What hardware does AI digital signage actually need?

Short answer: AI digital signage is a display built to hold a conversation rather than play a loop, and the difference is in the hardware, not the software. It needs a camera that can actually see the person standing in front of it, a microphone array that can hear them in a room with other noise, speakers aimed at where that person stands, and a real computer inside the screen — not a media player. Any screen can run AI software badly. The question to put to a supplier is not "does it support AI" but "what are the numbers", and below are the numbers worth asking for.

Why a normal digital signage screen cannot do this

A standard digital signage display is a one-way device. It has a panel, a small media player able to decode video and run a scheduling client, and no way of knowing whether anyone is in the room. That is the correct design for an advertising screen: content is pushed, the screen plays it, nobody expects a reply.

A conversational display inverts all of that. It has to detect that someone has approached, capture their speech cleanly enough for recognition, run or reach an AI model, and answer in a way the person can hear while standing in a public space. Each of those steps has a hardware requirement, and a media player meets none of them. Adding a USB webcam and a conference microphone to an existing screen produces a demonstration, not a deployable product — the camera ends up at the wrong height, the microphone picks up the whole concourse, and the audio comes out of a panel speaker firing backwards into a wall.

The four numbers that decide whether it works

For a playback screen, the specification that matters is the panel. For a conversational screen, what matters is the volume of space in which a person can actually be seen and heard. Call it the interaction envelope. Two units with identical panels, identical processors and identical camera resolutions behave completely differently depending on these four figures, and most datasheets do not publish any of them.

As a reference point, EKAA publishes these figures for its 75-inch AI conversation display: a 78° camera field of view, 15° downward tilt, camera at 1,700 mm from the floor, 60° horizontal microphone pickup and 1.5 m pickup distance. Those numbers describe a one-to-one conversation station for a person standing roughly within arm’s reach and a half — not a far-field screen that answers someone across a lobby.

That is the real planning constraint. Once you know the envelope, the floor layout matters more than any specification upgrade: an approach path, a floor marking, or simply placing the unit where people already slow down will do more for usable interaction than a better camera would.

Ask the supplier: "What is the camera field of view, mounting height and tilt, and what is the microphone pickup angle and distance?" A supplier who cannot answer is reselling a screen rather than building an interaction device.

A microphone array, not a microphone

A single microphone in a public space hears everything equally: the visitor, the air conditioning, the escalator, the conversation two metres behind. A multi-element array can steer towards the person in front and suppress the rest, which is the difference between speech recognition that works at lunchtime and speech recognition that works only when the building is empty.

Ask how many elements the array has and whether beamforming and noise reduction run on the unit. Then ask the harder question: what is the ambient noise level at which performance was measured? A quiet showroom demonstration tells you nothing about a station concourse at 17:30.

Android or Windows — and why it is not a preference

This is decided by the AI application, not by the buyer. A digital human avatar with a rendering engine, a Windows-only kiosk application, or a vision model that needs a specific runtime will push you to Windows. A lighter conversational assistant that talks to a cloud API, or an application already built for Android tablets, runs happily on Android and costs less.

What matters is that the choice exists before the order is placed. A unit with a soldered-in Android board cannot later host a Windows application, and swapping the computing module after installation is not a field operation on most hardware. Settle the application first, then specify the computer.

Ask the supplier: "Can the computing module be specified per project, and what is in it?"

Cloud or offline — ask before you design the content

Most conversational applications call a model over the network. That makes network quality a hardware-adjacent specification, and it is routinely discovered too late.

• Wired connection. Predictable, and the only sensible answer for a permanent installation. Requires a data point at the location, which in a finished building is a construction job rather than an IT job.
• Venue Wi-Fi. Works until the venue is busy, which is exactly when the unit is being used. Test at peak hour before committing.
• Mobile data. Fine for a free-standing unit in a mall, uncertain in a basement or a steel-framed building. Test the actual position, not the entrance.
• Offline operation. A model running locally needs a far more capable computer inside the unit, and that changes the specification and the price. If the site cannot provide reliable network, say so at quotation stage, not after delivery.

Decide this early because it also changes the content design. A unit that may lose its connection needs a sensible offline state — a touch menu that still answers the top ten questions — rather than a spinner.

