DEV Community: underway

How to remove background noise from video but keep voice

underway — Fri, 16 Oct 2026 10:40:03 +0000

if you are looking for how to remove background noise from video but keep voice, page one will hand you a dozen upload boxes: drop your file in, wait, download a cleaner one. they work. but if you record in a browser, there is a step that happens before any of them, and it was quietly ruining our founders' audio before a single cleaner got near it.

underway now fixes it at both ends. while you record, the browser is told to leave your voice alone. once the recording arrives, underway's own machines take the background noise out and level the volume, on every recording, before anything is made from it. nothing to upload, no slider, nothing to remember.

why every cleaner sounded the same

this started with a complaint about our own show. an audio episode of building underway, recorded at a desk with no headphones, sounded thin and far away, and not easy to listen to. so we ran the same minute through three things: the clean-up we already had, a free open source voice cleaner called deepfilternet, and a paid voice isolator. all three sounded the same. all three sounded bad.

when three different cleaners make no difference, the noise is not the problem. so we looked at the recording itself, as a picture of its sound: time along the bottom, pitch up the side, brightness for how loud. two things stood out. the voice stopped dead near 12 kHz, where a clear voice carries detail well above that. and the gaps between words were total silence, which no real room ever is.

twenty seconds of the recording behind episode 40 of building underway, as the browser handed it over: a hard ceiling near 12 kHz, and dead silence between the words.

both are the browser's doing. when a web page asks for your microphone, the browser offers to help: noise suppression, automatic volume and echo cancellation, all on by default in most video tools, because they were built for calls. on a call they are useful. on a recording they gate the room to nothing, pump the volume up and down, and strip the top off the voice. by the time a cleaner gets the file, that detail is already gone, and no cleaner can put back what is not there.

how to remove background noise from video but keep voice, in the right order

the order matters more than the tool. first, capture the voice whole, noise and all. second, take the noise out with something built to separate a voice from everything around it. third, level it, so a quiet answer and a loud one sit at the same volume. do it the other way round, cleaning while you record and again afterwards, and every step makes the voice thinner.

record with the browser's noise suppression and automatic volume off, so your voice arrives with all its detail
keep echo cancellation on only where something plays out loud while you talk: a co-host, a guest on a call, a clip you are talking over
take the noise out after, with a model trained to keep the voice and drop traffic, fans, hum and the room
level it to podcast loudness last, so nobody reaches for the volume between your answers

that is now exactly what underway does, without you choosing any of it.

what underway does now

every recorder that captures your voice for an episode asks the browser for echo cancellation only: the interview on camera, the co-host on voice and on camera, the guest's side of an interview, the take in front of a time-lapse, and a promote film. noise suppression and automatic volume are off.

then, when your episode's file is made, each recording is cleaned once on underway's own render machines with deepfilternet, an open source model that separates speech from noise, and levelled to podcast loudness. a six and a half minute recording takes about twenty five seconds. the voice is cleaned without moving it: we checked the start of every phrase against the picture, and it lands within two thousandths of a second of where it was, so your lips still match your words.

episode 72 of building underway after the new clean-up: the voice stands clear and the room between the words is taken down, not gated to nothing.

it keeps a trace of the room on purpose. removing every last bit of background leaves a voice that sounds processed, floating in nothing. a little air around you sounds like a person somewhere real, which is what a diary is.

how to remove background noise from video but keep voice, when it is already recorded

if the video already exists, the upload tools on page one are the right answer, and they are good. they all do the second and third steps above: separate the voice from the noise, then level it. some also try to rebuild a voice that was recorded badly, which can help with a thin laptop mic but can also make you sound like someone slightly different, which matters more in a diary than in an advert.

three things decide how well any of them works on your file. how loud your voice was against the noise when you recorded: a voice that sits well above the room cleans up well, and one buried in it does not. whether the detail is still there: a recording a browser or a call app already squeezed has lost its top end, and a cleaner can only work with what is left. and how much you ask it to take away: taking the noise all the way to zero sounds processed, taking most of it sounds like a quiet room.

the catch with any upload box is that it is one more step, every time. download, upload, wait, download again, put it back in the edit. once is fine. every day is a chore, and chores are what end a daily habit. that is why underway does it on every recording by itself, before you ever hear the episode.

