Every control that improves a recording for nothing is a physical one, and the largest of them is how far the microphone is from the thing making the sound. Move it and the ratio between the voice and the room changes. Buy a better one and the ratio stays exactly where it was.
The recording on this page is a real job with an invented cast: Aigerim recording her grandmother telling the story of a bread oven, at a kitchen table in Almaty, on a small microphone on a stand. Tiled walls, a bare table, a fridge in the corner, and twenty two minutes of material that has to be listened to by strangers.
The control that costs nothing
Sound from a source falls off with distance and the sound bouncing around the room does not, because the room is arriving from everywhere at once. So the two are not affected equally by moving closer, and that asymmetry is the whole technique: 40 cm to 20 cm adds 6 dB of her voice and none of the room.
Six decibels is a large amount of anything in this craft. It is the difference between a recording that sounds like a kitchen and one that sounds like a person in a kitchen, and it was bought by moving a stand. Every microphone comparison ever filmed has this variable uncontrolled in it, which is most of why the expensive one sounds better in the video.
There is a limit and it is not the technique failing. Closer than about 15 cm the low frequencies start to lift on most directional microphones, and the consonants, which live between 2 kHz and 8 kHz and carry most of whether a word is understood at all, start to burst. So the working range for a seated voice is roughly 20 cm to 30 cm, and capturing something worth mixing is mostly the business of living inside that range.
The table answers 0.7 ms later
Sound travels 34 cm in a millisecond, which is the number that turns a room from an atmosphere into arithmetic. Her voice reaches the microphone directly, and it also reaches the bare table and comes back up. From where she is sitting the reflected path is 24 cm longer, which is 0.7 ms later, and the microphone adds the two together because it has no way not to.
Two copies 0.7 ms apart cancel around 700 Hz, and 700 Hz is in the middle of a voice rather than off at an edge where nobody would care. That is what people are describing when they say a recording sounds hollow or boxy: not reverberation, which you would hear as a room, but a notch cut out of the voice by a surface half a metre away.
The repair is a folded blanket on the table, and it takes ten seconds. Nothing in making the parts sit together can undo it afterwards, because the notch is a cancellation and the frequencies inside it were never recorded. This is the general shape of the subject: the cheap fixes happen before the recording and the expensive ones happen after.
The order, and why it is strict
Each plank can destroy what the next one needs
None of these stages can restore what the one before it lost, which is why the order is not a preference here in the way it is in most crafts.
- 01DistanceFree, physical, and the largest single change available
- 02The roomA blanket and a switched off fridge. Ten seconds, before anything is recorded
- 03The levelHeadroom at the top, because the ceiling does not forgive
- 04The editWhere the twenty two minutes becomes worth the attention it asks for
- 05LoudnessThe last step, a measurement, and the only one with a right answer
The bridge is drawn for spoken recording. Music crosses the same gap in a different order, because a room that is a problem for a voice is often the instrument's best feature.
Where to put the level
Aim for peaks at -12 dBFS, not near zero. The temptation is to record as loud as possible, and it comes from tape, where the noise was in the medium and getting the signal well above it was the whole game. Digital noise is far below anything a room contributes, so the loudest part of the recording is no longer where the risk is.
The risk is at the top and it is absolute. A digital peak that goes past the ceiling is not distorted, it is missing: the samples above the line were never written and there is nothing to recover. Twelve decibels of headroom costs nothing and covers the laugh nobody expected, which on this recording was the grandmother's, and it arrived 14 dB above everything she had said in the previous four minutes.
The stall is the thing you have stopped hearing
After twenty minutes in that kitchen, Aigerim can no longer hear the fridge. Her hearing has adapted, which is a thing hearing does and cannot be trained out of. The microphone has not adapted, because it cannot, and the fridge is on all twenty two minutes at a level that will be obvious to every listener and was inaudible to the one person who could have switched it off.
That is the stall here, and it is not a knowledge problem: the thing wrecking a recording is reliably the thing the person in the room is guaranteed to stop noticing. The fix is mechanical rather than clever. Put the headphones on, record thirty seconds of the room with nobody speaking, and listen to it before anything else happens. The fridge is 50 Hz and unplugging it is the fix, and a high-pass at 80 Hz is not, because half of what a fridge makes is a broadband rattle sitting on top of the voice.
Where the trouble sits
The four places this recording went wrong, by frequency
Three of these were decided in the room and one of them was decided by a plug socket. None of them was decided by the microphone.
Frequency, low to high. The frequencies are this recording's own; where each one sits along the line was placed by eye rather than measured.
- The fridge50 Hz, plus a rattle spread across everything above it
- The high-pass80 Hz. Removes traffic and footfall, and not the fridge
- The table notch700 Hz, cut out of the middle of her voice by a surface 24 cm away
- Consonants2 kHz to 8 kHz, where intelligibility lives and where closeness starts bursting
A spectrum makes these look like four separate problems on one line. Three of them are one problem, which is that nobody listened to the room before the recording started.
From a phone on the table to something people finish
The order is unusually strict in this subject, because each stage can destroy what the next one needs and none of them can restore it. Distance, then the room, then the level, then the edit, then loudness at the end. Coming at it backwards is the common route and it has a signature: hours in a plugin chain trying to rebuild a voice that a blanket would have saved, which is also why the software everything happens in gets learned long before it can help.
Two of the clusters below sit outside that line rather than on it. Sound that is built rather than recorded starts where recording stops mattering, since a door that never existed has no room to fight. Writing, arranging and producing a track is a separate subject that happens to share a timeline, and it is the one place in this topic where the room is a creative choice rather than a tax.
What to open in 2026
Loudness is the one part of this that has a right answer, and since streaming platforms began normalising everything, the target for spoken word delivery is about -16 LUFS. Auphonic will measure a file and move it there, along with levelling the difference between two people at one table, and it replaced guessing at levels between one episode and the next; there is a free tier that covers a show of this size. ffmpeg does the same measurement free from a command line through its loudnorm filter, which implements the same standard, and it replaced exporting three versions to find one that matched.
iZotope RX is the repair application and it is genuinely remarkable at removing a fridge from a finished recording. It is also expensive, and the honest statement is the one nobody selling it makes: it replaced re-recording the session, at a quality cost you can hear on any voice it has worked hard on. Ten seconds with a plug socket beats it on every axis. Recording, editing and publishing a show covers what to do when the session cannot be recorded again.
What none of this settles is how loud the story should be, which is not a number. The grandmother drops to almost nothing at the part about her brother, and every tool above will read that as a fault and lift it. Levelling it makes the recording easier to hear and the story worse, and deciding which of the two matters is not an audio decision at all: it belongs to the craft under every photograph, film and edit, arrives on the timeline in moving pictures, from the shoot to the export, and is the same judgement a built soundtrack has to make deliberately in graphics and characters that move.