Define the destination first
“Real-time commentary” can describe several very different products. One system feeds private notes to a producer. Another suggests lines directly to a commentator. A third generates captions or synthetic speech for an audience. The closer the output sits to publication, the less room there is for a person to catch an error. Begin procurement by identifying the destination, not by comparing a headline speed claim.
For a human commentary team, assistance should reduce a specific burden. It might retrieve a verified player note, reconcile a substitution or flag an unusual match state. A stream of generic observations adds a second match for the commentator to watch. Useful support is selective, relevant and easy to dismiss without disrupting the call.
Measure the whole route
Latency includes capture, transport, event recognition, retrieval, generation, approval and delivery. A vendor's model response time measures only one part of that chain. Ask which start and end points are being reported and whether the result includes ordinary network variation.
Also record the delay between the source picture and the programme output. A commentator working from a delayed monitor may receive a data prompt ahead of the visible action. That is not automatically an advantage: it can spoil the sequence, confuse the call or reveal an outcome before the audience sees it.
Give every prompt an expiry
A historical fact may remain valid throughout a match, while an observation about an approaching milestone can expire on the next possession. Attach an event reference, creation time and expiry condition to each suggestion. Remove prompts whose premise no longer holds. The desk should not need to remember whether an old card is still live.
Priority should reflect editorial value, not merely model confidence. A confirmed correction to the scorer outranks a colourful but optional comparison. A changed score matters more than a repeated possession note. Allow producers to define categories and mute whole classes of low-value suggestions during intense passages of play.
Compare capabilities, not adjectives
The table describes platform patterns, not named products or test results. A single platform may combine several patterns. Request a demonstration using your sport, feed conditions and staffing model before treating a capability as available.
| Pattern | Useful contribution | Evidence to request | Key limitation |
|---|---|---|---|
| Feed-led assistant | Structured events and contextual statistics | Source IDs, corrections and expiry behaviour | Inherits feed delay and labelling errors |
| Vision-led analysis | Candidate actions from available pictures | Occlusion, replay and camera-change tests | Cannot observe outside the frame |
| Research copilot | Retrieval of prepared background notes | Document provenance and unsupported-query handling | Preparation may become stale |
| Generated audio workflow | Voiced output from approved information | Stop controls, delay budget and correction process | Errors can reach the audience immediately |
Design for a glance, not a reading session
A prompt card should show the suggested fact, its source and its current status without requiring a long explanation. Keep evidence available behind a deliberate expansion. Do not hide uncertainty in a tooltip that disappears when the operator moves a pointer. Distinguish confirmed, provisional and stale information with words as well as visual treatment.
Keyboard operation matters in a busy production environment. A producer should be able to hold, approve or dismiss a card without hunting for a small icon. Accessibility guidance from W3C, including WCAG, is a useful reference for focus order, contrast and readable controls. Those practices also help teams working under glare, fatigue or time pressure.
Test silence and recovery
Some of the most important tests contain no useful suggestion. A quiet system can be behaving correctly when the commentator already has the relevant information. Evaluate duplication, distraction and unnecessary interruption alongside accuracy. Ask the team whether a prompt changed a decision or simply occupied attention.
Then disconnect the feed, deliver old events out of order and simulate a restart. The interface should show degraded status, avoid replaying a backlog as live information and make recovery visible. A cached statistic needs an “as of” reference. A generator should not fill missing match data with plausible language merely to keep the interface active.
Desk caveat: fast averages hide difficult moments
A single average latency can conceal slow cases during peak traffic or complex events. Request distributions and inspect the occasions when the useful window was missed. Do not infer a service guarantee from a demonstration.
Rehearse with the actual team
Run a recorded match through the full workflow with the intended producer, commentator and editor. Preserve the timing and sequence of suggestions. Afterward, compare prompts with what the team actually said and ask which cards were ignored, mistrusted or too late. A clean transcript alone will not reveal cognitive overload.
Define a fallback before launch. The team should know who can mute the assistant, how a source outage is announced internally and whether any generated output can continue without approval. Broadcast practices discussed across the EBU ecosystem emphasise operational context; there is no universal interface setting that makes every production safe. A useful assistant fits the desk's habits without becoming an invisible authority.
Next: find the right voice