Annodex logo AnnodexNotes for people who make media
Stream Setup

Live Caption Setup for OBS Streams and What Breaks It

A charity stream I was asked to help with last winter ran nine hours with captions on. Three weeks later the organisers wanted a transcript for their report, and there was nothing to export.

A streaming computer screen showing OBS with a caption plugin panel open beside a preview where two lines of live caption text sit in a dark band above the webcam box

A charity stream I was asked to help with last winter ran nine hours with captions on. Three weeks later the organisers wanted a transcript for their report, and there was nothing to export. The captions had been rendered as a text source burned into the video, so what they actually had was 540 minutes of pixels shaped like words. We paid someone to retype the key segments from the VOD.

That is the mistake that sits underneath most live caption setups: the thing on screen looks like a caption, so nobody checks whether it is one. A caption is a timed text track that travels alongside the video and can be switched off, resized, searched and re-used. Text painted into the picture is a graphic. Both are useful. They are not interchangeable, and the choice you make at setup time decides what you still have a month later.

What OBS can and cannot send on its own

Out of the box, OBS has no caption pipeline. It encodes video and audio, and that is the whole list. Anything caption shaped has to be bolted on, and there are only three real shapes it can take.

The first is a genuine embedded caption track. Twitch accepts CEA-608 captions carried inside the RTMP stream, and a plugin can inject them. Viewers get a CC toggle in the player, which is the behaviour people expect. The second is platform side recognition: YouTube will generate live captions for a stream on its own, or accept a text feed posted to the caption ingestion endpoint it gives you in the stream settings. The third is a browser source or text source that draws words over your scene, which works on every platform and gives the viewer no control at all.

Plugin availability is the part that ages badly. The widely used OBS caption plugin drives recognition through a hidden Chrome instance, which means a Chrome update can quietly break your captions between one stream and the next. I check mine on a five minute private stream before anything that matters, because I have been caught twice.

The three routes side by side

Embedded CEA-608Platform recognitionBurned in text source
Viewer can switch offYesYesNo
Viewer can resizeYes, in the playerYes, in the playerNo
Works on every platformNoNoYes
Delay addedRecognition only, roughly 1 to 3 secondsRoughly 2 to 5 secondsRecognition only
Survives into a clip or a shortRarelyRarelyYes, including the cropping
Gives you a transcript afterwardsSometimesOn YouTube, usuallyNever
Fails whenPlugin breaks after a browser updateLanguage unsupported, or audio too hotScene changes and the text sits over a face

My own setup for a talk stream uses the embedded track as the real caption and keeps a burned in version switched off in a second scene, for the nights the plugin dies and I would rather have imperfect words than none.

The advice I would push back on

You will read, often from people with good intentions, that turning on automatic captions makes your stream accessible and that the job is then done. I spent six years subtitling for broadcasters and I think that sentence does real harm.

Live recognition on a gaming or interview stream lands somewhere around 88 to 95 percent of words correct in my experience, and the errors are not spread evenly. They cluster exactly where meaning lives: proper nouns, numbers, place names, the punchline of a story, the word someone says while laughing. A caption that reads "we are at four hundred" when you said "we are at fourteen thousand" is worse than silence, because a viewer relying on it has no way to know they were misled. Accuracy figures quoted by vendors are measured on clean read speech by a single speaker into a good microphone, which is not a stream.

What actually helps is unglamorous. Feed the recogniser a clean signal: a gate and a light compressor, not a heavily processed chain, since aggressive noise suppression eats consonants and consonants are what the model needs. We went through sensible values in OBS audio filter chain settings for a clean stream mic. Load a custom word list with the twenty terms your stream repeats, which on a game stream might be character names and on a podcast might be the guest's surname. Say numbers twice, once as digits and once in context. And post a corrected transcript afterwards rather than pretending the live pass was enough. The habits for cleaning up a messy automatic pass are the same ones in fixing auto-generated captions.

Position and pace, the settings nobody sets

If you are drawing captions yourself, you own two decisions the platform would otherwise make for you, and both get set wrong by default.

