{"resourceId":"nvidia-nemotron-3-5-transcribe-model-card-limitations-open-release-speed-20261028","versions":[{"version":"external-3b71819e9fb9b5041a701366fc1201336b5bfef9e610f75dfb3c6a61aa6d6215","resource":{"id":"nvidia-nemotron-3-5-transcribe-model-card-limitations-open-release-speed-20261028","title":"Nemotron 3.5 Transcribe model card lists hallucinated speech and speaker errors and says the open release is about two times slower than NVIDIA's NIM figures","organization":"NVIDIA","sector":"Speech AI platform (vendor model card)","geography":"Global","publishedAt":"2026-10-28","publicationDate":"2026-10-28","eventDate":null,"sourceName":"Hugging Face model card","sourceLabel":"nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B","sourceUrl":"https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B","evidenceClass":"vendor-claim","outcomeClass":"cautionary","topics":["data-security","governance-procurement","infrastructure","accessibility-workforce"],"finding":"The model card states that the 2,238-seconds-per-second throughput comes from optimized vLLM in the Nemotron Transcribe NIM, that these optimizations are not yet in the open-source release, and that the vLLM integration shipped on release day is expected to be roughly two times slower. Its limitations section lists substitutions, deletions, repetitions, hallucinated speech, incorrect language choice and speaker-attribution errors, recommends targeted testing on long recordings, overlapping speakers, noise, names, code-switching and misleading context, and says the training inventory does not establish proportional coverage across age, gender, ethnicity, disability, dialect, accent or recording environment.","sledRelevance":"These are the failure modes that matter most for public records: hallucinated words in minutes, wrong speaker labels in hearings and uneven accuracy across dialects and disabilities. The card gives agencies vendor-acknowledged grounds to require local testing and human review.","evidence":"Vendor model card, including limitations and fairness notes; descriptive, without error-rate breakdowns by demographic group.","architectureImplications":"Capacity planning should use open-release throughput (about half the NIM figure) unless the agency licenses the NIM.","governanceImplications":"Vendor-stated coverage gaps across dialect, accent and disability support an equity evaluation before use in resident-facing or legal records.","securityPrivacyImplications":"Hallucinated speech and misattribution can create false statements attributed to named people in public records, a reputational and legal risk.","caveats":"The card describes risks qualitatively; it reports no hallucination rate or subgroup error rates.","streamIds":["nvidia"],"roles":{"sales":"Bring the model card into early conversations: it is NVIDIA's own admission of hallucinated speech, speaker errors and unproven demographic coverage. That candour helps position a responsible pilot with built-in testing rather than a rushed rollout. For public bodies, the risk of a fabricated quote in official minutes outweighs speed. Offer an evaluation that measures hallucination and misattribution on local audio, including diverse speakers, and prices both the open path and the faster NIM path honestly.","engineering":"Engineering should size infrastructure on the open-release path, which the card expects to run roughly two times slower than the published NIM throughput. Design tests for each listed failure mode: hallucinated insertions in silence or noise, deletions in crosstalk, wrong-language output in code-switched speech and misleading context lists. Measure results by speaker subgroup where consented data allows, since the card says coverage is not established. Log model version and prompt with every transcript so errors can be traced.","delivery":"Delivery should turn the card's limitations into acceptance tests and review rules. Require human verification of any transcript that will be published or attributed to named speakers. Build a small, consented evaluation set covering accents, dialects and speakers with disabilities relevant to the community. Schedule re-testing after each model update. Publish a short notice describing AI assistance and review for any resident-facing transcripts, and keep a correction process for people who report errors."},"roleDetails":{"sales":{"customerProblem":"Public bodies need transcripts they can attribute to named speakers without fabricated or misassigned words.","stakeholders":["Clerk or records officer","Court administrator","ADA/accessibility coordinator","General counsel"],"discoveryQuestions":["How would a fabricated quote in minutes be caught today?","Which communities or dialects are hardest for current transcription?","Will you use the open release or the NIM?"],"valueHypothesis":"A test-first pilot reduces the risk of publishing hallucinated or misattributed text.","potentialEngagement":"Failure-mode evaluation on local audio with a priced comparison of open and NIM paths.","unsupportedClaims":["Any hallucination rate (none reported)","Equal accuracy across demographic groups","NIM-level speed on the open release"]},"engineering":{"fit":"Recorded-audio workloads where human review is feasible.","architectureIntegration":"Open vLLM path or licensed NIM; transcripts logged with model and prompt version.","prerequisites":["Consented evaluation audio across speaker groups","Capacity sized on open-release speed"],"constraints":["Open release about 2x slower than NIM figures","No reported hallucination or subgroup error rates","Card lists hallucinated speech and speaker errors"],"security":"Trace every transcript to model and prompt version; restrict publication to reviewed text.","proofOfValue":["Insertion/deletion tests in noise and crosstalk","Subgroup error comparison","Throughput on open path vs plan"]},"delivery":{"work":"Convert listed limitations into acceptance tests, review rules and a correction process.","dependencies":["Consented test audio","Review staff","Records policy"],"ownership":"Records or clerk office owns published transcripts; IT owns the model.","skillsAdoption":"Train reviewers on the card's failure modes.","governanceCheckpoints":["Human verification before publication","Re-test after each model update"],"acceptanceCriteria":["Proposed: zero unreviewed transcripts published","Proposed: subgroup error gap within an agreed threshold"],"risks":["Fabricated text in public records","Unequal accuracy for some speakers","Capacity shortfall on open release"]}},"retrievedAt":"2026-10-29T02:59:40Z","enrichedAt":"2026-10-29T03:06:28Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Uneven accuracy for speakers with disabilities or dialects could undermine accessibility goals; reviewers remain essential.","procurementImplications":"Agencies comparing the free open path with the NIM should price the speed difference; solicitations should require subgroup and hallucination testing on local audio.","operatingModelImplications":"Human review queues and periodic subgroup testing become ongoing operating costs.","sourceVerification":{"openedUrl":"https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B","referenceExcerpt":"These serving optimizations are not yet part of the open-source release","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}