Your fridge is in the recording

Play back the first ten seconds of your episode with headphones on and the volume up. That low steady rumble under your voice? The refrigerator in the next room. The faint hiss behind your co-host? A budget USB mic with the gain knob pushed too far. The sharp click when someone hit a spacebar? That is in the file forever.

Almost nobody records in a treated studio. You record in an apartment, a home office, a parked car, or a bedroom with a window AC unit. The goal is not silence — it is getting the noise quiet enough that a listener on cheap earbuds never notices it. Here is the order you should actually do things in.

First, identify which noise you have

Different noise needs different tools, and guessing wastes an hour.

The first two are easy wins. The third is where AI tools earn their keep. The fourth you mostly re-record.

Fixing steady hum and hiss in Audacity

1. Find a clean noise sample

Noise reduction needs a few seconds of the noise without your voice. If you recorded silence at the top of the session, use that. If not, find any 2–3 second gap where nobody is talking — before a question, mid-breath.

2. Get the profile, then apply it gently

Select those 2–3 seconds, then Effect → Noise Reduction → Get Noise Profile. Select the whole track and reopen the dialog. Start conservative: reduction 10–14 dB, sensitivity 5–6, frequency smoothing 3. Preview it. If your voice sounds underwater, or you hear a metallic swishing on the tail of every word, you went too far. Drop to 8 dB and try again. Two gentle passes at 8 dB beat one aggressive pass at 20 dB almost every time.

3. High-pass the rumble

Most of what people call background noise below 100 Hz is room rumble, desk thumps, and plosives. Apply a high-pass filter at 80 Hz for deeper male voices, 100–120 Hz for higher voices, with a 12 or 24 dB per octave slope. This alone can make a recording sound dramatically cleaner and it costs you nothing in quality.

4. Notch the hum if it is tonal

If you can hum along with the noise, it is tonal. Use Effect → Notch Filter at 60 Hz (or 50 Hz), then again at the harmonics — 120 Hz and 180 Hz, or 100 and 150. Keep the Q narrow, around 10–20, so you carve a thin slice rather than a canyon.

When the noise is uneven, manual tools break down

A noise profile assumes the noise is identical second to second. A motorcycle revving past your window at minute 14 is not identical to anything. Profile it and Audacity will either leave it untouched or eat half your consonants trying.

That is the problem AI noise removal actually solves. Instead of subtracting a static fingerprint, these models are trained to separate the human voice from everything else — traffic, fans, a coffee grinder, a barking dog, the neighbor's TV through the wall. They handle intermittent noise because they are not hunting for a pattern in the noise; they are hunting for the voice.

For a free browser-based option, ClearVoice does AI background noise removal for both video and audio: upload the file, let it process, download the cleaned version. That is handy if your podcast audio is really a Zoom recording or a video file you are also posting to YouTube, since it does not care about the container format.

How to get the best result from any AI cleaner

  1. Feed it the highest-quality source you have. The original WAV, not a 96 kbps MP3 you already exported three times.
  2. Trim obvious garbage first. Cut the 40 seconds of setup chatter, the phone ringing during the intro, the false starts.
  3. Listen on earbuds, not studio monitors. Artifacts that are inaudible on good speakers are obvious on the cheap earbuds your audience is using.
  4. A/B it against the untreated version. If the cleaned take sounds thin, hollow, or like your guest is calling from a submarine, dial it back or skip it.
  5. Watch breaths and sibilance. Aggressive models flatten breaths into nothing and turn s sounds into lispy hisses. If that happens, blend: use the cleaned track for noisy sections and the original for clean ones.

The processing order that actually works

  1. Edit for content — cut the ums, tangents, and dead air.
  2. High-pass filter at 80–100 Hz.
  3. Notch out any tonal hum.
  4. AI or profile-based noise removal, conservatively.
  5. Gentle EQ (a small cut around 200–400 Hz if it sounds muddy).
  6. Light compression, 2:1 to 3:1, to even out levels.
  7. Normalize to about -16 LUFS integrated with true peak at -1 dBTP — the sweet spot for Apple Podcasts and Spotify.

Running noise removal before the high-pass filter is the classic mistake. You end up fighting rumble that a free filter would have deleted.

Do not kill the room tone completely

A track that is dead silent between words sounds fake. Listeners cannot name it, but it feels off, like the audio is holding its breath. After cleaning, mix in a very quiet layer of the original room tone at around -55 to -60 dB, or use a setting that leaves a whisper of noise behind. Total silence is a tell.

Stop it at the source next time

Quick answers for common recordings

None of this requires a treated studio. It requires knowing which noise you have, using the gentlest tool that solves it, and resisting the urge to over-process. Clean it, check it on earbuds, and if it sounds natural, ship it.