NVIDIA Nemotron 3.5 Transcribe: 40 Languages, Local Audio, and the Real Limits

31 October 2026 | Global AI intelligence | Release date: 28 October 2026. What if your next meeting-transcription system did not have to send recordings to a third-party speech API? On 28 October, NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model designed to process recorded audio in 40 languages, with optional translation, speaker labels, timestamps and audio question answering. The model weights are downloadable today. But the most important implementation distinction is easily missed: this is an offline recording model, not a real-time streaming voice agent. NVIDIA's dated launch announcement and the official model card establish the release and its limits.

What NVIDIA actually released on October 28

Nemotron 3.5 Transcribe combines several operations that enterprise audio pipelines often assemble from separate components: automatic speech recognition, contextual transcription, translation, speaker-attributed transcription, word-level timestamp prediction, language identification and answering questions about recorded audio. Developers can supply natural-language instructions, such as preserving verbal fillers for a compliance transcript, removing fillers for readable notes, or supplying technical names as context. The model supports recordings up to one hour and a 65,536-token context according to NVIDIA. The model card's task matrix and prompt examples are the authoritative starting point; support varies by language and task.

NVIDIA says the model can transcribe 40 languages and translate between English and 31 languages. Its speaker attribution can distinguish up to eight speakers within a recording, but a speaker tag such as <spk:0> does not identify a real person. That matters for meeting minutes, interviews and audit evidence: diarization is not identity verification. Independent TechCrunch reporting from 28 October also identifies the key limitation: prerecorded audio rather than live streams.

NVIDIA also released Granary v2, an associated speech-data curation effort. Its public dataset manifests cover approximately 463,062 hours across 48 languages and 16 source corpora. These are dataset statistics, not the supported-language count for Transcribe and not a claim that every hour is licensed for every commercial purpose. See NVIDIA Research's dated project descriptions and the NVIDIA-authored technical release.

Open weights change deployment choices, not the laws of speech recognition

The practical opportunity is control. A team can evaluate a self-hosted deployment, a managed endpoint or a hybrid architecture without committing to one closed speech API. That is particularly relevant for recorded support calls, internal meetings, clinical documentation and multilingual research. Self-hosting can reduce the need to transfer raw audio outside a defined infrastructure boundary, but it does not automatically satisfy privacy law, employment rules, contractual consent, retention requirements or data-residency policies.

There is an important licensing distinction. NVIDIA describes the weights as open and lists OpenMDW-1.1 as the governing license, with separate third-party notices where applicable. Open-weight does not mean unrestricted open source. Procurement and legal teams should read the actual model license, notices and distribution conditions before embedding or redistributing a checkpoint. The official model card describes intended uses, limitations and licensing.

The 30B versus 34B detail is worth understanding

The released checkpoint is named NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, while NVIDIA's model-card architecture table reports 34 billion total parameters, approximately 4.3 billion active, including the audio encoder. NVIDIA's public messaging also describes the model as a 30-billion-parameter mixture of experts. These figures are not interchangeable memory requirements. The active-parameter count helps explain computation per token, but model storage, precision, KV cache, audio preprocessing and serving overhead determine deployment memory. Do not size hardware by multiplying 4.3 billion by bytes per parameter. Validate the exact checkpoint and precision on the target GPU. NVIDIA's model specifications and H100 serving example provide a better starting point than the marketing shorthand.

NVIDIA publishes BF16, FP8 and NVFP4 checkpoint variants. Its model card includes a single-H100 80 GB quick-start example, and the vendor discusses deployment from local systems to data-center GPUs. Hardware support depends on precision and serving stack; availability of a downloadable checkpoint is not proof that every laptop or GPU can run it acceptably. The model card reports that highly optimized NVIDIA NIM serving achieved up to 2,238 seconds of audio per second on a specified RTX PRO 6000 Blackwell Server Edition configuration, but explicitly notes those optimizations are not yet part of the open-source serving release. Treat that figure as a vendor-specific benchmark, not a throughput commitment for your deployment. See the exact hardware and efficiency qualifications.

How strong is the accuracy evidence?

