A maple-sugar shack wants NVIDIA Riva ASR to recognize local terms such as “spile,” “niter,” and “hydrometer.” Staff prepare a word list plus example utterances. How should that material be treated for Riva ASR customization?

A maple-sugar shack preparing a word list plus example utterances for local terms such as “spile,” “niter,” and “hydrometer” is assembling a site vocabulary speech-and-text customization dataset that NVIDIA Riva ASR can consume. That material is not a Domain 1 bake-off comparing recognizer architectures before any data is collected. It is not a CLIP English caption set for sugar-pan stills. A Kubernetes Helm values file deploys services; it does not replace spoken examples and vocabulary for ASR customization. Exam caveat: Domain 3 treats this as Riva ASR customization data, not vision captions or Helm-as-data. Operational check: store the civic term list with example utterances, validate transcript alignment, and register it as the ASR customization set before changing Helm charts or CLIP galleries. Version the maple vocabulary and example utterances with the Riva ASR customization release tag. Site vocabulary plus example utterances are Riva ASR customization data for maple terms—not CLIP stills or Helm-as-data. Package maple term lists with example utterances as the Riva ASR customization dataset and tag the release; do not substitute Helm values or CLIP pans. Close q136 with a municipal evidence note that records why the correct mechanism fits this civic multimodal scenario, why each of the three distractors fails by a distinct concept, and which operational check was observed in logs or dataset cards before promotion. File that note with the run packet so later maintenance cannot silently revive a rejected framing.