How to Fix Your ACX Noise Floor (Without Destroying Your Recording)

Most narrators discover they have a noise floor problem after the chapter is done — or three chapters, or the entire book. The rejection email arrives and there's a decision to make: fix it digitally or go back to the microphone.

The right answer depends on where you are on the meter and what's actually creating the noise.

Open your DAW and find a three-second patch of pure room tone where you aren't speaking, breathing, or moving. Don't look at a peak meter. Peak only catches instantaneous spikes — a cable thump, a mouth click. To measure noise floor accurately, you need an RMS meter.

If your DAW doesn't have a built-in RMS statistics tool, add a free metering plugin like Voxengo SPAN or Yooz Audio Juice to your master track. Set the metering standard to AES17 RMS if the option exists. Loop the three-second silence and hit play. Watch the integrated or average RMS value.

For ACX compliance, that number needs to sit at -60 dBFS or lower. A reading of -63 or -65 dBFS is clean. A reading of -55 dBFS is a problem — that's audible hiss filling every pause in your performance. A reading of -45 dBFS means your listeners aren't hearing silence; they're hearing your computer's cooling system.

The most common mistake when a narrator sees a failing number: reaching for a noise reduction plugin immediately.

A noise gate doesn't remove noise. It slams a digital door shut when you stop speaking, then swings open the moment you do — bringing the hiss right back alongside your voice. Before touching a plugin, put on closed-back headphones and listen to a few seconds of recorded silence. The fix depends entirely on what you hear.

Acoustic noise — a steady wide-band hiss or whir — is mechanical. Turn off HVAC entirely while recording. If your computer sounds like a jet engine, move it outside the room.

Electrical noise — a 60 Hz or 120 Hz hum — points to your cables or power supply. Unplug your laptop's charging brick and run on battery; cheap power adapters inject high-frequency interference into USB interfaces.

Preamp hiss — a smooth, continuous "shhh" when the room is genuinely quiet — means your hardware is struggling. This typically happens when a low-output microphone like the Shure SM7B is paired with a budget audio interface. To reach a healthy recording level, you crank the interface gain past 90%, which drags the preamp's inherent self-noise up with the signal. The hardware fix is an inline pre-preamp — a Cloudlifter CL-1, a Dynamite DM1, or a FetHead. These use the interface's phantom power to inject 20 to 28 dB of clean gain before the signal reaches the preamp, letting you back the gain dial down to 50% and drop the hiss below the audible threshold.

If you don't fix the physical source first, digital noise reduction has to work twice as hard — and when it does, it eats your consonants along with the noise.

There's a threshold where digital cleanup stops being audio engineering and starts sounding like bad science fiction.

• -60 dBFS or lower: Compliant. Proceed to mastering.
• -59 to -50 dBFS: The salvage zone. Transparent processing can nudge this down 5 to 8 dB without damaging vocal quality.
• -49 to -45 dBFS: Danger zone. Proceed carefully and listen closely to the tails of your words as you process.
• Above -45 dBFS: Re-record. Pulling a noise floor down 15+ dB forces the algorithm to strip energy from your consonants. The result sounds metallic and hollow, and ACX will reject the file for digital artifacts rather than noise floor.

If the problem only affected a few chapters — a fan that kicked on mid-session — re-record just those. If the entire book sits at -42 dBFS, fix the physical source first, then return to the microphone. Re-recording a clean take costs less time than fighting a digital battle you'll lose at quality control.

Between -59 and -50 dBFS, split the work across three stages rather than forcing a single plugin to carry it all.

Stage 1 — High-pass filter. Roll off everything below 50 to 60 Hz at a 12 or 18 dB-per-octave slope. Your voice carries no meaningful energy there, but your room does — HVAC rumble and building vibration live in the sub-bass range. This alone can drop your noise floor 2 to 3 dB without touching your vocal tone.

Stage 2 — Spectral de-noising. Use iZotope RX Spectral De-noise or Waves WNS. Select the three-second room tone patch and let the plugin learn the frequency fingerprint of your noise. Set the reduction to no more than 4 to 6 dB — you're softening the noise, not eliminating it. This precise, multi-stage balancing act is exactly why we built Greenlit Audio's processing pipeline with automated DeepFilterNet3 AI spectral cleanup — it separates vocal energy from background hiss without introducing the metallic artifacts that manual over-processing creates.

Stage 3 — Downward expansion. An expander like FabFilter Pro-G gently attenuates the signal when you stop speaking rather than cutting it abruptly. Set the threshold 3 to 5 dB above your raw noise floor, ratio at 1.5:1 or 2:1, attack around 1 to 2 ms, and release at 250 to 400 ms so word tails aren't clipped.

Let the de-noiser handle 5 dB and the expander handle 5 dB — you reach -60 dBFS without audible seams. What to avoid: Audacity's destructive noise reduction applied globally at 12 or 15 dB. That permanently strips high-frequency energy from your consonants, and the phase distortion it introduces often pushes true peaks past the -3 dBFS ceiling — trading one rejection reason for another.

Over-processed audio has a recognizable signature. Three things to listen for:

• The underwater quality — vowels sound hollow, phasey, or metallic, as if you're speaking through a PVC pipe.
• Chipped consonants — the word "track" starts to sound like "ra-ck."
• Gating artifacts — an audible pumping effect between sentences as room tone aggressively cuts in and out.

To catch these before submission: export a two-minute section, open it in a blank session with zero plugins active, and listen on closed-back headphones at a volume 15 to 20% louder than comfortable. Focus on the end of each sentence — the moment your voice fades into silence. Any metallic ringing, digital chirping, or abrupt snapping means your settings are too aggressive. Reduce spectral de-noise by 2 dB and test again.

The simplest fix runs before you record anything.

Slide a dense draft stopper against the base of your door. A one-inch floor gap on a hollow-core door lets in more ambient noise than most narrators expect — blocking it can reduce room noise by 3 to 4 dB before you've touched a single dial.

Unplug external hard drives (spinning platters generate a high-frequency whine) and run your laptop on battery. Turn your phone completely off — a phone searching for a cell signal near an unshielded microphone cable can register as a rhythmic tick on your RMS meter.

Then sit in your recording position in total silence for 30 seconds with the RMS meter running. Your body absorbs acoustic energy. The meter reading with you sitting still is your actual recording environment. If it reads -58 dBFS before you've spoken a word, you're already too close to the line. Adjust before recording two hours of audio you'll have to fight to salvage.

Ears adapt to steady background frequencies — that's why the room hiss stops bothering you after ten minutes of monitoring. A technical diagnostic doesn't adapt.

Before logging into ACX, run your finished file through Greenlit's free diagnostic. It catches three things ears reliably miss: true peak violations down to the tenth of a decibel, integrated RMS loudness averaged across the entire file, and format errors like a sample rate that drifted or a bit rate encoded at variable instead of constant.

If the report comes back clean, submit with confidence. If it flags something, you'll know exactly what to fix — before ACX tells you.

A clean noise floor doesn't make a great audiobook. But a broken one guarantees a rejection. Fix the source, process surgically when you have to, and verify the numbers before you submit.