NVIDIA's model card reports a 4.71% mean word error rate across its listed English Open ASR public sets, compared with 4.94% for the next open-weight result in its table. It reports a 4.62% mean WER on a seven-language multilingual evaluation and a 15.80% mean speaker-attributed error rate on selected multi-speaker benchmarks. These are vendor-published comparisons, with benchmark results captured on 12 October 2026. They do not establish best accuracy for every accent, noisy call, technical term, language or workplace. NVIDIA's evaluation tables and methodology show the exact datasets and metrics.

For a European HR or enterprise deployment, benchmark on actual permitted audio samples, not only English benchmark sets. Include accented speech, code-switching, overlapping speakers, acronyms, names, dates and amounts. Compare not only word error rate but also speaker attribution, critical-field accuracy, latency, cost per recorded hour and the frequency of confident but incorrect audio answers. Translation and audio question answering require separate evaluation; a strong transcript score is not a guarantee of a reliable summary.

Where I would test it first in an enterprise

Recorded meetings: Use a consented, non-sensitive meeting to compare readable notes, verbatim transcription and speaker attribution. Check whether the system accurately extracts decisions and owners; never silently turn an inferred action item into a confirmed commitment.

Customer-support recordings: Test domain-specific vocabulary and multilingual cases. Evaluate whether local processing and retention controls improve the governance model. A lower cost per hour is only meaningful after accounting for GPU utilization, operations, support, model updates and human correction.

HR and recruiting: Use synthetic or expressly authorized recordings to test interview-note preparation, not automated hiring decisions. Interview recordings can contain highly sensitive personal information. A transcription model should not be treated as a candidate-scoring or decision-making system. SAP SuccessFactors or SmartRecruiters integration is not announced as a native feature of this NVIDIA release; any integration would require separately validated APIs, permissions, consent, storage and review controls.

Seven acceptance tests before putting the model into production

1. Confirm the actual workload. Identify prerecorded audio under one hour, supported languages and whether live streaming is required. If real-time interaction is mandatory, this model alone is the wrong component.
2. Read the license and notices. Verify OpenMDW-1.1 terms, any downstream redistribution, third-party obligations and permitted commercial use with your legal team.
3. Test the exact checkpoint. Choose BF16, FP8 or NVFP4 only after checking GPU compatibility, memory headroom and accuracy impact. Do not use active-parameter count as a hardware-sizing shortcut.
4. Build a representative evaluation set. Measure WER, critical names and numbers, diarization, translation quality and audio-Q&A hallucinations on consented data.
5. Protect the audio lifecycle. Define access control, encryption, retention, deletion, logging and redaction. Check whether the deployment truly keeps raw audio within the intended boundary.
6. Measure end-to-end economics. Include audio ingestion, preprocessing, GPU idle time, serving, retries, human verification and downstream workflow integration.
7. Require human verification where stakes are high. Treat transcripts and summaries as assistive outputs, especially in employment, medical, legal and financial contexts.

Availability, roadmap and what remains unknown

Available now (as of 31 October 2026): NVIDIA lists the original model checkpoints, documentation and demonstrations, and describes options for self-hosting with vLLM or SGLang-Omni as well as hosted providers. The original release date is 28 October 2026. Not established by the release: a live-streaming speech service, guaranteed accuracy for all 40 languages, uniform performance on commodity hardware, a fixed enterprise total cost, or any bundled SAP integration. NVIDIA's deployment links, model card and independent launch reporting provide the evidence.

My practitioner takeaway: The breakthrough worth evaluating is not a headline about replacing human note-takers. It is the option to combine transcription, speaker labels, contextual vocabulary and audio understanding inside an infrastructure boundary you control. The test is whether that flexibility survives real accents, messy recordings, privacy controls and the full cost of operation. If it does, the next speech-AI architecture decision becomes much more interesting.

Source and status note: Primary sources: NVIDIA launch, 28 October 2026; NVIDIA Research projects, 28 October; official model card and license; NVIDIA technical release. Independent cross-check: TechCrunch, 28 October. Status: downloadable open-weight offline model, not a generally available streaming voice service. Reviewed 31 October 2026.