what to do on the day you record

sit close to the microphone, or wear a headset; distance matters more than any setting
face into the room rather than a bare wall, which throws your voice back at the mic
wear headphones when a co-host or a guest is talking, so their voice does not come back through your mic
turn off the fan or the dishwasher if you can, and if you cannot, record anyway: that part is now handled

what it cannot do

no clean-up makes a laptop's built in microphone sound like a studio one. the biggest single improvement is still distance: a headset, or a cheap clip-on mic a hand's width from your mouth, beats anything done afterwards. and an old recording that the browser already thinned only improves a little, because the missing detail was never recorded.

on the phone, recordings made with the camera never went through the browser's call processing, and they get the same clean-up on the way. a co-host call on the phone keeps the phone's call mode, because without it the co-host's voice from the speaker would come back through the microphone.

where to switch it

it is the clean up the dialogue switch in polish, on the web and the phone, and it is on unless you turn it off. it applies to everything you make, set once. episodes you already have pick it up the next time their file is made: press make the file again on any episode and it comes out cleaner.

the rest of what is set once for every episode: how to make your videos look professional.

faqs

do i need to upload anything or press anything?

no. every recording is cleaned when its episode's file is made. the switch is clean up the dialogue in polish, and it is on.

will it change how my voice sounds?

it takes the noise away and evens out the volume. it does not change your voice, its pitch or your words, and it leaves a little of the room so you still sound like you.

does it fix episodes i already recorded?

partly. press make the file again and the noise comes out, but detail the browser already removed when you recorded cannot come back. recordings made from 15 october 2026 keep their full voice.

what about echo from a room with no headphones?

the clean-up takes out a good deal of room sound, but echo is best fixed at the source: headphones when a co-host or a guest is speaking, and a mic close to your mouth.

is this on the phone too?

yes. phone recordings are cleaned the same way when their file is made.

Originally published at https://underway.life/how-to-remove-background-noise-from-video-but-keep-voice

Talking head video editing ai that draws what you say

underway — Fri, 16 Oct 2026 10:40:03 +0000

talking head video editing ai, as most tools sell it, does one of two jobs: it cuts your ums and long pauses, or it puts captions on. the edits that hold attention on youtube now do a third job as well. while you talk, what you're talking about appears beside you. the number lands on a card. the idea gets an icon. the two options you're weighing sit side by side. the eye always has something new to look at, and none of it covers your face.

underway now does that third job for you, on every wide episode, on series. you don't place anything, pick a template or open an editor. you talk, the episode is cut the way it always was, and the graphics come from what you said.

Episode 64 of Building Underway, as underway edited it. The card is his own words: cost to acquire the customer.

The same episode a few seconds later: if you get 200, only five convert, drawn as a funnel.

what talking head video editing ai adds to an episode

after your episode is cut, underway reads what you said, a minute or so at a time, and picks the moments worth seeing. a new one every two to four seconds, because that is the rhythm the attention-holding edits use. each one lands as you say the words it shows, sits there long enough to read, and slides away when the next thing you say needs the space.

a number you give, on a glass card, counting up: 300 signups from free.
a business idea with its own icon: cost to acquire the customer, with a money icon.
a funnel when you go from many to few: 200 signups, only five convert.
two options side by side when you weigh them: free against the direct route.
a short list that ticks in as you say each part.
a tool you name out loud, in its own logo, if underway connects to it.
small tags for a word said in passing, and the odd stamp for a verdict: not great. cheaper.

each card lands with a soft whoosh and each stamp with a low thump, so the sound tells you something changed even when you're half watching. the cards are dark glass with your accent colour on the thing that matters, the same colour your title, captions and episode label already use, so the graphics read as part of your series rather than a template laid over it.

your words, not an editor's

every word on every card is one you said. a number is one you said, and five and 5 count as the same number. the small label at the top of a card, the cost or the plan or two ways, only says what kind of thing the card holds. there are no invented app screens, no made-up charts, no stock people and no logos for tools you didn't name. if a card can't be built from your own words, it isn't shown.

that matters more for a diary than for an ad. the whole point of a build-in-public series is that it shows what happened, in the order it happened, in your voice. the graphics make it easier to follow. they never change what it says, they never move a line, and they never add a claim you didn't make.

