An aquarium dive-bell office tags species names in a transcript, then a speech synthesizer reads the tags aloud to a gallery. Visitors hear the wrong Latin name even though the photo wall is correct. Where should the next single-stage experiment look first?

A spoken error after token tagging can sit in the text-arm tags or in the later synthesizer. The next experiment should change only one stage so the failure can be localized. Blaming the correct photo wall, rewriting every speech stage at once, or shifting to facility cooling does not isolate the multimodal speech pipeline fault.