.

If the voice still sounds boxy after that, a gentle cut somewhere between 200 and 400 hertz can reduce the "recorded in a cupboard" quality. Small moves: minus two or three decibels, not minus ten.

Step three: reduce noise gently, or not at all

Noise reduction is a trade. It removes hiss and hum, and it also removes the quiet detail that makes a voice sound present. Push it far enough and the result is the underwater, robotic texture everyone recognises from bad conference calls.

Practical guidance:

• Apply it after the high-pass filter, so it has less to work on.
• Aim to reduce the noise floor, not to eliminate it. A little room tone is natural; absolute silence between words sounds synthetic.
• Compare against the original at matched loudness. If the processed version sounds thinner and the noise is still audible, undo it.

For steady tonal noise — a fan hum at one frequency, mains hum — a narrow notch is far less damaging than broadband reduction.

Step four: tame sibilance

Sharp "s" and "sh" sounds are exaggerated by phone microphones and by close placement. A de-esser, or a narrow cut around 5 to 8 kilohertz, fixes most of it.

Do not remove sibilance entirely. The difference between a voice with softened sibilance and one with none is the difference between a person and a text-to-speech engine. Aim for "not piercing" rather than "gone".

Step five: build the music bed properly

Music under speech is where most amateur clips fall apart, in one of two directions: the music is so quiet it adds nothing, or it competes with the voice.

Two practical rules:

• Duck the music under speech. Around 18 to 22 decibels below the voice is a reasonable starting point; the voice should be unmistakably louder at every moment.
• Give the music somewhere to be. If the music is only ever a background wash, let it come up in the gaps — before the first word and after the last. That contrast is what makes a bed feel intentional.

Also check the licence. A track that gets a video muted by a rights system costs more than the clip was worth.

Step six: trim silence without clipping the consonants

Cutting dead air tightens pacing, but automated silence removal has a predictable failure: it clips the tail of a word, so "right" becomes "righ". Listen to the first and last 200 milliseconds of every cut.

A workable compromise is to keep a short handle — a tenth of a second or so — on both sides of each cut, and to trim only gaps longer than about half a second. Silence shorter than that is pacing, not dead air.

What cannot be fixed

Worth knowing, so you do not spend an hour on it:

• Clipping. Distortion is in the file.
• Overlapping speech. Two voices at once cannot be separated cleanly by a filter.
• Heavy reverb in a large empty room. De-reverb tools exist and all of them sound like they are working hard.
• A microphone too far from the mouth. Distance costs clarity, and no processing adds back what was never captured.

In those cases, the fix is the next recording, not this one.

Record better next time, in four changes

1. Get closer. A phone held 15 to 20 centimetres from the mouth sounds dramatically better than one on a table a metre away.
2. Soften the room. A carpet, a curtain, a wardrobe full of clothes. Hard parallel walls are the enemy.
3. Turn off the noise. Air conditioning, fans, notifications, the fridge.
4. Watch the meters. Peaks should sit below the ceiling, not touching it.

Those four changes do more than any processing chain.

Monitor on the worst speaker you own

Mix on a phone speaker, not on headphones, because that is the device most viewers will use. If the voice is clear and the music does not muddy it on a phone speaker, the mix works. Headphones hide exactly the problems that a phone speaker exposes.

Then check once more with the volume at about a third. If you can still follow every word, the clip is done.