Overlay captions, automatic transcripts and burned-in subtitles: three different sources.

The short answer

Three things can put the Broker’s words on a screen, and they have nothing in common underneath. An overlay caption is the line he speaks, printed as text by the app at the moment he speaks it: the words are the generated line itself. An automatic transcript is a machine’s guess at the words on the audio track, made after the fact: it hears everything on that track, and in a Signal Broker recording that track carries the game as well as him. A burned-in subtitle is pixels: whatever words you handed a renderer, drawn into the picture at the times you gave it. So the words come from the overlay, the timing can come from the transcript, and burning in is what you do once both are right. That is how this site’s own clips were made.

The product facts behind it, from the app’s own pages: the Broker’s spoken line is shown as text in the app while he speaks; for a stream or recording, one transparent OBS Browser source shows his avatar and spoken captions on a transparent background; his voice reaches OBS through Desktop Audio, together with the game, and there is no separate audio track yet. His words and voice are generated by AI.

Three sources, three origins

Follow one line from the moment it exists. The app generates it as text, speaks it in his AI-generated voice and prints it in the overlay. The voice goes out through your Windows default output device, where OBS Desktop Audio picks it up together with the game and writes it to the recording’s audio track. A transcript is made later, from that track. A burned-in subtitle is made later still, from words and times somebody chose.

The overlay caption: the line itself

The overlay caption is the only one of the three that is not a copy of anything. The app shows the Broker’s spoken line as text while he speaks, and the OBS Browser source shows his avatar and spoken captions on a transparent background; the text in the bubble is the line his voice was generated from, printed, not a transcription of the voice. Dialogue size, in the app, sets how large the captions are. Whatever he actually said, the bubble says it, and that cuts both ways: a line that was wrong, repeated or badly timed is printed just as faithfully, because the caption prints the line, it does not judge it.

Holy hell, that Ruby Key’s a freaking ten grand ticket. I ain’t turning my pockets inside out for that unless it’s life or death. Three grand spare, though, not bad. Shop’s looking tempting but the price tag’s screaming in my ears. Brutality boost could mean I’m packing a bigger punch—nice. Alright, let’s see if the wallet’s just fat or if I’m playing smart today.

Two limits follow from what it is. The caption lives in the picture: once the recording is made, the words are pixels inside the Browser source area, and the only way to get them back as text is to read them, frame by frame, as the people who made this site’s clips did. And it exists only while the overlay is in the scene: a recording made without the Browser source has his voice, through Desktop Audio, and no caption at all.

The automatic transcript: words guessed from the audio

YouTube’s automatic captions are produced by speech recognition from the audio of the upload. YouTube’s own help page says their quality varies, that they can misrepresent what was said, that background noise, accents and overlapping speakers make them worse, and that creators should review them and edit what was transcribed incorrectly. None of that is a complaint. It is what a transcript is: a guess made from sound, after the fact, by something that has never seen the overlay.

In a Signal Broker recording the guess is made harder by the product fact above: his voice reaches the recording through Desktop Audio, together with the game, on the same track. The recognizer receives one mixed signal. A game’s own voices are speech and get transcribed; its music and effects are noise over his words. The recognizer also has to decide who is speaking, and it marks what it takes to be a change of speaker with >> at the start of a line. It cannot know which voice is the commentator.

The developer’s own uploads show every one of these. In the caption file of the DiRT 3 recording, between 4:02 and 4:30 the co-driver’s pace notes run through the transcript mixed with the Broker’s lines (“right four over crest”, “left four”, “right three”), with >> at each hand-over between the two voices. In the Dead Cells recording, the lines at 2:02 and 7:46 begin with >>; the overlay at those moments shows them to be his (“Full health, bright as hell…”; “Fully healed, huh?”), so the marker says that a voice started, not whose. And the words themselves slip: at 3:20 the transcript reads “Fatality boost” where the bubble shows “Brutality boost”; at 8:09 in DiRT 3 it reads “That sun lights a bastard” where the bubble shows “That sunlight’s a bastard”.

That sunlight’s a bastard right in my eyes—makes the road a guessing game. Audi’s gutsy, hanging low and fast, but that turn’s snapping at my heels. Gotta trust the grip, not my eyes.

