ATEAM Productions is a Singapore AV production company specialising in livestreaming, live camera production and LED walls for corporate, brand and community events. More and more of the events we stream need live captions that appear on screen as people speak. So we built our own live captioning tool, ATEAM Captions, and we run it as part of our livestream production.
It is a native macOS app that turns a live audio feed into captions burned into the picture, with full control over how they look, and it feeds straight into vMix or OBS. It can also show live English and Chinese captions together.
The problem: getting captions on a live stream without the headaches
Captioning a live stream usually means one of three things: booking a human writer, paying for a hosted service, or relying on whatever captions a platform offers. Each has trade-offs. Human writers are expensive and hard to schedule. Hosted services add a deployment and a recurring bill. Platform captions sit on a separate track that viewers have to switch on, and they don't appear in the recording.
We wanted a tool that:
- runs on a laptop the operator controls, with no server to maintain
- puts captions into the picture itself, so they appear in every stream and every recording
- can be styled to match each client's brand
- keeps a transcript of the event
- keeps working when the venue's internet doesn't
Used on a real show: a Great Eastern panel discussion livestream
We have already run ATEAM Captions on a live show. Great Eastern hired us for live English to Chinese translation of a panel discussion livestream. Five panellists, including a speaker from Great Eastern, spoke in English on stage, and their speech was captioned live in Simplified Chinese, burned into the programme. The screen behind the panel listed the panellists' organisations and the topics under discussion, such as Advance Medical Directive (AMD), Custodian Services and Annual Review.
The captions ran as two lines of Simplified Chinese across the lower part of the picture, clear of the panellists. Because they were part of the programme itself, every viewer of the livestream saw them without switching anything on.
How live captions get from the stage to your stream
- Audio in. The operator takes a feed from the mixing desk into a USB audio interface. The app listens to that line rather than the laptop's microphone, because room audio degrades recognition badly.
- Speech recognition. Speech goes to Microsoft Azure AI Speech, which returns text in near real time. The Azure key stays on the laptop, stored in the macOS Keychain.
- Settling the text. Azure revises its guess as it hears more audio. The app keeps words that have settled fixed on screen and only updates the unsettled tail, so the captions don't rewrite themselves while people are reading. Optionally, finished speech is translated into Simplified or Traditional Chinese.
- Overlay. The captions are drawn on a transparent web page that the laptop serves over the local network. vMix or OBS loads that page as a browser source and composites it over the programme.
- Record. Each finalised line is written to disk as the show runs, so a crash doesn't lose what came before. Afterwards the operator exports SRT subtitles or plain text.
Features
Captions that match your brand: look and placement
- Typeface, size, colour, contrast, position and backing, all set per client
- A transparent overlay by default, so vMix can key the alpha channel directly. This avoids the green fringe that chroma keying leaves around letter edges
- One to four lines, with either page-style or roll-up behaviour
- Languages on screen: the translation alone, or bilingual with the spoken English in a smaller strip below
Getting names and terms right: accuracy
- A vocabulary list of speaker names, products, sponsors and acronyms. This is a phrase list rather than a trained model, so it costs nothing extra and applies live
- A phrase-gap setting so that a speaker pausing for breath doesn't get a full stop in the middle of a sentence
- Punctuation control: automatic, spoken-only, or off
- Settling control for uncertain text: show it dimmed, show it solid, or hold it back until it is confirmed
Staying in control on air: show control
- A single Hide captions button that blanks the overlay instantly. Recognition keeps running, so the transcript isn't interrupted
- Clear-after-silence, so the last line doesn't stay burned into the stream after a speaker stops
- Live input meter with a NO SIGNAL alert if the feed goes quiet on air
- A sample-text preview for styling captions before anyone speaks. It cannot reach the live output
- Changes to recognition settings made on air are flagged, and the operator chooses when to restart, so captions aren't interrupted mid-sentence
Protecting the show: safety and continuity
- Offline fallback: if the internet or the cloud service drops mid-show, captions switch to Apple's on-device recognition and carry on in the source language
- Dead-man's switch: if the control window stops responding while on air, the overlay blanks rather than leaving a stale line on screen
- Blocklist: words that should never appear on screen are masked, and the raw text is kept in the session log for review
- Presets per client, with export and import so a backup laptop starts with the same looks and vocabulary
A record of the event: transcripts and subtitle files
- SRT and plain-text export after the event
- Timing comes from the recognition engine's own timestamps, not the time each caption arrived, so the subtitle file stays in sync with the recording
Why the captions behave the way they do: design choices
A few decisions shape how the captions behave, and we made them deliberately.
Captions shouldn't move. Text is left-aligned, because centring re-flows the whole line every time a word arrives. Each line is frozen once it is complete. Committed words are never rewritten, because an occasional wrong word is easier to forgive than text that changes while someone is reading it.
Nothing from the outside gets in. The overlay is the only thing the network can reach. The control panel is refused to any device other than the operator's own laptop, so someone on venue Wi-Fi can't push text onto the client's stream.
No stream delay is assumed. Captions trail speech by about one to two seconds. For live events, we'd rather accept that than add a delay mechanism that the production would have to manage.
What it isn't: captions for engagement, not a formal accessibility service
ATEAM Captions is built for engagement, and for producing a transcript after the event. It is not an accessibility deliverable. Automatic recognition is not a substitute for a human CART writer. If a client asks for captions as a formal disability accommodation, that request goes to a human writer.
What this means for your event
Captions in every view and every replay.
Because the captions are part of the picture, viewers don't have to switch anything on, and they stay in the recording you publish afterwards.
Bilingual audiences are covered.
For Singapore audiences, English and Chinese captions can run together on screen.
Your brand, not a default style.
Captions are set up in your typeface and colours, and positioned around your graphics and lower thirds.
Speaker names and sponsors spelt right.
Send us the run sheet, speaker list and sponsor names before the show and we load them into the vocabulary list.
Less risk on the day.
If the venue internet drops, captions carry on, and the operator can blank them instantly if needed.
A transcript when the show ends.
You get an SRT subtitle file for the replay and a plain-text transcript for minutes, press releases or follow-up content.
Part of one production.
Captions are run by the same ATEAM team that handles your cameras, sound and livestream, so there is one crew to brief.
Frequently asked questions
Can you add live captions to our event livestream in Singapore?
Yes. We run ATEAM Captions as part of our livestream production. The captions are composited into the programme in vMix or OBS, so they appear on every platform you stream to and in the recording.
Can the captions show English and Chinese at the same time?
Yes. Spoken English can be translated into Simplified or Traditional Chinese. The screen can show the translation alone, or both languages, with the English in a smaller strip below.
How accurate are live AI captions?
Accuracy depends on clean audio, which is why we take a feed from the mixing desk rather than a room microphone. We also load a vocabulary list of speaker names, products, sponsors and acronyms before the show. Expect an occasional wrong word: the captions are designed so that committed words are never rewritten while people are reading.
How far behind the speaker are the captions?
About one to two seconds. We don't add a stream delay to hide this, because that would be one more thing the production has to manage on a live show.
What happens if the venue internet goes down during the show?
Captions switch to Apple's on-device recognition and carry on in the source language. Translation relies on the cloud service, so while the connection is down the captions show the spoken language only.
Do we get a transcript after the event?
Yes. Every finalised line is saved as the show runs, and afterwards we export an SRT subtitle file and a plain-text transcript. The subtitle timing comes from the recognition engine, so it stays in sync with the recording.
Can these captions meet formal accessibility requirements?
No. ATEAM Captions is built for engagement and transcripts. If you need captions as a formal disability accommodation, that calls for a human CART writer.