The worst audio I have ever been handed was a two hour interview recorded on a phone lying face down on a glass table, in a cafe, next to an espresso machine, with a ceiling fan running. Both voices on one track. The fan hum sat right on top of the guest's voice. Every time the barista steamed milk, the phone's noise gate ducked the host mid word and then popped back up.
The host sent it over with one line: "Can you save this?"
I spent nine hours on that episode. It went out. It was not good. And it taught me more about podcast audio than any plugin chain I have ever built.
Bad audio is almost never one problem
People think of bad audio as a single villain. Too much echo. Too much hiss. Too quiet. In practice, a genuinely bad file is four or five problems stacked on each other, and the fixes fight.
On that cafe file I had:
- One mono track carrying two people at wildly different volumes
- Broadband noise from the fan, plus intermittent noise from the machine
- Room reflection off a glass table and a hard wall
- Automatic gain from the phone, constantly moving the noise floor
- Two clipped sections where someone laughed straight into the mic
Fix the noise first and you make the reverb more obvious, because you just removed the thing that was masking it. Fix the reverb and you thin the voices out. Level the two speakers and you bring the noise up under the quieter one. Every move costs you something somewhere else.
The order you fix things in matters more than the tools you use. That is the whole craft. I would rather hand a bad file to someone with three basic tools and a plan than to someone with a thousand dollars of plugins and no plan.
The order I actually work in
For anything genuinely rough, I go in this sequence and I do not skip ahead:
- Repair anything structural first. Dropouts, clipping, digital glitches. If a word is destroyed, decide now whether it gets cut or lived with.
- Split speakers, even from a mono file. I will manually separate host and guest onto two lanes so I can treat them differently. This is slow and boring and it is where most of the quality comes from.
- Subtractive EQ before anything else tonal. Take out what is wrong before you add what sounds nice.
- Noise reduction, gently, in two light passes rather than one heavy one.
- De-reverb last among the repair steps, and always less than I want to use.
- Then compression and levelling, then a final limiter to a sane loudness target.
The step people skip is number two. They run the whole file through one chain because it is one file. But the host was six inches from the phone and the guest was three feet away with a wall behind him. Those are two completely different acoustic problems wearing one filename.
The thing I will not do
I will not use aggressive AI speech enhancement on a whole episode and call it done.
I use those tools. They are genuinely useful in short bursts, on a single ruined sentence, under a heavier moment of noise. But run one across two hours of conversation and you get that underwater, plastic, slightly robotic voice that everyone recognises now and nobody can name. The breaths vanish. The room vanishes. The laughs get chewed. It sounds processed in a way that raw bad audio never does.
Listeners forgive imperfect audio. They do not forgive audio that sounds fake. A little room noise reads as real. A voice that has been rebuilt by a machine reads as a fraud, even when nobody can explain why.
This is the same instinct behind why we do not let software make edit decisions either. A machine can find silence. It cannot tell the difference between a dead pause and a person deciding whether to say the honest thing. I wrote more about that in a piece on why silence detection makes episodes feel wrong, and the audio version of the argument is identical. Automation is a tool inside a human decision, not a replacement for one.
What was actually salvageable, and what was not
Here is the honest breakdown of that nine hour rescue.
Salvageable: the fan. Steady, broadband, predictable noise is the easiest thing on earth to remove because it never changes. Two light passes and it was gone with almost no damage to the voices.
Salvageable: the level difference. Tedious, but a clip gain pass on the separated lanes fixed it properly. Not a compressor doing the work. Actual manual gain on individual regions, then compression on top of an already even signal.
Partly salvageable: the reverb. I got maybe sixty percent of it out before the voices started sounding like they were recorded inside a sock. So I stopped at forty percent and let the room stay in. The episode sounds like it was recorded in a cafe, because it was.
Not salvageable: the clipped laughs. When the waveform is a flat square, the information is gone. There is no plugin that invents it back. I cut around them and lost about eleven seconds of the episode.
Not salvageable: the espresso machine. Intermittent, loud, and overlapping speech. Anything strong enough to remove it also removed part of the words. Two moments stayed in. I told the host they were staying in and why.
That last part matters. A good editor tells you what cannot be fixed instead of quietly mangling it. If I have to choose between a slightly noisy episode and a clean episode where your voice sounds like a customer service bot, I will pick the noise every time, and I will tell you that is what I picked.
Nine hours is why some files get turned down now
That episode cost me a day and a half of real working time. It also cost the host nothing extra, because I had already quoted it. That was my mistake, not theirs.