it is the same rule the rest of underway already keeps. the cut never reorders you for a hook. a short never opens on words you struck. a thumbnail never shows a screen that didn't exist. motion graphics are one more thing made for you, held to the same line.

where the graphics go, and where they don't

cards sit in the space beside your head, never over your face, favouring whichever side of the frame has more room. a small tag can pop up on the other side while a card is up, so there is usually something on each side of you without anything in front of you.

they only appear while you're on camera, full frame. the intro, the outro, a preview copied in at the start and any screen you share are left clear, because those already have something to look at. if you share your screen to show a dashboard, the dashboard is the picture; the cards come back when you do.

the episode itself keeps no burned-in captions, so people watching on youtube and on your page can switch them on or off. the copy that goes to the feeds that play muted, tiktok, instagram, linkedin and x, carries bold short captions, one to three words at a time, with the key word in your colour, popping in as you say it.

talking head video editing ai versus doing it by hand

you can make an edit like this yourself. in premiere pro, davinci resolve or capcut it means listening back, marking every number and idea worth showing, building a card for each one, keyframing it in and out, adding a sound to each, and checking none of them sits on your face. for a five minute episode with a new visual every few seconds that is close to a hundred small decisions, and an afternoon if you're quick.

an editor can do it for you, and a good one will do it well. the catch is the same one that ends most build-in-public series: it has to happen every day. a daily diary that needs an afternoon of editing, or an editor on retainer, is a diary that stops in week three.

the ai tools on page one mostly start from footage you already have and ask you to choose a style, place the callouts and approve them. underway starts earlier. it asked you the questions, it knows what you were talking about, it already cut the episode, and the graphics go on as part of the same render. there is nothing to place and nothing to approve, and if you would rather an episode went out plain, it's one switch.

how long it takes and what it costs you

nothing extra on your side. the graphics are planned and drawn on underway's machines after the cut, so the episode takes a few minutes longer to finish than it did; for a five minute episode, about four minutes more. you'll see the counter on the episode while it's being made, the same as always.

it is part of series. there is no per-episode charge and no credit to buy. the founder plan keeps the plain edit: your cut, the title card, the episode label in the corner and captions as a track.

how to switch it on or off

motion graphics is on by default on series for wide episodes. it's a switch in polish, on the web and the phone, so it's set once for the whole series, the same as your title look and your branded intro. set it once and forget it; a series looks like a series because every episode gets the same treatment.

on a single episode, the cut has a motion graphics switch under how it opens and closes. turn it off and render it again, and that one episode goes out without it. turn it back on and it comes back. episodes you made before this change keep the file they have until you press render it again on them.

why a diary app does this at all

founders don't quit building in public because they've run out of things to say. they quit because the daily edit costs more than the day can give, and because posting into silence at episode three is demoralising. a talking head with nothing else on screen asks a stranger to do all the work of following you, and most of them won't. an edit with something moving every few seconds asks much less of them, so more of them stay for the next episode.

an editor would charge you for every minute of that. underway does it on the episode you already filmed, while you get on with the next thing. you film, it cuts, it draws, it posts where you've told it to.

music under the whole episode is the next part. the library underway has today is fine under an intro, and not good enough to sit under your voice for minutes at a time, so for now the episodes carry your voice and the sound of the cards, and nothing else.

is this talking head video editing ai available on every plan?

it's on series, for wide (landscape) episodes. the founder plan keeps the plain edit: your cut, the title card, the label in the corner and captions as a track.

will the graphics cover my face?

no. cards sit beside your head, never over it, and only while you're on camera full frame. a shared screen, the intro and the outro are left clear.

does it change what i said?

no. every word on a card is one you said. the cut is the same cut, in the same order; the graphics go on top of it.

can i turn it off for one episode?

yes. on the episode, the cut has a motion graphics switch. turn it off and render it again.

does it work on episodes i've already made?

new wide episodes get it as they're made. an older wide episode gets it when you press render it again on it.

does it work on portrait episodes?

not yet. the cards need the room beside you that a wide frame has. portrait episodes keep their usual edit.

Originally published at https://underway.life/talking-head-video-editing-ai