Position first. Put the caption block in the lower third and it collides with the thing every stream layout already has in the lower third: the webcam box, the now playing bar, the donation ticker. Worse, it sits in the region a vertical crop keeps, so your clips inherit half a caption. I keep captions in a band that starts about 12 percent up from the bottom of a 1080 canvas and leave a clear 48 pixel margin from the left and right edges, because mobile players overlay their own controls there.

Pace second. A plugin will happily draw every word the moment it arrives, which produces a line that reflows three times a second and is physically hard to read. Hold each line for a minimum of 1.2 seconds, cap at two lines, and cap line length near 37 characters. Those numbers come from broadcast subtitle practice rather than from streaming, and they hold up: past roughly 180 words a minute most readers lose the thread even though the words are correct. There is more on where those limits come from in caption reading speed and line length.

One more thing on contrast. White text with a thin drop shadow disappears over a bright game scene. A solid background box at around 70 percent opacity is uglier and far more reliable, and it is what every broadcaster uses for the same reason.

What you should have when the stream ends

Treat the live captions as a service to the people watching now, and plan separately for the record. The two jobs want different tools.

Record a local audio track while you stream, mono, separate from the video. That file is what you run a proper transcription pass over afterwards, offline, where the model can take its time and reach accuracy the live pass cannot. From that you get a timed file you can upload to the VOD, a text transcript for show notes, and a searchable record of a nine hour stream that cost you one checkbox at setup time.

Name the files the dull way: date first in year month day order, then the stream title, then the track. Future you is searching a folder with 200 of these in it, and the only thing that saves you is a name that sorts.

FAQ

Can OBS do live captions without a plugin?

Not as a real caption track. You can draw text into the scene with a browser source, and on YouTube you can let the platform generate captions from your audio, but OBS itself has no built in caption output. The CC toggle viewers expect comes from either an injected caption track or the platform.

How much delay do live captions add to a stream?

Recognition alone costs roughly one to three seconds, and platform side generation runs longer. If you need the words to match the lips, add a matching stream delay rather than hoping the gap closes. For conversation and commentary, most viewers tolerate the lag and prefer it to waiting.

Do live captions stay on the VOD afterwards?

Often not. Treat the live track as disposable and plan to attach a proper caption file to the archive, generated from a local audio recording after the fact. That is also the version worth correcting, since it is the one people will watch for years.

Which is better, burned in captions or a caption track?

A caption track, whenever the platform supports one, because the viewer controls size and visibility. Burned in text earns its place as a fallback for platforms with no caption support and for short vertical clips, where a separate track tends to get stripped out anyway.

PB
Philippa Brennagh

Philippa worked as a freelance subtitler for broadcasters and game studios. She writes about captions, transcripts and archives, and she believes good metadata is a form of kindness to your future self.

More posts by Philippa

More in Stream Setup

01doorganico 084goad 3hot8 4465456 after5internprogram ahandfulofblessings ahmadpc ahomcybersecurity allodialresources antiquesandperiodcollectibles aotc asianworldkashmir beedlectrics bloggertemplatefree bluefeatheragency bodyweightsolution campaign08blog carfixadvisor catstreetcollective cbstvsports cnfanspurchasing committeeof500yearsofdignityandresistance covertopsdjs cs-kaiyuangame cuestacomunicaciontotal cxoutlet dancebetgame deskgearhub destinyvillastay detectmagazine di33 diamondsteelbuildings donmo dyanadropship emarsysfinance emimattl escapechan eskitarih evautoplay everythingweddings-orlando farmzone feedviewerapp foxstudiosscratch gamevui123 gleasongrade greetingsday hacklabasu haskell iw-kaiyun jagoanproperti jilidskripsi joltbusinessfinance juegodepokergratis juegospucca kaeru-festival latestmehndidesigns lestanzediottaviano litemagazines localcommunityhero lustrsazan mariashebets menswheels newwebstechs1 nikkit ootheday oxigenvpn plurfitness prefetial probudgetcarpetcleaning psychicmrsmartin pyaten qvantor rejoignezlogis rexzbot riselayouts ruchin rwxqsgd5 sakaltime shibiaowang shop-bot skrtips spotlesshaven strcoding tetapatrium topselller uneplanteartificielle vipertechcr webtech1223411 wo-kygaming yamaha-mekar yeknava yoybuysheets zenunwa yamaha-mekar jilidskripsi ankertoto ankertoto ankertoto