Making a voice sound like a voice

Almost everything that makes a voice recording sound bad happened before you opened the editor. What follows is how to get the most out of what you have, and which problems you should stop trying to fix.

Cut before you boost

EQ moves frequency ranges up or down. Boosting is satisfying and usually wrong, because boosting raises the noise in that range along with the voice. Cutting what is in the way is quieter, cleaner, and does the same job.

Three moves handle most spoken word.

A high pass filter at 80 Hz. Nothing in speech lives below that, and traffic, air conditioning, footsteps, table bumps and microphone handling all do. This one filter improves more recordings than everything else combined.

A small cut between 200 and 400 Hz if the voice sounds boxy or muddy, which is what a room with parallel hard walls does to a close microphone. Two or three dB with a wide bell. Take out more and the voice goes thin.

A gentle lift above 3 kHz if the voice sounds dull. Two dB. This is where consonants live and where intelligibility comes from, and it is also where sibilance lives, so go carefully.

Sibilance and plosives

Harsh S and T sounds sit between 5 and 8 kHz. Close microphones make them worse and any high end lift makes them worse again. A de-esser is a compressor that only listens to that band and only ducks when the S arrives, which is far better than simply cutting the range and dulling every other word.

Plosives, the thumps on P and B, are a blast of air hitting the capsule. A four dollar foam screen prevents them and a high pass filter reduces them. Nothing removes one cleanly once recorded. If a plosive lands on an important word, the usual fix is a very short volume dip on that syllable alone.

Noise reduction has a limit and it is lower than you think

Noise reduction learns the shape of a steady background and subtracts it. It cannot tell the difference between room noise and the quiet tail of a word, so past a certain point it starts eating the voice.

Take out 6 to 10 dB. Live with the rest. Past that the voice develops a metallic, underwater quality that everyone recognises even if they cannot name it, and which sounds worse than the hiss you removed.

If the recording is truly unusable, the honest answer is to record it again rather than to process it into something artificial.

Mono for narration

A single speaking voice belongs in mono, centre. It is effectively louder for the same level, it has no phase problems, and nobody in history has complained that narration was not in stereo.

Watch for the file that is technically stereo with silence in one channel, which is what a single microphone into a two channel recorder produces. Played on a device that does not sum channels it comes out of one speaker at half the level. Check which channel your audio is really in before you build anything on it.

Music beds

A bed carries energy between points and covers the seams in an edit. It should be felt rather than listened to.

Instrumental only. A vocal under narration means the audience gets neither, because the brain cannot parse two streams of words at once and will not choose. Avoid anything with a strong hook or a dramatic build, since the audience will follow the music and stop listening to you.

Cut the music to the edit rather than fading it out arbitrarily. Ending on a downbeat sounds finished. Fading out mid phrase sounds like the file ran out.

Ducking, and how much

Ducking lowers the music automatically whenever somebody speaks. Set by hand it is a series of volume keyframes. Set by sidechain it is a compressor on the music listening to the voice track.

Nine to twelve dB of duck is the useful range. Less and the voice fights the bed, which is tiring to listen to even when every word is audible. More and the music pumps in and out distractingly between sentences.

The release time matters as much as the depth. Half a second or so, so the bed comes back gradually and does not jump up in the gap between two words. A fast release is what makes ducking sound like a machine doing it.

J cuts and L cuts

These are audio edits, and they are what makes a sequence feel like it was made by somebody rather than assembled.

An L cut keeps the outgoing speaker's voice running while the picture moves to somebody listening. It is the backbone of interview editing, because it lets you cut away from a talking head without cutting away from the point.

A J cut brings the next scene's sound in before its picture. The audience hears where they are about to go, which pulls them forward through the edit.

Both need the audio unlinked from the video, or a tool that lets you drag the audio edge independently. Relink afterwards so nothing slides out of sync later.

Fades on every audio cut

A hard cut through a waveform leaves a step, and a step is a click. Two or three frames of fade at every audio edit removes it and is completely inaudible as a fade. Some tools do it automatically. Most do not.

The same reason is why cuts snap to zero crossings in some editors: both sides start from silence, so there is no step to hear.

Order of work

1. Get every line of speech clear and consistent first, with nothing else playing.
2. Check it on a phone speaker.
3. Bring the music in underneath and set the duck.
4. Place effects and ambience last.
5. Measure the whole piece and set the master to your loudness target.

Doing this in any other order means redoing the speech after the music arrives, and the speech is the part that matters.