From the NVIDIA edition of October 28, 2026
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
NVIDIA · Speech AI platform (vendor model card) · Global
- Publisher
- Hugging Face model card
- Original publication
- 2026-10-28
- Source retrieved
- 2026-10-29
What happened
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.
Why it matters
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 and measured results
Vendor model card, including limitations and fairness notes; descriptive, without error-rate breakdowns by demographic group.
Limitations and uncertainty
The card describes risks qualitatively; it reports no hallucination rate or subgroup error rates.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-29; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- 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
- Discovery questions
- 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?
- Value hypothesis
- A test-first pilot reduces the risk of publishing hallucinated or misattributed text.
- Potential engagement
- Failure-mode evaluation on local audio with a priced comparison of open and NIM paths.
- Unsupported claims
- Any hallucination rate (none reported)
- Equal accuracy across demographic groups
- NIM-level speed on the open release
Pre-sales engineering
Role takeaway
- Fit
- Recorded-audio workloads where human review is feasible.
- Architecture and integration
- 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.
- Proof of value
- Insertion/deletion tests in noise and crosstalk
- Subgroup error comparison
- Throughput on open path vs plan
Delivery
Role takeaway
- 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.
- Skills and adoption
- Train reviewers on the card's failure modes.
- Governance checkpoints
- Human verification before publication
- Re-test after each model update
- Acceptance criteria
- 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
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Capacity planning should use open-release throughput (about half the NIM figure) unless the agency licenses the NIM.
Governance
Who approves, reviews and stays accountable for outcomes?
Vendor-stated coverage gaps across dialect, accent and disability support an equity evaluation before use in resident-facing or legal records.
Security and privacy
What data, permissions and controls need testing?
Hallucinated speech and misattribution can create false statements attributed to named people in public records, a reputational and legal risk.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Uneven accuracy for speakers with disabilities or dialects could undermine accessibility goals; reviewers remain essential.
Procurement
What should contracts, pricing and exit terms secure?
Agencies comparing the free open path with the NIM should price the speed difference; solicitations should require subgroup and hallucination testing on local audio.
Operating model
Which teams own the service once it runs?
Human review queues and periodic subgroup testing become ongoing operating costs.
Publication history
- 2026-10-28NVIDIA · Issue 539 resources
Stable resource ID: nvidia-nemotron-3-5-transcribe-model-card-limitations-open-release-speed-20261028