{"resourceId":"nvidia-nemotron-3-5-transcribe-open-speech-model-release-20261028","versions":[{"version":"external-3b71819e9fb9b5041a701366fc1201336b5bfef9e610f75dfb3c6a61aa6d6215","resource":{"id":"nvidia-nemotron-3-5-transcribe-open-speech-model-release-20261028","title":"NVIDIA releases Nemotron 3.5 Transcribe, an open-weight promptable speech model for transcription, speaker attribution and translation in 40 languages","organization":"NVIDIA","sector":"Speech AI platform (vendor model release)","geography":"Global","publishedAt":"2026-10-28","publicationDate":"2026-10-28","eventDate":null,"sourceName":"NVIDIA Research (Nemotron)","sourceLabel":"NVIDIA Nemotron 3.5 Transcribe: Intelligent Transcription with a Speech LLM","sourceUrl":"https://research.nvidia.com/labs/nemotron/NVIDIA-Nemotron-3.5-Transcribe/","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["knowledge-work","accessibility-workforce","developers-agents","infrastructure"],"finding":"NVIDIA's research page introduces Nemotron 3.5 Transcribe, a speech-augmented LLM (about 4.3B active parameters) that performs verbatim, written-form, contextual and speaker-attributed transcription, translation and language identification in 40 languages on recordings up to one hour, chosen per request by a natural-language instruction. NVIDIA reports a 4.71% mean WER across nine English Open ASR test sets, a six-fold cut in named-entity errors on ContextASR-Bench with entity lists, up to 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell via the NIM, and adaptation on 17 tasks at 0.5-2.3 GPU-hours each. Weights are released under OpenMDW-1.1.","sledRelevance":"State and local agencies, courts, school boards and universities record large volumes of meetings, hearings, lectures and calls. An open, self-hostable model that labels speakers and accepts term lists could support draft minutes, captioning and searchable archives on infrastructure the agency controls, but only after local accuracy validation.","evidence":"Vendor research page summarising NVIDIA-run benchmark results and a technical report; no public-sector deployment or independent replication is described.","architectureImplications":"Single checkpoint replaces separate ASR, diarization and translation stages; can be self-hosted on one GPU (BF16, FP8, NVFP4, GGUF variants) or served through the NIM, keeping audio on premises.","governanceImplications":"Prompt-selected transcript styles (verbatim vs cleaned) mean agencies must fix and document which style counts for records, minutes or captions.","securityPrivacyImplications":"Self-hosting keeps recordings off third-party APIs, which matters for closed sessions, student data and criminal-justice audio; the page itself makes no privacy claims.","caveats":"All figures are NVIDIA's own; the research page's 4.71% English WER differs from the 5.33% leaderboard figure NVIDIA cites on Hugging Face, and the headline throughput depends on NIM optimizations not in the open release (per the model card).","streamIds":["nvidia"],"roles":{"sales":"Frame this for clerks, court administrators and campus accessibility offices who pay per-minute transcription fees or face captioning backlogs. The credible opening is a bounded evaluation of an open, self-hosted model on the customer's own recordings, not a replacement promise. NVIDIA's numbers are vendor-run, and even its own figures differ between pages. Lead with control of sensitive audio and the ability to tune for local names and terms. Offer a short pilot that measures error rates on real council, hearing or lecture audio before any rollout decision.","engineering":"Engineering should see one model that does transcription, diarization and translation, deployable on a single GPU with vLLM or the NIM. Plan capacity on the open-release path, since the model card says the headline 2,238x speed relies on NIM optimizations not yet open-sourced. Build a test set from the agency's own recordings, including overlapping speakers, poor microphones, interpreters and local names. Measure word error, speaker-attribution error and named-entity accuracy against human transcripts. Keep the instruction prompt fixed and versioned, because output style changes with the prompt.","delivery":"Delivery should start with one recording type, such as council meetings or lecture capture, and a defined transcript style. Assemble at least 100 human-verified clips for evaluation, run the model alongside the current process and route every output through human review. Track correction time per hour of audio as the main operational measure. Document the prompt, model version and hosting location for records purposes. Decide in advance which outputs remain drafts and who approves anything published, and keep the existing vendor until acceptance thresholds are met."},"roleDetails":{"sales":{"customerProblem":"Agencies and campuses face transcription and captioning costs and backlogs, and some audio cannot go to third-party services.","stakeholders":["City or county clerk","Court administrator","Campus accessibility office","CIO/IT infrastructure lead","Records manager"],"discoveryQuestions":["What do you spend per hour of transcription or captioning today?","Which recordings cannot leave your environment?","What error rate does your current vendor deliver on your audio?","Who approves minutes or captions before publication?"],"valueHypothesis":"A self-hosted open model could lower per-hour cost and keep sensitive audio internal, if local error rates and correction time are acceptable.","potentialEngagement":"Four-to-six-week evaluation on the customer's recordings with measured error and correction time.","unsupportedClaims":["Guaranteed accuracy on local audio","Replacement of court reporters or certified transcripts","Live captioning (model is offline only)","Cost savings before GPU and review costs are counted"]},"engineering":{"fit":"Batch transcription of recorded meetings, hearings, lectures and calls.","architectureIntegration":"Self-hosted via vLLM or Nemotron Transcribe NIM on one GPU; output into records, captioning or search systems.","prerequisites":["GPU capacity (H100-class or Blackwell)","Local evaluation set with human transcripts","Fixed, versioned instruction prompts"],"constraints":["Recorded audio only, not live streams","Headline throughput relies on NIM optimizations not in open release","Vendor-run benchmarks only"],"security":"Keep audio and transcripts in agency-controlled storage; log prompts and model versions.","proofOfValue":["WER and speaker-attribution error on local audio vs current vendor","Named-entity accuracy with term lists","Correction minutes per audio hour"]},"delivery":{"work":"Pilot one recording type with human review, measured against the current process.","dependencies":["Human-verified evaluation clips","GPU hosting","Records and accessibility sign-off"],"ownership":"IT platform team operates the model; clerk or accessibility office owns transcript quality.","skillsAdoption":"Train reviewers on spot-checking names, speaker changes and numbers.","governanceCheckpoints":["Define which outputs are drafts versus official","Accessibility review before captions are published"],"acceptanceCriteria":["Proposed: WER on local audio at or below current vendor","Proposed: correction time per audio hour reduced from baseline"],"risks":["Hallucinated or dropped speech in noisy recordings","Speaker misattribution in public records","Prompt drift changing transcript style"]}},"retrievedAt":"2026-10-29T02:59:07Z","enrichedAt":"2026-10-29T03:06:28Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Could speed captioning and transcript availability for residents and students with hearing disabilities, and reduce clerk transcription load, if error rates on local audio are acceptable.","procurementImplications":"Open weights under OpenMDW-1.1 remove per-hour licence fees, shifting cost to GPUs, integration and validation; agencies should have counsel review the licence.","operatingModelImplications":"Moves transcription from a per-minute service to an internal platform with model updates, evaluation sets and human review queues to maintain.","sourceVerification":{"openedUrl":"https://research.nvidia.com/labs/nemotron/NVIDIA-Nemotron-3.5-Transcribe/","referenceExcerpt":"the model can be adapted to new languages, domains, and tasks on a single GPU, typically within 0.5–2.3 GPU-hours per task.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}