The Name That Went Missing

The grandmother called the crossing by its local name. The transcript renders a common word in its place, and the substitute looks plausible enough to pass. A search for the real name returns nothing. The story is on the tape and reachable in principle, and the person who remembers the phrase in their grandmother’s words cannot find it. The recording is still in the catalog. The route to the passage is gone (and nobody gets a notice saying so).

Speech recognition fails on names in a patterned way. In its baseline analysis of Whisper on basketball commentary, a domain dense with proper nouns, a study reports that phonetic substitution of names made up 35 percent of errors, with accent and dialect sensitivity accounting for another 22, concentrated on international names. The model heard “Kristaps Porzingis” and wrote “Christmas Por Zingas”. Those figures describe one corpus under one model. What the study cannot establish is how a community archive would fare. Whether a local place-name behaves like an international player name in transcription is one of the questions the trial exists to answer.

Whose stories go dark first

Most of the transcript could be accurate and the search could still fail. If the model substitutes a common word for the name of the crossing, it has lost the one word this listener brought to the archive. The catalog still shows the date and the interviewer. A volunteer who remembers the crossing might know where to listen.

The UtopAI record carries the same caution in its own words: transcripts may invent or omit content. Invention is the sharper problem for search. An omission leaves a gap a volunteer might bridge from memory. An invented name leaves a confident, wrong, searchable string that leads the next searcher away from the flood story and toward a word nobody ever said.

What the trial must watch

A trial that transcribes one interview and declares success has not tested the failure that matters. The test case is the name. The trial should include a phrase the speaker would recognize as mis-rendered, and should ask whether the person who knows the correct form can see the error and do something about it. Search that works only for the words the model already knows is search that deepens the existing order of findability: well-documented stories rise, and the local ones sink further beneath the index.

Try the name the listener remembers. See whether the flood story comes back.

Bibliography

khayali, ‘UtopAI’, possibility record and worked example, accessed November 2026.

Ron, Yonathan, Shiri Gilboa, and Tammuz Dubnov, ‘Whisper: Courtside Edition. Enhancing ASR Performance Through LLM-Driven Context Generation’, arXiv:2602.18966, submitted February 21, 2026, sections 1 and 3.2.