Levels, headroom, and loudness
Audio is where amateur work gives itself away fastest. People will forgive a soft picture. Nobody watches past ten seconds of bad sound.
Zero is the ceiling and everything is negative
The digital scale is dBFS, decibels relative to full scale. 0 dBFS is the largest number the format can hold. There is nothing above it. Drive a signal past it and the tops of the waveform are simply cut flat, which sounds like crackle on loud syllables and cannot be undone, because the shape was never recorded.
So you record with room to spare. Aim for peaks around -12 dBFS and let the loudest moments reach maybe -6. That leaves 12 to 18 dB of headroom, which is what saves you when someone laughs, slams a door, or leans into the microphone to make a point.
Recording quiet and lifting later is safe in 24 bit. Recording hot and clipping is not recoverable in any bit depth.
Peak tells you almost nothing about loudness
A snare hit and a sustained voice can peak at exactly the same number and sound nothing alike, because loudness is about energy over time and peak is about one instant.
This is why peak normalising is close to useless. Scaling a file so its single loudest sample lands at -1 dBFS tells you nothing about whether the piece is loud enough to hear in a car.
LUFS is the measurement that means something
LUFS averages level over time and weights the frequencies the way hearing does, so two files at the same LUFS sound equally loud. It is the number every platform now uses.
Platforms normalise. If you deliver louder than their target they turn you down, and all you have achieved is squashing your own dynamics for nothing. If you deliver quieter, some turn you up and some do not.
| Destination | Integrated target | True peak ceiling |
|---|---|---|
| YouTube | about -14 LUFS | -1 dBTP |
| Spotify, Apple Music | about -14 LUFS | -1 dBTP |
| Vertical social feeds | -14 to -16 LUFS | -1 dBTP |
| Podcast platforms | -16 LUFS mono, -19 to -16 stereo | -1 dBTP |
| Broadcast, EBU R128 | -23 LUFS | -1 dBTP |
For a spoken word piece going to the web, mixing to -16 integrated and holding true peak at -1 dBTP works everywhere and is what most people should do.
True peak, and the peak between the samples
A file can read -0.1 dBFS on every sample and still distort on playback. The samples are points, and reconstructing the smooth waveform through them can overshoot between two of them. Lossy codecs make this worse, because the decoded waveform is not the same shape as the one you encoded.
Leaving a dB of room at the top costs nothing audible and prevents it. That is why every delivery spec asks for -1 dBTP rather than -0.1.
Compression, so the quiet parts survive
Compression turns down the loud parts so the whole thing can be turned up. It has nothing to do with file compression despite the shared word.
For speech, a ratio around 3:1 with the threshold set so the compressor is working on the louder syllables is the standard starting point. What it buys you is a voice that stays present when the speaker turns their head or trails off at the end of a sentence.
Then a limiter last in the chain, set to catch true peaks at -1 dBTP. It does nothing most of the time and saves you from the one shout you did not plan for.
Dynamic range and where people actually listen
Wide dynamic range is beautiful in a quiet room with good headphones. It is useless in a car, on a bus, or through a phone speaker in a kitchen, where the quiet parts fall below the noise of the room and the listener gives up rather than reaching for the volume.
Check any mix on a phone speaker at half volume. If the softest line is unintelligible there, compress harder, even though it will feel over compressed on headphones. Your audience is not wearing headphones.
The noise floor decides how much you can lift
Every recording has a constant hiss and hum underneath: preamp noise, air handling, traffic, the room itself. That is the noise floor, and when you lift a quiet recording it comes up by exactly the same amount as the voice.
The distance between the voice and the floor is fixed at the moment of recording. Getting the microphone closer is the only real fix, and it is worth more than any piece of software. Every halving of the distance to the source gains you about 6 dB on the voice and nothing on the room.
Record room tone
Thirty seconds of the room making no particular sound, at every location. You lay it under gaps so that a cut does not drop into absolute digital silence, which sounds like the file broke. It takes half a minute at the shoot and it is unobtainable afterwards.