What the transcript is good for is the clock. Each cue carries a start and an end, and YouTube’s files add a timestamp to every word; the words may be wrong, but they are wrong at the right moment. That is why the site’s clips take their timing, and only their timing, from the audio-derived captions. The file itself is WebVTT: it opens with WEBVTT, carries Kind and Language header lines, and each cue is a timing line with --> followed by the text, blocks separated by a blank line, all as the W3C specification defines it. The specification is also why the marker is written &gt;&gt; inside the file: the characters < and > have a meaning in cue text, so a literal one is written as an escape. The caption retimer reads that format as it is and keeps every one of those details.

Burned-in subtitles: pixels, with whatever words you gave them

A burned-in subtitle is drawn into the video frames. Once rendered there is no text in the file: a player cannot switch it off, a search engine cannot read it, and a wrong word means a new render. In exchange it looks the same everywhere, in every player and after every re-upload, which is why short vertical clips use it.

This site’s clips are made that way. The renderer (scripts/clips/render_clip.py in the site’s repository) reads a small spec: the recording, a start and an end, a hook line, an optional rectangle around the bubble, and a list of captions, each a start time, an end time and the text. It writes those captions into a subtitle file and has ffmpeg draw them into every frame beneath the full game picture, below a close-up of the overlay bubble cropped from the same frame, with a fixed AI-disclosure line at the bottom. The words in the spec were read from the overlay; the times were taken from the audio-derived captions; nobody transcribed the audio by ear. The vertical clip planner writes the same spec shape, and the caption retimer page describes the timing half of the job.

The disclosure line is burned in for the same reason the words are: it has to survive every player, every crop and every re-upload. The Broker’s words and voice are generated by AI, and a clip of them should say so in the pixels.

Check a transcript against the overlay

The check needs the recording, the transcript and your eyes. It does not need your ears, and it should not rely on them: the overlay is the record of what he said.

1. 01Get the transcript as text

   Download the caption file from the caption editor of the platform you uploaded to, or use the .srt or .vtt you already have. Keep a copy: the check changes words, never times.

2. 02Mark the suspects

   Lines that start with >>; names, numbers and game terms; anything that reads oddly; anything you intend to quote. Game terms are the usual victims: “Brutality boost” is a Dead Cells upgrade, and the recognizer was never shown the shop.

3. 03Go to the cue’s time and read the bubble

   Open the recording at the cue’s start, pause, and read the overlay caption. If the bubble shows the words, the line is his: copy them exactly, punctuation included. If the bubble shows nothing, or something else, the transcript line is the game’s sound: delete it, or mark it as game audio if the transcript is meant to be complete.

4. 04Keep the clock

   Leave the cue times as they are unless the whole file is early or late. If it is, shift every cue by the same offset with the caption retimer; if the error grows across the file, use its stretch factor. It never touches the words.

5. 05For a clip, burn in the checked lines

   Put the corrected lines with their times into the clip spec (the planner exports the shape), render, and read the result in the finished video: the burn-in is the last step at which a wrong word can still be caught.

Limits, and the next step

What this page cannot do is make the line right. The caption prints what he said; what he said is AI-generated and can be wrong, repeated or badly timed, and he does not hear you or read chat (the app has no microphone input and no chat), so a line that misreads the scene stays misread in the overlay, in the transcript and in the burn-in alike. The two mishearings above are the ones checked for this page, not a measure of how often automatic captions fail; the recordings were made on earlier Founder Alpha builds, and caption wording and bubble layout change between builds.

Next: if your transcript is right on words and off on time, the caption retimer shifts and checks it in your browser; if you are cutting a vertical clip, the vertical clip planner gives you the layout and the spec. Watch the recorded sessions with the bubble in view and the captions on, and the three sources stop looking alike within a minute.

Limits

• This page explains where words come from; it does not grade the recordings. The mishearings named are the ones checked for this page, not a measurement of how often automatic captions fail.
• The overlay caption prints the line he speaks; the line itself is AI-generated and can be wrong, repeated or badly timed, and the caption prints it faithfully. The app has no microphone input and no chat.
• The recordings were made on earlier Founder Alpha builds; caption wording, bubble layout and voice behaviour change between builds. Nobody listened to the audio: words were read from the overlay, times from the audio-derived captions.
• Founder Alpha is early-access software; the commentary can be wrong, repeated or badly timed, and it is sold to customers in the United States only.