Compression does not make anything louder

Compression does not make anything louder

A compressor turns down anything above a threshold. That is the entirety of what it does. It has no mechanism for making sound louder.

What makes it louder is the make-up gain you add afterwards. Once the peaks have been pulled down, there is headroom, so you can raise the whole signal — and the quiet parts come up with it. Hence the only description of compression worth memorising: turning the loud bits down so the quiet bits can come up.

Understanding that changes how you set it, and it explains the failure mode at the end of this lesson.

Do it in this order

Clean the performance, then fix the tone, then control the dynamics, then set the loudness. In that order.

Out of order, the compressor reacts to noise you are about to remove, and your EQ shapes a signal you are about to change. People who cannot get a voice to sit right are usually sequencing wrong rather than doing any one step wrong.

Editing the performance

Cut the false starts, the long pauses and the ums. Do not cut every breath. A voice with no breaths sounds inhuman, and breaths mark where sentences begin and end — remove them all and listeners lose the structure of what you said. Reduce a distracting breath by 6–10 dB instead of deleting it.

Two mechanical points that account for most audible edits:

• Place a cut at a zero crossing, or put a 5–10 ms crossfade over it. Cutting through the middle of a waveform leaves a step, and a step is a click. Most editors handle this; Audacity does not always.
• Cut immediately after a hard consonant — a t, k or p — where the seam hides. Cutting in the middle of a vowel is audible even when the timing is right.

Then lay your room tone under the whole track. Silence between edited phrases sounds like the audio has dropped out. Continuous room tone is what makes heavily edited speech sound like one take.

EQ in three moves

1. 1High-pass filter. Around 80 Hz for a deep voice, 100–120 Hz for a higher one. Below that there is rumble, air conditioning, traffic and desk thumps, and no speech at all. Sweep it upward until the voice starts to thin, then come back a little.
2. 2Cut, do not boost, between 200 and 500 Hz if the voice sounds boxy, boomy or muddy. Two to four decibels with a fairly wide bandwidth. This is where a small untreated room lives.
3. 3A gentle lift of 2–3 dB between 3 and 6 kHz for clarity. If that makes the s sounds harsh, use a de-esser around 5–8 kHz rather than a static cut, which dulls the whole voice to fix a few moments.

The general rule: cuts are forgiving, boosts are not. If you find yourself boosting more than about 4 dB anywhere, the answer is at the recording end.

Setting a compressor

A workable starting point for speech:

• Ratio 3:1.
• Threshold set so the loudest phrases show 3–6 dB of gain reduction on the meter — and the quiet phrases show none.
• Attack around 10 ms. Fast enough to catch the peaks, slow enough to let the consonant transients through. A very fast attack dulls speech and is a common cause of a voice that sounds oddly soft.
• Release 100–200 ms, or auto.
• Make-up gain until the compressed version matches the uncompressed version in perceived level, so you are judging the sound and not just the volume.

The failure mode. If the background hiss rises and falls between sentences, you are over-compressing. During speech the compressor is pulling the level down; in the gap there is nothing above the threshold, so nothing is being pulled down, and the full make-up gain lands on the room noise. That swelling and receding is the sound of too much gain reduction. Back it off, or gate gently, or fix the noise floor at the source.

De-noising, honestly

It works by learning a profile of a steady sound and subtracting it in the frequency domain. So it handles hum, fans and hiss well, and anything that varies badly.

Push it hard and you get the characteristic artefacts: an underwater quality, and small warbling tones sometimes called birdies, which appear where the subtraction has removed most but not all of a frequency band. Six to ten decibels of reduction is usually inaudible. Twenty is usually obvious.

Take the least you can live with. A slightly hissy voice is far easier to listen to for ten minutes than a clean one that warbles.

Loudness, and why louder is pointless

Streaming platforms normalise. Push your file louder than roughly −14 LUFS and YouTube or Spotify will simply turn it back down, and all you have achieved is a squashed, airless dynamic range that arrives at the listener at the same volume as everyone else's.

Aim for around −14 LUFS integrated for video platforms, around −16 LUFS for podcasts, with true peak no higher than −1 dBTP. Free ways to measure: Resolve's Fairlight page has a loudness meter, ffmpeg's loudnorm filter measures and applies, and Youlean's free loudness meter plugin works in most hosts.

The free toolkit, and ten seconds today

• Audacity — free and open source, does every step above, including a usable noise reduction.
• DaVinci Resolve, Fairlight page — free, a proper mixer with compression, EQ, metering and a ducker.
• Reaper — an evaluation that never expires and a personal licence around $60, as of 2025.
• Ocenaudio, Ardour, Cakewalk on Windows — all free or near enough. On a phone, CapCut's audio enhancement is fine for speech.

Ten seconds today

Take one minute of your voice, apply a high-pass filter at 100 Hz, and A/B it. That one control is usually the largest single improvement available to you.