Automated captioning handles ordinary speech well and struggles noticeably with live soccer. The failures concentrate on exactly the words a viewer most needs.
Names are the hardest possible input
Player names come from many languages and are pronounced in ways that do not follow the pronunciation rules the recognizer learned from general speech.
Recognizers work by matching sounds against known words, so an unfamiliar name is replaced with whatever common word sounds closest. The output is confidently wrong.
Systems improve this by loading a match-specific vocabulary of the squads, officials and venue before kickoff, which raises accuracy substantially for a known fixture.
Crowd noise is not background noise
Noise reduction assumes background sound is steady and speech sits above it. Stadium noise is neither steady nor separable, rising sharply at exactly the moments commentary matters.
A goal produces peak crowd volume and peak commentary urgency simultaneously, so the recognizer faces its hardest acoustic conditions at the most important second.
Commentators also raise pitch and speed under excitement, which shifts their voice away from the patterns the system handles best.
Overlapping speakers defeat the model
Soccer commentary is typically two voices interacting, and they overlap constantly. Most recognizers assume one speaker at a time.
When voices overlap, the system produces a merged transcript that attributes fragments to the wrong speaker or drops one entirely.
Separating the voices before recognition is possible when each has an isolated microphone feed, and impossible when only a mixed program feed is available.
The vocabulary is unusual
Commentary uses terms that are rare in general speech and phrasing that is grammatically loose, with sentences abandoned mid-clause as play changes.
Language models trained on written text expect completed sentences, so they try to repair fragments into something grammatical and change the meaning doing it.
Domain-specific training helps considerably, which is why captioning quality varies by sport rather than by broadcaster.
Latency is a hard constraint
Captions must appear close to the audio, so the system cannot wait for later context that would resolve an ambiguous word.
Human captioners handle this by anticipating from knowledge of the game, an advantage that comes from understanding rather than from hearing.
The practical result is that automated captions serve as a supplement on live soccer rather than a replacement, with accuracy improving markedly once the same audio can be processed without a deadline.