What I learned is that rescue work is not editing. It is a different job with a different clock. A normal single camera episode at Starter is $199 and runs on a predictable pipeline: watch the whole thing, cut it, mix it, deliver a first cut in 48 hours. A rescue file breaks the pipeline. There is no way to promise 48 hours on something where I do not know until hour four whether the file is fixable at all.
So now I listen to five minutes before I say yes. Not the intro, because the intro is always the best recorded part of the episode. I ask for a chunk from the middle. If the middle is unfixable, I say so before anyone pays anything. That is part of a broader thing I have written about in why I turn down some shows, and audio quality is one of the two or three most common reasons.
I am not precious about it. I have taken plenty of imperfect files and been glad I did. But there is a difference between imperfect and unrecoverable, and I would rather find out on a five minute sample than on hour four of nine.
What you can fix for free before you hit record
Almost every disaster file I have received could have been prevented by a decision that cost nothing.
- Record each person on their own track. This single change fixes more problems than every plugin I own combined. If your platform offers local per speaker recording, turn it on.
- Turn off automatic gain control on your recorder or phone. It is trying to help. It is making the noise floor move, which makes noise removal much harder later.
- Get closer to the mic. Three inches instead of twelve is a bigger upgrade than a three hundred dollar microphone.
- Kill the room before you kill the noise. A rug, a curtain, and a bookshelf beat any de-reverb plugin ever written.
- Never record next to a fridge, an air conditioning vent, or a laptop with spinning fans. Steady noise is easy to remove but easier to just not have.
- Record thirty seconds of silence at the top. Yes, still. It gives whoever mixes your show a clean noise profile, and it takes half a minute.
- Do a ten second test and actually listen back with headphones. Most catastrophic files were audible as catastrophic in the first ten seconds.
None of that requires new gear. If you are setting up from scratch and want the full version of this argument, I put it in how I would start a podcast today, and the short version is that the room and the mic distance beat the equipment list every time.
Why audio is the thing that quietly kills shows
Hosts tell us they worry about visuals. Cameras, lighting, thumbnails, whether their setup looks professional. Visuals matter for whether someone clicks. Audio matters for whether they stay.
Nobody quits an episode because the framing is slightly off. People quit within ninety seconds because listening is work. Hiss, uneven levels, one speaker too quiet, a room that sounds like a stairwell. The listener does not think "this audio is poor". They think "I am not enjoying this" and they leave, and they never know why.
That is a slow bleed. It shows up as flat retention and a graph that never grows, and by episode seven the host concludes that nobody wants the show. Usually people wanted the show. They just could not comfortably hear it. I went deeper on that pattern in why most podcasts quit at episode 7, and bad audio is quietly underneath a lot of those quits.
The visual side still counts for getting the click, which is why we run a sister studio, ClickTheory, doing YouTube thumbnails at 8 for $250. But a great thumbnail on a rough sounding episode just gets more people to hear the problem.
What good sounds like, in practice
I am not chasing a mastering studio. For a conversation show the bar is fairly humble and completely achievable:
- Both voices at the same perceived loudness, all the way through
- A noise floor you have to strain to notice on headphones
- Breaths and small mouth sounds intact, because those are what make it sound like a person
- No pumping, no plastic artefacts, no ducking on consonants
- Consistent loudness across episodes, so nobody touches the volume dial between weeks
That is it. Not exotic. It just requires someone to sit with the whole file rather than run it through a preset. Our editors watch every episode end to end, which is why a mix problem in minute 54 gets caught. A preset never gets to minute 54, because a preset was never listening.
If you want the boring detail of how that works on our side, from upload to first cut in 48 hours to revisions the next working day, it is laid out in our editing packages.
That cafe episode still exists. I listened back to it last year. It is rougher than I remembered and better than it had any right to be, and the conversation in it is genuinely good, which is the only reason the rescue was worth doing at all.
That is the real lesson. The audio was worth nine hours because the content was worth nine hours. Bad audio on a great conversation is a repair job. Bad audio on a thin conversation is just a thin conversation that is also hard to hear.
So fix the recording side, because it is cheap and it is mostly free. Then spend your effort on the part no plugin touches, which is having something worth saying and someone worth saying it to.
If you have a file you are not sure about, send five minutes from the middle to samy@pixelsamy.com and I will tell you honestly whether it is worth rescuing. Sometimes the answer is to re-record. I would rather say that than take your money and hand you a robot.