{"resourceId":"ai-affairs-nemotron-transcribe-prompting-figures-differ-meeting-notes-check-20261028","versions":[{"version":"external-3b71819e9fb9b5041a701366fc1201336b5bfef9e610f75dfb3c6a61aa6d6215","resource":{"id":"ai-affairs-nemotron-transcribe-prompting-figures-differ-meeting-notes-check-20261028","title":"AI Affairs notes NVIDIA's prompting figures differ across evaluations and that NVIDIA advises checking names, speaker changes and answers against the audio","organization":"AI Affairs","sector":"AI news publication (independent reporting)","geography":"United States (San Francisco-based editor); global subject","publishedAt":"2026-10-28","publicationDate":"2026-10-28","eventDate":null,"sourceName":"AI Affairs","sourceLabel":"Nvidia releases promptable Nemotron 3.5 Transcribe and Granary v2","sourceUrl":"https://www.aiaffairs.com/technology/nvidia-nemotron-3-5-transcribe-granary-v2/","evidenceClass":"independent-reporting","outcomeClass":"mixed","topics":["knowledge-work","governance-procurement"],"finding":"AI Affairs summarises the release (40 languages, translation in 31, speaker attribution in 13, up to eight speakers, one-hour recordings, offline only) and points out that NVIDIA's figures for prompting differ across the evaluations it describes, citing SiliconANGLE's report of 30.8% falling to 20.8% with prompting. It explains that the 100-hours-in-under-three-minutes throughput used 30-second clips at high concurrency, and that audio question answering was trained in English. It relays NVIDIA's advice to check names, speaker changes and answers against the audio.","sledRelevance":"Using the model to answer 'what decisions were made in this meeting' is attractive for clerks and boards, but question answering is English-trained and the vendor itself advises verification, which matters for official minutes.","evidence":"Independent secondary reporting synthesising NVIDIA's Hugging Face post, SiliconANGLE and TechCrunch; no original testing.","architectureImplications":"Throughput figures reflect batch processing of short clips, not single long meetings; capacity planning should test full-length recordings.","governanceImplications":"AI-generated meeting summaries or decision lists should be treated as drafts subject to clerk review.","securityPrivacyImplications":"Not addressed.","caveats":"Secondary reporting; the 20.8% figure it cites from SiliconANGLE does not match the 23.8% in NVIDIA's own post, which illustrates the inconsistency it flags.","streamIds":["nvidia"],"roles":{"sales":"Position meeting-summary features carefully. AI Affairs shows that even published figures about prompting gains do not line up across sources, and NVIDIA itself tells users to check names and speaker changes. For clerks and school boards, offer AI-drafted decision lists as a time saver under clerk review, never as official minutes. Note that question answering is English-trained, which limits value for multilingual communities. A pilot should measure clerk editing time, not just raw accuracy.","engineering":"Engineering should test full-length meeting recordings, because the published throughput used 30-second clips at high concurrency. Separate evaluation of transcription from evaluation of question answering, which NVIDIA trained in English only. Keep one fixed scoring method so internal results are comparable, given the inconsistent public figures. Store generated summaries alongside source transcripts and audio timestamps so reviewers can verify each decision quickly.","delivery":"Delivery should define AI-generated decision lists as drafts that the clerk reviews and corrects before any publication. Measure editing time per meeting against the current minutes process. Restrict question answering to English-language meetings until other languages are tested. Keep links from each summary item to its audio timestamp. Document the scoring method used in the pilot so later model versions can be compared fairly."},"roleDetails":{"sales":{"customerProblem":"Clerks want faster minutes but remain accountable for accuracy.","stakeholders":["Clerk","Board secretary","Communications office"],"discoveryQuestions":["How long do minutes take per meeting today?","Which meetings are not in English?"],"valueHypothesis":"AI-drafted decision lists under review could cut clerk drafting time.","potentialEngagement":"Pilot measuring clerk editing time on drafted summaries.","unsupportedClaims":["Official minutes without review","Reliable question answering in non-English meetings","Consistent published prompting gains"]},"engineering":{"fit":"English-language meetings with recorded audio.","architectureIntegration":"Transcribe then question-answer; link outputs to audio timestamps.","prerequisites":["Full-length test recordings","Fixed scoring method"],"constraints":["QA trained in English only","Throughput measured on 30-second clips"],"security":"Store summaries with source audio under records controls.","proofOfValue":["Full-meeting throughput test","Decision-list accuracy vs clerk minutes"]},"delivery":{"work":"Pilot draft decision lists with clerk review and timing measurement.","dependencies":["Clerk participation","Recorded meetings"],"ownership":"Clerk office owns minutes; IT supports.","skillsAdoption":"Train clerks on verifying AI summaries against audio.","governanceCheckpoints":["Clerk approval before publication"],"acceptanceCriteria":["Proposed: reduced clerk drafting time vs baseline"],"risks":["Errors in decision lists","Misuse on non-English meetings"]}},"retrievedAt":"2026-10-29T02:59:08Z","enrichedAt":"2026-10-29T03:06:28Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Non-English meetings may get transcription and translation but not reliable question answering.","procurementImplications":"Buyers should request a single, documented evaluation set and scoring method from any vendor.","operatingModelImplications":"Clerks remain accountable for minutes even if AI drafts decision lists.","sourceVerification":{"openedUrl":"https://www.aiaffairs.com/technology/nvidia-nemotron-3-5-transcribe-granary-v2/","referenceExcerpt":"Nvidia’s figures for prompting differ across the evaluations it describes.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}