Voice fails more often than demonstrations suggest

Voice interaction fails in predictable ways: in noise, with unfamiliar accents, with people who do not want to speak out loud in public, and with anything involving a name, a number or an address. In a Japanese hotel lobby a visitor may be unwilling to say their room number aloud. In a retail store a customer may simply not want to be heard asking.

So touch is not a fallback feature, it is half the product. Every answer the unit can give by voice should be reachable by touch, and the touch surface has to work for the public: projected capacitive, ten-point, with an anti-glare finish so the fingerprints that accumulate in the first hour do not wreck the image. Optical bonding matters here too — it removes the air gap between the glass and the panel, which keeps the image readable at an angle and keeps the unit out of the failure mode where condensation or delamination shows as a black patch.

The panel still matters: brightness and position

Indoor AI units are usually specified around 500 cd/m², which is right for a lobby, a store interior or a corridor. It is not right for a position in direct daylight or opposite a glass frontage, where the image washes out and — more importantly for a conversational unit — strong backlight behind the visitor works against face detection. If the only available position faces a window, raise the brightness specification at the start; brightness cannot be retrofitted.

Resolution is a smaller decision than it looks. A 4K panel is worth it on a large portrait screen that people stand close to, because they are reading text at under a metre. Beyond that, content quality matters more than pixel count.

Where AI signage earns its place — and where it does not

It earns its place where a person would otherwise be asked the same question all day. Hotel and building lobbies, where a concierge answers the same twenty questions in several languages. Retail entrances and product walls, where the first round of questions is always the same. Exhibitions and showrooms, where the stand team spends the morning repeating a product introduction. Clinics, government offices and transport hubs, where the alternative is a staffed desk or a queue.

It does not earn its place on far-field announcement screens, on menu boards, in positions where nobody stops, or anywhere the honest answer is a well-designed sign. AI hardware costs more than a playback screen, and it only pays back where a conversation actually happens. Counting how many people stop at the intended position, before ordering, is a cheap piece of research.

What it costs you from year two

The hardware is the one-off. From the second year the recurring costs are the model or API usage, the knowledge base behind the answers, and the person who keeps that knowledge base current. A concierge unit whose answers went stale six months after launch is worse than no unit, because visitors learn to ignore it.

Budget for an owner, not just an installation. In practice the single best predictor of whether an AI signage deployment is still in use in year three is whether somebody inside the organisation was given the job of maintaining what it says.

On the hardware side, the question worth asking is which assemblies can be replaced individually — panel, touch layer, computing module, camera, microphone board — since a unit built as one sealed assembly means any fault takes the whole thing away.

Checklist before you order

1. Which AI application will run on it — and therefore Android or Windows: ______
2. Cloud or offline: ______ · network available at the position: wired / Wi-Fi / mobile / none
3. Camera field of view: ______ · height: ______ mm · tilt: ______°
4. Microphone elements: ______ · pickup angle: ______° · pickup distance: ______ m
5. Where will the person stand, and does that fit the envelope: ______
6. Ambient noise at peak hour: ______
7. Lighting behind the visitor: ______ · brightness specified: ______ cd/m²
8. Touch: ten-point PCAP / anti-glare / optically bonded: yes / no
9. Languages, and which one the unit opens in: ______
10. Screen size and orientation: ______ · fixed or movable between events: ______
11. Field-replaceable assemblies: panel / touch / computer / camera / microphone
12. Who owns the knowledge base after launch: ______

How EKAA answers these points

EKAA TECHNOLOGY designs and manufactures its own commercial displays, and builds AI interaction hardware on the same indoor interactive line. Because the mechanical design, assembly and testing happen in house, the camera position and angle, the microphone arrangement, the computing module, the screen size and the housing finish are set to the installation rather than taken from a fixed catalogue, and the interaction figures come from the drawings of the team that builds the unit rather than from a marketing sheet. Optical bonding is standard across sizes, construction is modular so individual assemblies can be replaced, and ODM and OEM work is taken alongside the standard range. The current AI range is listed under AI digital signage, and the published interaction envelope described above belongs to the 75" AI conversation digital signage.