Live dealer casinos: OCR and low-latency streaming
16 November 2026
A live dealer casino is a real dealer at a physical table, filmed and streamed in real time to players who bet from a screen. That much is familiar. The less visible part is the software that turns a card landing on felt or a ball dropping into a pocket into data the betting interface can act on. Optical character recognition, or OCR, does that job, and low-latency streaming keeps the picture in step with it. The two together are the engineering stack behind the format.
For insiders the interesting questions are about the seams. Where does the recognition happen? What keeps a player's video from running ahead of or behind the game state? And who decides, when the two disagree, what actually happened?
What OCR reads at the table
OCR in this setting reads physical cards and translates them into digital data in real time, according to an explainer on standard.co.uk. The practical effect is that the game interface updates when a card is dealt: the dealer turns over a card, the system identifies it, and the on-screen state of the hand changes without anyone typing anything in.
The same technology is used on roulette. Instead of cards, the system reads the position where the ball lands, and that reading becomes the result the game acts on. gamblingzone.com describes recognition technology in live casinos more broadly: it can help identify cards, roulette outcomes or other game information, and then transmit it to the casino interface.
One consequence is easy to miss. OCR can reduce reliance on manual result entry, as gamblingzone.com notes. Manual entry means a person keys in a result after the fact, which is slower and adds a point where human error can creep in. Automated recognition takes that step out of the critical path, or at least shrinks it. It does not remove the dealer. The dealer still deals, spins and handles the physical equipment, and the recognition layer watches what the dealer does.
The distinction matters when people talk about live games as "streamed video". The video is what the player sees. The OCR output is what the game uses. They come from the same table but travel as different things.
Inside the studio: cameras, light and a separate recognition feed
Live dealer setups commonly use high-definition cameras, professional lighting and low-latency streaming technology, according to betmgm.ca. None of that is exotic on its own. Together it explains why a live table looks more like a small television production than a casino floor.
A studio can also run several camera angles, such as wide, close-up and overhead views, per the same source. The wide shot gives the sense of a room and a dealer. The close-up lets a player see a card or a chip stack clearly. The overhead view is useful for what is happening on the table surface itself.
Lighting is part of the recognition story as well as the presentation. Cards and ball positions have to be read reliably, so a table that looks good to a viewer also has to be legible to software. The explainers reviewed do not publish lighting specifications, and it would be guesswork to supply them.
The architectural detail worth noting is that live dealer systems are often built so that the recognition camera is separate from the camera producing the player's video feed, as turbogeek.org describes it. That separation makes sense. The camera that serves recognition can be placed and tuned for reading a card or a pocket, while the broadcast cameras are framed for the viewer. Neither compromises for the other.
It also means the thing a player watches is not the thing the game trusts. The player sees one stream. The game logic is fed by another path entirely.
Synchronising video and data, and why latency is the constraint
Live casino systems work alongside cameras, game control systems and streaming infrastructure to connect the table to the player interface, according to gamblingzone.com. Within that, betmgm.ca identifies a specific component: a Game Control Unit, or GCU, used to synchronise the video stream with the OCR or recognition output at the table.
Think of the GCU as the piece that makes sure the data and the picture describe the same moment. If recognition reports a card instantly but the video arrives a few seconds later, the interface can show a result before the player has watched it happen. Run the other way, and the player sees a card on screen while the game state has not caught up. Either mismatch undermines the point of a live game. Keeping bets, outcomes and video aligned in real time is the stated purpose of the technology, per gamingzone.com.
Streaming delivery is the other half of that problem. Some live dealer systems use WebRTC to deliver ultra-low-latency streams, according to backfromthebrink.com. WebRTC is a real-time communications technology, and its attraction here is speed of delivery rather than image polish.
difficulties
There is a trade-off, and turbogeek.org frames it usefully: fast recognition against streaming delay, in service of short betting windows. A shorter window keeps the game moving. But the shorter it gets, the less room there is for the stream to lag behind the table, because a player who is still watching the previous moment is betting on stale information. Every second of delay is a second the player sees less of the round than the system already knows.
The published explainers do not put numbers on betting window lengths or on how many milliseconds of delay a given setup runs. Operators and studio providers presumably do hold those figures, but they are not in the material available here, so any claim about how short is short enough would be speculation.
One round, step by step
The basic design goal, as turbogeek.org puts it, is to keep the server as the source of truth while the stream reflects decisions already made. That is easiest to see by following a single round. In a typical round the sequence runs as follows:
- The betting window opens.
- The player places a bet.
- The server closes the window.
- The dealer plays the round.
- The server settles bets after the result is read.
Two points stand out. First, the window closes before the dealer acts. Bets are locked by the server, not by anything the player sees on screen, so there is nothing to be gained from watching the video for a cue and betting late. Second, settlement happens after the result is read. That is where OCR sits: the physical outcome is captured, converted to data, and handed to the server, which then pays or collects.
Players sometimes assume the stream is what drives the outcome, as if the video were the game and the numbers followed. The sequence says otherwise. The dealer plays a real round at a real table, the recognition layer reads it, and the server settles. The stream is a faithful window onto a process that is already underway.
Blackjack and roulette: the same stack, different inputs
Blackjack shows the card side of the technology. betmgm.ca describes live blackjack as an online version of the game broadcast from a live dealer casino game studio, with a professional dealer dealing eight-deck shoes. Eight decks mean a lot of cards moving through a lot of hands, and each one has to be read correctly as it appears. OCR translating physical cards into digital data in real time is what lets the interface show each hand as it builds, without a person transcribing the deal.
Roulette is the second example, and it is a different recognition problem. There are no faces to read, only a ball coming to rest. As standard.co.uk notes, OCR is also used to read roulette ball landing positions. The input changes while the architecture stays the same: a physical event happens, a recognition camera reads it, data goes to the server, and the server settles the bets.
That is the real argument for treating this as a stack rather than a set of game-specific tricks. Cards and ball positions are different objects, but both end up as the same kind of thing: a result the server can act on. gamblingzone.com makes the broader point that recognition can identify cards, roulette outcomes or other game information, which suggests the approach extends to other tables built on the same design.
Why the server, not the screen, decides
Everything above hangs on one rule. The server is the source of truth, and the stream shows decisions already made. A player's screen is a presentation layer. It can lag, it can buffer, it can drop frames, and none of that changes who won the hand.
That has a practical consequence for how to read a live game. If the video stutters at the moment a result is revealed, the bet has still been closed and the outcome still stands, because the server locked the window and settles on what the recognition system read, not on what a particular viewer managed to see.
It also frames the questions worth asking of any live studio. How is the recognition feed kept separate from the broadcast feed? How does the Game Control Unit hold data and video together? How short are the betting windows, and what does the stream do to respect them? The public explainers describe the architecture in general terms and stop short of operator-level specifics. Until studios and regulators publish more, those answers sit inside the operators' own systems.